GB2630383A - Method and apparatus for boiler failure prediction - Google Patents

Abstract

A fault may be predicted in a monitored environmental control system (e.g. a boiler) by performing an iterative monitoring process and performing a fault prediction. Monitoring includes receiving sensed data, and computing system state data based thereon. Previous state data is processed to have a lower dimensionality. Fault prediction is performed by providing inputs based on a monitored data set and system state data to a fault classifier. The fault classifier outputs a fault indication indicative of whether a future fault (e.g. part failure) of the environmental control system is expected to occur, and this fault indication is output.

Description

METHOD AND APPARATUS FOR BOILER FAILURE PREDICTION FIELD OF THE INVENTION

The present application relates to methods and apparatus for monitoring an environmental control system. In particular, the application relates to predicting the failure of an environmental control system, such as a boiler, and predicting which parts of the system will need replacing.

BACKGROUND OF THE INVENTION

The nominal operation of environmental control systems, such as heating systems, is essential for maintaining desired environmental parameters in an environment. For example, boilers must operate correctly in order to maintain a comfortable temperature in a household without excessive energy usage. However, due to regular use over an extended period of time, such systems can fail according to a variety of failure modes. Some of these failure modes can be critical, leading to downtime of the system (e.g. boiler) and thus complete loss of functionality.

Typically, environmental control systems such as boilers are either repaired following a failure, or serviced on a semi-regular basis. However, this leads to system downtime (e.g. loss of central heating and/or domestic hot water), as the repair only occurs after failure, while regular (e.g. yearly) services are too crude to detect failures in advance. Furthermore, these approaches also require in-situ inspection of the system in order to diagnose the fault and, if relevant, which parts will need replacing. This can increase the amount of time, cost and number of journeys travelled by a repair technician. Therefore, these approaches are often not suitable for preventing critical failures and can lead to system downtime and expensive, time-consuming repair processes.

SUMMARY OF THE INVENTION

Aspects of the invention are set out in the independent claims and preferable features are set out in the dependent claims.

There is described herein a computer-implemented method for predicting a fault in a monitored environmental control system, the method comprising the steps of: performing an iterative monitoring process comprising, at each of a plurality of iterations: receiving a monitored data set comprising a set of features relating to the environmental control system, including one or more features based on sensor data obtained from one or more sensors of the environmental control system; and computing system state data based on the monitored data set, wherein computing system state data comprises; obtaining a previous system state representation indicative of an operational state of the environmental control system at a previous iteration; calculating, based on the monitored data set and the previous system state representation, a current system state representation indicative of an operational state of the environmental control system at the current iteration, wherein the system state representations each have a lower dimensionality than the monitored data set; the method further comprising performing a fault prediction comprising: providing a set of inputs to a fault classifier, the inputs based on the monitored data set for a given iteration and system state data computed by the monitoring process, wherein the fault classifier is arranged to output based on the inputs a fault indication indicating whether a future fault of the environmental control system is expected to occur; and outputting the fault indication.

By predicting whether a future fault will occur, the method can enable an environmental control system such as a boiler to be fixed pre-emptively, thus reducing system downtime. In addition, by computing (for a given iteration) system state data by calculating a current system state representation based on sensor data and a previous system state representation, and using that system state data and the monitored data set as input to a fault classifier, the accuracy of the fault prediction can be improved, because both the latest readings and the historical state or health of the system can be taken into account. Moreover, by representing the operational state of the system with a representation having a lower dimensionality than the monitored data set, the computational efficiency of the method can be improved.

In addition, performing the iterative monitoring process at each of a plurality of iterations can enable the method to be used for live, on-the-fly monitoring of the environmental control system, which can enable a faster response in case a fault is predicted.

The terms "fault" and "failure" are generally used interchangeably herein and may be used to refer to complete failure of a system or component part or partial failure, such as a performance degradation of the system or component part. The term "environmental control system" refers to any system for controlling an environmental factor. For example, heating, ventilation and/or air conditioning (HVAC) systems are all examples of environmental control systems, since they control temperature, air flow and/or humidity of controlled environments. Specifically, a heating system can include a boiler as well as associated components/devices, such as connected sensors, meters, water/fuel/energy supplies, radiator networks etc. An environmental control system is typically associated with a specific premises, such as a building or other property.

The term "features" as used herein refers to data values determined based on sensor data and/or other data. For example, this can include raw sensor data or derived values computed from sensor data. Features may be characteristic of the environmental control system or of its operation, or of a building environment served by the environmental control system.

The term "system state" may also be referred to as a system health, system status, operating status or operating condition, and "system state data" thus refers to corresponding data. The term "system state representation" refers to a data representation indicative of an operational state of the system, for example dependent on whether the system is operating normally or abnormally. The system state data can include the obtained and/or calculated system state representations. The system state data can also include other types of data relating to the system state.

The terms "current", "previous" and "future" as used herein are relative to the point in time associated with the received monitored data set (in a given iteration), or the latest or most recently received monitored data set. For example, a "current" system state representation can be the system state representation at the time when the corresponding data in the monitored data set were recorded (not when the data set was received while performing the method). The term "previous" refers to points in time before the latest data were recorded. The term "future" refers to points in time after the latest data were recorded.

As will be recognised by those skilled in the art, the term "dimensionality" refers to the number of dimensions/attributes/features/types of data that exist in the data set. Thus, for time series data (e.g. a matrix of data), the dimensionality is the number of elements/features etc. corresponding to each time step.

Performing the fault prediction may comprise providing as an input to the fault classifier the system state representation calculated for the given iteration or an earlier, preferably immediately preceding, iteration. Inputs to the classifier may include the features of the monitored data set for the given iteration and/or the system state representation calculated for the given iteration.

The method may further comprise performing the fault prediction repeatedly for respective iterations. The fault prediction may be performed repeatedly for each iteration. The fault prediction may be performed in real-time at respective iterations of the monitoring process, or offline at a later time.

The system state representations may be two-dimensional representations. This can enable visualisation of the classifier's predictions and can achieve a significant dimensionality reduction to reduce computational load.

Calculating the current system state representation may comprise applying a mapping to the monitored data set and the previous system state representation. The mapping may be based on a self-organising map, SOM.

The fault indication may indicate whether a fault of the environmental control system is expected to occur within a future time period. The length of the future time period may be up to 100 days. The length of the future time period may be around 60 days. The fault indication may comprise a probability of a fault occurring within the future time period. The fault prediction may comprise determining an expected fault of the environmental control system by comparing the probability of a fault occurring to a threshold fault probability. Outputting the fault indication may be based on the probability of a fault occurring being greater than the threshold fault probability.

The method may further comprise performing a part failure prediction comprising: inputting the monitored data set and system state data computed by the monitoring process, preferably the current system state representation, into a part failure classifier, the part failure classifier arranged to output based on the inputs a replacement part indication indicating one or more parts of the environmental control system expected to experience a fault in the future; and outputting the replacement part indication. This can enable specific parts that will need replacing in the future to be identified. The replacement part indication may additionally or alternatively indicate one or more parts of the environmental control system expected to require replacement in the future. The terms "require replacement" and "experience a fault" in the context of a part of the environmental control system may be used interchangeably (in the sense that the method determines whether particular parts are expected to experience a fault or equivalently to require replacement).

The part failure prediction may be performed in addition to or instead of the fault prediction set out above. In an embodiment, the part failure prediction may be performed in response to the fault indication (produced by the fault prediction described above) indicating that a future fault of the environmental control system is likely to occur. The part prediction may be performed based on a probability of a fault occurring being above a threshold fault probability. This can avoid redundant data processing if a system fault is unlikely to occur in the first place.

The replacement part indication may indicate one or more parts of the environmental control system expected to experience a fault within a future time period. The length of the future time period may be up to 100 days. The length of the future time period may be around days. The replacement part indication may comprise for each of a predetermined set of parts a probability of the respective part experiencing a fault or needing to be replaced within the future time period. The one or more parts may include one or more parts of a boiler. The one or more parts may include one or more of: a circulation pump, a thermostatic radiator valve, an air vent, a heat exchanger, a flow sensor, a condensate pump, an expansion valve, a diverter valve and/or a pressure relief valve.

The sensor data may comprise time series data obtained from one or more sensors of the environmental control system. The time series data may include sensor readings from multiple points in time. This can enable sensor data which has a higher time resolution than the monitoring interval of the monitoring process to be aggregated or otherwise processed in order to be taken into account for the fault prediction and/or part failure prediction.

