Method and system for detecting faults in an electrolysis device having an electrolysis cell

Abstract

The present invention relates to a method, system and assembly for detecting a fault in an electrolysis device having a plurality of electrolysis cells. The method comprises: obtaining voltage measurements of the electrolysis cells during operation of the electrolysis device; generating a composite cell voltage of the electrolysis cells using a neural network architecture that takes into account normal cell degradation based on cell-specific parameters; comparing the voltage measurements of corresponding ones of the electrolysis cells with the composite cell voltage to obtain a voltage difference; and detecting a fault in the electrolysis device when at least one of the voltage differences reaches a threshold.

Description

Technical Field

[0001] The present invention relates generally to electrolysis devices and, more particularly, to fault detection in electrolysis devices. Background Art

[0002] Chlor-alkali electrolysis is the process of breaking down lower value chemicals (e.g. NaCl, KCl) into higher value chemicals (e.g. NaOH, Cl2, KOH) by applying a direct current. The reaction is carried out in an electrochemical cell. In an industrial setting, multiple cells are combined in series or parallel to carry out the reaction. This combination is called an electrolyzer.

[0003] A chlor-alkali electrochemical cell consists of an anode, a cathode, and a separator. The oxidation reaction occurs at the anode, and the reduction reaction occurs at the cathode. In some cases, an ion exchange membrane may be used to separate the anode reaction from the cathode reaction. For a chlor-alkali electrochemical cell, the main products of electrolysis are chlorine, hydrogen, and sodium hydroxide or potassium hydroxide, also known as "caustic soda."

[0004] There is a need for improved methods for detecting faults in electrolytic cells. Summary of the invention

[0005] According to one broad aspect, a method for detecting a fault in an electrolysis device having a plurality of electrolysis cells is provided. The method includes: obtaining voltage measurements of the electrolysis cells during operation of the electrolysis device; generating a composite cell voltage for the electrolysis cells using a neural network architecture that takes into account normal cell degradation based on cell-specific parameters; comparing the voltage measurements of corresponding ones of the electrolysis cells to the composite cell voltage to obtain voltage differences; and detecting a fault in the electrolysis device when at least one of the voltage differences reaches a threshold.

[0006] According to another broad aspect, a system for detecting a fault in an electrolysis device having a plurality of electrolysis cells is provided. The system includes a processing unit and a non-transitory computer-readable medium having program code stored thereon. The program code is executable by the processing unit to: obtain voltage measurements of the electrolysis cells during operation of the electrolysis device; generate a composite cell voltage for the electrolysis cells using a neural network architecture that takes into account normal cell degradation based on battery-specific parameters; compare the voltage measurements of corresponding ones of the electrolysis cells to the composite cell voltage to obtain a voltage difference; and detect a fault in the electrolysis device when at least one of the voltage differences reaches a threshold.

[0007] According to yet another broad aspect, there is provided an assembly comprising an electrolysis device having a plurality of electrolysis cells and at least one computing device operably connected to the electrolysis device. The at least one computing device comprises at least one processing unit and a non-transitory computer-readable medium storing program instructions. The program instructions are executable by the at least one processing unit for: obtaining voltage measurements of the electrolysis cells during operation of the electrolysis device; generating a composite cell voltage of the electrolysis cells using a neural network architecture that takes into account normal cell degradation based on battery-specific parameters; comparing the voltage measurements of corresponding ones of the electrolysis cells to the composite cell voltage to obtain a voltage difference; and detecting a fault in the electrolysis device when at least one of the voltage differences reaches a threshold.

[0008] The features of the systems, devices, and methods described herein may be used in various combinations according to the embodiments described herein. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Reference is made to the following drawings.

[0010] Figure 1 is a schematic diagram of an exemplary electrolytic cell;

[0011] Figure 2 is a block diagram of exemplary components for detecting faults in an electrolysis device;

[0012] Figure 3 is a block diagram of an exemplary embodiment for a fault detection system;

[0013] Figure 4 is a flow chart of an exemplary method for detecting a fault in an electrolysis device; and

[0014] Figure 5 is a block diagram of an example computer device.

