WO2016073725A1

WO2016073725A1 - Monitoring via neural network model

Info

Publication number: WO2016073725A1
Application number: PCT/US2015/059250
Authority: WO
Inventors: Manfred Battefeld
Original assignee: Hach Co
Current assignee: Hach Co
Priority date: 2014-11-07
Filing date: 2015-11-05
Publication date: 2016-05-12
Anticipated expiration: 2017-05-07

Abstract

A method can include processing input values for a set of parameters associated with operations of a water processing facility using a trained neural network to output an output value for a parameter that is not a member of the set of parameters; and controlling at least one of the operations of the water processing facility based at least in part on the output value.

Description

MONITORING VIA NEURAL NETWORK MODEL

RELATED APPLICATION

[0001] This application claims the benefit of and priority to a U.S.

Provisional Application having Serial No. 62/077,101 , filed 7 November 2014, which is incorporated by reference herein.

BACKGROUND

[0002] Sensors and associated equipment may be used to monitor a process and, for example, to control a process.

SUMMARY

[0003] A method can include processing input values for a set of parameters associated with operations of a water processing facility using a trained neural network to output an output value for a parameter that is not a member of the set of parameters; and controlling at least one of the operations of the water processing facility based at least in part on the output value. Various other methods, devices, systems, etc., are also disclosed.

BRIEF DESCRIPTION OF THE DRAWINGS

[0004] Features and advantages of the described implementations can be more readily understood by reference to the following description taken in conjunction with the accompanying drawings.

[0005] Fig. 1 is a series of diagrams including an example of an

environment and examples of equipment.

[0006] Fig. 2 is a diagram of an example of a scenario and an example plot of data.

[0007] Fig. 3 is a diagram of an example of a system.

[0008] Fig. 4 is a diagram of an example of a system.

[0009] Fig. 5 is a diagram of an example of a scenario and examples of sensor. [0010] Fig. 6 is a diagram of an example of a system and an example of a neural network.

[0011] Fig. 7 is a diagram of examples of neural networks.

[0012] Fig. 8 is a diagram of an example of a method.

[0013] Fig. 9 is a diagram of an example of a scenario that includes a recurrent neural network.

[0014] Fig. 10 is a diagram of an example of a method.

DETAILED DESCRIPTION

[0015] The following description includes the best mode presently contemplated for practicing the described implementations. This description is not to be taken in a limiting sense, but rather is made merely for the purpose of describing general principles of the implementations. The scope of the described implementations should be ascertained with reference to the issued claims.

[0016] As an example, a method may include determining a parameter value on the basis of a cognitive analysis from other parameters using one or more artificial neural networks, referred to as neural networks.

[0017] As an example, a trained neural network may operate as a "virtual" sensor, analyzer, etc., for example, for determination of a parameter value, otherwise determinable via measured signals (e.g., used directly or indirectly to calculate such a parameter value).

[0018] As an example, consider a sliding caliper that measures a length directly or, for example, an optical probe for suspended solids (e.g., indirectly using signal of stray light from suspended particles).

[0019] As an example, a method can include identifying a property based at least in part on a cognitive analysis using a neural network. As an example, a system may include circuitry that can determine an amount of phosphate in a wastewater treatment plant via a cognitive analysis of other parameters such as, for example, one or more of nitrate, pH, temperature of water and air, time, amount of oxygen in the water, turbidity, water flow, etc. [0020] As an example, a neural network-based sensor or analyzer may operate according to instructions stored in memory that can be executed by a processor. In such an example, the sensor or analyzer may determine a value for a parameter X by the analysis of values of other parameters Yi.

[0021] As an example, a virtual analyzer for a parameter X may be based at least in part on artificial intelligence, which may use information from other parameters Yi. As an example, a neural network may receive as input a vector of values of parameters Yi and, in turn, output a value or values as to a parameter X, optionally as a temporal sequence (e.g., values of the parameter X for times).

[0022] As an example, a method may include a learning or training phase and a working or operational phase (e.g., upon implementation). As an example, in a training phase, a virtual sensor or analyzer may access information about a parameter X (e.g. from a physical sensor or analyzer) and information about parameters Yi. In such an example, a neural network may be trained, for example, by generating weights associated with connections between neurons of the neural network. As an example, in an operational phase, a virtual sensor or analyzer may predict a value of the parameter X, for example, based on input of information about the parameters Yi (e.g., an input vector of parameter values).

[0023] As an example, a training phase may occur locally, remotely or locally and remotely. For example, a training phase may take place at a site, which may be a facility, etc. In such an example, a sensor or analyzer that can measure values for a parameter X may be available at that site (e.g., for generating a training set). Such a sensor or analyzer may optionally be available on a temporary basis. Or, for example, such a sensor or analyzer may optionally be available on a permanent basis. In the latter example, a facility may include performing "virtual" measurements or analyses, for example, to conserve power, conserve reagent, etc.

[0024] As an example, a training phase may be performed remote from a site where a trained neural network may be implemented. For example, training may occur at a supplier site, a laboratory, etc. At a supplier site, laboratory, etc., equipment can be available for measuring, for example, values of a parameter X and values of parameters Yi. As an example, once trained, a neural network may be implemented at one or more customer sites, for example, to determine a value or values for the parameter X given input values for the parameters Yi.

[0025] As an example, a training phase may be based on information from a plurality of sites. In such an example, a trained neural network may be implemented at one of those sites and/or at one or more other sites. As an example, training may be an on-going process. For example, information from one or more sites may be input as a training set to update a neural network where an updated, trained neural network may be implemented, distributed, etc.

