WO2025080250A1

WO2025080250A1 - System and method for sensor fault detection and mitigation in a machine or process

Info

Publication number: WO2025080250A1
Application number: PCT/US2023/035082
Authority: WO
Inventors: Weizhong Yan; Yiwei Fu; Guiju Song; Shanmuga-Priyan Subramanian; Tianyi Wang
Original assignee: General Electric Renovables Espana SL
Current assignee: General Electric Renovables Espana SL
Priority date: 2023-10-13
Filing date: 2023-10-13
Publication date: 2025-04-17
Anticipated expiration: 2026-04-13
Also published as: WO2025080250A8

Abstract

A system and method (200) for detecting, isolating, and accommodating a sensor fault in a machine or system define a group of the sensors (204) having a spatial-temporal sensor relationship. With a single neural network model (216) implemented by a controller, a measured value signal is received from each of the sensors (203). The value signal from a first sensor is masked and a predicted value is generated for the first sensor based on the spatial-temporal sensor relationships and the remaining value signals. The predicted value (218) is compared (220) to the received value signal from the first sensor and the first sensor is identified as faulty when a difference (222) between the predicted value and the received value signal exceeds a threshold value. For operations in the machine or system performed by the controller, the predicted value is substituted for the measured value signal from the first sensor when the first sensor is identified as faulty.

Description

SYSTEM AND METHOD FOR SENSOR FAULT DETECTION AND MITIGATION IN A MACHINE OR PROCESS

FIELD

[0001] The present disclosure relates in general to machine or process system controls that utilize sensor inputs, and more particularly to detection and correction of sensor faults in the machine or process system.

BACKGROUND

[0002] Sensor fault detection, isolation, and accommodation (FDIA) is a critical task in applications where sensors are used for collecting information to monitor and control performance of underlying systems, machines, or processes. Essentially any control system that automates a machine or industrial process relies on sensors for measurement and regulation of actuators. Sensors, however, experience a variety of failures, which results in measurement imprecision and unreliable process control. If sensor faults or failures are not identified in a timely manner, a variety of detrimental and cascading effects are likely, including plant or process shutdowns, safety risks, equipment failure, and so forth.

[0003] Wind turbines are an example of a system with related control processes that rely critically on accurate input from a variety of sensors for virtually every’ operational aspect of generating renewable energy from wind. The wind turbine industry is continuously seeking advancements in the sensor FDIA processes to improve the capacity, reliability, and safety of the w ind turbines, which is extremely important as these wind turbines grow’ in size and energy-production capability.

[0004] A number of methods have been developed or investigated for improved sensor FDIA. including model-based, data-driven, and hybrid approaches. Modelbased methods use physical models of the system and sensors to detect faults, but they require accurate modeling and may not be effective for all types of sensor faults. Data-driven methods use statistical techniques and machine learning algorithms to detect faults, but they may not be effective for fault isolation and accommodation. Hybrid methods combine model-based and data-driven approaches to improve effectiveness and efficiency. [0005] The most advanced data-driven method for sensor FDIA is generally known as the ‘“auto-associative neural network" (AANN) method, which uses artificial neural networks to learn sensor relationships and detect faulty sensors. AANN models are widely used in many fields for pattern recognition and signal validation purposes. AANN models have been used in wind turbine applications to learn (machinelearning) inter-sensor relationships and detect faulty sensors based on anomalies between actual sensor input values and predicted values generated from the learned relationships. The existing AANN models, however, are generally not effective for fault isolation and accommodation after detection of a sensor failure.

[0006] US Pat. No. 6,804,600 proposes a sensor error detection and compensation system and method for aircraft engines. The system includes an expected value generator and a sensor fault detector. The expected value generator is an AANN model and receives sensor data from a plurality of sensors in the turbine engine.

From the sensor data, the expected value generator generates an expected sensor value for a sensor under test. The expected sensor value is passed to the sensor fault detector, which compares the expected sensor value to a received sensor value to determine if a sensor error has occurred in the sensor under test. If an error has occurred, the error can be compensated for by generating a replacement sensor value to substitute for erroneous received sensor value. This system and method, however, are quite complex in that it relies on individual AANN models (expected value generators) for each sensor under test.

[0007] In view of the above, the present disclosure seeks to improve upon prior art AANN methods for performing sensor FDIA in machines or systems in general, and wind turbine systems in particular.

BRIEF DESCRIPTION

[0008] Aspects and advantages of the invention will be set forth in part in the following description, or may be obvious from the description, or may be learned through practice of the invention.

[0009] An embodiment of the invention is directed to a method for detecting, isolating, and accommodating a sensor fault in a machine or system that utilizes a plurality of sensors for carrying out an operation. The method has particular usefulness in a wind turbine that employs a number of sensors for detecting and measuring various parameters for operation of the wind turbine. Aspects of the present invention will be described herein with reference to a wind turbine for sake of example and ease of explanation. It should be appreciated, however, that the present methods (as associated system embodiments) are not limited to a wind turbine. The present invention is intended for any operation or control system that relies on sensor inputs.