[0015] It should be noted that throughout the drawings, like features are identified by like reference numerals. DETAILED DESCRIPTION

[0016] Chlor-alkali production is an energy consuming process and therefore the efficiency of each electrochemical cell is a concern during operation. Another concern in the operation of chlor-alkali electrochemical cells is how to prevent hazards such as gas release, fire, and liquid leakage. This article describes a method and system for early detection of faults occurring in a chlor-alkali electrolysis unit operating in series with other cells in the electrolysis unit. The method includes acquiring voltages and processing cell measurements that represent healthy behavior. A neural network model is constructed based on an encoder-decoder architecture and the neural network model is used to predict the voltage of the cell. The predicted voltage is compared to the measured voltage in order to detect abnormal deviations while taking into account normal cell degradation.

[0017] Figure 1 1 is a schematic diagram of an exemplary membrane cell 100 used in the chlor-alkali industry. The membrane cell can consist of two compartments. In the case of salt electrolysis (NaCl), the anode compartment 102 is filled with a saturated brine solution (NaCl), while diluted caustic soda passes through the cathode compartment 103. In a chlor-alkali plant, chlorine gas (Cl2) 104 is produced at a coated (e.g., titanium (Ti)) anode 105. The combination of hydroxide ions 106 and sodium ions 107 migrating across a selective membrane 108 produces caustic soda (NaOH) and hydrogen 109. The cathode 110 can be nickel with a catalytic coating to reduce the over-potential for hydrogen (H2) accumulation.

[0018] Voltage changes in the membrane cell 100 may be the result of physical changes within the cell components. The cell voltage drop is distributed between its components: anode 105, cathode 110, membrane 108 and electrical connections. A decrease or increase in the cell voltage can be considered a prerequisite for two types of degradation: (1) normal degradation and end of the cell life, and (2) abnormal / sudden degradation during the battery life. The cell may deteriorate abnormally due to the failure of the membrane 108 to properly separate the two compartments 102, 103, resulting in undesirable chemical reactions. The cell may also deteriorate abnormally due to the loss of its active coating by the electrodes 105, 110. The cell may also deteriorate abnormally due to a combination of a failed membrane 108 and electrodes 105, 110. The root cause of the failure of individual components may be due to (external) poor operating conditions of the electrolysis device or the cell itself. These conditions may be, but are not limited to: poor inlet flow, poor inlet contaminant control, electrical hazards, poor temperature control, poor equipment installation, etc.

[0019] A composite voltage is calculated for one or more of the cells 100 in the chlor-alkali electrolysis device for early detection of abnormal degradation of the cells. The composite voltage is a calculation of a mathematical function consisting of many estimated parameters. Some of these parameters are specific and identical operations for all cells performing electrolysis under the same operating conditions. Some parameters are cell-specific parameters. The calculation can take into account normal degradation of the cell 100 over time. An example mathematical function is illustrated below:

[0020]

[0021] where f is an arbitrary linear or nonlinear function, I is the main rectifier current, T is the catholyte outlet temperature, CC is the caustic concentration, t is the timestamp, d is the backward time delay, g is an arbitrary linear or nonlinear function, and I SU 、T SU 、CCSU 、V SU These are the current, temperature, caustic soda concentration and voltage at startup.

[0022] The voltage of each cell of the electrolysis device can be monitored during operation to detect abnormal changes that cause serious safety problems. However, if only abnormal degradation is considered, a failure is detected when the cell has reached the point of no return and needs to be replaced. To avoid this problem, normal cell degradation is also considered.

[0023] The voltage drop of the battery 100 is ohmic, so the relationship between the feed current and the battery voltage is proportional. A theoretical voltage estimated based on this ratio can be used to characterize a healthy battery. However, battery characterization using only the current-voltage ratio does not take into account the potentially complex effects occurring in the anode compartment 102 and the cathode compartment 103. Therefore, some other parameters are also considered, such as battery specificity caused by age, technology and location.

[0024] For the same operating conditions and normal degradation levels, the voltage of a given cell may be different from the voltage of other cells in the same electrolysis unit. Such differences may be due to inconsistencies in manufacturing, installation, and other factors that are difficult to quantify. Therefore, cell-specific parameters such as age, technology, and location may be used.

