Method for predicting maintenance operation and recommending maintenance of water treatment equipment

Abstract

The invention relates to an automated processing method for characterizing the state of a plurality of reverse osmosis membranes, the method comprising: acquiring (ACQ1) a first set of data (DATA1) from a sensor (20, 21, 22) arranged in the vicinity of a membrane (ENS1) receiving an incoming water flow (Qf) and producing a first effluent water flow (Qp), referred to as permeate water, and a second effluent water flow (Qc), referred to as concentrate; estimating a first operation lt of the first set of membranes (ENS1); sb / gt; an indicator (KPI1); an event timestamp (EVNi) related to maintenance of the membrane module; generating (GENA) an intermediate operating indicator (KPIiA) defining a point cloud (SERIE2A) corresponding to the value of the operating indicator (KPIi), for which point cloud (SERIE2A) the first set of membranes (ENS1) are considered to be new and / or clean; normalized operational indicators (KPIi ', KPIi0') are generated (GEN1) characterizing the state of the membrane used to filter the large volumes of liquid.

Description

Field of the Invention

[0001] The field of the invention is that of methods and systems for predicting and recommending maintenance operations in water production or treatment plants, such as desalination plants, and more generally, for filtering large volumes of water through dense membranes. More specifically, the invention relates to a method for predicting the maintenance and / or replacement date of a membrane module used in water treatment. The field of the invention relates to a method based on artificial intelligence algorithms, which aims to generate a normalized metric for predicting the maintenance and / or replacement date, in particular independently of the material or structural data of the components used. Background Art

[0002] Various techniques can be used to predict the maintenance and / or servicing of equipment such as membranes, especially reverse osmosis membranes, in plants such as desalination plants or other plants using such membranes. These predictions are necessary for the smooth operation of the plant and for maintaining system performance. Prediction is difficult, especially because of the life cycle of the membrane. The latter's life cycle is interrupted by external events related to operating conditions and is affected by environmental conditions such as temperature, water salinity, etc. These predictions become even more difficult because they also depend on the intrinsic properties of the membrane, which are not always known to the operator.

[0003] In terms of operating conditions, the membrane becomes clogged during operation and thus needs to be cleaned regularly. Cleaning is necessary to maintain good desalination performance, or more generally, to maintain good filtration performance. In fact, a contaminated membrane rapidly loses its filtration performance. Therefore, according to calculations, it is difficult to distinguish the causes of performance degradation related to membrane fouling or aging in the prediction results. Depending on the water flow pressure of the system, other parameters may affect the interpretation of aging, such as the pressure of the introduced water flow. These difficulties in analyzing the causes of performance degradation make prediction difficult and prevent the establishment of a reliable model.

[0004] In terms of environmental conditions, performance is highly dependent on changes in certain parameters, such as temperature or the concentration of water entering the system. Here again, current prediction systems also have difficulty distinguishing between temperature changes due to seasonal effects or performance degradation due to membrane aging.

[0005] In addition, membrane aging depends not only on operating configurations and environmental parameters but also on intrinsic properties. Each membrane has physical properties that ensure specific performance under given conditions. The new membrane installed after replacing the old membrane may not have the same intrinsic properties (e.g., due to changing membrane suppliers or introducing a new generation of membranes). These changes affect the prediction model based on manufacturer data. Therefore, current systems find that the models used are limited when the membrane type or even the entire operating phase changes.

[0006] Finally, the operating configuration itself also affects performance, depending on the size of the system, the number of processing stages, centralized routing and reprocessing circuitry, the conversion rate required by the system, and so on.

[0007] One difficulty in predicting the aging of a system, such as a membrane module or multiple stages consisting of different membrane modules, lies in the great heterogeneity of the components affecting the membrane service life and the multi-factorial causes of these components. In fact, the intrinsic properties of the membrane, water quality, temperature, the operating time of the membrane, and the operating configuration (including cleaning during the membrane life cycle) are all parameters that affect the membrane service life in different ways.

[0008] As a result, when certain operating configurations or hardware changes are made, current prediction systems find their reliability limited by the need to reconfigure the driving model. In fact, the specifications for replacing old membranes are not always described in a consistent manner, and testing and measurements are required to reconfigure the prediction model.

[0009] An object of the present invention is to overcome the above disadvantages. In particular, the present invention aims to propose a method for predicting the membrane replacement date that is automatic and does not require the prediction model to depend on the hardware configuration or physical characteristics of the equipment used.

[0010] To solve the problems that persist in the prior art modeling, the method of the present invention aims to standardize certain metrics for predicting membrane replacement and / or cleaning dates independently of operating or environmental conditions. Summary of the Invention

[0011] According to one aspect, the present invention relates to a method for automatically processing data characterizing the state of a plurality of membranes for filtering a large amount of liquid, the method comprising:

[0012] ■ Receiving a first set of data from state sensors arranged within or near a first set of membranes that receive an incoming water flow and produce a first outflow water flow called permeate water and a second outflow water flow called concentrate, the first set of data being related to external physical parameters, and the data acquisition being performed according to a plurality of first time series of data emitted by each sensor at a predetermined frequency;

[0013] ■ Determining at least one operating metric of the first set of membranes, the operating metric defining a second time series of calculated or estimated data;

[0014] ■ Recording the first data time series and the second data time series within a given acquisition time, each data time series defining a point cloud;

[0015] ■ Generate an intermediate operating metric that defines a point cloud corresponding to the value of the operating metric for which the first set of membranes is considered new and / or clean and / or pristine, the value of the intermediate operating metric being generated by applying a learned normalization model;

[0016] ■ Generate a normalized operating metric that characterizes the state of a plurality of membranes for filtering a large volume of liquid, the state characterizing fouling and / or aging of the state of the membranes and being independent of changes in environmental conditions, the normalized metric defining a third time series obtained from the operating metric and the intermediate operating metric.

[0017] An advantage is that a normalized metric can be generated to provide an indication of the state of the membranes in terms of aging and clogging, independent of any prior knowledge of their physical properties. Finally, the metric is independent of changes in environmental parameters. Such a metric can monitor changes in the membrane state and predict maintenance operations when appropriate.

[0018] In one embodiment, the intermediate operating metric value is generated by applying the learned normalization model and the operating metric value.

[0019] The value of the operating metric for which the first set of membranes is considered new and / or clean and / or pristine can be a value calculated from a function or a value measured from a sensor, or a value corrected by the learned normalization model. In the latter case, a new operating metric value is generated to produce the intermediate operating metric.

[0020] According to one embodiment, the normalization model is learned for each operating metric by regressing data of the second time series associated with the operating metric according to at least a first predetermined external physical parameter from the first time series, the regression being configured to determine a set of values within a factor of a minimum or maximum value of the second time series during a smoothing duration, the determined values corresponding to the configuration of new and / or clean and / or pristine membranes.

[0021] According to one embodiment, the normalization model is learned for each operating metric from a set of drive data of the operating metric during a smoothing duration that includes at least one maintenance and / or replacement operation of the first set of membranes.

[0022] In one embodiment, the drive data does not include data characterizing the physical properties of the membranes.

[0023] According to one embodiment, the determination of the operating metric includes determining a first operating metric that defines the pressure difference between the inlet and outlet of the membrane module and is represented in the form of a second time series of calculated or estimated data. The external physical parameters considered include at least one flow rate measurement value and one temperature measurement value, and these external physical parameters are used to calculate the value of an intermediate metric of the first operating metric based on a regression performed on the pressure difference value.

[0024] According to one embodiment, the determination of the operating metric includes determining a second operating metric that defines the pressure of the flow entering the first group of membranes and is represented in the form of a second time series of calculated or estimated data. The external physical parameters considered include at least one temperature measurement value, a concentration measurement value of the inflow entering the first group of membranes, the flow rate of the permeate flow from the first group of membranes, and the flow rate of the concentrate flow. These external physical parameters are used to calculate the value of an intermediate metric of the second operating metric based on a regression performed on the pressure value of the fluid entering the first group of membranes.

[0025] According to one embodiment, the determination of the operating metric includes determining a third operating metric that defines the permeate flow rate at the outlet of the first membrane module and is represented in the form of a second time series of calculated or estimated data. The external physical parameters considered include at least one temperature measurement value, a concentration measurement value of the inflow entering the first group of membranes, the flow rate of the permeate flow from the first group of membranes, and the flow rate of the concentrate flow. These external physical parameters are used to calculate the value of an intermediate metric of the third operating metric based on a regression performed on the flow rate value of the permeate flow leaving the first group of membranes.

[0026] According to one embodiment, the determination of the operating metric includes determining a fourth operating metric that defines the salt passage in the permeate water leaving the first membrane module and is represented in the form of a second time series of calculated or estimated data. The external physical parameters considered include at least one temperature measurement value, a concentration measurement value of the inflow entering the first group of membranes, the flow rate of the permeate flow from the first group of membranes, and the flow rate of the concentrate flow. These external physical parameters are used to calculate the value of an intermediate metric of the fourth operating metric based on a regression performed on the salt passage value in the permeate water leaving the first group of membranes.

[0027] According to one embodiment, the first set of data is related to the external physical parameters and includes:

[0028] ■ Measurements of the inflow rate, permeate flow rate, and / or concentrate flow rate, and / or

[0029] ■ Measurements of the conductivity of a certain volume of water, and / or

[0030] ■ Measurements of the total organic carbon, and / or

[0031] ■ The target value corresponding to the conversion rate of the feed water volume to the treated water volume, and / or

[0032] ■ The characteristic value of the influent flow rate (also referred to as "permeate flow rate"), and / or

[0033] ■ The characteristic value of the water permeability of the membrane.

[0034] According to one embodiment, the third time series corresponds to:

[0035] ■ The time series obtained by subtracting from the second time series the time series corresponding to the correction value generated by the learning normalization model; and / or

[0036] ■ The time series obtained by subtracting from the second time series the time series corresponding to the correction value generated by the learning normalization model, and a reference component has been added to the time series, the reference component corresponding to the time series of the operating metrics corresponding to the state of a new and / or clean and / or pristine membrane, the component being calculated under average or standard environmental conditions.

[0037] According to one embodiment, a regression of the operating metrics is performed based on a plurality of external physical parameters on which the operating metrics depend.

[0038] According to one embodiment, the regression is implemented by a first learning function, the first learning function including a machine learning model having parameters learned by implementing a loss function.

[0039] According to one embodiment, the regression is an expected value regression, the regression being performed based on an expected loss function and an error function between the value of the operating metrics and the value estimated by the regression model, the regression model being for operating metric values considered within a given expectation of the operating metric value distribution. An advantage is that the value of the operating metrics for which the state of the membrane is new, clean or pristine can be determined.