[0026] As an example, where a "virtual" sensor or analyzer may be implemented using existing on-site equipment, its cost may be less than that of an actual sensor or analyzer. And, as an example, based on performance of a virtual sensor or analyzer, a site operator may determine that its performance is suitable or, for example, that installation of an actual sensor or analyzer may be recommended (e.g., budgeted, etc.). Where installation of an actual sensor or analyzer occurs, such a site may optionally be an informational site that may provide information that may be used to train one or more neural networks.

[0027] As an example, a system can include a controller that utilizes a trained neural network that receives input information and that can estimate a parameter value as an output. In such an example, the system may not include a sensor that can measure the parameter value or, for example, may not include a sensor that is utilized as frequently for measurements as the controller may be utilized for estimations (e.g., via the trained neural network). In such an approach, the controller may operate as a "virtual" sensor with respect to the particular parameter; noting that, as an example, a controller may output a plurality of parameters. Such an approach may allow a system to optionally operate without a particular physical sensor or, for example, to utilize such a sensor less frequently (e.g., to conserve reagent, technician time, etc.).

[0028] Fig. 1 shows an example of an environment 100 that includes equipment that may be implemented for monitoring the environment 100. As shown, the equipment can include a network 101 , electronic equipment 105-1 and 105-2 that can be operatively coupled to the network 101 and sondes 1 10-1 and 1 10-2 that can be operatively coupled to the electronic equipment 105-1 and 105-2. As an example, one or more of the sondes 1 10-1 and 1 10-2 may include an interface or interfaces that can operatively couple to electronic equipment (e.g., the electronic equipment 105-1 and 105-2), a network (e.g., the network 101 ), etc.

[0029] In the example of Fig. 1 , the sonde 1 10-1 is implemented to monitor land-based conditions while the sonde 1 10-2 is implemented to monitor sea- based conditions. As an example, a land-based implementation may monitor for leakage from a tank, a pipe, etc. As an example, a land-based implementation may monitor a river, a lake, or other aqueous environment. As an example, in a sea or other water body, a sonde may be operatively coupled to a flotation device, an anchor, a watercraft, etc. As an example, one or more sondes may be part of a control system. For example, a sonde may transmit information that may be used in making a control decision, adjusting a parameter, etc. As an example, a control system may be open loop, closed loop, etc.

[0030] As an example, environmental monitoring may involve one or more processes, activities, etc. that aim to characterize and/or monitor quality of an environment. Environmental monitoring may occur prior to, during or after preparation of an environmental impact assessment. Environmental monitoring may be implemented to establish a current status of an environment or to establish a trend in one or more environmental parameters. Results of monitoring may be reviewed, analyzed statistically and reported. A monitoring program may consider use of data prior to monitoring. Monitoring may be subject to rules, regulations, etc.

[0031] Fig. 1 shows an example of a sonde 1 10, which may include one or more types of circuitry 120. As an example, the sonde 1 10 can include a controller 122, memory 124 and one or more interfaces 126. In such an example, the controller 122 may be a microcontroller (e.g., ARM, ARC, etc.) that may be powered by a power source (e.g., battery, power cable, etc.). Such a controller may interact with other circuitry such as one or more of pH circuitry 132, resistivity circuitry 134, salinity circuitry 136, pressure circuitry 142, oxygen reduction potential (ORP) circuitry 144, dissolved oxygen (DO) circuitry 146, reference potential (RP) circuitry 152, clock circuitry 154, power/battery circuitry 156, specific conductivity (SC) circuitry 162, specific gravity (SG) circuitry 164, total dissolved solids (TDS) circuitry 166, depth circuitry 172, temperature circuitry 174 and other circuitry 176.

[0032] Fig. 2 shows an example of a scenario 210 that includes a system 220 that includes one or more processors 222, memory 224 accessible by at least one of the one or more processors 222, one or more modules 226 that includes one or more of hardware, software, firmware, etc., and one or more other system components 228 (e.g., one or more interfaces, one or more displays, one or more power supplies, etc.).

[0033] In the example scenario 210, the system 220 is implemented to monitor and/or control a process 230. As shown in Fig. 2, the process 230 includes a primary clarifier 232, an aeration basin 234 and a secondary clarifier 236. The process 230 may be a wastewater treatment process, for example, for a wastewater treatment plant.

[0034] As an example, a process such as the process 230 may be implemented with an aim to reduce phosphorous concentration in an input stream or streams. Phosphorous can occur in waters and wastewaters as phosphates. The forms of phosphates arise from a variety of sources. As an example, orthophosphate may be added to a water supply during treatment. Other sources can include water used for laundering or other cleaning. As an example, surface waters can include phosphorus (e.g., in various compounds). In natural conditions, phosphorus concentration in water may be subject to balancing processes where, for example, accessible mass is related to requirements of an ecological system. When the input of phosphorus to waters is higher than it can be assimilated by a population of living organisms, excess phosphorus content can occurs. [0035] Domestic wastewater may be a predominant source of phosphorus in municipal discharges, particularly in densely populated regions. For example, in a country such as the Netherlands, about 20 million tons per year of phosphate phosphorus comes from domestic wastewater while about 4 million tons per year comes from industrial discharges. Domestic wastewater flow may vary depending on time of day, rainwater, melting of snow and/or ice, etc. Agricultural flows may also impact input to a wastewater treatment facility or facilities.

[0036] As an example, one wastewater treatment facility may be fitted with equipment that differs from another wastewater treatment facility. For example, a geographic region may include two or more facilities (e.g., plants) where people living within the geographic region have similar habits and where, for example, the geographic region experiences similar weather (e.g., rainfall, etc.). In such an example, the facilities may include different equipment. For example, one facility may include one or more sensors that are not installed at another facility. As to the example process 230 of Fig. 2, it may be associated with a facility that may include one or more sensors and/or other technology (e.g., equipment, modules, etc.) that may not be installed at another facility (e.g., due to cost, labor, etc.).