[0025] Due to its degradation, the voltage drift of the battery may occur slowly over time. In addition, there may be a delay between the change in the operating conditions of the electrolyzer and the response of the battery. To account for this delay, the operating conditions at previous time steps can be taken into account by the predictive model.

[0026] In order to predict abnormal degradation early, the predicted voltage is independent of the measured voltage. As such, the measured voltage is not used as an input to the prediction model.

[0027] In some embodiments, a synthetic voltage is calculated for the cells of an electrolysis device using a neural network modeling architecture. The model is deployed in a production environment where operating conditions are not predefined and each cell has a different normal degradation level. The neural network model can be based on an encoder-decoder architecture, in which the decoder is replaced by a predictor (a subnetwork for predicting cell voltage). A neural encoder is a neural architecture whose purpose is to take an input vector and reduce its dimensionality to a desired dimensionality. It can be paired with a decoder. The decoder receives the output of the encoder and transforms it to minimize an objective function.

[0028] The neural encoder can be used to discover features that are specific to the battery at the operating cycle and therefore represent normal degradation. The predictor can take into account the time delay of the measured values. The predictor does not use the measured voltage as input, but is able to predict a different voltage for each battery despite using the same operating conditions as input. It does this by taking the output of the encoder as input, which is unique for each battery. Therefore, the voltage prediction value will not have a bias caused by the measured voltage of the battery.

[0029] refer to Figure 2 , shows an example embodiment of an assembly 200 including an electrolysis device 214 and a fault detection system 218. The fault detection system 218 is comprised of at least one computing device operably connected to the electrolysis device 214. In the illustrated embodiment, the fault detection system 218 includes a battery data acquisition device 252, a synthetic battery voltage generation device 272, and a fault detection device 262.

[0030] The cell data acquisition device 252 is configured to obtain voltage measurements of multiple electrolysis cells during operation of the electrolysis device 214. The voltage measurements can be obtained in real time or pseudo real time. A single unit or multiple units can be connected to each cell to obtain voltage measurements. In some embodiments, voltage measurements are obtained for each cell of the electrolysis device 214.

[0031] The synthetic battery voltage generation device 272 is configured to generate a battery-specific synthetic battery voltage using a neural network architecture. The neural network takes into account normal battery degradation based on battery-specific parameters in order to estimate the battery-specific synthetic battery voltage. In some embodiments, the neural network includes an encoder subnetwork and a predictor subnetwork. The encoder subnetwork can be configured to determine normal battery degradation based on battery-specific parameters. The predictor subnetwork can be configured to predict the synthetic battery voltage using the normal battery degradation output by the encoder subnetwork. In some embodiments, the predictor subnetwork is further configured to apply a time delay when predicting the synthetic battery voltage.

[0032] Various neural network architectures may be used to generate a battery-specific synthetic battery voltage. For example, the encoder subnetwork may include a masking layer, a LSTM layer, and two dense layers, and output a two-dimensional vector representing normal battery degradation specific to the battery. The predictor subnetwork may include two LSTM layers and two dense layers. Other embodiments are also applicable depending on the actual implementation.

[0033] In some embodiments, the neural network is trained using historical data from batteries of electrolyzer 214. In some embodiments, the neural network is trained using historical data from batteries of multiple electrolyzers. Various techniques can be used to train the neural network to learn healthy battery behavior while taking into account normal battery degradation based on battery-specific parameters.

[0034] The fault detection device 262 compares the voltage measurement from the battery data acquisition device 252 with the composite cell voltage from the composite cell voltage generation device 272 for the corresponding one of the electrolytic cells to obtain a voltage difference. For example, a measured voltage and a composite voltage can be assigned to each electrolytic cell, and the difference between the measured voltage and the composite voltage corresponds to the voltage difference. In some embodiments, voltage measurements are obtained for a subset of cells in the electrolysis device 214. In some embodiments, voltage measurements are obtained for multiple subsets of cells in an alternating manner. Other embodiments may also be applicable.