[0040] According to one embodiment, a regression is performed on the data of the second data series of the operating metrics based on a plurality of predefined external physical parameters of a plurality of first time series obtained by a plurality of sensors, the regression being performed based on a generalized additive model modeling function, within a range of predefined expected values and for a given value of the external physical parameter, by seeking to optimize the parameters of the generalized additive model through an expected loss function between the calculated value of the operating metrics and the estimated value of the operating metrics, the regression further simulating the error function, and the regression running over a so-called smoothing duration, the regression generating a set of values defining a point cloud of the intermediate metrics, the set of values corresponding to the new and / or clean and / or pristine state of the first set of membranes.

[0041] According to one embodiment, the smoothing duration is determined to include a plurality of event markers related to the maintenance of the membrane module, and the smoothing duration is less than the acquisition duration.

[0042] According to one embodiment, the method includes time markers for events related to membrane module maintenance, the events corresponding to cleaning and / or replacement, and each time marker is executed according to a marker time reference within the acquisition duration. An advantage is that maintenance operations can be considered in the training of the learning function.

[0043] According to one embodiment, the method includes generating a value of a predicted operation metric by applying a second learning function, the second learning function being trained from the values of the normalized operation metrics corresponding to a third time series considered within a prediction duration, and the second learning function generates prediction data for the evolution of the normalized metrics. An advantage is that maintenance operations intended to clean the membrane are anticipated.

[0044] According to one embodiment, the method includes predicting a set of values for each normalized metric, the training data being selected between the last two event timestamps respectively associated with two consecutive cleanings, and new training of the second learning function is triggered after each new event associated with cleaning. An advantage is to create a training data set unaffected by maintenance operations in order to obtain only the evolution related to the state of the membrane.

[0045] According to one embodiment, the method includes comparing at least one predicted value of the normalized metric with at least one predetermined threshold, and the comparison enables the generation of a cleaning date. An advantage is that an alert can be generated for the operator.

[0046] In one embodiment, the predetermined threshold is a variable threshold, and the value of the variable threshold is generated by executing a function dependent on predetermined parameters.

[0047] According to one embodiment, the second learning function is a function implementing a second generalized additive model.

[0048] According to one embodiment, it includes using a third learning function to calculate a set of aging indices of the membrane, the third learning function including a set of training data, the set of training data including values extracted from the first set of data for estimating the metric, the training data being selected during the acquisition period and taking into account the timestamps of events occurring during the acquisition period. Analysis of the aging indices is used to determine the membrane replacement date.

[0049] According to one embodiment, the third learning function is a recurrent neural network and has a regression function based on an autoregressive method.

[0050] According to another aspect, the present invention relates to a data processing system, which includes a computer, a memory, a clock, and a communication interface for receiving data in the form of a time series. The data processing system includes a communication interface for transmitting data to a server, and the data processing system includes a server for performing the method of the present invention to calculate the steps of the normalized operation index.

[0051] In one embodiment, the system includes a display for generating a real-time representation of at least one normalized index. One advantage of this is that the evolution of the normalized index can be tracked in real time.

[0052] According to another aspect, the present invention relates to a system for treating a certain volume of feed water into a certain volume of filtered treated water through a plurality of membrane modules. The water treatment system includes an inlet for receiving the water flow into at least one given membrane module, a first filtered water outlet called permeate water, and a second residual water outlet called concentrate. The water treatment system further includes a set of external parameter status sensors, and the set of external parameter status sensors includes a water temperature sensor and at least one pressure sensor. The water treatment system includes a data processing system according to the present invention.

[0053] In one embodiment, the system for treating a certain volume of feed water includes a plurality of membranes organized into a plurality of membrane modules, and each membrane module defines a stage (French: étage) for treating the input volume of water and generating an output stream. One advantage is that the normalized index control can be configured for the sub-components of the system. In one embodiment, the system includes a composite index, and the composite index includes different components related to the different normalized operation indexes of these different sub-components.

[0054] In one embodiment, the system for treating a certain volume of feed water includes at least one second set of membranes arranged at the outlet of the first set of membranes, wherein the concentrate of the first set of membranes defines the inlet of the second set of membranes.

[0055] In one embodiment, the system for treating a certain volume of feed water includes at least one third membrane module arranged in parallel with the first membrane module, and the concentrates from the first membrane module and the third membrane module define the inlet of the second membrane module.

[0056] According to one embodiment, the first learning function includes the implementation of a first generalized additive model that takes into account a third time series exposed according to a plurality of predefined external physical parameters.

[0057] According to one embodiment, the smoothing duration is selected to correspond to the maximum value of the aging index of at least one membrane or the time specific to its lifespan.

[0058] According to one embodiment, the estimation of the first operation index of the first set of membranes results in:

[0059] ■At least one calculation based on data from the first data set;

[0060] ■At least one measurement value from a sensor located near the membrane module.

[0061] According to one embodiment, the normalization index further includes:

[0062] ■Normalization index characteristics of the feed or output flow rate; and / or

[0063] ■Normalization index characteristics of the conductivity of a certain volume of water at the inlet and / or outlet of the first group of membranes; and / or

[0064] ■Normalization index related to the measurement value of the salt passage amount between the inlet and outlet of the first group of membranes; and / or

[0065] ■Normalization index characteristics of the supply pressure.

[0066] According to one embodiment, the normalization index characteristics of the feed flow to at least the first group of membranes include:

[0067] ■A specific feed flow rate parameter, which represents the volume of water entering the first membrane module per unit time and unit membrane area when pressure is applied to the described volume of water, and / or;

[0068] ■A parameter related to the first output flow rate, which represents the volume of water filtered at the outlet of the first membrane module per unit time and unit membrane surface area when pressure is applied to the described amount of water at the inlet, called permeate water; and / or;

[0069] ■A parameter related to the second output flow rate, which represents the residual volume of water per unit time and unit membrane surface area at the outlet of the first group of membranes when pressure is applied to the described amount of water at the inlet, called concentrate; and / or;

[0070] ■A parameter related to the inlet supply pressure of the diaphragm module.

[0071] According to one embodiment, the normalization index characteristics of the conductivity of a certain volume of water entering or leaving at least the first group of membranes include:

[0072] ■A parameter related to the conductivity of the permeate water leaving the first group of membranes;

[0073] ■A parameter related to the permeability coefficient of the first group of membranes,

[0074] ■A parameter related to the measurement value of the difference in conductivity of a certain volume of water at the inlet and outlet of the first group of membranes.

[0075] According to one embodiment, the method includes suppressing raw data from at least one sensor when an acquired value is below a predetermined threshold.

[0076] According to one embodiment, the first pre-training of the first learning function is performed based on a set of training data and correction parameters of a seasonal or temperature threshold, and the second training of the first learning function is performed when defining the membrane module, the number of membrane modules, and operating parameters for a smoothing duration. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] With reference to the accompanying drawings, other features and advantages of the present invention will become apparent from the following detailed description, the drawings being described as follows:

[0078] Figure 1 : A schematic diagram showing various stages of the method of the present invention;

[0079] Figure 2 : An example of membrane fouling in a system with a series of continuous cleaning, and showing the effect of long-term membrane aging;

[0080] Figure 3 : An example of the system of the present invention, the system including a plurality of sensors and data processing means for generating a normalized index according to the method of the present invention;

[0081] Figure 4 : An example of a set of membranes simulating the stages of processing an input stream and generating two output streams including permeate and concentrate;

[0082] Figure 5 : An example of a water treatment system according to the present invention, including a plurality of treatment stages in which different membrane modules are implemented;

[0083] Figure 6 : An example of the evolution of the operating metrics of a set of membranes, representing a first evolution curve of the differential pressure of a set of membranes, wherein the seasonal effects and the evolution trend of aging are concerned;

[0084] Figure 7 : A representation of the operating metric (here the differential pressure) according to the expected graph, such that the operating metric can be represented according to a physical parameter (here the temperature); note the curve representing the minimum envelope in the figure, which curve enables the generation of a set of correction values used in the method of the present invention;

[0085] Figure 8 : A diagram of a second curve showing the evolution of the corrected operating metric of the set of membranes considered when the membrane set is in a substantially clean state; and showing a first curve,

[0086] Figure 9 : Figure 9Representation of the difference between the two curves shown in [FIGURE], in which a reference curve is added to show the corrective operating indicators independent of the operating and environmental conditions, in particular to restore the effect of the aging of the membrane module over time. Detailed implementation

[0087] Definition

[0088] In the remainder of this specification, a "data processing system" is referred to as a system that includes the means required to perform the method steps, i.e., at least one computer and a memory. However, according to a preferred mode, the system includes:

[0089] ■ Local software means, which are co-located in the processing plant to process the data obtained from the sensors, and which may be a computer configured as a local server;

[0090] ■ Remote means, such as at least one remote data server, for performing the method steps that result in the generation of normalized indicators and the prediction of the plant intervention date.

[0091] Hereinafter, we refer to a "plant (French: usine)" or a "water treatment system" as a set of physical means for treating a certain volume of water and measuring operating or environmental parameters. The plant includes at least one dense spiral membrane module, for example, for reverse osmosis applications. A set of membranes is used to treat a large volume of influent water delivered through an inlet and to produce at least two water streams, called "permeate (French: "perméat", also translatable as "permeate liquid")" and "concentrate (French: "concentrat", also translatable as "concentrate", "concentrate liquid")", delivered through outlets. Generally, the water treatment system includes sensors, hydraulic means, and a data processing system. According to the configuration, the hydraulic means include valves, such as balance valves, shut-off valves or throttle valves, control valves, possible turbines or microturbines, or any other device for controlling, regulating, and guiding the flow of a fluid such as water. The water treatment system also includes hydraulic equipment, such as pumps, water tanks, containers, filters, pipes, and any other equipment required for the operation of the plant.

[0092] The water treatment plant or system may include one or more filtration channels for treating the introduced water volume, and may include one or more stages that define an arrangement of membrane modules installed in series or in parallel and have a conversion rate for treating the introduced water volume. Generally, one stage treats 40% to 50% of the pumped water volume. To increase the proportion of water treated, several stages may be installed in series to achieve a conversion rate of 75% to 85%.

[0093] The filtration channel (French: “passe de filtration”) filters a given volume of water at a given characteristic operating pressure. In a water desalination system, filtration is typically carried out at least twice.