[0010] The method includes defining a group of the sensors in the machine or system having a spatial-temporal sensor relationship such that a parameter detected and/or measured by each sensor (and thus the value signal generated by each sensor) is affected by a parameter detected and/or measured by at least one other sensor in the group within a relevant time interval. For example, a first sensor in the group may detect/ measure temperature of a component or a space, wherein the temperature also affects another parameter, such as power or current flow through a component that has been affected by the temperature change, that is detected/measured by another sensor.

[0011] The method utilizes a neural network model that is common to all of the sensors to perform the following functions: receive the value signal from each of the sensors; mask the received value signal from a first one of the sensors and generate a predicted value for the first sensor based on the spatial-temporal sensor relationships and the remaining value signals; compare the predicted value to the received value signal from the first sensor; identify the first sensor as faulty when a difference between the predicted value and the received value signal exceeds a threshold value; and for operations in the machine or system performed by a controller or control system, substitute the predicted value for the received value signal of the first sensor when the first sensor is identified or determined to be faulty.

[0012] The method may include repeating the functions for the sensors in the group individually or in groups on a continuous or periodic basis.

[0013] The single neural network may be any suitable neural network capable of handling time series data. In a particular embodiment, the single neural network model is an auto-associative neural network (AANN) model. The neural network model can be any neural network model capable of handling sensor data, including at least one of (i) a Neural Network (“NN’") model, (ii) a Feed-Forward Multilayer Perceptron ( “FF MLP"), (iii) a One Dimensional Convolutional Neural Network (‘ I D CNN⁷’), (iv) a Temporal Convolutional Network ("TCN”), (v) a Recurrent Neural Network (“RNN”), (vi) a Long Short-Term Memor /Gated Recurrent Unit (“LSTM/GRUGRU”), or (vii) a Transformer model.

[0014] The functions of comparing the predicted value to the received value signal and identifying the first sensor as faulty may include generating a residual value between the predicted value and the received value signal.

[0015] The method may further include training the neural network model in an offline phase by repeatedly and randomly masking one or more of the received value signals and computing the predicted value only for the masked value signals. This training is repeated a sufficient number of times with “good” (reliable) sensor data to drive a loss value betw een the masked value signals and the predicted values to a minimum based on the same spatial -temporal sensor relationships. The threshold value is based on the results of the training such that the threshold value is lower as the loss value approaches zero. In other words, the more accurate and complete the training is, the less the threshold value is in the actual sensor data analysis phase. [0016] In a particular embodiment, the spatial-temporal sensor relationships between the sensors in the group are such that a value signal from each sensor is affected by a respective different parameter measured by a plurality, or all, of the other sensors in the group.

[0017] The method is not limited to the number of sensors in the group. In and embodiment wherein the method is practiced in a wind turbine, the group may include at least two sensors selected from the group including: a wind speed sensor, a rotor speed sensor, a nacelle temperature sensor, a pitch angle sensor for each rotor blade, a tow er fore-aft acceleration sensor, and a tower side-side acceleration sensor.

[0018] The present invention also encompasses a machine or system, such as a wind turbine, having a control system and a plurality of sensors in communication with the control system, wherein each sensor measures a parameter used by the control system for operation of the wind turbine. The control system is configured to define a group of the sensors (w hich may be all of the sensors) having a spatial-temporal relationship such that a value signal from each sensor in the group is affected by the parameter measured by at least one other sensor in the group. The control system is also configured to support a neural network model that is common to all of the sensors (i. e. , a single neural network model) to: (a) receive the value signal from each of the sensors; (b) mask the received value signal from a first one of the sensors and generate a predicted value for the first sensor based on the spatial-temporal sensor relationships; (c) compare the predicted value to the received value signal from the first sensor; (d) identify the first sensor as faulty when a difference between the predicted value and the received value signal exceeds a threshold value; and (e) for operations in the machine or system performed by the control system, substitute the predicted value for the received value signal of the first sensor when the first sensor is identified as faulty. The controller outputs the predicted value to a component or process in the machine or system so that operation of the component or system is dependent on the predicted value.

[0019] The control system and neural network may be configured to perform any of the functions discussed above with respect to the method embodiments.

[0020] These and other features, aspects and advantages of the present invention will become better understood with reference to the following description and appended claims. The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and, together with the description, serve to explain the principles of the invention.

BRIEF DESCRIPTION OF THE DRAWINGS

[0021] A full and enabling disclosure of the present invention, including the best mode thereof, directed to one of ordinary skill in the art, is set forth in the specification, which makes reference to the appended figures, in which:

[0022] Fig. 1 depicts a wind farm with wind turbines configured in accordance with the present disclosure;

[0023] Fig. 2 depicts a wind turbine according to the present disclosure;

[0024] Fig. 3 is diagram of a controller configuration for wind turbines according to the present disclosure;

[0025] Fig. 4 is a diagram of a method for detecting a faulty sensor in accordance with aspects of the present disclosure; [0026] Fig. 5 is a diagram of the offline training phase associated with the neural network; and

[0027] Fig. 6 is a flow chart of a method for detecting, isolating, and accommodating sensor faults in accordance with aspects of the present disclosure. [0028] Repeat use of reference characters in the present specification and drawings is intended to represent the same or analogous features or elements in the invention.