[0037] As shown in the example scenario 210, the system 220 is operatively coupled to a phosphate module 250 ("P-Module") that includes interfaces operatively coupled to a flow meter 252 (e.g., "Q", representing flow rate), a phosphate measurement unit 254 (e.g., "PO₄", representing phosphate ion concentration or other phosphate species concentration) and a chemical feed unit 256 (e.g., "Al, Fe", representing a type of chemical feed such as an aluminum based chemical and/or an iron based chemical).

[0038] As an example, the phosphate measurement unit 254 may include analyzer components for monitoring of phosphate in an aqueous sample (e.g., in water being treated, having been treated, etc.). As an example, an analysis may include measuring ammonia via a vanadomolybdo phosphoric acid colorimetric technique. As an example, the unit 254 may include analyzer components that can compensate for sample variations, for example, consider compensation for background color of a sample.

[0039] As to measurement of phosphorous in a sample, in dilute

orthophosphate solution, ammonium molybdate reacts under acid conditions to form a heteropoly acid, molybdo phosphoric acid. In the presence of vanadium, yellow vanadomolybdo phosphoric acid can be formed. In such an example, intensity of the yellow color can be proportional to phosphate concentration in the sample.

[0040] As an example, measurement of phosphate concentrations can be or can include measurement of orthophosphate concentrations. As an example, measurements of phosphate concentrations can include measurements of one or more types of phosphates. As an example, orthophosphates can be inorganic forms of phosphate that can include forms used in, for example, fertilizers. As an example, organically bound phosphates can be types found in human and animal wastes and/or in decaying organic matter. As an example, condensed

phosphates (e.g., polyphosphates) can be added to water supplies and industrial processes, for example, to help to reduce formation of scaling and to help to inhibit corrosion. As an example, a value for total phosphates may be a sum of values for one or more types of phosphates.

[0041] As an example, phosphate may be considered a nutrient that is to be removed to an extent to meet a desired concentration in effluent from a process. In such an example, to determine whether the concentration has been met, a measurement technique may be employed that is sensitive enough to detect a maximum desired concentration. As an example, in Fig. 2, the unit 254 may provide for detection limits of phosphate in samples in a range from about 0.05 mg/L to about 50 mg/L. As an example, a process may include an effluent target that may be measured using the unit 254, optionally with dilution of a sample.

[0042] As an example, the unit 254 may include a response time, for example, a delay from receipt of a sample to output of a value. Such a response time may be of the order of minutes. [0043] As an example, the unit 254 may include a filter probe. As an example, the unit 254 may consume one or more reagents (e.g., for purposes of analysis, etc.). As an example, the unit 254 may include a cleaning process, a calibration process, one or more self-diagnosis processes, etc.

[0044] As an example, a phosphorus removal scheme may aim to eliminate the excess phosphorus content from wastewater discharged to receiving waters. In such an example, the excluded phosphorus load may be handled, for example, as a separate stream (e.g., of solids, liquids, etc.). As an example, a process may aim to diminish eutrophication-related issues resulting from transport of phosphorus.

[0045] As an example, a portion of phosphorus in an inlet stream to a facility may be settleable as particles. For example, consider a process where about 15% of total phosphorus received is amenable to removal as settleable particles via a primary settling (e.g., sedimentation) process that may operate without addition of one or more metal salts (e.g., flocculating agents).

[0046] As an example, a process may include transfer of soluble phosphorus to a solid phase, which may be complemented by solid-liquid separation. As an example, transfer to a solid phase may be performed via one or more of chemical precipitation and adsorption (e.g., via a trivalent metal salt), biological uptake (e.g., via nonexchangeable phosphorus or/and enhanced uptake by bacteria), and ion exchange and adsorption.

[0047] As an example, chemical phosphorus removal (e.g., removal by salt addition) may be applied, for example, as one or more of direct precipitation, pre-precipitation, simultaneous precipitation, and post-precipitation. As an example, in a plant, a process may include addition of one or more metal salts at one or more points, which may result in two or more kinds of precipitated phosphorus.

[0048] As an example, a process may include addition of one or more trivalent metal salts for removal of phosphorus (e.g., from municipal wastewater, etc.). For example, as to some examples of metal salts, consider aluminum sulfate, ferric chloride, and ferrous(bivalent) and ferric sulfates; noting that ferrous ions, under certain conditions, may be subject to electron removal to become ferric ions. As an example, for aluminum sulfate, a weight ratio may be about 0.87 to 1 Al to P while, for example, for ferric sulfate, a weight ratio may be about 1 .8 to 1 Fe to P. Such ratios may be theoretical as, for example, in actual practice the weight ratio applied may differ.

[0049] As an example, choice of metal salt or metal salts and one or more addition points may depend on various factors. For example, ferrous sulfate can be an inhibiting factor to Nitrosomonas bacteria, which may influence where it may be added, particularly where nitrogen removal via such bacteria digestion is desired.

[0050] As an example, a guideline may recommend about 20% excess of aluminum ions (by weight) to achieve about an 80% reduction of phosphorus while, for example, about 100% excess may be recommended for about 95% P removal. As an example, a process may include adding one or more metal salts along with a polymer or polymers. For example, a polymeric material may carry one or more metal salts.

[0051] As an example, a process may optionally include trivalent iron salt production, for example, where oxidation of a copperas (ferrous sulfate) is achieved via hydrochloric acid (e.g., a pH dependent process). As an example, a mixture of ferric sulfate and ferric chloride may be applied as flocculants (e.g., to precipitate phosphorus).

[0052] In the example of Fig. 2, the process 230 may be applied to a facility with a capacity of about 40,000 population equivalent (PE). Such a facility may include multiple lanes, for example, intermittently aerated via circular tanks. Such a facility may employ chemical precipitation and simultaneous aerobic sludge stabilization.