[0094] In the remainder of this specification, we will refer to the concentration as the concentration of salts in the fluid, i.e., for example, the concentration of all minerals present in a given volume of water. When measuring the salinity of a given volume of water, a concentration measurement or calculated value can be obtained by measuring the conductivity of the given volume of fluid under discussion. In the following description, conductivity is a measure of the fluid conductivity, reflecting the presence of conductive elements in the fluid.

[0095] The concentration can be obtained from the conductivity measurement of the fluid because they correspond to equivalent quantities. Concentration and conductivity measure equivalent physical properties of the fluid. One of these quantities can be obtained from the measurement of the other quantity by a simple ratio (e.g., a coefficient or a constant). For this purpose, constants characterizing the water quality (e.g., surface water or seawater or tap water) can be used to convert the conductivity measurement into a concentration value. In addition, when a parameter (e.g., temperature) affects the conductivity, this effect can be compensated for by using a model (e.g., a linear model) to infer the concentration.

[0096] Figure 5 An example of a structure including different stages ET1, ET2, and ET3 is shown, which uses a plurality of membrane assemblies ENS arranged in different configurations in parallel or in series, depending on the stage to which they belong. 11 、ENS 12 、ENS2, and ENS3. In this embodiment, it is a 3-stage series connection, and the first stage includes two groups of membranes in parallel. At the outlet of each membrane assembly, the concentrated streams Qc1, Qc2 are reinjected into the next stage. The permeate streams Qp1, Qp2, Qp3 can be directed to other treatment stages or directly to operation.

[0097] Sensors can be used to measure environmental data, such as water temperature, water pressure, water volume salinity, or any other water volume quality parameter. Other sensors can be used to measure physical plant parameters, such as equipment consumption level, inflow or outflow flow rate, pressure difference, or event detection sensors.

[0098] In the remainder of this specification, we will refer to the plant stage as a plant sub-assembly that includes at least one membrane assembly having one connection or channel for receiving the input stream of water to be treated, and two output connections or channels for producing the following two output streams: permeate water, which is a treated stream corresponding to a salinity lower than the salinity of the input stream; and concentrated water, which is a water stream corresponding to a salinity at least equal to the input stream.

[0099] Figure 1shows the main stages of the method of the present invention. According to an example of the structure of a water treatment system capable of implementing the method of the present invention, the first step includes obtaining a set of data ACQ1 from Figure 3 the various sensors 20, 21, 22 shown.

[0100] The sensors are preferably arranged inside or near the water treatment system so as to be as close as possible to the membrane and measure the environmental conditions to which the membrane is subjected.

[0101] According to one embodiment, the data processing system of the present invention includes software means (such as a computer and a memory) for executing a computer program that implements the steps of the method of the present invention. The computer program includes software instructions that, when executed, are capable of implementing the steps of the method of the present invention.

[0102] Application

[0103] According to a first application, the method of the present invention relates to the field of reverse osmosis for filtering a certain volume of water. The method relates both to the field of membranes for reverse osmosis in the case of filtering the salt content of a certain volume of water and to so-called low-pressure reverse osmosis for uses other than the desalination of a certain volume of water.

[0104] The method of the present invention is particularly applicable to generating a normalization indicator for organic membranes, that is, membranes made of organic polymers (such as polyamides), called "dense membranes" in a spiral wound configuration. These membranes are particularly used in reverse osmosis or nanofiltration applications. However, the present invention is not limited to dense membranes.

[0105] According to a second application, the method of the present invention relates to the field of membrane nanofiltration. In this case, the process is applicable to membranes configured to separate molecules in a certain volume of liquid (such as water or blood).

[0106] According to a third application, the method of the present invention relates to the field of membrane ultrafiltration using dense membranes.

[0107] In the remainder of the specification, the present invention will be described for the application of reverse osmosis. However, the method of the present invention relates to any other field related to the separation of particles or elements from a certain volume of liquid using a membrane.

[0108] Modeling

[0109] Figure 2 shows a schematic example of the reasons affecting the evolution of the state of one or more membranes. For this purpose, the evolution of characteristic indicators is shown on a graph, namely the differential pressure DP. This graph illustrates the various reasons for the change in this indicator, including:

[0110] ■ The fouling of these membranes is denoted as Fo, and the maintenance operations aimed at cleaning them are represented here by the NET1 cleaning operation;

[0111] ■ Intrinsic aging of the membrane, which degrades its physical properties over time, represented by the line marked Tr and showing the aging trend, and finally;

[0112] ■ Effects related to external physical parameters, including operating components related to the plant architecture, operating variables, and environmental components related to water quality and temperature.

[0113] This representation helps to better understand the different cycles that the membrane undergoes and the reasons leading to fouling. One of the aims of the method according to the invention is to isolate some of these causes in order to better predict future maintenance operations related to membrane fouling and replacement.

[0114] Figure 2 The differential pressure of a set of membranes is shown on the ordinate and time is shown on the abscissa. The line marked BLNS represents the evolution of the differential pressure DP of a set of membranes as a function of external physical parameters. The line marked Fo represents the change in differential pressure due to membrane fouling, and the line marked Tr represents the change in differential pressure due to membrane aging. It is this last component that enables us to establish a reliable predictive model of real membrane aging.

[0115] Figure 6 The differential pressure value DP defining the first operating indicator KPI1 is shown. The differential pressure value DP is recorded as raw data over several months or years. Due to the influence of the water temperature varying over time, this graph allows for the observation of large variations in DP over the years. This influence gives the data a wavy pattern. However, on average, the operating indicator of the differential pressure DP increases from the first year to the fifth year, and even the Tr line increases, indicating a trend of membrane aging and / or blockage. This fouling trend is not obvious during operation because the temperature-related variations are much larger than those caused by irreversible fouling.

[0116] Environmental data acquisition

[0117] The method includes the step of receiving data from sensors arranged in or near the facility, the data being represented as AQC1 in Figure 1 in order to measure the value of an environmental parameter (i.e., the so-called external physical parameter) related to the installation of a plurality of membranes for treating or filtering an incident volume of water. The sensors can include temperature probes, probes such as conductivity probes, flow meters, and pressure sensors, etc.

[0118] In one embodiment of the present invention, the data acquired by the sensor is stored in a memory. The data is collected and stored in the form of a time series. Therefore, the data is preferably timestamped. The method of the present invention involves a first step that includes reading the recorded data from the sensor. However, according to one embodiment, the method of the present invention may include a preliminary step of acquiring the sensor. In the case where the method of the present invention is implemented by a computer or a plurality of computing units, the method does not have to include the preliminary acquisition step, which can be separated from the implementation of the method of the present invention, because the latter can be performed a posteriori at a certain time period after the acquisition.

[0119] The recorded data regarding the external physical parameters at least includes the temperature value Tf of a certain volume of water entering the stage including a set of membranes, and the pressure value Pf of the same volume of the incoming water. These values are preferably measured regularly by at least one temperature sensor.

[0120] These values can also be measured at the permeate water {Tp, Pp} or the concentrate water {Tc, Pc} level.

[0121] According to one embodiment, other external physical parameter values are recorded and evaluated by the method of the present invention. In particular, the values of the inlet flow rate Qf and the concentration Cf of the introduced volume are measured. The inlet flow rate is also referred to as the feed flow rate Qf. Figure 4 The inlet flow rate Qf of the ENS1 membrane module and the outlet flow rates Qp and Qc of the permeate and the concentrate are shown respectively.

[0122] According to one embodiment, the external physical parameter values at the outlet of the membrane module are recorded and evaluated by the method of the present invention. These can be the external physical parameter values of the permeate water or the concentrate water. The physical parameters of the permeate water are denoted as Tp, Qp, Cp, Pp respectively, and the physical parameters of the concentrate water are denoted as Tc, Qc, Cc, Pc.

[0123] The general measured values of the external physical parameters Ti, Qi, Ci, Pi are recorded below, and these values can be specified by the indices f, p, c according to their measurement points, depending on whether the parameter is measured upstream of the membrane module or downstream at the permeate water or the concentrate water level.

[0124] An advantage is to measure the same parameter at different measurement points and calculate specific differences, such as the pressure difference DP or the salt flux or retention (also referred to as the concentration difference).

[0125] It should be noted that the time series can be collected, recorded, and evaluated at different frequencies and at different acquisition times. In one embodiment, the method of the present invention includes any preliminary steps aimed at oversampling or undersampling the time series so that when they are used jointly by mathematical operations or algorithms, the number of values of each time series is made uniform.

[0126] Each time series containing all values of each external physical parameter is called SERIE1. Thus, there are as many SERIE1 time series as there are time series of external physical parameter values.

[0127] When acquisitions are made during different acquisition periods, operations can be performed to evaluate the data for the same period.

[0128] Determine metrics

[0129] The method of the present invention enables the storage of measured or calculated data from a second SERIE2 time series corresponding to the KPI A during a time period of the acquisition period i Value of the operating metric. In its simplest implementation, the method of the present invention involves the operation of a single operating metric, such as the first operating metric KPI1. In other embodiments, the method of the present invention is implemented to utilize multiple operating metrics, such as the four operating metrics mentioned above: KPI1, KPI2, KPI3, KPI4. According to various embodiments, combinations of these metrics are utilized, such as the first and third KPI1, KPI3 or other combinations. The selection of operating metrics can be advantageously utilized in order to produce a normalized metric that will be used to train a learning function (i.e., a machine learning model) for the purpose of short-term or medium-term prediction. The metrics are selected based on the assumed impact of plant configuration, environmental conditions, and changes in operating parameters. Thus, different variants of the method of the present invention can be implemented according to the plant configuration.

[0130] The various stages of the method of the present invention include using operating metrics to achieve several objectives:

[0131] - Define a normalization model;

[0132] - Determine intermediate metrics and normalized metrics;

[0133] - Define a prediction model;

[0134] - Determine short-term and / or long-term predictions to predict the date of membrane maintenance operations.

[0135] To meet all these objectives, various metrics are defined as part of the inventive process.

[0136] KPI i : Represents an operating metric obtained with the actual environmental conditions obtained during acquisition, such as temperature Ti, pressure Pi, flow rate Qi, etc.

[0137] KPI i ’: Represents a normalized operating metric that is independent of environmental conditions and mainly reflects the fouling or aging state of the membrane for predicting maintenance operations.

[0138] KPI i0 ’: represents the normalized operation index, corresponding to the index calculated for the reference environmental conditions, mainly reflecting the pollution or aging state of the membrane to predict maintenance operations.

[0139] KPI i ^: represents the membrane operation index estimated by the loss function in the regression step, which is obtained from the actual environmental conditions during the acquisition process, such as temperature Ti, pressure Pi, and flow rate Qi, etc.