The features and/or sensor data may comprise one or more of: an internal temperature of a building environment served by the environmental control system, an external temperature of the building environment, a target temperature of the building environment, a temperature on the flow pipe of a boiler, a temperature on the return pipe of a boiler, an electrical power consumption of a boiler, and/or a combustible fuel consumption of a boiler. A combustible fuel consumption may be a gas consumption. The one or more sensors of the environmental control system may comprise one or more of: a temperature sensor, a water flow sensor, a thermostat, a smart meter, an electrical energy sensor and/or a gas consumption sensor.

The set of features of the monitored data set may include features based on premises data indicative of properties of a building environment served by the environmental control system. The premises data may comprise data indicative of the thermal behaviour of the building environment. The premises data may comprise one or more of: a type of building, a number of rooms, a number or type of windows, and/or an amount or type of insulation.

The fault classifier and/or the part failure classifier may comprise a decision tree or random forest classifier. This can further improve the interpretability of the predictions. The fault classifier and/or the part failure classifier may be trained using a training data set. The monitoring process may further comprise, for a given iteration: identifying a type of data in the training data set that is not present in the monitored data set; suppressing, when computing the system state data, a dimension of the monitored data set corresponding to the identified type of data; and identifying a replacement value corresponding to the identified type of data, wherein the replacement value is learned based on the training data set. The method may further comprise combining the monitored data set for the given iteration, system state data computed by the monitoring process and the replacement value to form an input for the fault classifier and/or the part failure classifier. Suppressing the dimension of the monitored data set may be performed when calculating the current system state representation, and the calculated current system state representation may be combined with the monitored data set for the given iteration and the replacement value to form the input.

The mapping of the SOM may be learned using a training data set. The method may further comprise: generating a visual representation of the SOM based on the training data set and the fault classifier and/or part failure classifier, wherein the visual SOM representation comprises regions shaded based on the probability of the system experiencing a fault and/or the probabilities of one or more parts experiencing a fault; and displaying the visual SOM representation via a user interface.

There is also described herein a computer-implemented method for identifying one or more parts of a monitored environmental control system to be replaced, the method comprising the steps of: receiving a current monitored data set comprising a set of features relating to the environmental control system, including one or more features based on sensor data obtained from one or more sensors of the environmental control system; obtaining a system state representation encoding a history of operational states of the system corresponding to previous monitored data sets, wherein the system state representation has a lower dimensionality than the current monitored data set; performing a part failure prediction comprising: inputting the current monitored data set and the system state representation into a part failure classifier, the part failure classifier arranged to output based on the inputs a replacement part indication indicating one or more parts of the environmental control system expected to experience a fault in the future; and outputting the replacement part indication.

The method may further comprise receiving an indication that a future fault of the environmental control system is likely, and performing the part failure prediction in response to receiving the fault indication. The fault indication may be generated based on applying a fault prediction model (e.g. as set out above).

Obtaining the system state representation may comprise: obtaining a previous system state representation indicative of a previous operational state of the environmental control system; and calculating, based on the current monitored data set and the previous system state representation, a current system state representation indicative of a current operational state of the environmental control system.

The fault indication may include a probability of a fault occurring within a future time period and the part failure prediction may be performed based on the probability of a fault occurring being above a threshold fault probability. The replacement part indication may indicate for each of a plurality of parts (e.g. a predetermined set of parts of the environmental control system) whether the part is expected to experience a fault within a future time period.

The length of the future time period may be up to 100 days. The length of the future time period may be around 60 days. The replacement part indication may comprise for each of the plurality of parts a probability of the respective part experiencing a fault or needing to be replaced within the future time period. The plurality of parts may include one or more parts of a boiler.

The plurality of parts may include one or more of: a circulation pump, a thermostatic radiator valve, an air vent, a heat exchanger, a flow sensor, a condensate pump, an expansion valve, a diverter valve and/or a pressure relief valve.

Outputting the replacement part indication may comprise: aggregating the replacement part indication with one or more replacement part indications corresponding to one or more other monitored environmental control systems to form an aggregated replacement part indication; and outputting the aggregated replacement part indication. The aggregated replacement part indication may comprise a list of parts for stocking a vehicle to service the environmental control systems. This can improve the efficiency of repairing multiple systems.

The list may indicate a total number of each part identified in the aggregated replacement part indications. The method may further comprise any of the further steps or features (or any combination thereof) of the method set out above or of any of the methods described herein.

Any of the methods described herein may be applied to a plurality of environmental control systems (e.g. associated with respective premises/building environments), to receive and process monitoring data from a plurality of such systems and perform system fault prediction and/or component part failure predictions(s) for the plurality of systems and possibly generate alerts for systems where faults/part failures are predicted.

There is also described herein apparatus for monitoring a monitored environmental control system, the system comprising: one or more sensors for obtaining sensor data from the monitored environmental control system; a server comprising an external data source; one or more processors; and a memory storing instructions which, when executed by the one or more processors, cause the apparatus to perform any of the methods described herein.

Also described is a system for monitoring one or more environmental control systems, the system having means, optionally including a processor with associated memory, for performing any method as set out herein, the system optionally further including one or more sensors for obtaining sensor data from the one or more monitored environmental control systems and/or a database for storing premises data relating to the one or more environmental control systems and/or one or more associated building environments.

There are also described herein a computer program, computer program product and non-transient computer readable medium comprising software code adapted, when executed by a data processing system, to perform any method as described herein.

Any system feature as described herein may also be provided as a method feature, and vice versa. As used herein, means plus function features may be expressed alternatively in terms of their corresponding structure.

Any feature in one aspect of the invention may be applied to other aspects of the invention, in any appropriate combination. In particular, method aspects may be applied to system aspects, and vice versa. Furthermore, any, some and/or all features in one aspect can be applied to any, some and/or all features in any other aspect, in any appropriate combination.

It should also be appreciated that particular combinations of the various features described and defined in any aspects of the invention can be implemented and/or supplied and/or used independently.

BRIEF DESCRIPTION OF THE FIGURES

Methods and systems for monitoring an environmental control system are described by way of example only, in relation to the Figures, wherein: Figures 1a-1b show (a) a diagram of a monitoring system and (b) a diagram of a process for predicting failure using the monitoring system; Figure 2a shows a diagram of an example method for predicting failure in a monitored environmental control system; Figure 2b shows a diagram of an example method for predicting a required replacement part in a monitored environmental control system; Figure 2c shows a diagram of an example method for predicting failure and a required replacement part in a monitored environmental control system; Figure 3a-3b illustrate (a) a prediction model for a boiler and (b) an example trained self-organising map; Figures 3c-3e show plots of heating behaviour for six example heating systems; Figure 3f shows versions of the self-organising map of Figure 3b for verifying replacement part prediction; Figure 4 illustrates an implementation of a prediction model; Figures 5a-5c show screenshots of an example user interface for managing an environmental control system; and Figure 6 shows a diagram of an example server device.

DETAILED DESCRIPTION

Figure la illustrates an example monitoring system 100 for a boiler 104. The boiler 104 is connected to a heating system or environmental control system requiring heated water. The boiler 104 receives water from a domestic cold water supply 106 and produces a domestic hot water (DHVV) output 105. The boiler is also connected to the heating system (e.g. a central heating system comprising a series of radiators) through supply and return pipes 108, 109. The pipe network in the heating system forms a closed-loop water circulation system. The skilled person would understand that variations in the boiler design and operation are possible, including further features and boilers which do not feed radiators directly. For instance, different boiler types may have a hot water output 105 which feeds directly to domestic supply outlets or may feed into a hot water storage tank from which outlets are then supplied.

The boiler 104 itself includes a plurality of components for heating and/or circulating water in the circulation system. For example, the boiler 104 can include one or more pumps (e.g. circulation pump and/or condensate pump), one more valves (e.g. thermostatic radiator valve (TRV), expansion valve, diverter valve and/or pressure relief valve), an air vent, a heat exchanger and/or a flow sensor.

The monitoring system of Figure la includes sensors attached to various system components, including heating supply flow temperature sensor 111 and heating return flow temperature sensor 112. Heating supply flow temperature sensor 111 obtains measurements indicative of the temperature of heated water being supplied into the circulation system via supply pipe 108 after heating by the boiler, referred to herein as the "supply flow temperature". Heating return temperature sensor 112 obtains measurements indicative (during normal operation) of the temperature of water returning to the boiler for heating via return flow pipe 109 after passing through the heating system (e.g. heating pipes and radiators), referred to herein as the "return flow temperature".