[0035] When one or more voltage differences reach a threshold, the fault detection device 262 detects a fault. In some embodiments, detecting a fault includes sending an alarm signal (which may be an audio signal, a visual signal, or other signal) to alert one or more operators to the fault. In some embodiments, detecting a fault includes sending a shutdown signal to the electrolysis device 214 or to another device or system configured to shut down the electrolysis device 214. A combination of an alarm signal and a shutdown signal may be used. For example, a first threshold may be associated with an alarm signal, and a second threshold may be associated with a shutdown signal. When the voltage difference is within a first range, an alarm signal is sent, and when the voltage difference is within a second range greater than the first range, a shutdown signal is sent. In another example, the nature of the fault signal may be related to the number of batteries that exhibit a voltage difference that has reached a threshold. For example, if the voltage difference of a first number of batteries reaches a threshold, an alarm signal is sent, and if the voltage difference of a second number of batteries greater than the first number of batteries reaches a threshold, a shutdown signal is sent. Other variations may also be applicable.

[0036] Reference Figure 3, a specific but non-limiting example of a fault detection system 218 is shown. The data acquisition and transmission unit 202 measures the differential voltage from cathode to cathode or from anode to anode of the cells in the electrolysis device 214 with a given accuracy (e.g., + / -1 millivolt or other accuracy level). The electrolysis device 214 can be any available industrial chlor-alkali electrolysis device consisting of membrane cells connected in series. In some embodiments, the electrolysis device 214 includes up to 160 cells, but the total number of cells can vary. The protected metal line 213 connects the input of the data acquisition and transmission unit 202 to the cathode or anode terminal of the adjacent cell. In some embodiments, each unit 202 can measure up to 32 voltage inputs, but the total number can vary. The unit 202 can include an analog-to-digital converter, a digital filter, a memory buffer and / or a microcontroller to perform the acquisition and transmission routine. The unit 202 can be powered by the power supply 203 using the protected metal line 215.

[0037] Data sent from unit 202 can be transmitted to processing and communication unit 204. In addition to processing the data transmission routine to the main computer server unit 209, unit 204 can also execute and send an emergency stop signal to the shutdown relay unit 206. Unit 204 can receive transformer rectifier shunt current measurements from unit 205 using, for example, 4-20mA converter terminal 217. Unit 204 can broadcast voltage and current data streams sampled at a given rate (e.g., one sample point per second) to unit 209. Ethernet communication unit 207 can broadcast process data streams not measured by units 202 and 204 from computing device 208 to computer server unit 209. Computing device 208 can be a third-party computer server, sometimes referred to as a distributed control system. Some example process data include, but are not limited to: cathode electrolyte outlet temperature, caustic outlet concentration, and inlet and / or outlet pH. Units 202, 204, 209, 207, and 208 can be connected using fiber optic line loops 216 or other connecting materials. In some embodiments, one or more of the connections are wireless. According to the illustrated embodiment, unit 209 is a main computer server that receives and processes individual cell voltage, current and process data, stores all data for one or more electrolyzers, executes a series of steps, and sends electrolyzer shutdown commands to unit 204, if any.

[0038] Unit 209 may be responsible for the collection and conversion of historical data. A similar setup may be deployed on multiple electrolysis units 214 operating in multiple chlor-alkali plants. Unit 209 may store all voltage data, current measurements, and process data from each cell. Since the sampling rates of voltage measurements and process data may be different, unit 209 may downsample the voltage observations to a given sampling rate. In some embodiments, the collected process data include outlet caustic concentration and electrolysis unit outlet cathode electrolyte temperature. Unit 109 may transform and align the data in a tabular form, where each row represents a different timestamp of an observation and each column represents a different measured variable. Unit 109 may store all data from weeks of deployment of a fault detection system 218 in multiple electrolysis units in a database.

[0039] The list data may be processed to select an operating cycle. An operating cycle is the time interval between successive starts and stops of the electrolyzer. Its length may vary from a few hours to a few weeks, depending on production constraints, changing demands, work shifts or maintenance requirements. Each cycle may be divided into two phases of varying length: a startup phase and an operating phase. The startup phase occurs when the current increases from zero to the maximum value allowed for each cell in its full production condition. The rate of current increase may vary at each startup due to variations in operating practices determined by the plant operator. The length of the startup phase may vary, for example, from 20 minutes to 12 hours. The operating phase may include the remainder of the cycle. The current may vary during this phase, for example, between 50% and 100% of the total range.