[0140] KPI iA : represents the operation index of a new or clean membrane under the actual environmental conditions obtained during acquisition, such as temperature Ti, pressure Pi, flow rate Qi, etc. It is also called the intermediate index.

[0141] KPI i0 : represents the membrane operation index obtained using the reference environmental conditions obtained with reference operation parameters, such as temperature T0, pressure P0, and flow rate Q0, etc.

[0142] KPI iA0 : represents the operation index of a new or clean membrane obtained under the reference environmental conditions obtained from the reference operation parameters, such as temperature T0, pressure P0, and flow rate Q0, etc.

[0143] KPI ip1 : represents the membrane operation index predicted by the first prediction model, which is trained based on the normalized operation index, and this prediction is a short-term prediction.

[0144] KPI ip2 : represents the membrane operation index predicted by the second prediction model trained on the normalized operation index, and this prediction is a long-term prediction.

[0145] The method of the present invention includes determining at least one operation index (denoted as KPI i ), and this step is denoted as EST1 in Figure 1 . This step is preferably executed by a computer, which can be the local computer K1 shown in Figure 1 , or a computer on a remote server represented by SERV1. A memory is shown to store the data generated during the calculation of the operation index KPI i and the normalized operation index KPI i ’, as well as possible intermediate correction values or prediction values, such as the value of the reference curve operation index KPI iA or the estimated value of the operation index KPI i ^, or the prediction values KPI ip1 of the normalized index monitoring prediction model, KPI ip2 .

[0146] An indicator can be defined based on a time series defined over a collection period different from the first SERIE1 time series. In this case, the SERIE1 and SERIE2 series can be processed within the same time period. This may involve, for example, defining a common time window.

[0147] KPI1(DP)

[0148] According to a first example, the first operating indicator KPI1 corresponds to a differential pressure labeled DP.

[0149] This differential pressure DP is generated by the pressure difference between the volume entering the first membrane module and the outlet volume (e.g., the concentrate). The pressure difference between the inlet and the permeate can also be measured.

[0150] According to the first example, the differential pressure DP can be calculated from certain operating parameters, especially those measured by sensors. For this purpose, the function f1 can be used to model this differential pressure as a function of temperature and the average flow rate Qn, where Qn = (Qc + Qf) / 2N. Here, N corresponds to the number of tubes including the membrane modules, and each tube forms an arrangement of a spiral wound membrane. Qc is the concentrate flow rate, and Qf is the feed flow rate. KP1 = DP = f1(Qn, Ti).

[0151] The method of the present invention includes steps capable of obtaining the operating indicator KPI1 from the measurements of temperature Ti and flow rates Qc and Qf.

[0152] According to other examples, other external physical parameters can be considered to calculate or simulate the influence of these parameters on the evolution of the indicator.

[0153] According to a second example, the differential pressure DP can be directly measured from at least one differential pressure sensor or a plurality of pressure sensors arranged upstream and downstream of the diaphragm module.

[0154] The first indicator KPI1 can advantageously take the form of a SERIE2 time series. In this case, each calculated or measured value of the differential pressure DP is associated with a date. The date of each operating indicator value can be set to be equal to the date of each operating parameter value of the first SERIE1 time series used in calculating the first operating indicator KPI1.

[0155] When the differential pressure DP is directly measured by a sensor, in order to maintain the date consistency between the first SERIE1 time series and the second SERIE2 time series, a common clock can be used, which is used to timestamp other physical parameters measured.

[0156] The method of the present invention generates an indicator KPI representing an equivalent differential pressure DP’ 1A , which indicator KPI 1Ais a reconstructed indicator corresponding to the differential pressure of a new membrane or a membrane bank cleaned by a cleaning operation. This indicator is obtained under the actual measurement conditions of external physical parameters. Then, this indicator will be used to obtain the normalized indicator KPI i0 ’, and this normalized indicator KPI i0 ’ gives the membrane state under reference environmental conditions.

[0157] One objective is to calculate the value of the first indicator KPI1 for different measurements of temperature Ti and average flow rate Qn, so as to estimate the evolution of the first intermediate indicator KPI N1 through the first learning normalization model MOD 1A . The first normalization model MOD N1 is learned, for example, by regression. Due to the present invention, this first intermediate indicator KPI 1A can then be used to calculate the first normalized indicator KPI1’ or KPI 10 ’, which can more reliably reproduce the effects of membrane aging and fouling.

[0158] One advantage of this first KPI i ’ indicator is that it helps to monitor the longitudinal blockage of the membrane module.

[0159] The method of the present invention enables the definition of multiple operating indicators {KPI i} i[1 ; k] and the related normalized operating indicators {KPI i '} i[1 ; k] to provide information about the membrane state, especially information about membrane blockage and aging. To obtain reliable, robust and environment data-independent (especially season impact) indicators, the method of the present invention enables the normalization of {KPI i '} i[1 ; k] operating indicators.

[0160] KPI2(Pf)

[0161] The second indicator KPI2 is defined by operating parameters related to the incident flow pressure Pf (also called feed pressure Pf) applied to the first group of ENS1 membranes. The incident flow pressure Pf can be calculated according to the measurement of the model and the physical parameters of the model, or directly measured from at least one pressure sensor arranged at the inlet of the first group of ENS1 membranes.

[0162] When calculating this second KPI2 from the model, the incident flow pressure Pf can be modeled according to the function f2 of the following environmental parameters: the temperature Ti of the first set of ENS1 membranes (e.g., one stage or channel of a processing device including the first set of ENS1 membranes), the incoming flow concentration Cf, the permeate water flow rate Qp, and the concentrate flow rate Qc. According to other examples, other external physical parameters can be considered to calculate or simulate the influence of these parameters on the evolution of this metric.

[0163] We obtain the following expression: KPI2 = Pf = f2(Ti, Cf, Qp, Qc).

[0164] The example described refers to 4 parameters selected to influence the second metric, but other environmental parameters can also be considered within the framework of the present invention. The present invention is capable of considering at least one parameter that influences the evolution of the second metric.

[0165] Thus, the method of the present invention makes it possible to obtain the second operating metric KPI2 from the measurements of the temperature Ti, concentration Cf, permeate flow rate Qp, and concentrate flow rate Qc of the first ENS1 membrane module.

[0166] According to a second example, the pressure of the incoming flow Pf can be directly measured from one or more pressure sensors arranged upstream of the first set of ENS1 membranes.

[0167] The second KPI2 metric can advantageously take the form of a second SERIE2 time series. It should be remembered that each time series corresponding to an operating metric is labeled SERIE2, although the time series may vary according to the metric selected. In this case, each calculated or measured value of the incoming flow pressure Pf is associated with a date. The date of each operating metric value can be considered equal to the date of each operating parameter value of the first SERIE1 time series used in the calculation of the second operating metric KPI2.

[0168] Similar to the first KPI1 metric, when the incoming flow pressure Pf is directly measured by at least one sensor, a common clock can be used, which is used to time-stamp the other measured physical parameters in order to maintain the date consistency between the first SERIE1 time series and the second SERIE2 time series.

[0169] The method of the present invention generates an intermediate indication KPI 2A ′ representing an equivalent incoming flow pressure Pf′, and this intermediate indication KPI 2A ′ is a reconstructed indication whose value corresponds to the incoming flow pressure when the membrane module is new or clean or has been cleaned by a cleaning operation.

[0170] One objective is to calculate the value of a second operating key performance indicator KPI2 for different measured values of external physical parameters, in order to estimate an intermediate key performance indicator KPI N2 through a second learning normalization model MOD 2A for its evolution. For example, through regression, learn the second normalization model MOD N2 . Then, this intermediate key performance indicator KPI 2A can be used to calculate a normalized key performance indicator KPI2' or KPI 20 ', which more reliably reproduces the effects of membrane aging and fouling.

[0171] One advantage of this second KPI2 is that it helps monitor the energy consumption of all ENS1 membranes.

[0172] KPI3 (permeate flow rate)

[0173] The third KPI3 is defined by operating parameters related to the permeate flow rate Qp at the outlet of the first group of ENS1 membranes. The permeate flow rate Qp can be calculated from the measurements of the model and the physical parameters of the model, or directly measured by a sensor arranged at the permeate outlet of the first group of ENS1 membranes.

[0174] This indicator can also be expressed by the specific flux SP. It represents the volume of water produced per unit time and per unit membrane surface area when a given pressure is applied to the feed water. It is an indicator of the ability of the membrane to produce a given volume of output water. As the membrane ages or degrades, the SP specific flux tends to decrease. A decrease in SP is an indicator of membrane material fouling or deterioration. However, like the indicator corresponding to the permeate flow rate Qp, this indicator is also affected by temperature, water salinity, and water temperature. According to other examples, other external physical parameters can be considered as factors affecting the evolution of this indicator.

[0175] When calculating this third KPI3 indicator from the model, the permeate flow rate Qp can be modeled as a function f3 of the following environmental parameters: the temperature Ti of the first group of ENS1 membranes (e.g., a stage or channel of a treatment device including the first group of ENS1 membranes), the inlet stream concentration Cf, the permeate flow rate Qp, and the concentrate flow rate Qc. According to another embodiment, other external physical variables can be considered in the KPI3 modeling.

[0176] We obtain the following expression: KPI3 = Qp = f3(Ti, Cf, Qp, Qc).

[0177] The examples described refer to 4 parameters selected as affecting the third indicator, but other environmental parameters can also be considered within the framework of the present invention. The present invention enables at least one parameter affecting the evolution of the third indicator to be considered.

[0178] Thus, the method of the present invention enables the third operating indicator KPI3 to be obtained from the measurements of the temperature Ti of the first set of ENS1 membranes, the concentration Cf of the incoming stream, the flow rate Qp of the permeate stream, and the flow rate Qc of the concentrate stream.

[0179] According to the second example, the permeate flow rate Qp can be directly measured by at least one sensor of a device arranged downstream of the first set of ENS1 membranes.

[0180] The third indicator KPI3 can advantageously take the form of a second time series SERIE2. In this case, each calculated or measured value of the permeate flow rate Qp is associated with a date. The date of each value of the third operating indicator KPI3 can be considered equal to the date of each operating parameter value of the first time series SERIE1 used in the calculation of the third operating indicator KPI3.

[0181] Similar to the first and second KPI1, KPI2 indicators, when the permeate flow rate Qp is directly measured by at least one sensor, a common clock can be used, which is used to time-stamp other measured physical parameters to maintain the date consistency between the first SERIE1 time series and the second SERIE2 time series.