[0053] As an example, a facility may be subject to fluctuating influent conditions (e.g., heavy rain, purged channels, etc.) where associated nitrogen and phosphorus loads may vary and challenge effluent goals (e.g., compliance with one or more discharge consent values). [0054] In the example of Fig. 2, the module 250 and the unit 254 can be implemented to measure orthophosphate concentrations in an outlet from the aeration basin 234. Where a facility includes multiple lanes, one or more of the module 250 and the unit 254 may be implemented to measure orthophosphate concentrations in one or more lanes (e.g., one or more aeration basin outlets). As an example, the module 250 may be a controller that can transmit a signal or signals (e.g., commands, etc.) for control of addition of one or more precipitant agents, for example, based at least in part on a phosphorus load (e.g., for one or more tanks). Such an approach may aim to ensure reliable compliance with a Ptotai limit value(s) (e.g., a total phosphates value). Such an approach may help to reduce the amount of one or more precipitant agents dosed to the process (e.g., via one or more lanes).

[0055] In Fig. 2, the plot 260 shows influent flows, precipitant agent dosing and orthophosphate concentration versus time where at a particular time, the module 250 is implemented along with the unit 254. As shown, the module 250 and the unit 254 reduced precipitant agent dosing while providing for stable and compliant effluent values (e.g., below a limit value). The data in the plot 260 may be representative of strongly fluctuating influent conditions (e.g., where influent volume may double and halve over a period of days) that can present challenges for control schemes that aim to meet discharge target limits as to effluent.

[0056] As an example, information acquired via the system 220 of Fig. 2 may be used in a method that includes training one or more neural networks. As an example, such a method may include receiving information germane to geography, weather, habits, etc. For example, information acquired from a facility may be used to train one or more neural networks where the one or more trained neural networks may be implemented to control a process at another facility (e.g., and/or optionally at the same facility).

[0057] As an example, consider a facility that is at a distance of about 100 kilometers from the facility illustrated via the process 230 of Fig. 2. Where the distant facility does not include the module 250 and the unit 254, information from one or both of the facilities may be input to a computing system that can train one or more neural networks. In such an example, one or more of the trained neural networks may receive input from one or both of the facilities to at least in part enhance control of the distant facility that does not include the module 250 and the unit 254. As an example, a trained neural network may be a "virtual" module, a "virtual" sensor, etc.

[0058] While the example scenario 210 of Fig. 2 pertains to particular equipment, an approach that includes training and/or implementing one or more neural networks may be applied to a scenario such as a scenario that

implements a sonde such as the sonde 1 10 of Fig. 1 . As an example, a trained neural network may be implemented to control one or more sonde related processes. For example, consider timing of a measurement, frequency of a measurement, etc.

[0059] Fig. 3 shows an example of a system 300 that includes a network 301 (e.g., the Internet, etc.), data storage equipment 302 (e.g., a storage area network (SAN), a network attached storage (NAS), etc.), a cellular network 303, a device 305 that includes an interface that can couple to the cellular network 303, one or more devices 310, a system 320, a services system 330 and a system 340.

[0060] As shown, the system 320 includes one or more processors 322, memory 324 accessible by at least one of the one or more processors 322, one or more modules 326 that includes one or more of hardware, software, firmware, etc., and one or more other system components 328 (e.g., one or more interfaces, one or more displays, one or more power supplies, etc.).

[0061] As shown, the services system 330 includes one or more

processors 332, memory 334 accessible by at least one of the one or more processors 332, one or more modules 336 that includes one or more of hardware, software, firmware, etc., and one or more other system components 338 (e.g., one or more interfaces, one or more displays, one or more power supplies, etc.).

[0062] As shown, the system 340 includes one or more processors 342, memory 344 accessible by at least one of the one or more processors 342, one or more modules 346 that includes one or more of hardware, software, firmware, etc., one or more other system components 348 (e.g., one or more interfaces, one or more displays, one or more power supplies, etc.) and one or more neural networks 350.

[0063] As an example, the services system 330 may receive information from one or more of the devices 310, the system 320, the storage 302, the device 305 operatively coupled to the cellular network 303 and the system 340 and train one or more neural networks such as the one or more neural networks 350 of the system 340.

[0064] Fig. 4 shows an example of a system 400 that includes systems 320-1 , 320-2 to 320-N (see, e.g., the system 320 of Fig. 3) with corresponding modules 326-1 , 326-2 to 326-N, the services system 330 of Fig. 3, one or more neural networks 350-1 , 350-2 to 350-N and systems 340-1 , 340-2 to 340-N (see, e.g., the system 340 of Fig. 3) with corresponding modules 346-1 , 346-2 to 346- N. As shown, the modules 326-1 , 326-2 to 326-N may include particular modules which may, for example, be mimicked as virtual modules, for example, as included in the modules 346-1 , 346-2 to 346-N.

[0065] Fig. 4 also shows an example of a method 490 that includes a reception block 492 for receiving information, a train block 494 for training one or more neural networks and an operation and/or distribution block 496 for operating and/or distributing the one or more trained neural networks. The method 490 is shown along with blocks 493, 495 and 497, which may be computer-readable media blocks, for example, of non-transitory storage media that are not carrier waves. Such blocks may store instructions such as processor-executable instructions that may be executed by one or more processors, for example, to implement a method such as the method 490 of Fig. 4.

[0066] Fig. 5 shows an example of a scenario 510 that includes a system 540 that includes one or more processors 542, memory 544 accessible by at least one of the one or more processors 542, one or more modules 546 that includes one or more of hardware, software, firmware, etc., one or more other system components 548 (e.g., one or more interfaces, one or more displays, one or more power supplies, etc.) and one or more neural networks 550.