The supply and return flow temperature sensors 111, 112 may measure the temperature of the water itself or may provide indirect measurements. In a preferred embodiment, to avoid the need for invasive sensors, the sensors are installed externally, being placed against the pipes so as to measure the temperatures of the exterior pipe surfaces, which are used as representative temperatures, since the pipe surface temperatures are expected to track the temperature of the water flowing within. In this approach, sensor 111 may thus be affixed to an exterior of the supply flow pipe 108 and sensor 112 may be attached to the exterior of the return flow pipe 109.

Similarly, hot water sensor 110 associated or attached to hot water output pipe 105 may be a temperature sensor indicating flow of domestic hot water from the boiler 104 e.g. to taps, showers and/or a hot water tank. The sensors 111, 112, 110 may alternatively be other sensors capable of providing indications of temperature and/or flow by measuring flow or related phenomena. For example, the sensors can include a TRV temperature sensor and/or a pressure sensor for measuring the pressure within a pipe (of the boiler). Additionally or alternatively, an electrical energy sensor 114 may be attached to or associated with a power connection of the boiler 104 and is configured to detect the amount of electrical energy being used by the boiler. The electrical energy sensor 114 may be a current sensor or voltage sensor, or other suitable means for measuring electrical energy.

One or more of the sensors may be integral to the boiler and communicate with a controller 102. Additionally or alternatively, one or more of the sensors are separate from the boiler (e.g. they are retrofitted to pipes/connections of the boiler). The controller 102 may be a microcontroller or other control device. The controller is connected to the sensors by a connection means 113 which may be wired (e.g. a wire) or wireless (e.g. Wi-Fi TM Bluetooth TM or Zigbee Tm) for each sensor. In typical embodiments a wired connection is used to allow the controller 102 to power the sensors. In some examples, the controller 102 is provided within the boiler. The controller communications with one or more server(s) 101 to provide sensor data to the servers and/or receive control signals from the server(s).

In the example of Figure 1a, the system 100 also includes a smart meter 115 and a thermostat 116 configured to communicated with the server(s) 101 and provide further sensor data. For example, the smart meter 115 is connected (via a wired or wireless connection) to a gas consumption sensor 117 or meter which is configured to measure a gas consumption (e.g. a total volume consumed or volumetric flow) for the property/premises associated with the system 100. This gas consumption data can be indicative of the amount of gas being used by the boiler. The electrical energy sensor 114 and/or the gas consumption sensor 117 may be referred to herein collectively as energy consumption sensors. In some examples, the smart meter 115 communicates with both the gas consumption sensor 117 and the electrical energy sensor 114 (or an electrical energy meter associated with the whole property) to obtain energy consumption data.

The thermostat 116 is connected to or has included therein a temperature sensor (not shown) for measuring a temperature of the environment controlled by the system 100. For example, the thermostat 116 can be wirelessly connected to one or more thermometers/temperature sensors which provide temperature readings from one or more locations at the premises/property, such as indoor and/or outdoor temperature(s). In other examples, depending on the type of environment being controlled, the system can include other types of sensors for measuring different environmental characteristics, such as humidity. The thermostat may be a network-connected smart thermostat.

According to one example, the controller 102 communicates with the sensors 110, 111, 112 via wired connections in order to obtain sensor readings and other boiler parameters, while the smart meter 115 communicates with the energy consumption sensors (e.g. electricity and/or gas meters) in order to obtain energy consumption data and the thermostat 116 is configured to provide thermostat readings (e.g. temperature sensor readings and temperature setpoints).

The controller 102, smart meter 115 and/or thermostat 116 are preferably linked via wired or wireless connection to communicate with the servers 101 (either directly or through one or more intermediate devices and/or networks e.g. the Internet). The controller 102 is configured to relay sensor data from the boiler sensors to the one or more servers 101. The smart meter 115 is configured to relay energy consumption data from energy consumption sensors to the one or more servers 101, and the thermostat 116 is configured to relay thermostat readings (e.g. temperature sensor data and/or temperature setpoint data) to the server(s). The server may also be in communication with a user device 103, e.g. via the Internet, although it is also possible for a user device to connect on a local area network (including to the controller/local device(s) or an intermediary device such as a hub). The user device 103 (for instance a personal electronic device such as a mobile telephone or smartphone, or other user interactive device) may allow a user to interact with the controller, local device(s) or server and/or to send instructions to the controller, local device(s) or server and/or to receive notifications from the controller, local device(s) or server, for instance regarding a diagnostic problem or boiler fault. The notification may be by text message or notification in an application on the user device 103. The notification may prompt the user to take an action to control the boiler or may inform the user of an action already taken by a boiler controller. Alternatively, the server may control, or send instructions to, the boiler directly based on a fault determination.

The server(s) 101 are configured to store data received from the controller 102, smart meter 115, thermostat 116 and/or the user device 103. The controller 102, smart meter 115 and/or thermostat 116 are configured to collect measurements from the various sensors and send them to the server(s) 101, where the data is processed by a diagnostic algorithm. The controller 102 and/or thermostat 116 may also send other operational data (or control data) to the server, such as setpoint data or a demand signal used to control the boiler (or processed/summarised data describing the demand signal, e.g. specifying ON/OFF transitions). For example, the server 101 can receive setpoint data relating to one or more controlled environmental characteristics (e.g. temperature or humidity) from the thermostat 116. When the server detects a fault based on the supplied data, the user is notified, e.g. via an application on the user device.

In some examples, the server(s) 101 are also configured to communicate with one or more external data sources 120 (which themselves may be implemented on one or more remote servers) via a communications network (e.g. the internet). The server 101 can obtain from the external data source 120 data which is indicative of a type and/or characteristic of the premises associated with the boiler 104. For example, external data obtained from the data source 120 can include a property type which the boiler 104 serves (e.g. "Bungalow", "Flat", "Maisonette") and/or a characteristic of the property (e.g. number of bedrooms, number/type of windows, amount/type of insulation). Such data may also be referred to herein as "premises data".

Figure lb illustrates a process 150 for predicting the failure of a monitored environmental control system (e.g. a boiler) using the monitoring system 100. In this example, household data 154 are obtained from various sources 152. The sources 152 include a smart meter (e.g. smart meter 115) which provides energy (e.g. gas) consumption data, a sensor kit (e.g. including sensors 110, 111, 112 and/or 114) which provides boiler parameters and other sensor readings, a thermostat kit (e.g. thermostat 116) which provides thermostat readings, and a web site which can be used by a user to provide property details (e.g. premises data).

These various types of household data are then transmitted to the one or more servers 101 of the system 100 to be processed. In particular, the household data 154 are aggregated by a readings aggregator 156, which is implemented using one or more processors at the server(s). This aggregation can include formatting the household data into a predetermined form for input to a predictor model. In some examples, the readings aggregator 156 can additionally determine features based on the raw household data, which can then be used as input for a predictor model.

The aggregated readings are then sent to a predictor 158 which evaluates the likelihood of the control system failing in the future. If the predictor 158 determines that there will be a failure, various output actions 160 can be taken. For example, as shown in Figure 1 b, based on this determination by the predictor, a user can be contacted to schedule a maintenance appointment, and a list of replacement parts can be generated and used to stock a vehicle appropriately, allowing the control system (e.g. boiler) to be fixed pre-emptively.

Figure 2a illustrates a method 200 for predicting failure in a monitored environmental control system. In preferred examples, the method 200 is implemented using the system 100 of Figure la, and the monitored environmental control system is (or includes) a boiler, such as the boiler 104 described above. The method 200 is preferably implemented as part of the process 150 of Figure lb. In the following examples, the method 200 is performed by one or more servers (e.g. servers 101) in communication with the monitored environmental control system (e.g. via a controller 102).