[0182] The method produces a third intermediate indicator KPI 3A ’, to represent an equivalent permeate flow rate Qp equivalent Qp’, which is a reconstructed indicator whose value corresponds to the permeate flow Qp entering the membrane module when the membrane module is new or clean or cleaned by a cleaning operation.

[0183] An object is to calculate the value of the third operating indicator KPI3 for different measured values of external physical parameters, so as to estimate the evolution of the intermediate indicator KPI N3 through the third learning normalization model MOD 3A . For example, the third normalization model MOD N3 is learned by regression. Then, this intermediate indicator KPI 3A is used to calculate the normalized indicator KPI3’ or KPI 30 ’, which more reliably reproduces the effects of membrane aging and fouling.

[0184] One advantage of this third KPI3 (which may be combined with other data) is that it helps to monitor the transmembrane blockage of the ENS1 membrane module.

[0185] KPI4 (salt passage)

[0186] The fourth operating indicator KPI4 is defined by operating parameters related to the salt passage rate SP, expressed as the percentage of salt filtered into the concentrate SPc or the residue in the permeate SPp at the outlet of the first set of ENS1 membranes. The salt passage rate can be expressed, for example, as the concentration ratio between the inlet and the outlet. In this example, we consider the salt passage rate in the permeate SPp. The salt passage rate SPp can be calculated from the measurements of the model and the model physical parameters, or directly measured by a sensor arranged at the permeate outlet of the first ENS1 membrane module. This indicator can also be expressed by the conductivity of the permeate.

[0187] According to other examples, other external physical parameters can be considered to calculate or simulate the influence of these environmental parameters on the evolution of this indicator.

[0188] When calculating this fourth indicator KPI4 from the model, the salt passage rate in the permeate SPp can be modeled as a function f4 of the following environmental parameters: the temperature Ti of the first set of ENS1 membranes (for example, the stage or channel of the treatment plant (French: l’usine de traitement) including the first set of ENS1 membranes), the inlet stream concentration Cf, the permeate flow rate Qp, and the concentrate flow rate Qc.

[0189] We obtain the following expression: KPI4 = SPp = f4(Ti, Cf, Qp, Qc).

[0190] The described example refers to 4 parameters selected to affect the fourth indicator, but other environmental parameters can also be considered within the framework of the present invention. The present invention enables at least one parameter affecting the evolution of the fourth indicator to be considered.

[0191] Therefore, the method of the present invention enables the fourth operating indicator KPI4 to be obtained from the measurements of the temperature Ti, the inlet stream concentration Cf, the permeate flow rate Qp, and the concentrate flow rate Qc of the first set of ENS1 membranes.

[0192] According to the second example, the salt passage rate in the permeate SPp can be directly measured from at least one sensor or device arranged downstream of the first ENS1 membrane module.

[0193] The fourth indicator KPI4 can advantageously take the form of a second time series SERIE2. In this case, each calculated or measured value of the salt passage rate in the permeate SPp is associated with a date. The date of each value of the fourth operating indicator KPI4 can be considered equal to the date of each operating parameter value of the first time series SERIE1 used in the calculation of the fourth operating indicator KPI4.

[0194] Identical to the first, second, and third KPIs KPI1, KPI2, KPI3, when the salt passage in the permeate water SPp is directly measured by at least one sensor or device, a common clock can be used, which is used to time-stamp other measured physical parameters in order to maintain the date consistency between the first SERIE1 time series and the second SERIE2 time series.

[0195] The method of the present invention generates a fourth intermediate KPI representing the salt passage in the equivalent permeate water SPp' 4A ', which is a reconstructed index whose value corresponds to the salt passage Sp obtained by the membrane module when the membrane module is new or clean or cleaned by a cleaning operation.

[0196] One objective is to calculate the value of the fourth operating KPI4 for different measured values of external physical parameters in order to estimate the evolution of the intermediate KPI N4 through a fourth learning normalization model MOD 4A For example, the fourth normalization model MOD N4 is learned through regression. Then, this intermediate KPI 4A is used to calculate the normalized index KPI4' or KPI 40 ', which more reliably reproduces the effects of membrane aging and fouling.

[0197] One advantage of the fourth KPI4 is that it helps to monitor the quality of the drinking water produced by all ENS1 membranes.

[0198] Other operating KPIs can be used. In particular, the fourth KPI can be replaced by a substantially equivalent KPI (i.e., the permeate concentration Cp).

[0199] The measurement data from the first SERIE1 time series and the second SERIE2 time series are stored in the system memory. This step is labeled as ENR1 in Figure 1

[0200] Other examples of KPIs can be implemented. According to one example, the fifth operating KPI5 corresponds to the concentration of the permeate or the concentrate. The latter KPI can be a function of the temperature Ti and the pressure difference DP.

[0201] Normalization

[0202] Normalization includes two steps: the first step is to automatically define a normalization function or a normalization model MOD Ni from the history of data of new, clean, or cleaned membranes through a learning algorithm, and the second step is to apply the learned normalization model MOD Ni to the actual recorded data to generate the intermediate KPI iA and the normalized KPIi ’ or KPI i0 ’. This last step is represented as GEN Figure 1 in A .

[0203] The first learning method is used to create a MOD Ni normalized model to generate a normalized metric for predicting membrane replacement and / or cleaning. In this case, the objective is to obtain an aging metric. In this case, the training data is preferably selected at the beginning of the life cycle of a membrane or a set of membranes. The aim is to train the normalized model during the first few weeks, months or even years of operation of the membrane or membrane module.

[0204] This learning method can be used to generate metrics indicating aging and fouling and thus a metric for fouling or scaling of a set of membranes.

[0205] The second learning method is used to create a MOD Ni normalized model to generate a normalized metric specifically for predicting membrane replacement. In this case, the objective is to obtain a fouling metric. In this case, the training data does not have to be selected from the first stage of the life cycle of a membrane or a set of membranes. In one embodiment, the objective is to train the normalized model during a period including several maintenance operations (e.g., cleaning a membrane or a set of membranes).

[0206] This second learning (French: second apprentissage) produces a metric that gives an indication of fouling or scaling of a set of membranes regardless of their replacement.

[0207] Normalization consists of modeling the metric according to the operating environmental conditions or reference environmental conditions and according to the state of the membrane, depending on whether they are considered to be in their operating state or in their new, clean or cleaned state.

[0208] The following relationship should be noted:

[0209] ■ KPI i = KPI iA + TC, where TC is a correction term related to membrane wear, aging and fouling, and KPI iA is the term representing the operating metric when the membrane is new and / or clean and / or cleaned and is taken from the normalized model.

[0210] This equation still holds under reference conditions, so we get:

[0211] ■ KPI i0 = KPI iA0 + TC

[0212] Note that at this point, the initial KPI can be written with the TC term i 'Normalization, which provides an indication of aging or clogging.

[0213] ■TC = KPI i ' = KPI i -KPI iA

[0214] Note that the correction term calculated according to the reference conditions is adjusted by deferral to generate a normalized index with the same order of magnitude as the KPI i index:

[0215] ■TC0 = KPI i0 ' = KPI iA0 +(KPI i -KPI iA )

[0216] In another example, additional KPIs iA0 components can be added. The addition of constants can also be achieved in another way.

[0217] Learn the normalization function, regression

[0218] To calculate the KPI iA value when the membrane is new and / or clean and / or pristine, the present invention includes learning a learning function (also referred to as a normalization function or normalization model). This learning method aims to define the parameters of the model by regression. Considering at least one external physical parameter, the present invention advantageously implements regression based on the expected value of the operating index value. Considering multiple external physical parameters, the regression can be multi-factor.

[0219] Therefore, the hypothesis of modeling the loss function based on the expected value is to learn the normalization model to generate the operating index:

[0220] ■There are sufficient explanatory variables: the residuals are entirely due to contamination, aging, and measurement noise;

[0221] ■The measurement points of the KPIi for the new and / or clean and / or pristine membrane basically correspond to the extreme values: in particular, the minimum value of the operating index: the maximum values of DP, Cp, Pf, and specific flux Sf; in this case, "apparently" means that the measurement points are within a factor of the expected value of the distribution, especially when several external physical parameters affect the index value;

[0222] ■The difference between the measured KPI and the modeled KPI does not depend on this explanatory variable.

[0223] To normalize these indices, the method includes calculating the selected operating index KPI iThe point cloud of the correction value is used to generate the intermediate operation index KPI iA The purpose of normalization is to generate one or more normalized indices representing wear or fouling independently of changes in external physical parameters. In other words, the method of the present invention attempts to generate normalized indices KPI1' or KPI0', which are insensitive to changes in environmental conditions such as the water temperature or salt concentration of the water entering the first ENS1 membrane module, or physical operation parameters (e.g., feed flow rate, permeate flow rate, and concentrate flow rate, or permeate pressure).

[0224] To standardize the operation index, the first step is to calculate the intermediate operation index KPI iA . This intermediate index is represented by the reference curve C REF1 . The reference curve C REF1 includes all the points of the new point cloud generated by the MOD Ni normalization model. This point cloud can be represented in the Figure 7 desired graph (represented here by a single external physical parameter temperature), or in the Figure 8 as the second intermediate time series SERIE 2A . This intermediate time series is represented as SERIE 2A and consists of the points of the reference curve C REF used to obtain this curve. The first KPI1 is associated with the first reference curve C REF1 , the second KPI2 is associated with the second reference curve C REF2 , the third KPI3 is associated with the third reference curve C REF3 , and the fourth KPI4 is associated with the fourth reference curve C RElF4 . Generally speaking, we call each KPI i operation index the reference curve C REFi .

[0225] Regression

[0226] The reference curve C REF or the second intermediate time series SERIE 2A is obtained due to the intermediate index obtained by applying the normalization model, which is generated due to the regression operation. The regression operation consists of obtaining the operation index KPI iParameterized values of the normalization model for the second SERIE2 time series. Each intermediate operating metric is generated by applying its own normalization model trained according to a given regression. For this purpose, we consider the influence of the external physical parameters used in the modeling of each external metric (especially in functions f1, f2, f3, f4). It should be noted that these functions f1, f2, f3, f4 may or may not be explicit. They reflect the influence of the external physical parameters on the operating metrics under consideration. For each set of values of the external physical parameters PARA i , the regression enables us to select KPIs located within a given expected range of the KPI value distribution i values.

[0227] Figure 7 Shows the representation of the first KPI1 metric, i.e., the pressure difference DP as a function of the temperature Ti. This representation provides a simple illustration of the desired function for a single variable, but is unrealistic in the case of two variables (e.g., the temperature Ti and the average flow rate Qn of the first KPI1). In the multi-factor case, another representation is necessary.