[0067] In the example scenario 510, the system 540 is implemented to monitor and/or control a process 530. As shown in Fig. 5, the process 530 includes a primary clarifier 532, an aeration basin 534 and a secondary clarifier 536. The process 530 may be a wastewater treatment process, for example, for a wastewater treatment plant. As an example, a process such as the process 530 may be implemented with an aim to reduce phosphorous concentration in an input stream or streams.

[0068] As shown in the example scenario 510, the system 540 is operatively coupled to a virtual phosphate module 551 ("Virtual P-Module") that includes interfaces operatively coupled to a flow meter 552 (e.g., "Q",

representing flow rate), one or more sensor units 554 and a chemical feed unit 556 (e.g., "Al, Fe", representing a type of chemical feed such as an aluminum based chemical and/or an iron based chemical).

[0069] As an example, the one or more sensor units 554 can include one or more of an oxygen sensor, a conductance sensor, an outlet turbidity sensor, a nitrate sensor, a pH sensor, an inlet flow rate sensor, a suspended solids sensor and an ammonium sensor.

[0070] As an example, the one or more sensors 554 may be received as input to one or more of the one or more neural networks 550. As an example, one or more of the one or more neural networks 550 may receive input from another source, sources, system, network, etc. As an example, weather, time of day, etc. may be received as input to one or more neural networks.

[0071] In the example scenario 510 of Fig. 5, the one or more neural networks 550 may be implemented as a virtual module such as the virtual P- module 551 . In such an example, the process 530 may be monitored and/or controlled without operation of a local phosphate, specific sensor and/or without a local P-Module. As an example, information provided via one or more of the one or more sensors 554 and optionally other information may be input to one or more of the one or more neural networks 550 to determine an output, which may be, for example, an approximated phosphate concentration or a surrogate thereof. As an example, an output may be a value such as a dosing rate for dosing of one or more flocculating agents (e.g., one or more metal salts, etc.).

[0072] As an example, the system 540 can be a controller where the one or more modules 546 can be stored in the memory 544 and include processor- executable instructions (e.g., executable by one or more of the one or more processors 542) to instruct the controller to process input values for a set of parameters associated with operations of a water processing facility (e.g., per the process equipment of the process 530, etc.) using a trained neural network (see, e.g., the one or more neural networks 550) to output an output value for a parameter that is not a member of the set of parameters; and to control at least one of the operations of the water processing facility based at least in part on the output value. As an example, a controller can include instructions to instruct the controller to output a phosphate concentration value as the output value where, for example, instructions are included to instruct the controller to control at least one of the operations to control an aqueous phosphate concentration (e.g., phosphate concentration in water) of the water processing facility based at least in part on the phosphate concentration value. In such an example, at least one of the operations may include a flocculating agent dosage operation where a dosage value of a flocculating agent of the flocculating agent dosage operation depends at least in part on the phosphate concentration value.

[0073] As an example, a neural network may receive information such as weather information, information from a facility within a distance from a facility where the neural network may provide useful output, etc. For example, where rainfall impacts a facility with a P-Module and an associated phosphate sensing unit, the response of that facility may be used as input to a neural network, for example, for purposes of training and/or purposes of predicting an output. For example, the response of a remote facility may be input information along with local sensor information of a facility to control dosing of one or more flocculating agents at the local facility. In such an example, information from one facility may benefit one or more other facilities which may not include certain equipment (e.g., due to cost, labor, etc.).

[0074] As an example, a facility may include valves, flow meters, controllers, etc. As an example, a valve may be controlled to control pressure, flow rate, etc. As an example, a facility may include equipment that can redirect material (e.g., fluids, solids, slurries, etc.). As an example, a facility may include one or more electric motors. For example, consider an electric motor that drives a pump, an electric motor that drives a mixer, etc. As an example, a facility may include equipment driven by one or more sources of energy, optionally via a source derived from operation of the facility (e.g., gas, solids, etc.). As an example, one or more trained neural networks may be implemented for purposes of control, prediction, etc. of one or more processes at a facility.

[0075] Fig. 6 shows a system 600 where the data analysis block 604 may include or be based at least in part on a technique 605. In the example of Fig. 6, the technique 605 illustrated is an artificial neural network (ANN or "neural network") that includes at least one input layer, at least one hidden layer and at least one output layer.

[0076] ANNs find use in recognition scenarios such as handwriting recognition for determination of individual characters, and for speech recognition for determination of individual sounds and words, etc. In the example of Fig. 6, the technique 605 can recognize statuses based on input. As an example, the technique 605 can involve training, for example, using actual information and/or synthetic information.

[0077] For example, the technique 605 may be defined by a set of input neurons which may be activated by receipt of information. After being weighted and transformed by a function or functions (e.g., as may be determined via historical analyses, training, etc.), the "activations" of these neurons are then passed on to other neurons. Such a process may be repeated until finally, an output neuron is "activated". The output neuron can correspond to a status, which may optionally be represented as a code. [0078] As an example, one or more algorithms may be defined based at least in part on inputs and outputs of an ANN. For example, consider an algorithm that includes variables where values for the variables may be associated with outputs. Such an approach may optionally operate via a look-up such as in a look-up table (LUT). As an example, a LUT approach may be implemented using memory and a controller that can execute instructions.

[0079] As an example, a predictive model may be generated based at least in part on historical information. A predictive model may be generated using predictive analytics, for example, consider use of one or more statistical techniques in combination with machine learning and data mining. As an example, a method may include building a predictive model using historical information and then using the predictive model to make predictions about future, or otherwise unknown, events, states, etc. (e.g., based on current information and/or historical information).