The method begins at step 205 by receiving a monitored data set. The monitored data set includes a set of features relating to the control system (e.g. boiler), including features which are determined based on sensor data (e.g. from sensors 110, 111, 112, 114) and premises data indicative of characteristics of the premises (e.g. building) associated with the control system. In other examples, the features of the monitored data set may be based on sensor data alone, since the premises data may not vary over time. Note "features" refer to data values used as input to the classification models and can comprise raw sensor or other data (i.e. as received, without processing) and/or can include derived data values computed from the raw data or other data. Examples of derived classification features could include average temperature values or other aggregated sensor data, a temperature difference between boiler inlet and outlet computed from the raw inlet/outlet sensor readings, normalised values etc.) Sensor data can include one or more of: an internal/interior temperature; a target or setpoint temperature; a temperature on the flow pipe of the boiler; a temperature on the return pipe of the boiler; a temperature of one or more TRVs; a pressure inside the flow pipe and/or return pipe of the boiler; energy consumption, such as an electricity consumption (e.g. current or power consumption, number of current spikes) or a gas consumption (e.g. volumetric, power or energy consumption) from the boiler or a smart meter.

Premises data, which may also be referred to as building data, household data or environmental condition data, for a heating system is indicative of the thermal behaviour/properties of the premises (e.g. building) associated with the heating system. Premises data may be constant over time, or at least constant over the course of the monitoring time period. The premises data can include one or more of: a property type (e.g. terraced house, bungalow, flat, maisonette); a size of the property (e.g. a footprint area and/or a number of storeys of a house); a number of rooms (e.g. number of bedrooms); a number or type of windows (e.g. single, double glazing); an orientation of the property and/or of the windows; or an amount or type of insulation.

Additionally or alternatively, the data set can include other types of data, such as maintenance records (e.g. an amount of time since the last servicing or maintenance, details of any previous repairs/diagnoses, an average length of time between maintenance appointments etc.) and/or failure history (e.g. an amount of time since the last failure of the system, the type of failure, which parts were responsible for the failure/needed replacing, an average length of time between failures etc.).

At step 210, the monitored data set is pre-processed, specifically by being used to determine a current system state representation. The current system state representation is indicative of the current operational state (e.g. health) of the environmental control system. This representation uses fewer dimensions than the sensor and premises data (e.g. 2 dimensions). Accordingly, determining the current system state representation at step 210 is equivalent to applying a dimensionality reduction to the data, examples of which are described in more detail below. The system state representation encodes, at each point in time, the sensor and/or premises data in the received data set.

In particular, step 210 involves calculating the current representation of the system state based on the monitored data set and a previous system state representation. The previous representation can be obtained from a previous iteration of step 210 or from an external source, e.g. a user or operator can provide a previous system state representation.

Accordingly, the system state representation is updated based on the most recent sensor data and/or premises data.

Following the pre-processing of the data at step 210, a failure prediction model is applied to the data in order to determine, at step 220, whether the monitored data set is indicative of a future failure of the environmental control system. The input to the model includes the monitored data set as well as the current system state representation. Advantageously, since the current system state representation is based on the previous system state representation, this enables the prediction to take into account the system state history as well as the latest sensor readings.

The failure prediction model or failure classifier is configured to relate features derivable from sensor data and premises data to future failures of an environmental control system. Specifically, based on a system state representation, the model is configured to produce an output which indicates whether the environmental control system (e.g. boiler) will fail within a future time window. The future time window for which this prediction is made has a predetermined size (length) of up to around 100 days, typically no more than around 80 days or around 60 days. The output of the failure prediction model can be a binary prediction and/or can include a probability of failure within the future time window. For example, if the model outputs a probability of boiler failure within the future time window, the model can additionally provide a binary output (a classification indicative of whether a boiler failure is predicted) based on a probability threshold -for instance, if the output probability of boiler failure is above 50%, then the model can output an affirmative determination that a failure will occur within the future time window. If at step 220 it is determined that a failure of the environmental control system will (or is likely to) occur within the future time window, an output is generated at step 230. The generated output includes alerting a user (e.g. via user device/app 103) of the predicted failure. In some examples, the generated output can include contacting a user to schedule a service appointment.

If at step 220, it is not determined that a failure of the environmental control system will (or is likely to) occur within the future time window, the method 200 returns to step 205 to gather a further monitored data set. Thus, steps 205, 210 and 220 are repeated over multiple monitoring intervals (e.g 10-minute or 30-minute intervals) to form an iterative loop. Accordingly, the method 200 can be used to continuously monitor the performance of an environmental control system and provide live predictions as to whether a failure of the system will occur within a rolling future time window in real-time. In the example method of Figure 2a, the failure prediction at step 220 is performed upon each iteration of the loop (e.g. at every 10-minute monitoring interval). However, in other examples the system state representation is updated every monitoring interval (by repeating steps 205 and 210) but the failure prediction is performed less often, for example only every 24 hours. In some examples, steps 205 and 210 may need to be repeated for a minimum number of iterations (or for a minimum period of time) before a prediction is made, e.g. for at least around 4000 iterations or around 1 month. This can allow the system state representation to stabilise and for the impact of initialising parameters on the accuracy of the failure prediction to be reduced. In some examples, the steps 205 and 210 are repeated to update the system state representation based on data sets corresponding to at least a month of sensor and/or premises data for a system (boiler) which has been running in a "healthy" or "nominal" state for heating and hot water purposes.

Equally, in other examples, step 205 may be repeated before moving to step 210 such that multiple monitored data sets are received and aggregated (e.g. so as to form a single data set of time series data) before the current system state representation is determined.

The monitored data set received at each iteration of step 205 can include features based on sensor and premises data which are time series data. Typically, premises and sensor data readings are taken around every 10 minutes, which approximately matches the monitoring interval defining how often the loop of Figure 2a is repeated. However, some or all of the sensor and premises data may be obtained with a different frequency/time interval between readings. In this case, the data can be processed (e.g. aggregated) before proceeding to step 210. For example, some types of sensor data may be available more often than the monitoring interval (i.e. the time resolution of some of the received data can be higher (less time between each new value) than the monitoring time interval). In that case, the raw data may be pre-processed to generate representative data for the monitoring interval (e.g. by computing average or cumulative values or selecting a most recent value).

For example, the monitored data set may include the average readings over the last minutes for the internal temperature, boiler flow pipe temperature, boiler return pipe temperature and boiler current consumption readings. Equally, the monitored data set can include the average readings over the last 30 minutes for current and gas consumption of a boiler. Alternatively, cumulative values may be calculated for current and gas consumption over the monitoring interval.

Some types of sensor data may be available at irregular times and/or at larger time intervals than the monitoring interval, for example readings from a smart meter may only be available every 30 minutes and thermostat readings/setpoints may be provided at irregular user-defined intervals. Accordingly, in these cases, the most recent values for such low-resolution or irregular data types are used in the monitored data set at step 205.

Figure 2b illustrates a method 250 for predicting a replacement part required for repair of a monitored environmental control system when a fault is predicted. The method 250 may be used in conjunction with method 200, for example as part of or following step 220. In particular, following the failure prediction at step 220, the method 250 may be performed in order to determine which parts of the system (e.g. boiler) will cause that predicted failure. In preferred examples, the method 250 is implemented using the system 100 of Figure la, and the monitored environmental control system is a boiler, such as the boiler 104 described above. In the following examples, the method 250 is performed by one or more servers (e.g. servers 101) in communication with the monitored environmental control system (e.g. via a controller 102).

The method 250 begins at step 255 by receiving an indication that a failure will occur in the system within a future time window. For example, this indication may be the output of a failure prediction model, such as that used in the method 200 described above.

At step 260, a determination is made as to which part(s) of the system will need replacing within the future time window. In particular, the determination at step 260 is made by applying a part prediction model (part classifier) to the same data used for failure prediction (according to the Figure 2a process). Thus, the input to the model includes at least the current system state representation indicative of the current operational state of the environmental control system (e.g. boiler), computed as set out above, as well as the monitored data set comprising a set of classification features based on sensor data and premises data.

The part prediction model is configured to relate features derivable from sensor data and premises data to future failures of specific parts or components of the environmental control system (or one or more parts of the system that will need replacing in the future). The model may be conditioned (e.g. during a training process) on the assumption that a failure of the system will occur during a future time window. In particular examples, the part prediction model is configured to produce an output indicative of one or more boiler parts that will fail during the future time window. The future time window for which this prediction is made has a predetermined size of up to around 100 days, typically no more than around 80 days or around days. The output of the part prediction model can be, for each of a defined set of parts of the control system (e.g. boiler), a binary prediction and/or can include a probability of failure. For example, if the model outputs a probability of failure within the future time window for a given boiler part, this can be converted to a binary output (indicative of whether failure of that part is predicted) based on a probability threshold -for instance, if the output probability of the part failure is above 50%, then the model can output an affirmative determination that the specific part will fail (or will need replacing) within the future time window.