DETAILED DESCRIPTION

[0029] Reference now will be made in detail to embodiments of the invention, one or more examples of which are illustrated in the drawings. Each example is provided by way of explanation of the invention, not limitation of the invention. In fact, it will be apparent to those skilled in the art that various modifications and variations can be made in the present invention without departing from the scope or spirit of the invention. For instance, features illustrated or described as part of one embodiment can be used with another embodiment to yield a still further embodiment. Thus, it is intended that the present invention covers such modifications and variations as come w ithin the scope of the appended claims and their equivalents.

[0030] The terms “coupled,” “fixed,” “attached to,” and the like refer to both direct coupling, fixing, or attaching, as well as indirect coupling, fixing, or attaching through one or more intermediate components or features, unless otherwise specified herein. [0031] Approximating language, as used herein throughout the specification and claims, is applied to modify any quantitative representation that could permissibly vary' without resulting in a change in the basic function to which it is related. Accordingly, a value modified by a term or terms, such as “about”, “approximately”, and “substantially”, are not to be limited to the precise value specified. In at least some instances, the approximating language may correspond to the precision of an instrument for measuring the value, or the precision of the methods or machines for constructing or manufacturing the components and/or systems. For example, the approximating language may refer to being within a 10 percent margin.

[0032] While examples herein refer to wind turbines for illustrative purposes, it is to be appreciated that the novel concepts disclosed herein can be employed for any type of system or machine application that utilize input from sensors to perform control operations.

[0033] The present disclosure is applicable to essentially any type of sensor. In general, sensors can be used to sense light, motion, acceleration, temperature, magnetic fields, gravitational forces, humidity, vibration, pressure, electrical fields, current, voltage, sound, and other suitable physical aspects of an environment through one or more sensing elements. Non-limiting examples of sensors can include acoustic sensors, vibration sensors, air data sensors (e g., air speed, altimeter, angle of attack sensor,), inertial sensors (e.g., gy roscope, accelerometer, inertial reference sensor), magnetic compass, navigation instrument sensor, electric current sensors, electric potential sensors, magnetic sensors, radio frequency sensors, fluid flow sensors, position, angle, displacement, distance, speed, (e.g., inclinometer, position sensor, rotary' encoder, rotary /linear variable differential transducers, tachometer, etc.), optical, light, imaging sensors (e.g., charge-coupled device, infra-red sensor, LED, fiber optic sensors, photodiode, phototransistors, photoelectric sensor, etc.), pressure sensors and gauges, strain gauges, torque sensors, force sensors piezoelectric sensors, density sensors, level sensors, thermal, heat, temperature sensors (e.g., heat flux sensor, thermometer, resistance-based temperature detector, thermistor, thermocouple, etc.), proximity/presence sensors, or any other suitable type of sensor.

[0034] Physical components can comprise any hardware component of a system or machine, non-limiting examples of which can include, gears, bearings, motors, generators, fittings, fans, joints, or any other suitable hardware component of a system or machine, such as the components of a wind turbine.

[0035] It is to be appreciated that sensor devices as described in examples herein can be located within a suitable proximity to a physical component to perform a primary' sensing function of the physical component, and can also be located within a suitable proximity to another physical component to perform a backup sensing function of the other physical component.

[0036] Generally, the present disclosure is directed to a machine-learning neural network model for detecting, isolating, and accommodating a sensor fault in a system (e.g., a wind turbine) that utilizes a plurality of sensors for carry ing out one or more operations. The neural network model uses supervised machine learning on sensor data together with various pre-processing steps and continuous learning to create an analytic tool capable of detecting faulty or erroneous sensor operation with minimal model complexity, while also isolating and accommodating the faulty sensor data. This model can be continuously improved over time through continuous learning techniques to capture spatial-temporal sensor relationships and may continuously add new analytics as they become available.

[0037] The present methods and systems can provide the following benefits: improved system performance and reduced maintenance costs by ensuring that faulty sensor measurements are detected and corrected quickly; improved safety by detecting faulty sensors in critical systems, such as in aerospace, energy, defense, and healthcare industries; reduced downtime and increased uptime of industrial assets and processes by providing timely and accurate sensor measurements; reduced need for manual sensor inspection and troubleshooting; integration into existing industrial control and automation systems, which can enhance their capabilities and provide more advanced functionality; and identify and diagnose underlying issues in industrial systems, which can lead to more targeted and effective maintenance and repairs.

[0038] As mentioned, aspects of the present disclosure are explained by reference to wind turbines for illustrative purposes only.