[0228] Use the desired regression to determine the normalization function for generating the intermediate operating metric, which can be represented according to the reference curve C REF1 . However, when the regression is multi-criteria, i.e., performed by considering different external physical parameters PARA i , it results in an envelope of the values obtained by considering the representation of the KPI i operating metric in a 2D space.

[0229] It can be recalled that the expected value is the distribution function of the variable Y, which is the value of the KPI i operating metric in this case. The expected value describes the distribution function of the variable.

[0230] Therefore, each KPI i operating metric is represented in an N-dimensional space, with each dimension associated with the external physical parameter PARA i . Implement the regression to generate the normalization model, using MOD Ni to generate a point cloud. This point cloud can be represented by the lower envelope or upper envelope of the KPI i values, the reference curve C REFi . According to what is considered to be the KPI iOperating parameters, where the lower or upper envelope corresponds to the operation when the first set of ENS1 membranes is new and the membranes are not fouled, or when the components related to membrane fouling are very low or even non-existent. The advantage of considering the points on the indicator where the membranes are not significantly aged for regression purposes is to obtain a normalized model that provides an indication of membrane aging. When these points are identified (especially within the lower or upper envelope), they correspond to values that are only sensitive to external parameters and no longer sensitive to membrane fouling, as the membranes are considered clean at these points.

[0231] According to an example, expectile regression (French: régression d’expectile) can be implemented to retain the KPIs corresponding to given expected values for given physical parameter values. i Part of the operating indicator points. In other words, for a given temperature Ti and a given average flow rate, this method can be used to maintain the KPIs by considering a predetermined percentage of the expected values. i Operating indicator values. Then regression is performed to consider that the retained values are the operating values of new and / or clean and / or cleared membranes after a cleaning operation within this part of the expected values.

[0232] The method of the present invention enables the configuration of expected values, for example, characteristic values of the distribution relative to 2% expected value or 4% expected value or 6% expected value, or even expected value functions of operating indicators with higher or lower percentages. The method of the present invention enables the determination of an expectile regression configuration that defines a good compromise between obtaining good regression accuracy with the lowest possible percentage of expected values to obtain a stable algorithm with optimized error capabilities and obtaining a regression with fast convergence capabilities with the maximum number of points. In fact, the lower the expected value, the fewer the number of points and the less accurate the regression.

[0233] In one embodiment, the observed external physical parameters are associated with the KPIs iA The function for correlating the indicators is configured based on the Generalized Additive Model for Expectiles (GAM, French: modèle additif généralisé GAM d’expectile). Then, the GAM function corresponds to the set of functions that we are trying to define according to the optimization criteria achieved through loss functions and error modeling during regression. In other embodiments, other functions can be implemented within the framework of the present invention.

[0234] According to an example, based on the KPIs i Observable values and the KPIs considered within the selected percentage of expected values iThe regression is modeled using a loss function between the estimated values of the operating metrics. This loss function is used to determine the optimal normalization function, i.e., the optimal normalization model, i.e., the parameters of the normalization model. The normalization model includes coefficients or parameters calculated by regression such that for different values of the external physical parameter PARA i the estimated operating metric KPI i corresponds to the measured or calculated operating metric value KPI within a selected desired percentage iA . Then, the loss function converges the error to reduce the estimated KPI i and the calculated KPI i values.

[0235] The estimation error takes into account different weightings for overestimation and underestimation.

[0236] The loss function can be modeled as follows:

[0237] F LOS = ∑W(y observed - y estimated )·(y observed - y estimated ) 2

[0238] where:

[0239] W(y observed - y estimated ) = Alpha, if (y observed - y estimated ) > 0

[0240] W(y observed - y estimated ) = 1 - Alpha, if (y observed - y estimated ) < 0

[0241] This is called the Alpha expectile (French: expectile Alpha).

[0242] Error modeling can include various embodiments. According to one example, error modeling including least squares minimization can be implemented.

[0243] One advantage of the GAM model is that it can perform regression regardless of the number of external physical parameters. Thus, one advantage is the ability to consider models where the KPI i operating metric may be affected by several external physical parameters (e.g., 2 to 5 external physical parameters).

[0244] Another advantage of using the GAM model is that it eliminates the need to map each KPI i operating metric to the external physical parameter PARAi The need for a related function, regardless of the relationship between the metrics and external physical parameters. In fact, the GAM model is applicable regardless of whether the relationship is linear or non - linear.

[0245] GAM is particularly suitable for eliminating the influence of the variation of each physical parameter value on the KPI being considered i value, all other factors being equal, that is, considering the same evolution or value of other parameters when eliminating the influence of a given parameter.

[0246] When modeling the GAM model, the data generated at the output of the model includes, on the one hand, the parameterization of the dependent variables {x, y} of each physical parameter {PARA1, PARA2} with respect to the observable values (i.e., the KPI i operational metric), and on the other hand, the values generated from the observable values (i.e., the estimated KPI i ^).

[0247] In other words, the GAM model can be used to generate the following model: where the KPI i metric x depends on the first PARA1 parameter and y depends on the second PARA2 parameter. Then, these factors can be used to weight each physical parameter value (these physical parameter values are considered in relation to the dependence on the KPI1 metric) to predict the KPI i ^ new estimated value of the metric. Then, the regression converges the error between the known value of the KPI i and the estimated value of the KPI i ^.

[0248] According to another example, a regression based on a quantile GAM generalized additive model can also be configured. According to another example, a logistic regression can be configured.

[0249] The regression is preferably carried out within a so - called smoothing duration D L which takes into account the data starting from the membrane lifetime. L

[0250] One advantage is to obtain training values for a normalized model that has not yet been affected by membrane aging. In this way, the intermediate operational metric KPI iA and the reference curve C REF can be used to generate a normalized metric KPI i ’ or KPI i0 ’ to show the evolution of the membrane condition since its initial state. One advantage is that it can better monitor the evolution of membrane degradation or clogging. However, for a stable and convergent regression - normalized model, a minimum data set is required. Therefore, the time range considered can vary from a few days to several years, or from a few months to a year. These durations depend on the plant size, data volume, etc.​

[0251] Alternatively, the entrainment values (French: valeurs d’entrainement) need not be concentrated at the start of the membrane life cycle, but rather in a part of the life cycle that includes several maintenance operations such as cleaning. One advantage of this is that a model of membrane fouling trends can be established without having to precisely measure aging.

[0252] Using the learning model outside the smoothing duration

[0253] According to one example, the normalization model has been learned within a smoothing duration D L The smoothing duration D L corresponds to a learning period during which regression is used to calibrate the normalization model. In one embodiment, the cleaning cycle / frequency can be incorporated into this training such that the training data includes several cleaning cycles.

[0254] Then, the normalization model learned by regression can be used for the entire D A acquisition period and / or to calculate in real time, based on newly acquired data, the KPI REFi value corresponding to the reference curve. Each normalization model can be learned according to a given smoothing duration D iA depending on the KPI i operating metric being considered. According to a preferred mode, for each constructed normalization model MOD L associated with a given operating metric KPI i , the learning period D Ni is the same. Thus, the method generates as many learned normalization models as there are calculated operating metrics. L

[0255] Then, the value of the normalized metric KPI i ' or KPI iA ' is obtained through an operation between the calculated operating metric KPI i and an intermediate operating metric KPI i0 . According to the embodiments described in detail below, the normalized operating metric KPI i ' is used to train a machine learning model called the prediction model MOD Pi or MOD PLTi to predict the evolution of the normalized operating metric KPI i ' or KPI i0 ' after the acquisition period D A .

[0256] By applying the MOD Ni normalization model learned within the smoothing duration, the method of the present invention can generate each intermediate metric KPIiA The value, thus representing an indicator of a set of membranes considered to be new and / or clean and / or pristine over a period exceeding the DL smoothing duration (e.g., during the acquisition period). Thus, the input to the normalization model is the operational indicator KPI i and possibly external physical parameter PARA i . The output of the normalization model is the intermediate operational indicator KPI iA .

[0257] To obtain the normalized operational indicator KPI i ’, the method of the present invention includes an operation aimed at combining the value of the operational indicator KPI i with the value of the intermediate operational indicator KPI Ni obtained by the learning normalization model MOD iA corresponding to a set of new and / or clean and / or pristine membranes, so as to generate a new time series SERIE 2A .

[0258] The third time series SERIE3 refers to the time series values of the normalized operational indicator KPI i ’ or KPI i0 ’ that the method attempts to obtain.

[0259] According to one example, each normalization model has been learned within the D L smoothing duration, and each normalization model is used to calculate the values of these intermediate operational indicators KPI i within the entire acquisition period D A based on the value of the considered external physical parameter PARA A and the values of the first, second, third, and fourth indicators KPI1, KPI2, KPI3, KPI4 at the input of the normalization model learned within the entire acquisition period D 1A , KPI 2A , KPI 3A , KPI 4A . The values of the intermediate indicators KPI 1A , KPI 2A , KPI 3A , KPI 4A are used to represent an indicator of a set of membranes considered to be new and / or clean and / or pristine.

[0260] To obtain a normalized operational indicator KPI i ’ or KPI i0 that reduces the effects caused by the influence of external physical parameters, the method of the present invention includes an operation aimed at combining the operational indicator KPI iThe values and the operations obtained from the intermediate operation metrics KPI through a learning normalization model corresponding to a new and / or clean and / or purified membrane are combined in order to generate a new SERIE3 time series. iA The operations obtained from the intermediate operation metrics KPI through a learning normalization model corresponding to a new and / or clean and / or purified membrane are combined in order to generate a new SERIE3 time series.

[0261] Third time series KPI iA , normalized operation metric: KPI i ’

[0262] When the intermediate operation metric KPI REFi represented by the reference curve C iA is generated for each operation metric KPI i under consideration, the method of the present invention generates a normalized operation metric KPI i ’ or KPI i0 ’. To this end, the SERIE 2A time series generated by the learned normalization model is used to obtain a new SERIE3 time series for defining the normalized operation metric KPI i ’ from the operations of another time series (e.g., the second SERIE2 time series).

[0263] In Figure 1 , the step of generating the normalized operation metric KPI i ’ is denoted as GEN1.

[0264] Figure 8 Shows a first curve representing the value of the first KPI1 and a second curve representing the value of the first KPI 1A which is obtained using the learned normalization model.

[0265] According to the first example, the third time series SERIE3 is generated by applying an operation (e.g., by subtracting them) between the second time series SERIE2 and the time series SERIE 2A corresponding to the correction value generated by the learned normalization model.