The output of the model determined at step 260 is then used to generate a list of replacement parts at step 270. For example, where the model provides as output a probability for each part, the list may be generated by selecting the parts for which the probability of failure (or needing to be replaced) within the future time window is greater than 50% (while the various examples herein use 50% probability thresholds, these thresholds can be altered as needed).

The generated list of parts is stored at the server, and communicated to a user (e.g. via user device/app 103) at step 280.

The generated list can enable the user to schedule a service appointment, load a vehicle with identified replacement parts, and thus fix the system (e.g. boiler) before the predicted failure actually occurs. This can eliminate system downtime that occurs after a system failure. In some embodiments, fault predictions and/or generated lists of replacement parts from multiple systems can be aggregated (e.g. according to geographic location and/or failure type) in order to improve the efficiency of repairing multiple systems. In some examples, instead of communicating the list of parts to a user, the (aggregated) list of parts can be output to a part management service, for example to enable instructions for stocking a vehicle with the necessary replacement parts to be generated and communicated to a warehouse or other part storage location. Equally, the part management service may receive multiple lists of replacement parts (e.g. generated according to the method 250), and aggregate lists according to geographic location. Accordingly, the service can provide instructions to one or more vehicles associated with system management for a particular geographic region to stock a total number of each type of part, determined from the aggregated lists. This can allow a technician to travel from the premises of each system (e.g. household) to the next without having to return to a depot/warehouse to collect more replacement parts. Thus, the efficiency of the process for (pre-emptively) repairing environmental control systems can be improved.

Figure 2c illustrates an example method 200' for predicting the failure of an environmental control system and for predicting which parts of the system will need replacing in event of failure. Figure 2c demonstrates an implementation of how the methods of Figures 2a and 2b may be combined into a single prediction method. The method 200' is preferably implemented using the system 100 and logical arrangement 150 of Figures la-1b.

The method 200' begins at step 205' by collecting new sensor readings (sensor data) from one or more sensors of the control system. The sensor data are then aggregated at step 206' which includes adapting the data to the required time resolution (e.g. by aggregation as described above) and formatting the sensor data into a predetermined format (e.g. vector format) for subsequent processing. Then, at step 210', the previous system state representation is obtained by evaluating a previous position on a self-organising map (SOM), as explained in more detail below. This previous system state representation is then used to build at step 215' a state vector, including the previous system state representation and the current sensor and premises data. This state vector is used to determine a current system state representation according to the SOM. The current system state representation and the latest sensor and premises data are then used as input at step 220' for a failure prediction model. By evaluating the model based on these inputs, a failure prediction in the form of a percentage probability of expected failure of the control system is determined.

At step 230', the output failure probability is compared with a threshold probability of failure. In the example of Figure 2c, this threshold is 50% and if the output probability is less than or equal to this threshold, then the method returns to repeat steps 205'-230' based on new data. This preferably corresponds to the loop of method 200 illustrated in Figure 2a. If however, the output probability is greater than the threshold, then an indication is provided to a user that a failure of the environmental control system (e.g. boiler) will occur in the future (e.g. within the next 60 days). This also triggers a replacement part prediction at step 260', which includes providing the current system state representation and the latest sensor and premises data as input to a replacement part prediction model (part classifier), to determine which of a predetermined set of parts (e.g. boiler parts) will need replacing due to the predicted failure. Based on this prediction, a user can be contacted at step 265' in order to schedule an appointment to (pre-emptively) repair the environmental control system (e.g. boiler). As shown in Figure 1 b, in some examples a list of parts can be generated for stocking a technician's van efficiently in preparation for a repair appointment. Following this, the method returns to step 205' to continue monitoring the environmental control system.

As indicated by corresponding reference numerals, in some implementations the steps 205', 210', 220', 230' and 260' of Figure 2c can correspond to the steps 205, 210, 220, 230 and 260 respectively of Figures 2a-2b.

Figures 3a-3b illustrate an example model 300 for predicting the future failure of a boiler and for predicting which parts will need replacing. The model 300 may be used in the methods 200, 250, 200' described above.

The model 300 is configured to apply a dimensionality reduction, denoted as F(x), to a received data set. The data set x can include sensor data, 1, and premises data (or household data), h. The data 1, h are time series data as described above (though values of h may be essentially constant over long periods), and have a combined total of n dimensions (at each time step).

The dimensionality reduction maps the data set to a lower-dimension space in order to provide a representation of the state of the system at a given point in time.

In the example of Figure 3a, F (x) is implemented as a self-organising map (SOM). The SOM ultimately maps higher-dimension data (including the sensor/premises data) to a lower-dimension vector, preferably to a two-dimensional vector. In particular, the SOM is a model including a mesh of linked nodes (or cells) each with an associated weight vector. The mesh of nodes defines a two-dimensional plane, and any given heating system at a point in time can be mapped to a node by comparing the sensor/premises data with the node weights and selecting the node having the nearest weight vector (e.g. in a Euclidean sense). Due to the topology-preserving properties of an SOM, systems with similar operating behaviours are mapped to the same node or nearby nodes in the SOM. The SOM also enables non-linear modelling (classification) of the thermal properties and sensor data at the heating system premises (e.g. a building). As explained in more detail below, the mapping provided by the SOM (specifically the weights associated with each node) are learned during a training process.

Examples of a processing schema suitable for the structured data of the monitored data set can be found in "Fuzzy clustering of structured data: Some preliminary results," (G. Vettigli and A. Ciaramella, 2017 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), Naples, Italy, 2017, pp. 1-6, doi: 10.1109/FUZZ-IEEE.2017.8015648). Further details concerning implementation of SOMs can be found in "Self-organizing maps" (Kohonen T., 2001, Vol. 30. Springer Science & Business Media) and in "MiniSom: minimalistic and NumPybased implementation of the Self Organizing Map" (Vettigli G., 2018, hitps corn/ 4.19t.Glowitmitnin:sora).

In order to map a data set of sensor and/or premises data using the SOM, a state vector xt at a time step t (where t = 1,2, ... T) is defined recursively as: It x= ht

F

The function F returns the coordinates (a, b) of the cell of the SOM to which the system is mapped, and thus provides a representation of both the state of the system at a given point in time and (due to the recursive definition of x) how that state has progressed over time.

Using the SOM in this manner therefore encodes how the system has evolved over time but using far fewer dimensions (compared to the raw source data set), improving the computational efficiency of the model 300. In addition, using a two-dimensional SOM in particular aids visualisation of the system state.

When calculating the state vectors xt, the position on the SOM is initialised by selecting a position on the SOM, F(x0) = (i, j). The initialised representation F(x0) may be a nominal predetermined position on the SOM (e.g. the centre coordinates of the SOM, or at a position associated with a healthy system/low probability of failure) or selected randomly. The current state vector xT is then used to determine the current (or most recently updated) system state representation, F(xT), or in other words the SOM coordinates (a, b) to which the most recent system state is mapped.

Preferably, the example data processing scheme shown in Figure 3a represents a single iteration of the loop shown in Figure 2a. At each monitoring interval / iteration of the loop, the data set includes a single value for each feature (input dimension) corresponding to different types of sensor/premises data. As explained above, if the sensor/premises data have a higher time resolution than the monitoring interval, the value for each type of data may be based on multiple readings (e.g. by aggregating or pre-processing the readings). Accordingly, the data set has a total of rt dimensions corresponding to 71 different types of data / features.

An example trained SOM of 26x26 cells (each with associated high-dimension weight vectors) is shown in Figure 3b. The current system state representations F (x T) of six example heating systems a-f are shown on the SOM, and corresponding heating behaviour plots for each of these systems are illustrated in Figures 3c-3e. These plots show the measured indoor temperature, outdoor temperature and target temperature of the system over a time period of several days. The vertical dashed line on each plot indicates the "current" points in time Tat which the representations are calculated.

Systems a and b (shown in Figure 3c) are two systems which struggle to reach the target temperature and also exhibit frequent sharp drops in internal temperature. These systems are therefore very likely to fail within a future time window, so are assigned high probabilities of failure. In addition, the systems behave similarly, so are mapped to coordinates close to each other in the SOM. Accordingly, this region of the SOM can be associated with a higher probability of system failure.

Systems c and d (shown in Figure 3d) are two systems that can maintain the internal temperature near the target temperature accurately and perform short periods of heating in order to do so. Thus, these systems perform well so are assigned low probabilities of failure and are mapped to nearby locations in the SOM. Accordingly, this region of the SOM is associated with a lower probability of system failure.