[0080] As an example, a predictive model may be, in part, a virtual machine of a device such as a sensor, a module, etc. For example,

characteristics of a sensor may be modeled in software (e.g., a software emulation of a sensor). In such an example, time may be accelerated such that the virtual machine performs various actions that can establish states, which may be possible states of a real machine (e.g., a sensor, a controller, etc.). Where a possible state is uncovered that may be problematic, a method can include taking one or more corrective actions. For example, a signal may be transmitted to a sensor, a controller, an actuator, etc. As an example, an alert may be transmitted to a device, an account, etc., for example, consider an email alert, a text alert, etc. Such an alert may include a link (e.g., URL) to instructions associated with the alert (e.g., adjust controller, adjust actuator, dose flocculating agent or agents, etc.). As an example, one or more security mechanisms such as an encryption mechanism that can encrypt information may be implemented, for example, as to instructions, actions, etc.

[0081] Fig. 7 shows example neural networks 710 and 730, which may be referred to as recurrent neural networks (RNNs). In the examples of Fig. 7, weights in feedforward channels can be modified and arrows show the direction of the information flow. Such networks may be implemented to learn temporal sequences. For example, consider a temporal sequence of events related to control of a process such as, for example, the process 230 of Fig. 2, the process 530 of Fig. 5, etc. As an example, RNNs may provide an output vector which depends on the temporal order of the input vectors. As an example, for a given input vector, a series of output vectors may be produced.

[0082] As an example, a method can include receiving input information and output information for one or more facilities and training a recurrent neural network using such information to provide a trained recurrent neural network. In such an example, the trained recurrent neural network may receive input information (e.g., an input vector) and output information for one or more times (e.g., as a series of output vectors). Such output information may be used, for example, to control one or more processes.

[0083] As to the neural network 710, which may be referred to as an Elman RNN, it includes an input layer, a hidden layer, and an output layer that are connected in a feedforward manner. The hidden layer, however, is not only connected to the output layer but also, in a 1 :1 connection, to a further layer called a context layer. To form recurrent connections, the output of this context layer is also inputted to the hidden layer. Except for these 1 :1 connections from hidden to context layer, the weights of which are fixed to 1 , all other layers may be fully connected and all weights may be modifiable. The recurrent connections of the context layer provide the system with a short-term memory. In such an example, the hidden units observe actual input and, via the context layer, obtain information on their own state at the last time step. Since, at a given time step, hidden units have already been influenced by inputs at the earlier time steps, this recurrency includes a memory which depends on earlier states (e.g., though their influence may decay with time).

[0084] During operation, the input of the Elman RNN 710 may be provided with a temporal series of input vectors. To change the weights, the output can be compared with a desired output vector, and the generalized delta rule, for example, can be applied to change the weights. Thereby, the RNN 710 can learn to attribute an output directly to the actual input and also to the temporal sequence of several subsequent input vectors.

[0085] In Fig. 7, the RNN 730 may be referred to as a Jordan RNN. As shown, recurrent connections start from the output, rather than the hidden layer as in the RNN 710. Also, the layer corresponding to the context (e.g., a "state" layer) includes a recurrent net itself, for example, with 1 :1 connections and weights (e.g., fixed weights).

[0086] As an example, for the RNN 730, a constant input vector may be given as input to the net where the output of the net may provide a temporal sequence of vectors. As an example, variation in time may be produced by two types of recurrent connections, namely those from the output layer to the state layer and those within the state layer. As an example, for each input vector, another temporal sequence may be produced.

[0087] As an example, the RNN 730 may be provided an input vector and output a series of temporal vectors. For example, given inputs as to one or more sensor measurements, the series of temporal output vectors may show how a system may respond over time to the given inputs. In such an example, a controller may "intervene" in a physical process, for example, to avoid a particular output (e.g., an effluent concentration value of a wastewater treatment plant exceeding a desired value).

[0088] As an example, a system may implement an at least partially recurrent neural network. In such an example, the RNN may be used to learn temporal sequences such that it can produce an output vector for a given sequence of input vectors, or, for example, to produce a series of output vectors for a given input vector.

[0089] Fig. 8 shows an example of a method 800 that includes a reception block 810 for receiving input information (e.g., an input vector), a prediction block 820 for predicting one or more output sequences (e.g., one or more output vectors) via one or more neural networks (e.g., one or more trained neural networks) and a control block 830 for controlling a process based at least in part on the one or more output sequences.

[0090] In the example of Fig. 8, the blocks 810, 820 and 830 are shown with computer-readable media (CRM) blocks 81 1 , 821 and 831 . A CRM block may include instructions executable by a controller, a processor, etc. to cause a device, a system, etc. to perform one or more actions such as one or more of the actions of the method 800.

[0091] Fig. 9 shows an example of a scenario 900 that includes two facilities 910 and 920 that may be separated by a distance and, for example, subject to weather conditions such as, for example, rain.

[0092] As an example, the facility 910 may provide at least a partial training set 915 for training a RNN to generate a trained RNN 930. As an example, other information for training may include weather information, information from the facility 920, information from one or more laboratory models, etc. In the example of Fig. 9, an input vector 925 may be received by the trained RNN 930, which may output a sequence 935 (e.g., a temporal sequence of output vectors). In such an example, the output 935 may include dosing information, for example, to dose one or more flocculating agents at the facility 920. As an example, an output sequence may include information from which one or more control parameters may be determined. For example, where an output vector includes one or more phosphate concentrations, a dose or doses of flocculating agents may be determined and delivered. As an example, a dose determined for one point in time may be an input that can be used to determine one or more output vectors (e.g., that may account for the dose).