[0266] According to this second example, the third time series SERIE3 corresponds to the subtraction of these two series SERIE2 and SERIE 2A and results in a deviation of the first metric KPI i attributed to membrane fouling. In other words, the previously introduced correction term TC = KPI i - KPI iA .

[0267] It follows that the first operation metric: KPI1' = TC = KPI1 - KPI 1A .

[0268] Normalized Indicator KPI i ' represents the term related to the first indicator and is related to the wear and blockage of the membrane of the first ENS1 membrane module. Assume the normalized operating indicator KPI i ' is independent of operating and environmental conditions.

[0269] According to the second example, the third SERIE3 time series can correspond to the time series obtained by subtracting two SERIE2 and SERIE 2A sequences, where the KPI 10 component has been added under average or standard environmental conditions. The latter component also takes the form of a time series.

[0270] This solution makes it possible to reintroduce standard environmental conditions to obtain normalized operating indicator values of the usual order of magnitude, thus creating comparability between them. In this case, the indicator values obtained according to the standard reference conditions of physical parameters can be added, such as the standard operating temperature T0 and the standard average operating flow rate Qn0.

[0271] For each operating indicator, we have the equation:

[0272] KPI i0 ' = KPI i - KPI iA + KPI iA0

[0273] Figure 9 Shows the first normalized indicator KPI 10 ' in the form of this third time series SERIE3 obtained according to the second example, that is, by subtracting the correction parameter value KPI 2A in the SERIE 1A sequence from the KPI1 value of the second time series SERIE2 and summing the KPI 1A0 values of the first indicator, and this KPI 1A0 value corresponds to the state of a new and / or clean and / or pristine membrane calculated with standard environmental conditions (i.e., the standard operating temperature T0 and the standard average operating flow rate Qn0).

[0274] Figure 9 Displays that the normalized indicator KPI 10 ' can be used to evaluate the evolution of the membrane module independently of changes in environmental conditions. This makes it possible to observe changes in the membrane conditions, especially aging and fouling, independently of changes in environmental conditions.

[0275] In particular, one advantage of generating a normalized metric is that it provides a monitoring metric independent of variables such as water temperature Ti, feed flow rate Qf, inlet pressure Pf, and inlet conductivity Cf. Such a metric has the advantage of varying primarily as a function of wear and fouling conditions, in order to prevent membrane replacement and cleaning.

[0276] Advantageously, the KPI i0 ' data can be displayed to generate a metric that evolves over time. Figure 3 The AFF1 display screen is shown to illustrate an example of an operator console.

[0277] Event timestamp

[0278] According to one embodiment, the plant includes a set of maintenance operations, namely membrane cleaning and / or replacement operations. These operations improve water filtration and enable the membrane to be used within its optimal operating range. The object of the present invention is to define a method for predicting maintenance operations. To this end, the method of the present invention includes a step of learning a learning function. This learning can correspond to a regression operation of a normalization step or to the learning of a prediction model described in detail below. In order to obtain a high-quality training data set, the data related to the maintenance operations are time-stamped according to the time reference that can be used for processing time series (in particular the first and second SERIE1 and SERIE2 time series). The timestamp step is represented as HOR1 in Figure 1 .

[0279] Thus, when such an operation is performed in the plant, the method of the present invention enables these events to be time-stamped and marked in order to generate a set of timestamp data, which can then be used to implement a learning function (e.g., a machine learning model) to perform a prediction step for future maintenance operations.

[0280] The method of the present invention includes a step of recording timestamp data according to the same time reference in order to match the timestamp events with the time reference of the timestamp relative to each time series.

[0281] One advantage of timestamping events is that it enables a machine learning model to take into account maintenance events that explain discontinuities in the acquired and recorded raw data. This is particularly relevant in the case of long-term operation, as the aim of long-term operation is to measure the evolution of the normalized operation metrics over a long period of time.

[0282] The maintenance data timestamp can be used in a short-term prediction algorithm to define a set of training data between two timestamps such that the data collected during the maintenance operation does not interfere with the prediction.

[0283] The step of maintaining event timestamps is not necessary for the process of implementing the present invention (the sole purpose of the process of the present invention is to generate normalized metrics), but it can better explain and / or learn the operations that generate the normalized metric values.

[0284] Short-term prediction

[0285] In one embodiment, the normalized KPI i ’ data is used to learn the learning function FA2 (e.g., a machine learning model), i.e., the so-called prediction model MOD Pi . Such a prediction model MOD Pi is defined by coefficients or parameters learned during a training period called the first prediction period D P . In the case of short-term prediction, the method of the present invention enables a reliable prediction of the evolution of the normalized metric KPI i ’ or KPI i0 ’. This predicted value is denoted as KPI ip1 .

[0286] Figure 1 The GEN2 step in represents a short-term prediction step and / or a long-term prediction step.

[0287] The advantage of this prediction function is that it enables us to predict the next cleaning of the membranes in the first set of ENS1 membranes.

[0288] According to various embodiments, the prediction model MOD Pi can be a machine learning model, which is constructed by a function implementing a second generalized additive model (denoted as GAM2) and training data corresponding to the calculated values of the normalized metric KPI i ’ or KPI i0 ’.

[0289] According to another example, the prediction model MOD Pi can be constructed from a regression performed by a support vector machine model SVM on training data corresponding to the calculated values of the normalized metric KPI i ’ or KPI i0 ’.

[0290] According to one embodiment, the input vector includes the calculated data of the normalized metric KPI1’ or KPI0’, which is collected over a time window, which can be a sliding window including the most recently collected data. As the output of the prediction model MOD Pi , prediction data of the KPI ip1 operational metric values over several days, for example, can be generated.

[0291] Advantageously, these short-term and / or long-term prediction data can be displayed to generate an indication of the evolution over time and enable the planning of maintenance operations. Figure 3 An AFF1 display screen is shown to illustrate an example of an operator console.

[0292] Long-term prediction

[0293] According to one embodiment, the original data received, obtained or read from the memory of the operating metric KPI i and the calculated operating metric KPI i ’ or KPI i0 ’s normalized data are used. This data can be used to learn the learning function FA3 (e.g., a machine learning model). Such a prediction model MOD PLTi is defined by the coefficients or parameters learned during a training period (referred to as the second prediction period D PLT ). In the case of long-term prediction, the method of the present invention enables the obtaining of a prediction of the evolution of the normalized metric KPI i ’ or KPI i0 ’. This predicted value is denoted as KPI ip2 . The short-term prediction model MOD Pi and the long-term prediction model MOD PLTi are obtained from different trainings and thus correspond to different models.

[0294] The advantage of such a prediction function FA3 is to predict the next replacement of the first group of ENS1 membranes.

[0295] According to various embodiments, the machine learning model can be a function implementing a recurrent neural network, which has a regression function based on an autoregressive method. According to another example, the regression is performed by an LSTM-type model.

[0296] According to an implementation example, the input vector includes the calculated data of the normalized metric KPI1’ or KPI0’, which is obtained on the time window defined between two maintenance events and the time series for encoding the nature of the maintenance operations and the timestamps associated with these operations. KPI ip2 The prediction data of the operating metric value can include a long prediction period.

[0297] Advantageously, this data can be displayed to generate an indication of the evolution over time and enable the planning of maintenance operations.

[0298] The present invention also relates to a system for treating a certain volume of feed water into a certain volume of filtered treated water through a plurality of membrane components. The water treatment system includes a data processing system, which includes hardware means for performing the method steps of the present invention.

[0299] Thus, the present invention relates to a first data processing system and a second water treatment system, also called a "plant" (French: usine).

[0300] The water treatment system includes means for delivering a volume of water to an inlet to receive a water flow into at least a first set of membranes, such as a hydraulic pipe.

[0301] The water treatment system further includes a first filtered water outlet (called permeate) and a second residual water outlet (called concentrate). The water treatment system also includes a set of external parameter status sensors, the set of external parameter status sensors including a water temperature sensor and at least one pressure sensor.

[0302] The water treatment system further includes a data processing system, the data processing system including a computer, a memory, a clock, and a communication interface, so as to receive data in the form of a time series from various sensors. The data processing system also includes a communication interface for transmitting data to a server, the system including a server for performing steps of calculating a normalized metric according to the method of the present invention.

Claims

1. A method for automatically processing data representing the state of a plurality of membranes for filtering a large volume of liquid, characterized in that, The method comprises: ■ receiving (ACQ1) a first set of data (DATA1) from state sensors (20, 21, 22) arranged within or near a first set of membranes (ENS1), the set of membranes receiving an incoming water flow (Qf) and generating a first outgoing water flow (Qp), known as permeate water, and a second outgoing water flow (Qc), known as concentrate, the first set of data (DATA1) being related to external physical parameters (Pi, Qi, Ti, Ci), the data acquisition (ACQ1) being performed according to a plurality of first time series (SERIE1) of data transmitted by each sensor at a predetermined frequency; ■ Determine (EST1) at least one Key Performance Indicator (KPI) of the first set of membranes (ENS1) i ), said Key Performance Indicator (KPI i ) defining a second time series (SERIE2) of calculated or estimated data; ■ Record (ENR1) the first and second data time series (SERIE1, SERIE2) within a given acquisition time (D A ), each data time series (SERIE1, SERIE2) defining a point cloud; ■ Generation (GEN A ) Intermediate Key Performance Indicator (KPI iA ), the Intermediate Key Performance Indicator (KPI iA ) defining a point cloud (SERIE i ) corresponding to the value of the Key Performance Indicator (KPI 2A ), for which point cloud (SERIE 2A ), the first set of membranes (ENS1) is considered new and / or clean and / or sanitized, the value of the Intermediate Key Performance Indicator (KPI iA ) being generated by applying a learned normalization model (MOD Ni ) and from the Key Performance Indicator (KPI i ); ■ Generate (GEN1) normalized operating metrics (KPI i ’, KPI i0 ) that characterize the state of a plurality of membranes for filtering large volumes of liquid, said state characterizing fouling and / or aging of the state of said membranes and being independent of changes in environmental conditions, said normalized operating metrics (KPI i ’, KPI i0 ) defining a third time series (SERIE3), said normalized operating metrics (KPI i ’, KPI i0 ) being obtained from said operating metrics (KPI i ) and said intermediate operating metrics (KPI iA ).

2. The method according to claim 1, characterized in that, By regressing (REG i ) the data of the second time series (SERIE2) related to the operation metric (KPI i ) according to at least first predetermined external physical parameters (Pi, Qi, Ti, Ci) from the first time series (SERIE1), learning the normalization model (MOD i ) for each operation metric (KPI Ni ), the regression (REG1) being configured to determine a set of values within a smoothing duration (D L ) that are substantially within a factor of the minimum or maximum value of the values of the second time series (SERIE2), the determined values corresponding to the configuration of a new and / or clean and / or pristine membrane.