Systems e and f (shown in Figure 3e) are two systems which exhibit a small number of long, discrete heating periods and frequently do not reach the target temperature (e.g. because it takes a long time for the system to heat the controlled environment). Therefore, these systems are assigned a medium probability of failure and are mapped to nearby locations in the SOM.

Once a system is mapped to a cell of the SOM, the coordinates of that cell (a, b) (the current system state representation) along with the most recent sensor data IT and premises data hr from the data set are used as input to a failure prediction model, g. The failure prediction model can be any type of statistical classifier, such as a random forest classifier or neural network, and provides a prediction yr as to whether the system will fail within a future time window. Accordingly, this prediction can be expressed as: yr = g(lr,hr,F(xr)). The output yr may be a binary value (e.g. 1 if the system will fail within a future time window from t = T to T + W, and otherwise 0) and/or may include a probability (of the system failing within the future time window t = T to T + W).

In some examples, a previous system state representation is used as input to the model,g. For example, a system state representation from an immediately preceding iteration of steps 205 and 210 may be used as input instead of an updated or current system state representation (determined based on the current sensor/premises data), along with the most recent sensor and/or premises data. Accordingly, in this case the prediction may be expressed as yr = g(lr,hr, F(xr_i)) = g (xr). In principle, the system state representation used as input to the model could be an even earlier state representation, computed during an earlier iteration (though likely at a cost of reduced accuracy of the predictions), or could be based on the values from multiple recent iterations (e.g. averaged), but for predictive accuracy using the most recently available system state representation (F(xT)) is typically preferred.

The cells of the SOM shown in Figure 3b are shaded according to the result of inputting the weight vector of each cell to the failure prediction model g. Accordingly, the shading is indicative of the approximate probability of failure for a system which is mapped to different regions of the SOM (although the actual probability of failure for a system will also depend on its sensor/premises data, in addition to its mapped cell). For example, the high-dimension current state vector xT of system a is closest to the weight vector of the SOM cell (5,19), so the system is mapped to that cell, meaning F(xT) = [1q1. This cell is shaded according to the (high) failure probability value obtained by inputting the weight vector for cell (5,19) to the failure prediction model g. This approach can enable the operational state or health of a heating system to be easily interpreted. For example, the shaded SOM cell to which the system is mapped (or a shaded version of the whole SOM) can be provided to a human user via a user interface, which can enable the user to interpret and/or manually verify the prediction of the failure prediction model (e.g. random forest).

If the failure prediction model indicates that a heating system is likely to fail within the future time window (e.g. if the output yr is greater than a threshold value such as 0.5 or 50%), the model 300 can proceed to apply a part prediction model, q. The part prediction model q can receive as input the most recent sensor data IT and premises data hr of the data set as well as the current system state representation F(xT) (SOM coordinates (a, b)). The failure prediction model can be any type of statistical classifier, such as a random forest classifier or neural network, and provides a prediction pr indicative of which parts of the heating system will need to be replaced within a future time window. Accordingly, this prediction can be expressed as: Pr = q (IT, hT,F(xT)). The output pr includes values (e.g. arranged as a vector) corresponding to different parts of the heating system. These values may be binary values (e.g. equal to 1 if a corresponding part will need replacing within a future time window from t = T to T + W, and otherwise 0) and/or may include a probability (of the corresponding part failing within the future time window t = T to T + W). For example, as shown in Figure 3a, the output of the part prediction model may have a value corresponding to the circulation pump of 0.7, indicative that there is a 70% probability that the circulation pump will need to be replaced within the future time window. The output of the part prediction model may be indicative of one or more parts that are responsible for the future system failure predicted by the failure prediction model g.

As described above for the failure prediction model g, in some examples an earlier system state representation could be used as input to q (e.g. F(xT_,)) . The cells of the SOM can also be shaded according to the part prediction model q to indicate the probability of needing to replace a given part of the heating system when a failure occurs in the future. For example, versions of the SOM are shown in Figure 3f shaded according to the probability of needing to replace (i) a circulation pump or (ii) one or more TRVs when a failure happens in the future for a system which is mapped to a given cell. In particular, the shading (probability) of each cell of the SOM is determined by inputting the learned weights of the cell to the part prediction model q and using the output corresponding to (i) the circulation pump or (ii) one or more TRVs. The darker the shading, the higher the probability of needing to replace the respective parts. Shading the SOM in this way can enable the probability of needing certain replacement parts to be easily interpreted. For example, once the system has been mapped to a cell of the SOM, a human user can use the shading (corresponding to one or more parts of the system) to better interpret/understand the prediction of the part prediction model. For instance, in the example of Figure 3f, the lower right region of the SOM is associated with a relatively high probability of needing a replacement circulation pump, whereas the upper left region is associated with a relatively high probability of needing to replace one or more TRVs. Therefore, different regions of the SOM can be associated with different types of parts that will need replacing. Accordingly, different regions of the SOM can be associated with different types or modes of failure. For example, the SOM shading of Figures 3b and 3f may be used to indicate (or verify) that a heating system which is mapped to an upper left region of the SOM (e.g. system a or b) is highly likely to fail within a future time window (e.g. the next 60 days), and that failure will likely require one or more new TRVs. For example, if this matches predictions made by the failure and part prediction models (e.g. communicated to the user in an alert), then the boiler of the heating system can be serviced and the one or more TRVs replaced pre-emptively. Accordingly, this preventative repair can provide improved overall operation of the heating system by avoiding the future failure.

Figure 4 shows an implementation of the model 300 according to one example. In this example, the sensor/premises data received from a household does not contain all of the types of data (features) on which the SOM, failure prediction model and part prediction model are trained. For example, as shown in Figure 4, data is received from a household which includes internal temperature, a temperature setpoint, the electricity (e.g. current) consumption of the boiler and the property type. However, the SOM and the prediction models are additionally trained using other features (expected data types), such as boiler flow pipe temperature data and the number of bedrooms in the property. Accordingly, certain expected data types are missing in the received data set. In response to receiving such a data set, the state vector xt is constructed by indicating the missing types of data in the sensor and premises data vectors (e.g. with a predetermined symbol or character). In order to determine the system state representation F (x), the components of the weight vectors (of the SOM cells) corresponding to the indicated missing data types are ignored or supressed, and only those weights for which data has been provided are used to determine the nearest cell for the SOM mapping. For example, as shown in Figure 4 only the (circled) weights in the weight vectors w1 to w, which correspond to the internal temperature, temperature setpoint, electricity consumption, property type and the system state representation are considered when determining which weight vector is closest to the state vector. Once a current system state representation (SOM coordinates (a, b)) has been determined, the elements of the corresponding weight vector w which correspond to the missing data types are used (along with the provided data types and current system state representation) as input to the failure and part prediction models g,q. Due to the fact that the SOM maps systems with similar behaviours near to each other, the weights for the missing data types that are used with the received data as input to the prediction models are therefore representative of the missing data types for similar systems. Thus, the SOM allows the model 300 to make predictions even if certain data types are not provided.

The failure prediction method according to an embodiment can be summarised as follows: 1) Consider t as current time-step and ingest a new set of sensor readings It 2) Evaluate the state of the boiler F(xt) at time t using the SOM, where xt is a state vector based on the sensor data It and the premises data ht at the current time step, and the system state F(xt_l) at the previous time step 3) Evaluate g(lt,ht,F(xt)) to see if there will be a failure, where g is the failure prediction model 4) If step 3 reports the possibility of a failure, also evaluate q(1,, ht, F(x,)) and raise an alert, where q is the parts prediction model 5) Go to step 1.

Training The training process for the models which predict failure in an environmental control system and predict which parts will need replacing can be summarised as follows: 1) Train a self-organising map (SOM) which encodes the states of systems at different points in time 2) Train a classifier to predict a future failure 3) Train a classifier to predict parts to be replaced in case of failure In particular, at step (1) a self-organising map (SOM) for providing a system state representation of an environmental control system (such as that described above) is trained using a training data set of sensor data and/or premises data. The training data set can include sensor and premises data from multiple systems, e.g. multiple households. For example, each of a number of households may provide sensor data and premises data over one or more training time periods. The SOM is then trained using unsupervised learning techniques to derive weight vectors for each cell/node of the SOM based on the received training data. In preferred examples, the SOM comprises a two-dimensional map having N nodes. The SOM is initialised by randomly selecting a weight vector % for each node. Then, the weights of each node are updated by repeating the following steps for each system (e.g. household) in the training data set, for each time step t = 1,2, ... , T of the data provided by each system, and for different learning rates k = 1, 2, ... K: * Construct xt = h, F (x,-1)) using sensor data it and premises data it, from the household (using randomly/predetermined initialised coordinates for F(x0)).