[0040] In one specific non-limiting example, pseudo code for detecting a possible cycle is as follows:

[0041] For each observation n in the database:

[0042] Diff n =Time n –Time n-1

[0043] If Diff n >10 minutes

[0044] Save Time n-1 and Time n

[0045] Each possible period is surrounded by pairs of dates corresponding to:

[0046] (0,Time n-1 ),(Time n ,Time n+1),(Time n+2 ,Time n+3 )…

[0047] In one specific non-limiting example, pseudo code for a confirmation cycle is as follows:

[0048] For every possible cycle of length k n :

[0049] If any observation for this period has "Current > 16kA":

[0050] Then idx = the first observation of "current > 16kA" if idx <= 12 hours:

[0051] Then start n = cycle [0: idx]

[0052] Operation n = cycle [idx: k]

[0053] If length(operation)>=length(startup):

[0054] Cycle n Is the effective period

[0055] Once the data is structured in cycles, a unit-based normalization scaling method can be applied to each of the data columns. This will scale the data linearly so the individual values ​​are in the range of [0,1]. This scaling helps improve the training time for building a neural network model. The data columns can be scaled using the following formula:

[0056]

[0057] Minimum (min) and maximum (max) values ​​can be used to scale the measurement column.

[0058] In some embodiments, the neural network model is established and trained using collected and transformed historical data. The encoder subnetwork infers the following features during the startup phase, which are used to characterize the behavior of the battery voltage in a certain period. These features illustrate the specificity and normal degradation of the battery. This is a self-supervised method because it does not require labeled degradation data. This step can be regarded as performing dimensionality reduction. However, the encoding vector (reduced dimensionality vector) is not the result of a statistical process, but the best representation that makes learning the voltage predictor subnetwork easier. The length of each startup phase may be different and can be limited to an upper limit value, such as a maximum value of 12 hours. For example, if each time step represents one minute, the length of the vector provided to the subnetwork is 720 time steps. Three input features can be used: the cathode electrolyte outlet temperature of the electrolyzer, the caustic soda concentration of the factory, and the current and voltage of the battery. The input features can be used to normalize the voltage to the specific operating conditions of each startup. In some embodiments, the shape of the input vector is [720 time steps, 4 features]. The masking layer forces the successive layers to ignore a time step if all features of the time step are equal to the masked value, which can be set to "-1" to filter out the time step previously added during padding. To account for the temporal nature of the sequence, the next layer can be a long short-term memory (LSTM) layer. Afterwards, two dense layers can be chained for a smoother transition to the final two-position encoding result. The output can be a vector of coordinates [X, Y] for each cell and launch, with a shape of (1 time step, 2 features). Furthermore, these coordinates can be represented graphically, providing insight into the decision process taken by the network.

[0059] The predictor subnetwork can be responsible for predicting the synthetic battery voltage. In some embodiments, the two inputs for prediction are a window of time steps from the operating phase and an encoded representation of the battery startup phase. According to this embodiment, a window of four observations is sufficient to represent the dynamic characteristics of the chemical phenomena after the battery response. The encoded battery startup can be repeated four times and connected to the window of operating characteristics. Two LSTM layers can be used to find the temporal correlation between the observations of each window. Two dense layers can be followed to output the predicted voltage. The output layer can have a Sigmoid activation function because the output voltage may have been previously scaled to the range [0,1]. In this step, the entire model is trained by minimizing the loss between the voltage predicted by this subnetwork and the measured voltage. The Adam optimizer and back-propagation algorithm can be used. In order for this subnetwork to obtain good accuracy in voltage prediction, the encoder should learn a faithful representation of the characteristics of the battery.