3. The method according to claim 1 or 2, characterized in that, During a smoothing duration (D L ) that includes at least one maintenance and / or replacement operation of the first set of membranes (ENS1), learn the normalization model (MOD i ) for each operation metric (KPI i ) from a set of training data of the operation metrics (KPI Ni ).

4. The method according to claim 1, characterized in that The determination of the operating indicator (KPI i ) includes determining a first operating indicator (KPI1) that defines the pressure difference (DP i ) between the inlet and the outlet of the membrane module (ENS1) and is represented in the form of a second time series (SERIE2) of calculated or estimated data. The external physical parameters considered include at least one flow rate measurement (Qi, Qp, Qf, Qn) and one temperature measurement (Ti). The external physical parameters (DATA1) are used to calculate the value of an intermediate indicator (KPI 1A ) of the first operating indicator (KPI1) based on a regression performed on the pressure difference value (DP).

5. The method according to any one of claims 1 to 4, characterized in that, The determination of the operating indicator (KPI i ) includes the determination of a second operating indicator (KPI2), which defines the inlet pressure (Pf) into the first membrane module (ENS1) and is represented in the form of a second time series (SERIE2) of calculated or estimated data. The external physical parameters considered include at least one temperature measurement (Ti), a measurement of the concentration (Cf) of the inflow into the first group of membranes (ENS1), the flow rate (Qp) of the permeate flow from the first group of membranes (ENS1), and the flow rate (Qc) of the concentrate flow. The external physical parameters (DATA1) are used to calculate the value of an intermediate indicator (KPI 2A ) of the second operating indicator (KPI1) based on a regression performed on the value of the inlet pressure (Pf) in the first membrane module (ENS1).

6. The method according to any one of claims 1 to 5, characterized in that, The determination of the operating indicator (KPI i ) includes determining a third operating indicator (KPI3), which defines the permeate flow rate (Qp, SF) at the outlet of the first membrane module (ENS1) and is represented in the form of a second time series (SERIE2) of calculated or estimated data. The external physical parameters considered include at least one temperature measurement value (Ti), a measurement value of the concentration (Cf) of the inflow entering the first group of membranes (ENS1), the flow rate (Qp) of the permeate water flow from the first group of membranes (ENS1), and the flow rate (Qc) of the concentrate flow. The external physical parameters (DATA1) are used to calculate the value of an intermediate indicator (KPI 3A ) of the third operating indicator (KPI1) based on a regression performed on the value of the permeate water flow rate (Qp) at the outlet of the first membrane module (ENS1).

7. The method according to any one of claims 1 to 6, characterized in that The determination of the operating indicator (KPI i ) includes determining a fourth operating indicator (KPI4), which defines the salt passage in the permeate water (SPp) leaving the first membrane module (ENS1) and is represented in the form of a second time series (SERIE2) of calculated or estimated data. The external physical parameters considered include at least one temperature measurement value (Ti), a measurement value of the concentration (Cf) of the inflow entering the first group of membranes (ENS1), the flow rate (Qp) of the permeate flow from the first group of membranes (ENS1), and the flow rate (Qc) of the concentrate flow. The external physical parameters (DATA1) are used to calculate the value of an intermediate indicator (KPI 4A ) of the fourth operating indicator (KPI1) based on a regression of the value of the salt passage in the permeate water (SPp) at the outlet of the first membrane module (ENS1).

8. The automated data processing method according to any one of claims 1 to 7, characterized in that The first data set (DATA1) is related to external physical parameters (PARA1) and includes: ■ measurements of the inflow (Qf), the permeate flow (Qp) and / or the concentrate flow (Qc), and / or ■ measurements of the electrical conductivity (Ci) of a given volume of water, and / or ■ measurements of the total organic carbon (TOC1), and / or ■ a target value corresponding to the conversion rate of the feed water volume to the treated water volume, and / or ■ an eigenvalue of the inflow rate, also known as the "permeate flow", and / or ■ an eigenvalue of the water permeability of the membrane.

9. The method according to any one of claims 1 to 8, characterized in that, The third time series (SERIE3) corresponds to: ■ By subtracting from the second time series (SERIE2) the time series corresponding to the correction value generated by the learning model (MOD Ni ) to obtain the time series (SERIE 2A ); and / or ■ By subtracting from the said second time series (SERIE2) the time series (SERIE Ni ) corresponding to the correction value generated by the said learning model (MOD 2A ), and having added to the said time series a reference component (KPI iA0 ), the said reference component (KPI iA0 ) corresponding to the time series of the operating indicators corresponding to the state of a new and / or clean and / or pristine membrane, the said component being calculated under average or standard environmental conditions.

10. The method according to any one of claims 2 to 9, characterized in that, Perform a regression on the operation metric (KPI i ) based on multiple external physical parameters (PARA i ) on which the operation metric (KPI i ) depends.

11. The method according to any one of claims 2 to 9, characterized in that, The regression is achieved by a first learning function (FA1), the first learning function (FA1) comprising a machine learning model having parameters learned by implementing a loss function.

12. The method according to claim 11, wherein The regression is an expected value regression and is based on an expected loss function and an error function between the value of the operating metric (KPI i ) and the value estimated by a regression model (KPI i ^) for the operating metric (KPI i ), and the value of the operating metric (KPI i ) is considered within a given expected range of the distribution of the value of the operating metric (KPI i ).

13. The method according to any one of claims 2 to 12, characterized in that, Based on multiple predefined external physical parameters ({PARA 1i )} i ) of multiple first time series ({SERIE i} i ) obtained by multiple sensors, perform regression on the data of the second data series (SERIE2) of the operation metric (KPI i ). The regression is performed based on a generalized additive model (GAM) modeling function. Within a predefined range of expected values and for a given value of the external physical parameters ({PARA i} i ), seek to optimize the parameters of the generalized additive model through an expected loss function between the value of the calculated operation metric (KPI i ) and the value of the estimated operation metric (KPI i ^). The regression further simulates an error function, and the regression runs on a so-called smoothing duration (D L ). The regression (REG) generates a set of values defining a point cloud of the intermediate metric, and the set of values corresponds to a new and / or clean and / or pristine state of the first set of membranes (ENS1).

14. The method according to any one of claims 1 to 13, characterized in that, The smoothing duration (D L ) is determined to include a plurality of event markers (EVN i ) related to the maintenance of the membrane assembly, and the smoothing duration (D L ) is less than the acquisition duration (D A ).

15. The method according to any one of claims 1 to 14, characterized in that, The method includes time stamps (HOR1) for events (EVN i ) related to membrane module maintenance, the events corresponding to cleaning (NET1) and / or replacement (REP1), each time stamp (HOR1) being performed according to a marked time reference within the acquisition duration (D A ).

16. The method according to any one of claims 1 to 15, characterized in that, The method includes generating values of predicted operating key performance indicators (KPIs ip1 , KPIs ip2 ) by applying a second learning function (FA2), the second learning function (FA2) being trained from values of normalized operating key performance indicators (KPI1’, KPI0’) corresponding to a third time series (SERIE3) considered within a prediction duration (D P ), the second learning function (FA2) generating prediction data (KPI ip1 ) for the evolution of the normalized indicators (KPI1’, KPI0’).

17. The processing method according to claim 16, wherein, The training data for training the second learning function (FA2) is selected between the last two event timestamps respectively associated with two consecutive cleanings, and new training of the second learning function (FA2) is triggered after each new event associated with a cleaning.

18. The processing method according to any one of claims 16 to 17, characterized in that The method includes comparing at least one predicted value of a normalized metric (KPI 1p , KPI 2p , KPI 3p , KPI 4p ) with at least one predetermined threshold (S1, S2, S3, S4), the comparison enabling the generation of a cleaning date.

19. The processing method according to claim 18, wherein The predetermined thresholds (S1, S2, S3, S4) are variable thresholds, the values of which are generated by executing a function depending on predetermined parameters.

20. The processing method according to any one of claims 16 to 19, characterized in that, The second learning function (FA2) is a function implementing a second generalized additive model (GAM2).

21. The processing method according to claims 1 and 18, characterized in that, It includes calculating an aging index of a set of membranes (ENS1) according to a third learning function (FA3), the third learning function (FA3) including a set of training data, the set of training data including values extracted from the first set of data (DATA1) for estimating the metrics (KPI1, KPI2, KPI3, KPI4), the training data being selected during a collection period (D A ) and taking into account the timestamps of events that occurred during the collection period (D A ).

22. The processing method according to claim 18, characterized in that, The third learning function (FA3) is a recurrent neural network comprising a regression function based on an autoregressive method.

23. A data processing system, characterized in that, It comprises a computer, a memory, a clock and a communication interface for receiving data in the form of time series, the data processing system comprising a communication interface for transmitting data to a server, the data processing system comprising a server for performing the steps of calculating a normalized operating metric according to any one of claims 1 to 22.

24. The system according to claim 23, wherein It includes a display for generating a representation of at least one normalized key performance indicator (KPI i ’, KPI i0 ’) in real time.

25. A system for treating a given volume of feed water into a given volume of filtered treated water by means of a plurality of membrane modules, the water treatment system comprising an inlet for receiving a water flow into at least one given membrane module, a first filtered water outlet for the so-called permeate water and a second residual water outlet for the so-called concentrate, the water treatment system further comprising a set of external parameter state sensors, the set of external parameter state sensors including a water temperature sensor and at least one pressure sensor, the water treatment system comprising a data processing system according to any one of claims 23 to 24.

26. The system according to any one of claims 23 to 25, characterized in that, It includes a plurality of membranes organized according to multiple sets (ENS1, ENS2, ENS3) of membranes, each set of membranes defining a stage for treating the water of the input volume and generating an output stream.

27. The system according to any one of claims 23 to 26, characterized in that It includes at least one second set of membranes (ENS2), the at least one second set of membranes being arranged at the outlet of the first set of membranes (ENS1), and the concentrate (Qc) of the first set of membranes (ENS1) defining the inlet of the second set of membranes (ENS2).

28. The system according to any one of claims 23 to 26, characterized in that, It includes at least one third membrane assembly (ENS3) arranged in parallel with the first membrane assembly (ENS1), and the concentrates (Qc) from the first membrane assembly (ENS1) and the third membrane assembly (ENS3) define the inlet of the second membrane assembly (ENS2).

Images