[0093] Fig. 10 shows an example of a method 1000 along with the network 301 , the services system 330 and the system 340 of Fig. 3. As shown, the method 1000 includes a train block 1010 for training one or more neural networks, an identification block 1020 for identifying one or more systems that may implement at least one of the one or more trained neural networks, a decision block 1030 for deciding whether to implement the at least one trained neural network and an implementation block 1040 for implementing the at least one trained neural network based on an affirmative decision of the decision block 1030 and, for example, a schedule block 1032 for scheduling a future inquiry (e.g., a reminder as to implementation of a trained neural network).

[0094] As an example, a RNN may be trained to predict one or more parameters of a facility such as, for example, a water processing facility where at least one parameter may be predicted based on at least one other parameter. For example, where training uses information of a facility that includes a particular sensor and/or module and where another facility does not include that sensor and/or module, a parameter output by such a sensor and/or module may be predicted using one or more other parameters. As an example, in a learning phase, consider X, Yi where time t may be one of the input parameters Yi. After the learning phase, a trained NN (e.g., a RNN, RTNN) may output X via input Yi where time t may be one of the input parameters Yi. As an example, a trained neural network may output temporal information. For example, time may be a dimension, a parameter, etc. of a neural network.

[0095] As an example, a wastewater facility may aim to meet requested limits of ingredients of water that leaves the facility. For example, consider phosphate concentration. As an example, a technique to reduce concentration of phosphate can include adding a flocculation agent (e.g., e.g. ferric sulfate). In such an example, the amount of the flocculation agent may be calculated according to a highest expected phosphate concentration at the facility as, for example, a small facility may not include an online analyzer for phosphate control (e.g., online in that it measures on-site without interruption of water flow, etc.). As an example, a constant dosing rate of a flocculation agent that is based on a highest expected phosphate concentration can increase cost, particularly where actual phosphate concentration may be well below an estimated maximum value.

[0096] As an example, a wastewater treatment facility may include sensors that can sense oxygen, pH, conductance, flow rate of an inlet, turbidity at an outlet and, for example, in an activated basin, concentration of suspended solids, nitrate and ammonium concentration. As an example, a "virtual" analyzer for phosphate (e.g., phosphate sensing and/or control) may be based on a trained neural network (e.g., a recurrent neural network, etc.). In such an example, control of dosing rate of a flocculating agent that can reduce phosphate may be based at least in part on output of a trained neural network. In such an example, dosing may vary with respect to time (e.g., based on input of local information and/or remote information).

[0097] As an example, a system may provide for cloud-based services. As an example, services may include training and/or implementation services with respect to one or more neural networks.

[0098] As an example, a method can include collecting sensor data and optionally other data and analyzing the data via one or more neural networks. In such an example, where results may be germane to one or more sensors, sensor systems, control systems, etc., a method may include generating one or more notifications, for example, to notify equipment, operators, etc.

[0099] As an example, a trained neural network may be implemented to output information germane to tuning a sensor (e.g., frequency of measurements, extending battery life, conserving chemicals/reagents, maintenance routines, inference of non-measured parameter(s) based on one or more measurements.

[00100] As an example, a system may gather information from a plurality of sites and may train one or more neural networks, which may be implemented, for example, remotely, in part locally and in part remotely, etc. As an example, where a system implements a trained neural network, a remote system may act to revise, update, etc. the trained neural network (e.g., via wire, wirelessly, etc.).

[00101] As an example, a method may include training at least one neural network and, for example, integrating instructions (e.g., software, firmware, etc.) in a sensor, a controller, etc.

[00102] As an example, a method may include training one or more neural networks and implementing the one or more trained neural networks for improving measurement/prediction/estimation of phosphate concentration, dosing of chemicals, mixing rates, flow rates, etc.

[00103] As an example, a method can include processing input values for a set of parameters associated with operations of a water processing facility using a trained neural network to output an output value for a parameter that is not a member of the set of parameters. In such an example, the method may also include controlling at least one of the operations of the water processing facility based at least in part on the output value (e.g., controlling one or more water treatment operations, etc.).

[00104] As an example, input values for a set of parameters can

correspond to sensor measurements of sensors such as, for example, sensors of the water processing facility.

[00105] As an example, a method can include processing input values for a set of parameters associated with operations of a water processing facility using a trained neural network to output an output value for a parameter that is not a member of the set of parameters, for example, where the output value for the parameter is a phosphate concentration.

[00106] As an example, a method can include processing input values for a set of parameters associated with operations of a water processing facility using a trained neural network to output an output value for a parameter that is not a member of the set of parameters and controlling dosage of a flocculating agent based at least in part on the output value for the parameter. As an example, a method can include controlling dosage of an iron compound (e.g., ferric sulfate, ferric chloride, etc.). As an example, a method can include controlling dosage of an aluminum compound (e.g., aluminum sulfate, etc.).

[00107] As an example, a method may train a recurrent neural network to provide a trained recurrent neural network. As an example, a recurrent neural network (RNN) may be a Jordan RNN.

[00108] As an example, a neural network may be trained using a set of parameters that includes a flow rate parameter. As an example, a neural network may be trained using a set of parameters that includes a weather parameter. As an example, a weather parameter may pertain to rainfall. For example, consider rainfall that may alter flow of fluid to a facility such as a water processing facility. As an example, a neural network may be trained using a set of parameters that includes a population parameter. For example, consider a municipal water treatment facility that may treat water where flow of water to the facility and/or out of the facility may correspond to a population of a region (e.g., a municipality, etc.).

[00109] As an example, a method can include receiving input values for a set of parameters associated with operations of a water processing facility and an output value for a parameter that is not a member of the set of parameters; and training a neural network based at least in part on the input values and the output value to provide a trained neural network. In such an example, the method may include storing the trained neural network to memory of a control unit. As an example, a method may include implementing a stored, trained neural network by a control unit to predict at least one output value for a given set of input values.