[0060] The following pseudocode represents a specific and non-limiting example of a training loop for one observation per forward pass-backward pass:

[0061] For each cycle:

[0062] For each cell in the electrolysis unit:

[0063] Define BATTERY_STARTING and BATTERY_OPERATIONAL

[0064] Single battery in window_operation

[0065] For each window:

[0066] Operating_Conditions = Window (excluding voltage)

[0067] Battery_Voltage = the last voltage observation in the window

[0068] #Forward pass

[0069] Encoder_Battery_Start = Encoder(Battery_Start)

[0070] predicted_voltage = predictor(operating_condition,encoding_battery_start)

[0071] Loss = MSE(prediction_voltage, battery_voltage)

[0072] #Backpropagation

[0073] update_network_weights(loss)

[0074] In order to improve the computational efficiency of this training cycle, mini-batch training can be used to parallelize the computation. In this training mode, many observations can be grouped into batches to be processed in parallel. According to this embodiment, one or more graphics processing units (GPUs) are used in unit 209 to perform this operation. In addition, three operations can be used to make the network converge efficiently: padding, shuffling, and window striding. Not every cycle's launch has the same duration. However, all batches fed to the GPU should have the same number of time steps. This problem is solved by padding the sequence, that is, adding a "-1" value at the end of the corresponding launch of each observation. In this way, all launches have the same duration of 720 minutes, which is equal to the maximum duration of the launch. This padding value will be ignored by the shielding layer of the encoder subnetwork later, so it will not affect the results. In order to reduce the time required for the network to converge to the optimal solution, a shuffling procedure can be used, in which each batch has observations from different batteries, cycles, and electrolytic cells. The encoder task of inferring the characteristics of the battery during the launch process is more complex than the task of the predictor. Therefore, it will be more efficient to train the network with fewer observations per cycle and more different starting sequences. The stride of the window function can be increased to address this issue. The stride is a number that defines how much of the sequence window is ignored between two consecutive training observations. At the end of the training cycle, the constructed neural network predictor can be stored in a file, such as in unit 209.

[0075] In some embodiments, the fault detection system 218 is deployed in one or more electrolysis units of the plant. The voltage of multiple cells is measured, and process data from third-party devices can be obtained. Alignment and cycle detection routines can be performed in real time or pseudo real time in unit 209. The reconstructed neural network voltage predictor can be used to calculate the composite voltage of the cell. The measured voltage (V t ) and the composite voltage When one or more batteries reach a threshold, an emergency shutdown signal may be sent to unit 204 and then to unit 206.

[0076] Figure 4 is a flow chart of an example method 400 for fault detection. In some embodiments, the method 400 comprises Figure 2 The method 400 is performed by the fault detection system 218 of the component 200. In some embodiments, the method 400 is performed by multiple computing devices.

[0077] In step 402, a voltage measurement of an electrolysis cell is obtained. In step 404, a composite voltage of the electrolysis cell is generated, as described in more detail above. Steps 402 and 404 may be performed simultaneously. In some embodiments, step 404 may be performed before step 402. In step 404, a neural network trained to learn healthy electrolysis cell behavior takes into account normal degradation of the electrolysis cell. The composite cell voltage is predicted using cell-specific parameters so that the composite cell voltage of two or more cells in the same electrolysis device can be different.

[0078] In step 406, the voltage measurements obtained in step 402 are compared to the composite cell voltage generated in step 404 for the corresponding cell. In other words, the composite cell voltage of a given cell is compared to the measured cell voltage of the same cell. This comparison is performed on a cell-by-cell basis for any number of cells included in the electrolysis. In some embodiments, this comparison is performed simultaneously or sequentially for all cells of the electrolysis using any random or predetermined order. If one or more voltage differences obtained when comparing the measured voltage of a given cell to the composite voltage reaches a threshold, a fault is detected in step 408.

[0079] Method 400 may be repeated any number of times at random or predetermined intervals. In some embodiments, method 400 is performed continuously until at least one fault is detected. Other embodiments may also be applicable.