* Select the node i ("the winning node") whose weight vector wi is nearest to xt, defined as argmin, = lixt -WWII * Update the weights of each node n as % = w, + ri (k),8 (i) (xt -w").

In this example, ?? is a learning rate function yielding a vector with lower values for the elements corresponding to the representation of the previous node, and)6 is a neighbourhood function yielding a vector with higher values the nearer a node is to the selected node/. Using this approach, the weights of each node are determined. The training process at step 1 can be implemented (e.g. at a server processor) using an algorithm summarized by the following pseudo-code: Initialize a bidimensional SOM where each node has weights w, picked randomly; For each household sending data: For t = 1,2,...,T: For k = 1,2,...,K: For each node n in the map: Build x, -(1_, h_, 7(xL-1) with data from the household; Find the winning node i as argmini = I Ix Update the weights of each node n as wfl=w,-En(k)P( )(xt-w.); At steps (2) and (3) of the training process, models (e.g. random forest models, such as those described above) for predicting the future failure of an environmental control system and for predicting parts of such a system that will need replacing can be trained using the training data set described above. Training samples for training these classification models are constructed based on the training data set, and comprise the sensor and premises data at each time step of the training time periods as well as the corresponding system state representations F (xt) (determined using the trained SOM) of each training time period (for each household). The training data set additionally includes an indication as to when each system/household experienced a failure (e.g. a date and time of failure). For example, for each training sample (the data at each time step of the training time period) may have an associated (e.g. binary) label indicative of whether a boiler of the corresponding environmental control system failed within a time window W (e.g. 60 days) after that time step, and/or which parts of the system needed replacing within that window. The label for each training sample within a received training data set (for a given system/household) may be determined based on an indication in the training data set as to when one or more failures occurred (e.g. date/time of failure). The length of the time window W is the same for all of the households/systems in the training data set.

In some examples, the value of W may be determined from the training data set. For example, the value of W can be chosen (e.g. based on a number or frequency of failures of the environmental control systems) such that a predetermined number or proportion of the training samples are deemed to have failed within the time window W. The model(s) are then trained based on the indicated/determined value of W, so the predictions made by the models during inference are based on the same value of W. For example, if the label of each training sample indicates whether a system failure and/or specific system part failure(s) occurred within a period of 60 days (or up to 100 days) from the corresponding time step, then the model(s) trained on this basis would predict system failure and/or replacement parts within a future time window of the same length W. However, different models could be trained from the training data to provide predictions for different values of W. Once the training samples have been derived from the training data set, the classifier models g and q can be trained using standard techniques for training random forest classifiers as known to those skilled in the art.

Figures 5a-5c show an example user interface 500 for managing an environmental control system, such as a boiler. These figures illustrate example screenshots of a smartphone application in which the interface 500 is provided for monitoring operation of a boiler. In particular, Figure 5a shows a currently selected "Event Log" tab which lists alerts generated by the described methods for failure prediction. Here, the top entry shows an alert for a predicted issue. By selecting the alert, the user may be taken to a further interface providing information and instructions for validating the predicted issue and/or attempting to resolve the issue. As illustrated in Figure 5b, this may include a series of information messages, queries and instructions, requesting that the user carry out diagnostic/corrective steps and/or provide information. In some cases, these steps and interactions may lead to avoidance of the predicted issue/failure, in which case the process may terminate and the issue may be marked as "resolved".

If the issue cannot be resolved/prevented by the user, the interface may prompt the user to perform additional steps, such as creating a booking for a maintenance technician to visit the user's property and perform further diagnostics/pre-emptive repair. An example is illustrated in Figure 5c. For example, this interface may link to a further automated booking service provided as part of the application, or a separate booking application or web service, to allow a technician visit to be booked. A technician then attends the property to perform in situ diagnostics (e.g. to verify the system failure/replacement part predictions) and repair on the boiler as needed.

Prior to attending the property, the technician can use the output of the part prediction model to ensure that the correct parts are available and that they bring them to the property. The technician can also use the output of the part prediction model to only bring the parts which will be required, allowing for better use of space within their vehicle. This can enable more repair/diagnostic tasks to be performed without having to return to a depot or storage facility to pick up more parts.

While illustrated as a smartphone application, a similar interface could also be provided as a web application accessible via a web browser on any Internet-connected computer or device.

In some embodiments, instead of (or in addition to) notifying the user, a control signal may be transmitted to the boiler, controller or another system component in response to prediction of a future failure, for example to deactivate or change a mode of operation of the heating system/boiler (e.g. to prevent damage due to a leak or low pressure, or to extend the amount of time until the predicted failure occurs), or to reconfigure the system/boiler or take some other control action. As another example, the control signal could be to change the target temperature to see if the "health" of the system can be improved to reduce the probability of a failure with a future time window by alleviating the load on the boiler. Where system failure and the need to replace specific replacement parts of the system/boiler are predicted, automatic control actions (e.g. adjusting operating parameters or modes) may be taken for some types of parts, while other predicted replacement parts may only result in user notification as previously described. This can correspond to predictions of critical/non-critical failures and parts.

Figure 6 illustrates a server device 101 in more detail. The server includes one or more processors 602 together with volatile / random access memory 604 for storing temporary data and software code being executed. A network interface 606 is provided for communication with other system components (e.g. controller 102 and user device 103) over one or more networks 616 (e.g. Local and/or Wide Area Networks, including the Internet).

Persistent storage 608 (e.g. in the form of hard disk storage, optical storage and the like) persistently stores software for performing the described functions. This includes the prediction models, in particular the state representation (self-organising map) model 610 and a learning algorithm 611 (for training the SOM and random forest classifiers), along with a live prediction process 612 for real-time failure/part prediction using the learned SOM and random forest models. Persistent storage also stores data used by these processes such as a database 614 of sensor data, premises data and derived feature data and training data sets. The persistent storage also includes other server software and data (not shown), such as a server operating system.

The server will include other conventional hardware and software components as known to those skilled in the art, and the components are interconnected by a data bus (this may in practice consist of several distinct buses such as a memory bus and I/O bus).

User device 103, e.g. in the form of a smartphone or other personal computing/communications device, connects to the central server via the network and may run an application 618 for boiler monitoring/management (as described in relation to Figures 5a to 5c). User device 103 is representative of a population of such devices associated with different households/customers and controller 102 is representative of a population of such controllers associated with different households/heating system installations. Sensors 620 correspond, for example, to the various sensors shown in Figure la, for measurement of various operating characteristics of heating systems.

Sensor data (including operational data such as demand signals) is received from controllers 102 of multiple households/heating system installations and stored in the database 614. Premises data is received from the controllers 102 or from one or more databases 622 at the external data source(s) of Figure la (e.g. via a web service), and stored in the database 614. The stored sensor data and premises data are processed to produce training samples for the learning algorithms, the samples including sets of feature values derived from the raw sensor data. The training samples are input to the state representation model 610 and learning algorithm 611 to obtain one or more trained random forest classifiers. Further sensor data and premises data (which may be from the same population of boilers or from a different population of boilers) are then received and processed in the same way to generate samples of monitored data sets. The live prediction process 612 applies the trained random forest classifiers to the sample data sets to predict future failures and/or replacement parts. Alerts/notifications are sent to user devices and/or other system components when a failure is detected and/or when one or more replacement parts are identified.

The prediction process 612 may operate essentially in real-time (to monitor sensor data and premises data as it is received from the controllers) or alternatively, prediction may be performed periodically on collected sensor and premises data. For example, the failure prediction may be run once a day on data collected for a particular boiler over the past 24 hours.

While a specific system is shown, any appropriate hardware/software architecture may be employed. For example, external communication may be via a wired or wireless network connection, or the system may include a different selection or arrangement of sensors, controllers and/or servers.

The above embodiments and examples are to be understood as illustrative examples.

Further embodiments, aspects or examples are envisaged. It is to be understood that any feature described in relation to any one embodiment, aspect or example may be used alone, or in combination with other features described, and may also be used in combination with one or more features of any other of the embodiments, aspects or examples, or any combination of any other of the embodiments, aspects or examples. Furthermore, equivalents and modifications not described above may also be employed without departing from the scope of the invention, which is defined in the accompanying claims.