[0039] Referring now' to the drawings, Fig. 1 illustrates an exemplary embodiment of a wind farm 100 containing a plurality of wind turbines 102 that may be operated in accordance with aspects of the present disclosure. The w ind turbines 102 may be arranged in any suitable fashion. Typically, wind turbine arrangement in a wind farm is determined based on numerous optimization algorithms such that AEP (annual energy production) is maximized for corresponding site wind climate. It should be understood that any wind turbine arrangement may be implemented, such as on uneven land, w ithout departing from the scope of the present disclosure.

[0040] In addition, it should be understood that the w ind turbines 102 of the w ind farm 100 may have any suitable configuration, such as the embodiment shown in Fig. 2. As shown, the wind turbine 102 includes a tower 114 extending from a support surface, a nacelle 1 16 mounted atop the tower 1 14, and a rotor 118 coupled to the nacelle 1 . The rotor includes a rotatable hub 120 having a plurality of rotor blades 112 mounted thereon, which is, in turn, connected to a main rotor shaft that is coupled to the generator housed within the nacelle 116 (not shown). Thus, the generator produces electrical power from the rotational energy’ generated by the rotor 118. [0041] As shown generally in the figures, each wind turbine 102 of the wind farm 100 may also include a turbine controller 104 communicatively coupled to a farm controller 108. Moreover, in one embodiment, the farm controller 108 may be coupled to the turbine controllers 104 through a network 110 to facilitate communication between the various wind farm components. The wind turbines 102 may also include one or more sensors 105, 106, 107 configured to monitor various operating, wind, and/or loading conditions of the wind turbine 102. For instance, the one or more sensors may include blade sensors for monitoring the rotor blades 112; generator sensors for monitoring generator loads, torque, speed, acceleration and/or the power output of the generator; wind sensors 106 for monitoring the one or more wind conditions; shaft sensors for measuring loads of the rotor shaft and/or the rotational speed of the rotor shaft; temperature sensors for monitoring the temperature of a component or space. Additionally, the wind turbine 102 may include one or more tower sensors for measuring the loads transmitted through the tower 114 and/or the acceleration of the tower 114 in a fore/aft or side/side direction. In various embodiments, the sensors may be any one of or combination of the following: accelerometers, pressure sensors, angle of attack sensors, vibration sensors. Miniature Inertial Measurement Units (MIMUs). camera systems, fiber optic systems, anemometers, wind vanes, Sonic Detection and Ranging (SOD AR) sensors, infra lasers, Light Detecting and Ranging (LIDAR) sensors, radiometers, pitot tubes, rawinsondes, other optical sensors, and/or any other suitable sensors.

[0042] Referring now to FIG. 3, a block diagram is provided of an embodiment of suitable components that may be included within the farm controller 108, the turbine controller(s) 104, and/or other suitable controller according to the present disclosure. As shown, the controller(s) 104, 108 may include one or more processor(s) 150 and associated memory device(s) 152 configured to perform a variety of computer- implemented functions (e.g.. performing the methods, steps, calculations and the like and storing relevant data as disclosed herein). Additionally, the controller(s) 104, 108 may also include a communications module 154 to facilitate communications between the controller(s) 104, 108 and the various components of the wind turbine 102. Further, the communications module 154 may include a sensor interface 156 (e.g., one or more analog-to-digital converters) to permit signals transmitted from one or more sensors 105, 106. 107 (such as the sensors described herein) to be converted into signals that can be understood and processed by the processors 150. It should be appreciated that the sensors 105, 106, 107 may be communicatively coupled to the communications module 154 using any suitable means. For example, as shown, the sensors 105, 106. 107 are coupled to the sensor interface 156 via a wired connection. However, in other embodiments, the sensors 105, 106, 107 may be coupled to the sensor interface 156 via a wireless connection, such as by using any suitable wireless communications protocol know n in the art.

[0043] As used herein, the term "‘processor” refers not only to integrated circuits referred to in the art as being included in a computer, but also refers to a controller, a microcontroller, a microcomputer, a programmable logic controller (PLC), an application specific integrated circuit, and other programmable circuits. Additionally, the memory device(s) 152 may generally include memory element(s) including, but not limited to, computer readable medium (e.g.. random access memory (RAM)), computer readable non-volatile medium (e.g., a flash memory ), a floppy disk, a compact disc-read only memory⁷ (CD-ROM), a magneto-optical disk (MOD), a digital versatile disc (DVD) and/or other suitable memory⁷ elements. Such memory device(s) 152 may generally be configured to store suitable computer-readable instructions that, when implemented by the processor(s) 150, configure the controller(s) 104, 108 to perform various functions as described herein.

[0044] Moreover, the network 110 that couples the farm controller 108, the turbine controllers 104, and/or the wind sensors 106 in the wind farm 100 may include any known communication network such as a wired or wireless network, optical networks, and the like. In addition, the network 1 10 may be connected in any known topology, such as a ring, a bus, or hub, and may have any known contention resolution protocol without departing from the art. Thus, the netw ork 110 is configured to provide data communication between the turbine controller(s) 104 and the farm controller 108 in near real time.