[0080] Figure 5 is an example embodiment of a computing device 500 for implementing the method 400 for detecting a fault in an electrolysis device as described above. In some embodiments, the fault detection system 218 is implemented using one or more computing devices 500. The computing device 500 includes a processing unit 502 and a memory 504 storing computer executable instructions 506. The processing unit 502 may include any suitable device configured to cause a series of steps to be performed, such that the instructions 506, when executed by the computing device 500 or other programmable device, may cause the functions / actions / steps specified in the method 400 described herein to be performed. For example, the processing unit 502 may include any type of general purpose microprocessor or microcontroller, a digital signal processing (DSP) processor, a CPU, an integrated circuit, a field programmable gate array (FPGA), a reconfigurable processor, other appropriately programmed or programmable logic circuits, or any combination thereof.

[0081] The memory 504 may include any suitable known or other machine-readable storage medium. The memory 504 may include a non-transitory computer-readable storage medium, such as, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any suitable combination thereof. The memory 504 may include a suitable combination of any type of computer memory located internally or externally to the device, such as a random access memory (RAM), a read-only memory (ROM), an electro-optical memory, a magneto-optical memory, an erasable programmable read-only memory (EPROM), and an electrically erasable programmable read-only memory (EEPROM), a ferroelectric RAM (FRAM), etc. The memory 504 may include any storage device (e.g., a device) suitable for retrievably storing machine-readable instructions 506 that can be executed by the processing unit 502.

[0082] It should be noted that the techniques described herein may be performed by computing device 500 substantially in real time.

[0083] The method and system for detecting a fault in an electrolysis device as described herein may be implemented in a high-level procedural or object-oriented programming or scripting language or a combination thereof to communicate with a computer system (e.g., computing device 500) or to assist in the operation of a computer system. Alternatively, the method and system for detecting a fault in an electrolysis device may be implemented in component or machine language. The language may be a compiled or interpreted language. Program code for implementing the method and system for detecting a fault in an electrolysis device may be stored on a storage medium or device, such as a ROM, a disk, an optical disk, a flash drive, or any other suitable storage medium or device. When the computer reads the storage medium or device to perform the process described herein, the program code may be read by a general or special programmable computer to configure and operate the computer. Embodiments of the method and system for detecting a fault in an electrolysis device may also be considered to be implemented by a non-transitory computer-readable storage medium having a computer program stored thereon. The computer program may include computer-readable instructions that cause a computer, or more specifically a processing unit 502 of a computing device 500, to operate in a specific and predefined manner to perform the functions described herein.

[0084] Computer executable instructions can have many forms, including program modules that are executed by one or more computers or other devices. Typically, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. Typically, in various embodiments, the functionality of program modules can be combined or distributed as needed.

[0085] The embodiments described herein provide non-limiting examples of possible implementations of the present technology. After reading this specification, those of ordinary skill in the art will recognize that the embodiments described herein can be changed without departing from the scope of the present technology. For example, software modules can be combined or separated in different ways to perform the steps of method 400, or specific equipment for obtaining various measurements of electrolytic cells and / or electrolytic devices can be changed. In view of this specification, those of ordinary skill in the art can implement further modifications, which will fall within the scope of the present technology.

Claims

1. A method for detecting a fault in an electrolysis device having a plurality of electrolysis cells, the method comprising: obtaining voltage measurements of the electrolysis cell during operation of the electrolysis device; The synthetic cell voltage of the electrolysis cell is generated using a neural network architecture trained by historical data from a plurality of electrolysis devices, wherein each of the historical data constructed over a plurality of operating cycles includes a startup phase and an operating phase, the neural network architecture comprising: The encoder subnetwork is configured as: determining cell degradation indicative of end of life of each electrolysis cell based on parameters describing the specificity of each electrolysis cell, the parameters including, for each electrolysis cell, at least one of the age of the electrolysis cell, the technology of the electrolysis device, and the location of the electrolysis cell in the electrolysis device, and outputting a two-dimensional vector representing the cell degradation of each electrolytic cell during the startup phase, and a predictor subnetwork configured to predict the resultant battery voltage using the two-dimensional vector output by the encoder subnetwork; comparing the voltage measurements of corresponding ones of the electrolytic cells to the combined cell voltage to obtain a voltage difference; and When at least one of the voltage differences reaches a threshold, a fault in the electrolysis device is detected.

2. The method according to claim 1, in, The predictor subnetwork is also configured to impose a time delay when predicting the composite battery voltage.