Claims (27)

1. CLAIMS1. A computer-implemented method for predicting a fault in a monitored environmental control system, the method comprising the steps of: performing an iterative monitoring process comprising, at each of a plurality of iterations: receiving a monitored data set comprising a set of features relating to the environmental control system, including one or more features based on sensor data obtained from one or more sensors of the environmental control system; and computing system state data based on the monitored data set, wherein computing system state data comprises: obtaining a previous system state representation indicative of an operational state of the environmental control system at a previous iteration; and calculating, based on the monitored data set and the previous system state representation, a current system state representation indicative of an operational state of the environmental control system at the current iteration, wherein the system state representations each have a lower dimensionality than the monitored data set; the method further comprising performing a fault prediction comprising: providing a set of inputs to a fault classifier, the inputs based on the monitored data set for a given iteration and system state data computed by the monitoring process, wherein the fault classifier is arranged to output based on the inputs a fault indication indicating whether a future fault of the environmental control system is expected to occur; and outputting the fault indication.
2. 2. The method of claim 1, wherein performing the fault prediction comprises providing as an input to the fault classifier the system state representation calculated for the given iteration or an earlier, preferably immediately preceding, iteration.
3. 3. The method of claim 1 or 2, further comprising performing the fault prediction repeatedly for respective iterations, optionally for each iteration.
4. 4. The method of any preceding claim, wherein the system state representations are two-dimensional representations.
5. 5. The method of any preceding claim, wherein calculating the current system state representation comprises applying a mapping to the monitored data set and the previous system state representation, optionally wherein the mapping is based on a self-organising map, SOM.
6. 6. The method of any preceding claim, wherein the fault indication indicates whether a fault of the environmental control system is expected to occur within a future time period, optionally wherein the length of the future time period is up to 100 days, preferably around 60 days.
7. 7. The method of claim 6, wherein the fault indication comprises a probability of a fault occurring within the future time period.
8. 8. The method of any preceding claim, further comprising performing a part failure prediction comprising: inputting the monitored data set and system state data computed by the monitoring process, preferably the current system state representation, into a part failure classifier, the part failure classifier arranged to output based on the inputs a replacement part indication indicating one or more parts of the environmental control system expected to experience a fault in the future; and outputting the replacement part indication.
9. 9. The method of claim 8, wherein the part failure prediction is performed in response to the fault indication indicating that a future fault of the environmental control system is likely to occur, preferably based on a probability of a fault occurring being above a threshold fault probability.
10. 10. The method of claim 8 or 9, wherein the replacement part indication indicates one or more parts of the environmental control system expected to experience a fault within a future time period, optionally wherein the length of the future time period is up to 100 days, preferably around 60 days; optionally wherein the replacement part indication comprises for each of a predetermined set of parts a probability of the respective part experiencing a fault or needing to be replaced within the future time period.
11. 11. The method of any of claims 8 to 10, wherein the one or more parts include one or more parts of a boiler, preferably one or more of: a circulation pump, a thermostatic radiator valve, an air vent, a heat exchanger, a flow sensor, a condensate pump, an expansion valve, a diverter valve and/or a pressure relief valve.
12. 12. The method of any preceding claim, wherein the sensor data comprises time series data obtained from one or more sensors of the environmental control system, optionally wherein the time series data includes sensor readings from multiple points in time.
13. 13. The method of any preceding claim, wherein the features and/or the sensor data comprise one or more of: an internal temperature of a building environment served by the environmental control system, an external temperature of the building environment, a target temperature of the building environment, a temperature on the flow pipe of a boiler, a temperature on the return pipe of a boiler, an electrical power consumption of a boiler, and/or a combustible fuel consumption of a boiler.
14. 14. The method of any preceding claim, wherein the one or more sensors of the environmental control system comprise one or more of: a temperature sensor, a water flow sensor, a thermostat, a smart meter, an electrical energy sensor and/or a gas consumption sensor.
15. 15. The method of any preceding claim, wherein the set of features of the monitored data set includes features based on premises data indicative of properties of a building environment served by the environmental control system, preferably comprising data indicative of the thermal behaviour of the building environment, optionally comprising one or more of: a type of building, a number of rooms, a number or type of windows, and/or an amount or type of insulation.
16. 16. The method of any preceding claim, wherein the fault classifier and/or the part failure classifier comprise a decision tree or random forest classifier.
17. 17. The method of any preceding claim, wherein the fault classifier and/or the part failure classifier are trained using a training data set, wherein the monitoring process further comprises, for a given iteration: identifying a type of data in the training data set that is not present in the monitored data set; suppressing, when computing the system state data, a dimension of the monitored data set corresponding to the identified type of data; and identifying a replacement value corresponding to the identified type of data, wherein the replacement value is learned based on the training data set; wherein the method further comprises combining the monitored data set for the given iteration, system state data computed by the monitoring process and the replacement value to form an input for the fault classifier and/or the part failure classifier.
18. 18. The method of claim 5 or any claim dependent thereon, wherein the mapping of the SOM is learned using a training data set, and wherein the method further comprises: generating a visual representation of the SOM based on the training data set and the fault classifier and/or part failure classifier, wherein the visual SOM representation comprises regions shaded based on the probability of the system experiencing a fault and/or the probabilities of one or more parts experiencing a fault; and displaying the visual SOM representation via a user interface.
19. 19. A computer-implemented method for identifying one or more parts of a monitored environmental control system to be replaced, the method comprising the steps of: receiving a current monitored data set comprising a set of features relating to the environmental control system, including one or more features based on sensor data obtained from one or more sensors of the environmental control system; obtaining a system state representation encoding a history of operational states of the system corresponding to previous monitored data sets, wherein the system state representation has a lower dimensionality than the current monitored data set; performing a part failure prediction comprising: inputting the current monitored data set and the system state representation into a part failure classifier, the part failure classifier arranged to output based on the inputs a replacement part indication indicating one or more parts of the environmental control system expected to experience a fault in the future; and outputting the replacement part indication.
20. 20. The method of claim 19, further comprising: receiving an indication that a future fault of the environmental control system is likely, optionally wherein the fault indication is generated based on applying a fault prediction model; and performing the part failure prediction in response to receiving the fault indication.
21. 21. The method of claim 19 or 20, wherein obtaining the system state representation comprises: obtaining a previous system state representation indicative of a previous operational state of the environmental control system; and calculating, based on the current monitored data set and the previous system state representation, a current system state representation indicative of a current operational state of the environmental control system.
22. 22. The method of any of claims 19 to 21, wherein the fault indication includes a probability of a fault occurring within a future time period and the part failure prediction is performed based on the probability of a fault occurring being above a threshold fault probability; and/or wherein the replacement part indication indicates for each of a plurality of parts whether the part is expected to experience a fault within a future time period, optionally wherein the length of the future time period is up to 100 days, preferably around 60 days, optionally wherein the replacement part indication comprises for each of the plurality of parts a probability of the respective part experiencing a fault or needing to be replaced within the future time period; and/or wherein the plurality of parts includes one or more parts of a boiler, preferably one or more of: a circulation pump, a thermostatic radiator valve, an air vent, a heat exchanger, a flow sensor, a condensate pump, an expansion valve, a diverter valve and/or a pressure relief valve.
23. 23. The method of any of claims 19 to 22, wherein outputting the replacement part indication comprises: aggregating the replacement part indication with one or more replacement part indications corresponding to one or more other monitored environmental control systems to form an aggregated replacement part indication; and outputting the aggregated replacement part indication.
24. 24. The method of claim 23, wherein the aggregated replacement part indication comprises a list of parts for stocking a vehicle to service the environmental control systems, preferably wherein the list indicates a total number of each part identified in the aggregated replacement part indications.
25. 25. The method of any of claims 19 to 24, further comprising the further steps or features of any of claims 1 to 18.
26. 26. A system for monitoring an environmental control system, the system having means, optionally including a processor with associated memory, for performing a method according to any of claims 1 to 25, the system optionally further including one or more sensors for obtaining sensor data from the monitored environmental control system and/or a database for storing premises data relating to the environmental control system and/or an associated building environment.
27. 27. A computer program, computer program product or non-transient computer readable medium comprising software code adapted, when executed by a data processing system, to perform a method according to any of claims 1 to 25.