[0045] Fig. 4 depicts a method 200 that incorporates aspects of the present disclosure. The method 200 is applicable to any ty pe of machine, system, or operation employing multiple sensors, such as a wind turbine. The method 200 includes defining a group 204 of the sensors 203 in the system having a spatial- temporal sensor relationship. Those skilled in the art of artificial intelligence (Al) understand that a spatial-temporal relationship represents how items or data fit together in a space. Spatial-temporal reasoning and databases host data that is collected across both space and time to describe a phenomenon or relationship in a particular location within a period of time. Spatial-temporal analysis is particularly well-suited to examiner patterns in data over time to detect anomalies. In the present use, the group of sensors are related spatially and temporally in that a measured value (MV) from each sensor 203 is affected by a parameter measured by at least one other sensor 203 in the group 204 within a relevant time period. For example, a first sensor

203 in the group 204 may detect/measure temperature of a component or a space, wherein the temperature also affects another parameter, such as power or current flow, that is detected/measured by another sensor 203 that has been affected within temporal proximity to the temperature change. It should be appreciated that the group

204 need not include all sensors in the overall system, machine, or operation, and that multiple groups 204 may be defined within a single system, machine, or operation. [0046] In the depicted embodiment of Fig. 4, the group 204 includes five sensors 203, with each sensor 203 generating a measured value signal MV1 through MV5, respectively. In this example, sensor 3 (highlighted) is a faulty sensor.

[0047] The method 200 utilizes a neural network model 216 that is common to all of the sensors to perform multiple functions. The measured value signals MV1-MV5 are received from the respective sensors 203 in the defined group 204. The group 208a indicated in Fig. 4 represents the first analysis of the group 204 wherein one of the received value signals (e.g., from sensor 1 in this example) is "masked”, as indicated at 210. The signal is “masked” in that the actual measured value of the signal is not transmitted downstream. The actual value signal MV1 may be blocked, overwritten, or substituted with a signal that presents a different value (or altogether different type of signal) as compared to the actual value signal MV1 measured by the sensor to generate the masked signal 210.

[0048] Still considering the first group 208a of Fig. 4, the masked signal value 210 for sensor 1 and the measured value signals MV2-MV5 for the remaining sensors 203 in the group are received by the neural network 216, which may, for example, an AANN neural network.

[0049] At 218 in Fig. 4, the neural network 216 generates a predicted value (PV1 ) for sensor 1 (the masked sensor in the group 208a) based on the measured value signals MV2-MV5 and the spatial-temporal sensor relationships between all of the sensors in the group 208a. This predicted value (PV1) for the masked sensor is then compared to the actual measured value MV 1 received from the first sensor. This comparison may be done, for example as indicated in Fig. 4, by generating a residual value (RV1) at the junction 220 as the difference between the predicted value (PV1) and the measured value (MV1) and storing the residual value (RV1) at 222.

[0050] A threshold value is established for the residual values particular to sensor 1 and, at 224, a comparison is made between the residual value (RV1) and the threshold value. If a difference between the predicted value (PV1) and the measured value (MV1) (represented as the residual value (RV1)) exceeds the threshold value, then the first sensor is identified as faulty and appropriate isolation and corrective actions are taken.

[0051] As depicted in Fig. 4, the method 200 may include repeating the functions discussed above for each of the remaining sensors 2-5 in the initial group 204 of sensors. For example, the group 208b represents the process wherein sensor 2 is the masked sensor 210b. the group 208c represents the process wherein sensor 3 is the masked sensor 210c, and so on, until all of the sensors 1-5 in the group 204 have been tested as described. This testing sequence may be conducted continuously by the controllers or on a timed periodic basis when the system/machine is operational or in a shutdown or standby mode.

[0052] Neural networks are known in the art. The single neural network model of this disclosure may be any conventional neural network capable of handling time series data, including at least one of (i) a Neural Network (“NN”) model, (ii) a Feed- Forward Multilayer Perceptron (“FF MLP”), (iii) a One Dimensional Convolutional Neural Network (“ID CNN”), (iv) a Temporal Convolutional Network ("TCN”). (v) a Recurrent Neural Network (“RNN”), (vi) a Long Short-Term Memory/Gated Recurrent Unit (“LSTM/GRUGRU”), or (vii) a Transformer model. [0053] In a particular embodiment, the neural network is embodied as an auto- associative neural network (AANN) model. As generally understood, an AANN model is a feedforward net trained to produce an approximation of the identity mapping between network inputs and outputs using backpropagation or similar learning procedures. A key feature of an AANN is a dimensional bottleneck between input and output. A typical AANN model is a five-layer perceptron feed-forward network that is divided into two neural networks of three layers each connected in series. The input and output vectors of an AANN are identical. An input side of the model includes an input layer, a mapping layer, and a bottleneck layer. The output side of the model includes the bottleneck layer, a de-mapping layer, and an output layer.