[00110] As an example, a method can include distributing a trained neural network via a network to a plurality of control units. For example, a neural network may be trained using information from a facility that includes a set of instruments (e.g., sensors, one or more controllers, etc.). In such an example, the trained neural network may be implemented at a facility that includes a different set of instruments. For example, such a trained neural network may operate as a "virtual" sensor and/or "virtual" controller (e.g., where a facility may not include a corresponding physical sensor and/or corresponding physical controller).

[00111] As an example, a control unit can include control circuitry that controls phosphate concentration based at least in part on output of a trained neural network.

[00112] As an example, a method can include training a recurrent neural network.

[00113] As an example, a method can include receiving input values for a set of parameters associated with operations of a water processing facility and an output value for a parameter that is not a member of the set of parameters where the input values and the output value correspond to a particular time. In such a method may include training a neural network using the values that correspond to the particular time. In such an example, the method may include receiving input values for the set of parameters and an output value for the parameter that is not a member of the set of parameters where the input values and the output value correspond to a different time.

[00114] As an example, a method can include training a neural network based at least in part on input values and output values for a plurality of times to provide a trained neural network that, for a given set of input values, predicts output values with respect to time.

[00115] As an example, a controller can include a processor; memory operatively coupled to the processor; and one or more modules stored in the memory that include processor-executable instructions to instruct the controller to process input values for a set of parameters associated with operations of a water processing facility using a trained neural network to output an output value for a parameter that is not a member of the set of parameters; and to control at least one of the operations of the water processing facility (e.g., consider one or more water treatment operations, etc.) based at least in part on the output value. In such an example, the instructions can include instructions to instruct the controller to control dosage of a flocculating agent and/or instructions to instruct the controller to control phosphate concentration.

[00116] As an example, a controller can include a processor; memory operatively coupled to the processor; and one or more modules stored in the memory that include processor-executable instructions to instruct the controller to process input values for a set of parameters associated with operations of a water processing facility using a trained neural network to output an output value for a parameter that is not a member of the set of parameters; and to control at least one of the operations of the water processing facility based at least in part on the output value. In such an example, the instructions can include instructions to instruct the controller to output a phosphate concentration value as the output value. In such an example, the instructions can include instructions to instruct the controller to control at least one of the operations to control an aqueous phosphate concentration of the water processing facility based at least in part on the phosphate concentration value. In such an example, the at least one of the operations can include a flocculating agent dosage operation where a dosage value of a flocculating agent of the flocculating agent dosage operation depends at least in part on the phosphate concentration value.

[00117] Although various examples of methods, devices, systems, etc., have been described in language specific to structural features and/or methodological acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described. Rather, the specific features and acts are disclosed as examples of forms of implementing the claimed methods, devices, systems, etc.

Claims

CLAIMS What is claimed is:

1 . A method comprising:

processing input values for a set of parameters associated with operations of a water processing facility using a trained neural network to output an output value for a parameter that is not a member of the set of parameters; and

controlling at least one of the operations of the water processing facility based at least in part on the output value.

2. The method of claim 1 wherein the input values for the set of parameters correspond to sensor measurements of sensors of the water processing facility.

3. The method of claim 1 wherein the output value for the parameter comprises a phosphate concentration.

4. The method of claim 1 wherein the controlling controls dosage of a flocculating agent.

5. The method of claim 1 wherein the controlling controls dosage of an iron compound.

6. The method of claim 1 wherein the controlling controls dosage of an aluminum compound.

7. The method of claim 1 wherein the trained neural network comprises a recurrent neural network.

8. The method of claim 1 wherein the set of parameters comprises a flow rate parameter.

9. The method of claim 1 wherein the set of parameters comprises a weather parameter.

10. A method comprising:

receiving input values for a set of parameters associated with operations of a water processing facility and an output value for a parameter that is not a member of the set of parameters;

training a neural network based at least in part on the input values and the output value to provide a trained neural network; and

storing the trained neural network to memory of a control unit.

1 1 . The method of claim 10 further comprising implementing the stored, trained neural network by the control unit to predict at least one output value for a given set of input values.

12. The method of claim 10 further comprising distributing the trained neural network via a network to a plurality of control units.

13. The method of claim 12 wherein the control units comprise control circuitry that controls phosphate concentration based at least in part on output of the trained neural network.

14. The method of claim 10 wherein the neural network comprises a recurrent neural network.

15. The method of claim 10 wherein the input values for the set of parameters and the output value for the parameter that is not a member of the set of parameters correspond to a particular time and further comprising receiving input values for the set of parameters and an output value for the parameter that is not a member of the set of parameters that correspond to a different time.

16. The method of claim 15 further comprising training the neural network based at least in part on the input values and the output values for the particular time and the different time to provide a trained neural network that, for a given set of input values, predicts output values with respect to time.

17. A controller comprising:

a processor;

memory operatively coupled to the processor; and

one or more modules stored in the memory that comprise processor- executable instructions to instruct the controller

to process input values for a set of parameters associated with operations of a water processing facility using a trained neural network to output an output value for a parameter that is not a member of the set of parameters; and

to control at least one of the operations of the water processing facility based at least in part on the output value.

18. The controller of claim 17 wherein the instructions comprise instructions to instruct the controller to output a phosphate concentration value as the output value.

19. The controller of claim 18 wherein the instructions comprise instructions to instruct the controller to control at least one of the operations to control an aqueous phosphate concentration of the water processing facility based at least in part on the phosphate concentration value.

20. The controller of claim 19 wherein the at least one of the operations comprises a flocculating agent dosage operation where a dosage value of a flocculating agent of the flocculating agent dosage operation depends at least in part on the phosphate concentration value.