3. The method according to claim 1, in, The encoder subnetwork includes a shielding layer, a long short-term memory layer, and two dense layers.

4. The method according to claim 1, in, The predictor subnetwork includes two long short-term memory layers and two dense layers.

5. The method according to claim 1, in, Detecting the fault in the electrolysis device includes outputting an alarm signal.

6. The method according to claim 1, in, Detecting the fault in the electrolysis device includes outputting a signal for shutting down the electrolysis device.

7. The method according to claim 1, in, The electrolysis cell is a chlor-alkali electrolysis cell.

8. A system for detecting a fault in an electrolysis device having a plurality of electrolysis cells, the system comprising: Processing unit; and A non-transitory computer readable medium having program code stored thereon, the program code being executable by the processing unit for: obtaining voltage measurements of the electrolysis cell during operation of the electrolysis device; The synthetic cell voltage of the electrolysis cell is generated using a neural network architecture trained by historical data from a plurality of electrolysis devices, wherein each of the historical data constructed over a plurality of operating cycles includes a startup phase and an operating phase, the neural network architecture comprising: The encoder subnetwork is configured as: determining cell degradation indicative of end of life of each electrolysis cell based on parameters describing the specificity of each electrolysis cell, the parameters including, for each electrolysis cell, at least one of the age of the electrolysis cell, the technology of the electrolysis device, and the location of the electrolysis cell in the electrolysis device, and outputting a two-dimensional vector representing the cell degradation of each electrolytic cell during the startup phase, and a predictor subnetwork configured to predict the resultant battery voltage using the two-dimensional vector output by the encoder subnetwork; comparing the voltage measurements of corresponding ones of the electrolytic cells to the combined cell voltage to obtain a voltage difference; and When at least one of the voltage differences reaches a threshold, a fault in the electrolysis device is detected.

9. The system according to claim 8, in, The predictor subnetwork is also configured to impose a time delay when predicting the composite battery voltage.

10. The system according to claim 8, in, The encoder subnetwork includes a shielding layer, a long short-term memory layer, and two dense layers.

11. The system according to claim 8, in, The predictor subnetwork includes two long short-term memory layers and two dense layers.

12. The system according to claim 8, in, Detecting the fault in the electrolysis device includes outputting an alarm signal.

13. The system according to claim 8, in, Detecting the fault in the electrolysis device includes outputting a signal for shutting down the electrolysis device.

14. The system according to claim 8, The electrolysis cell is a chlor-alkali electrolysis cell.

15. An assembly comprising: An electrolysis device having a plurality of electrolysis devices; and at least one computing device operably connected to the electrolysis apparatus, the at least one computing device comprising at least one processing unit and a non-transitory computer-readable medium storing program instructions executable by the at least one processing unit for: obtaining voltage measurements of the electrolysis cell during operation of the electrolysis device; The synthetic cell voltage of the electrolysis cell is generated using a neural network architecture trained by historical data from a plurality of electrolysis devices, wherein each of the historical data constructed over a plurality of operating cycles includes a startup phase and an operating phase, the neural network architecture comprising: The encoder subnetwork is configured as: determining cell degradation indicative of end of life of each electrolysis cell based on parameters describing the specificity of each electrolysis cell, the parameters including, for each electrolysis cell, at least one of the age of the electrolysis cell, the technology of the electrolysis device, and the location of the electrolysis cell in the electrolysis device, and outputting a two-dimensional vector representing the cell degradation of each electrolytic cell during the startup phase, and a predictor subnetwork configured to predict the resultant battery voltage using the two-dimensional vector output by the encoder subnetwork; comparing the voltage measurements of corresponding ones of the electrolytic cells to the combined cell voltage to obtain a voltage difference; and When at least one of the voltage differences reaches a threshold, a fault in the electrolysis device is detected.

16. The assembly of claim 15, wherein: The at least one computing device comprises: a plurality of data acquisition and transmission units connected to said electrolysis cells to obtain said voltage measurements during operation of said electrolysis device; a processing and communication unit connected to said plurality of data acquisition and transmission units; and A host computer server is connected to the processing and communication unit.

Images