[0054] The method may include training the neural network model in an offline phase using self-supervised training methods. Referring to Fig. 5, the offline training phase of the neural network may include repeatedly and randomly masking one or more of the received spatial-temporal value signals (X_n(i)/ti) and computing the predicted value for the masked input only based on the spatial-temporal sensor relationships learned by the model. This training is conducted with a sufficient number of random masking of the input signals to ensure that the model is trained on all of the sensor inputs. The training is conducted on “good” of valid input sensor data and seeks to drive a loss value between the predicted values for the masked sensor signals to zero.

[0055] The threshold value against which the residual value is compared (as discussed above with respect to Fig. 4) is established based on the training of the neural network model. The threshold value is lower as the loss value in the training phase approaches zero.

[0056] The method 200 is not limited to the number of sensors in the group 204. For example, in the wind turbine embodiment, the group 204 may include at least two sensors selected from the group including: a wind speed sensor, a rotor speed sensor, a nacelle temperature sensor, a pitch angle sensor for each rotor blade, a tower fore-aft acceleration sensor, and a tower side-side acceleration sensor. Multiple groups 204 may be defined for each machine or system. [0057] Fig. 6 depicts a flowchart of a method embodiment 300 that is particularly configured for a wind turbine. Step 302 depicts a wind turbine having a plurality of sensors (l ... n).

[0058] Step 304 depicts defining a group of the sensors having a spatial-temporal sensor relationship, as discussed above. The wind turbine may have multiple such groups, and the number of sensors within each group may vary.

[0059] Step 306 depicts that each of the sensors in the group generates a measured value signal (MVi).

[0060] Step 308 depicts that at least one of the measured value signals in the group is masked by a blocked, substitute, or overwritten signal such that the actual measured value detected by the sensor is blocked from downstream processing. More than one measured signal may be blocked in this step.

[0061] Step 310 depicts that the neural network (e.g., an AANN model) receives the masked value signal(s) and the remaining measured value signals from the group of sensors.

[0062] Step 312 depicts that the neural network may be trained in an offline training process, as discussed above.

[0063] Step 314 depicts that the neural network generates a predicted value (PVi) for each masked value signal based on its learned spatial-temporal sensor relationships and the actual measured values from the non-masked sensors.

[0064] Step 316 depicts that a residual value (RVi) is computed for the masked sensor signal based on a difference between the predicted value (PVi) and the measured value (MVi) for the respective sensor.

[0065] At step 318, the residual value (RVi) for the masked sensor is compared to a threshold value.

[0066] At step 320, the residual value (RVi) is determined to exceed the threshold value, indicating that the respective sensor is faulty.

[0067] At step 322, the faulty sensor is isolated in the system at least to the extent that signals from the sensor are not used for operational purposes in the wind turbine. The sensor may be physically replaced or repaired during a subsequent maintenance procedure. [0068] At step 324, accommodation is made for the faulty sensor by substituting the predicted value (PVi) for the measured value (MVi) from the sensor in downstream control processes in the wind turbine.

[0069] Step 326 indicates that the residual value (RVi) is below the threshold value, thereby indicating that the sensor is functioning properly (within specifications).

[0070] Step 328 indicates that the testing process proceeds for each of the sensors in the initial group of sensors. The testing process can proceed continuously or according to a timed periodic schedule on a repetitive basis during operation of wind turbine or any other time.

[0071] The skilled artisan will recognize the interchangeability of various features from different embodiments. Similarly, the various method steps and features described, as well as other known equivalents for each such methods and feature, can be mixed and matched by one of ordinary skill in this art to construct additional systems and techniques in accordance with principles of this disclosure. Of course, it is to be understood that not necessarily all such objects or advantages described above may be achieved in accordance with any particular embodiment. Thus, for example, those skilled in the art will recognize that the systems and techniques described herein may be embodied or carried out in a manner that achieves or optimizes one advantage or group of advantages as taught herein without necessarily achieving other objects or advantages as may be taught or suggested herein.

[0072] This written description uses examples to disclose the invention, including the best mode, and also to enable any person skilled in the art to practice the invention, including making and using any devices or systems and performing any incorporated methods. The patentable scope of the invention is defined by the claims, and may include other examples that occur to those skilled in the art. Such other examples are intended to be within the scope of the claims if they include structural elements that do not differ from the literal language of the claims, or if they include equivalent structural elements with insubstantial differences from the literal languages of the claims.

[0073] Further aspects of the invention are provided by the subject matter of the following clauses: [0074] Clause 1. A method for detecting, isolating, and accommodating a sensor fault in a machine or system that utilizes a plurality of sensors, the method comprising: defining a group of the sensors in the machine or system having an spatial-temporal sensor relationship such that a value signal from each sensor is affected by a parameter measured by at least one other sensor in the group within a relevant time interval; with a common neural network model, performing the following functions: receive the value signal from each of the sensors; mask the received value signal from a first one of the sensors and generate a predicted value for the first sensor based on the spatial-temporal sensor relationships and the remaining value signals; compare the predicted value to the received value signal from the first sensor; identify the first sensor as faulty when a difference between the predicted value and the received value signal exceeds a threshold value; and for operations in the machine or system performed by the controller, substitute the predicted value for the received value signal from the first sensor when the first sensor is identified as faulty.

[0075] Clause 2: The method as in clause 1. comprising repeating the functions for the sensors in the group individually or in groups on a continuous or periodic basis. [0076] Clause 3: The method as in clause 1 or 2, wherein the neural network model is any neural network model capable of handling sensor data, including at least one of (i) a Neural Network (“NN”) model, (ii) a Feed-Forward Multilayer Perceptron (“FF MLP’’), (iii) a One Dimensional Convolutional Neural Network (‘TD CNN”), (iv) a Temporal Convolutional Network (“TCN”), (v) a Recurrent Neural Network (“RNN”), (vi) a Long Short-Term Memory / Gated Recurrent Unit C’LSTM/GRUGRU"). or (vii) a Transformer model.

[0077] Clause 4: The method as in any one of clauses 1-3, further comprising training the neural network model in an offline phase by repeatedly and randomly masking one or more of the received value signals and computing the predicted value only for the masked value signals, wherein the training drives a loss value between the masked value signals and the predicted values to a minimum, and wherein the threshold value is based on results of the training such that the threshold value is lower as the loss value approaches zero. [0078] Clause 5: The method as in any one of clauses 1-4, wherein the steps of comparing the predicted value to the received value signal and identifying the first sensor as faulty comprise generating a residual value between the predicted value and the received value signal.

[0079] Clause 6: The method as in any one of clauses 1-5 wherein the spatial- temporal sensor relationships between the sensors in the group are such that a value signal from each sensor is affected by a respective different variable measured by a plurality of other sensors in the group.

[0080] Clause 7: The method as in any one of clauses 1-6, wherein the value signal from each sensor is affected by the respective different variable measured by all of the sensors in the group.

[0081] Clause 8: The method as in any one of clauses 1-7, wherein the machine or system is a wind turbine.

[0082] Clause 9: The method as in any one of clauses 1-8, wherein the sensors in the group comprise at least two sensors selected from: a wind speed sensor, a rotor speed sensor, a nacelle temperature sensor, a pitch angle sensor for each rotor blade, a tower fore-aft acceleration sensor, and a tower side-side acceleration sensor.

[0083] Clause 10: A machine or system, comprising: a control system; a plurality of sensors in communication with the control system, wherein each of sensors measures a parameter used by the control system for operation of the machine or system; the control system configure to perform the following functions: define a group of the sensors having an spatial-temporal sensor relationship such that a value signal from each sensor in the group is affected by the parameter measured by at least one other sensor in the group within a relevant time interval; implement a common neural network model to: (a) receive the value signal from each of the sensors; (b) mask the received value signal from a first one of the sensors and generate a predicted value for the first sensor based on the spatial-temporal sensor relationships; (c) compare the predicted value to the received value signal from the first sensor; (d) identify the first sensor as faulty when a difference between the predicted value and the received value signal exceeds a threshold value; and (e) for operations in the machine or system performed by the control system, substitute the predicted value for the received value signal of the first sensor when the first sensor is identified as faulty ; the controller outputting the predicted value to a component or process in the machine or system so that operation of the component or system is dependent on the predicted value.

[0084] Clause 11 : The machine or system as in clause 10, wherein the neural network model repeats the functions (a) through (e) for the sensors individually or in a group on a continuous or periodic basis.

[0085] Clause 12: The machine or system as in one of clauses 10 or 11, wherein the neural network model is a neural network model capable of handling sensor data, including at least one of (i) a Neural Network (“NN”) model, (ii) a Feed-Forward Multilayer Perceptron (“FF MLP”), (iii) a One Dimensional Convolutional Neural Network (“ID CNN”), (iv) a Temporal Convolutional Network (“TCN”), (v) a Recurrent Neural Network (“RNN”), (vi) a Long Short-Term Memory/Gated Recurrent Unit (“LSTM/GRUGRU”), or (vii) a Transformer model.

[0086] Clause 13: The machine or system as in any one of clauses 10-12, wherein the neural network model generates a residual value between the predicted value and the received value signal.

[0087] Clause 14: The machine or system as in any one of clauses 10-13, wherein the spatial-temporal sensor relationships between the sensors in the group are such that a value signal from each sensor is affected by a respective different variable measured by a plurality’ of other sensors in the group.

[0088] Clause 15: The machine or system as in any one of clauses 10-14, wherein the value signal from each sensor is affected by the respective different variable measured by all of the sensors in the group.

[0089] Clause 16: The machine or system as in any one of clauses 10-15, wherein the machine or system comprises a wind turbine.

[0090] Clause 17: The machine or system as in any one of clauses 10-16, wherein the sensors in the group comprise at least two sensors selected from: a wind speed sensor, a rotor speed sensor, a nacelle temperature sensor, a pitch angle sensor for each rotor blade, a tower fore-aft acceleration sensor, and a tower side-side acceleration sensor.

Claims

WHAT IS CLAIMED IS:

1. A method for detecting, isolating, and accommodating a sensor fault in a machine or system that utilizes a plurality of sensors, the method comprising: defining a group of the sensors in the machine or system having a spatial- temporal relationship such that a value signal from each sensor is affected by a parameter measured by at least one other sensor in the group within a relevant time interval; with a common neural network model, performing the following functions: receive the value signal from each of the sensors; mask the received value signal from a first one of the sensors and generate a predicted value for the first sensor based on the spatial-temporal relationships and the remaining value signals; compare the predicted value to the received value signal from the first sensor; identify the first sensor as faulty when a difference between the predicted value and the received value signal exceeds a threshold value; and for operations in the machine or system performed by the controller, substitute the predicted value for the received value signal from the first sensor when the first sensor is identified as faulty.

2. The method as in claim 1, comprising repeating the functions for the sensors in the group individually or in groups on a continuous or periodic basis.

3. The method as in claim 1. wherein the neural network model is a neural network model capable of handling sensor data, including at least one of (i) a Neural Network ("NN") model, (ii) a Feed-Forward Multilayer Perceptron (“FF MLP”), (iii) a One Dimensional Convolutional Neural Network (“ID CNN”), (iv) a Temporal Convolutional Network (“TCN”), (v) a Recurrent Neural Network (“RNN”), (vi) a Long Short-Term Memory /Gated Recurrent Unit (“LSTM/GRUGRU”), or (vii) a Transformer model.

4. The method as in claim 1 , further comprising training the neural network model in an offline phase by repeatedly and randomly masking one or more of the received value signals and computing the predicted value only for the masked value signals, wherein the training drives a loss value between the masked value signals and the predicted values to a minimum, and wherein the threshold value is based on results of the training such that the threshold value is lower as the loss value approaches zero.

5. The method as in claim 1, wherein the steps of comparing the predicted value to the received value signal and identifying the first sensor as faulty comprise generating a residual value between the predicted value and the received value signal.

6. The method as in claim 1, wherein the spatial-temporal relationships between the sensors in the group are such that a value signal from each sensor is affected by a respective different variable measured by a plurality of other sensors in the group.

7. The method as in claim 6, wherein the value signal from each sensor is affected by the respective different variable measured by all of the sensors in the group.

8. The method as in claim 1, wherein the machine or system is a wind turbine.

9. The method as in claim 8. wherein the sensors in the group comprise at least two sensors selected from: a wind speed sensor, a rotor speed sensor, a nacelle temperature sensor, a pitch angle sensor for each rotor blade, a tower fore-aft acceleration sensor, and a tower side-side acceleration sensor.

10. A machine or system, comprising: a control system; a plurality of sensors in communication with the control system, wherein each of sensors measures a parameter used by the control system for operation of the machine or system; the control system configure to perform the following functions: define a group of the sensors having a spatial-temporal relationship such that a value signal from each sensor in the group is affected by the parameter measured by at least one other sensor in the group within a relevant time interval; implement a common neural network model to:

(a) receive the value signal from each of the sensors;

(b) mask the received value signal from a first one of the sensors and generate a predicted value for the first sensor based on the spatial-temporal relationships;

(c) compare the predicted value to the received value signal from the first sensor;

(d) identify the first sensor as faulty when a difference between the predicted value and the received value signal exceeds a threshold value; and

(e) for operations in the machine or system performed by the control system, substitute the predicted value for the received value signal of the first sensor when the first sensor is identified as faulty; and output the predicted value to a component or process in the machine or system based so that operation of the component or system is dependent on the predicted value.

11. The machine or system as in claim 10, wherein the neural network model repeats the functions (a) through (e) for the sensors individually or in groups on a continuous or periodic basis.

12. The machine or system as in claim 10, wherein the neural network model is a neural network model capable of handling sensor data, including at least one of (i) a Neural Network (“NN”) model, (ii) a Feed-Forward Multilayer Perceptron (“FF MLP”), (iii) a One Dimensional Convolutional Neural Network (“ID CNN”), (iv) a Temporal Convolutional Network (“TCN”), (v) a Recurrent Neural Network (“RNN”), (vi) a Long Short-Term Memory/Gated Recurrent Unit (“LSTM/GRUGRU”), or (vii) a Transformer model.

13. The machine or system as in claim 10, wherein the neural network model generates a residual value between the predicted value and the received value signal.

14. The machine or system as in claim 10. wherein the spatial-temporal relationships between the sensors in the group are such that a value signal from each sensor is affected by a respective different variable measured by a plurality of other sensors in the group.

15. The machine or system as in claim 14, wherein the value signal from each sensor is affected by the respective different variable measured by all of the sensors in the group.

16. The machine or system as in claim 10, wherein the machine or system comprises a wind turbine.

17. The machine or system as in claim 16, wherein the sensors in the group comprise at least two sensors selected from: a wind speed sensor, a rotor speed sensor, a nacelle temperature sensor, a pitch angle sensor for each rotor blade, a tower fore-aft acceleration sensor, and a tower side-side acceleration sensor.