A method and apparatus for predicting failure of an electrolytic cell

Abstract

The application belongs to the technical field of hydrogen production by water electrolysis, and specifically discloses an electrolytic tank fault prediction method and device, which comprises the following steps: multi-source data synchronous monitoring, edge real-time feature extraction, self-adaptive correction of dynamic working conditions, double-channel parallel diagnosis, fault early warning, residual life prediction, and generation of operation and maintenance decision suggestions. The eight-dimensional perception system is formed, the selection and combination logic of the sensing module are designed based on the electrolytic tank fault propagation mechanism and the multi-physical field coupling characteristics, the monitoring logic of "one fault, multi-source verification" is realized, and the fault accuracy is greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of water electrolysis hydrogen production technology, and particularly relates to a method and device for predicting electrolyzer failures. Background Technology

[0002] Water electrolysis for hydrogen production, as a primary method for green hydrogen production, is a key support for achieving a low-carbon transformation of the energy structure. The electrolyzer, as the core equipment of the water electrolysis hydrogen production system, directly determines the hydrogen production efficiency, hydrogen purity, and overall plant safety. In actual industrial operation, the electrolyzer operates under harsh conditions of a strongly alkaline (or acidic) electrolyte environment, high current density, and alternating thermal stress, making it highly susceptible to various types of performance degradation and structural failures.

[0003] Typical failure modes include: (1) electrolyte leakage and gas mixing due to gasket aging or improper installation; (2) blockage of flow channels and diaphragms due to impurities deposited in the influent or accumulation of electrode dissolution products; (3) decreased electrode activity and coating peeling due to potential fluctuations and chemical corrosion; and (4) local dry areas or hot spots caused by uneven electrolyte distribution. If these failures are not detected in time, they will not only lead to a significant increase in DC power consumption (usually an increase of more than 5% will trigger the overhaul threshold), but may also cause excessive concentrations of oxygen in hydrogen or hydrogen in oxygen, resulting in serious explosion safety hazards.

[0004] Currently, the main problems and shortcomings of anomaly identification and fault prediction technologies for electrolytic cells are as follows: I. Existing monitoring methods are mostly offline and post-event detection, lacking online real-time early warning capabilities. Currently, most hydrogen production plants still rely on periodic shutdowns for maintenance and offline testing to assess the condition of their electrolyzers. For example, during annual overhauls, electrolyzers are disassembled to check for areas of nickel plating peeling off the electrodes (repair is required if more than 20% is missing), damage to the asbestos cloth in the diaphragm, and aging cracks in the gaskets. This "reactive" approach is significantly delayed, failing to provide early warnings of potential problems (such as minor gasket leaks or initial localized blockages). Even if a few online parameters (such as chamber voltage, cell temperature, and gas purity) are monitored during operation, these parameters are typically only used to trigger over-limit alarms, lacking in-depth analysis of subtle parameter trends. Consequently, chronic compression damage to gaskets caused by improper oil pressure control is often only discovered after a leak has occurred.

[0005] Second, the single threshold alarm mechanism leads to frequent false alarms and missed alarms, and makes it difficult to identify complex faults. Existing technologies mostly use a "fixed threshold" method to monitor electrolyzers. This mechanism has two major drawbacks: First, it ignores the dynamic nature of operating conditions. During load fluctuations (especially frequent partial load operation caused by renewable energy grid connection), the concentration of gas impurities will naturally increase due to the decrease in gas production rate. If the fixed threshold under rated operating conditions is still used, it is very easy to cause frequent "unplanned shutdowns". Second, a single parameter is difficult to distinguish the root cause of the fault. For example, an increased dispersion in the chamber voltage may be caused by electrode corrosion or uneven distribution of liquid inlet flow. Simply relying on voltage alarms cannot pinpoint the specific cause, let alone handle the coupling effect of multiple faults occurring simultaneously (such as gasket leakage combined with electrode contamination).

[0006] III. Model-based diagnostic methods are limited by the complexity of multiphysics modeling in electrolyzers. Academics and industry have attempted to diagnose faults in electrolyzers by establishing mechanistic models (such as electrochemical and thermodynamic models). However, the internal workings of electrolyzers involve complex multi-physics couplings (electric, flow, temperature, and concentration fields), and exhibit severe aging and nonlinear time-varying characteristics with increasing operating time. Constructing accurate analytical models is extremely difficult, resulting in poor robustness of model-based residual analysis methods in practical applications. Especially for early faults such as microleakage caused by gasket aging or blockage of diaphragm micropores, the resulting changes in model parameters are extremely subtle, making it difficult for existing models to effectively capture these changes.

[0007] Fourth, there is a lack of means to perceive the microscopic state of key components (such as gaskets and electrodes). Corrosion and damage to gaskets and electrodes are the most common causes of electrolytic cell failures, but current technologies lack effective online sensing methods. For gaskets, damage is complex, including material aging, chlorate corrosion, and excessive assembly stress (such as cut-off due to excessive locking pressure). However, current monitoring technologies cannot directly measure the compressive stress distribution or residual life of the gasket. For electrodes, although it is known that the detachment of the nickel plating layer leads to exposure of the base iron and electrochemical corrosion to form rust, this corrosion initially manifests only as microscopic iron ion dissolution, with minimal changes in macroscopic electrical parameters (such as cell pressure). Only when corrosion products accumulate to a certain extent does a sharp performance drop occur, at which point the loss is irreversible.

[0008] In summary, due to issues such as offline latency, single threshold, modeling difficulties, data contamination, and blind spots in microscopic sensing, existing technologies struggle to provide timely and accurate early warnings for critical faults in electrolyzers, such as gasket damage, flow channel blockage, electrode corrosion, and efficiency decline. Therefore, there is an urgent need for a method for electrolyzer anomaly identification and fault prediction that can integrate multi-source data, possess early and subtle fault identification capabilities, and predict performance degradation trends, in order to improve the operational reliability and economy of water electrolysis hydrogen production systems. Summary of the Invention

[0009] The purpose of this invention is to provide a method and apparatus for predicting electrolytic cell faults, which can integrate multi-source data and has the ability to identify early and subtle faults.

[0010] To achieve the above objectives, the technical solution of the present invention is: a method for predicting electrolytic cell failures, comprising the following steps: Multi-source data synchronous monitoring: The cell voltage monitoring module collects the voltage of all individual cells; the electrochemical noise monitoring module collects the potential fluctuations of the anode and cathode, as well as the inter-electrode current fluctuations; the fiber optic grating strain monitoring module collects the strain values ​​of the end plate bolts; the acoustic emission sensing module collects the structural wave signals on the outside of the end plate; the temperature sensing module collects the temperature on the side of the electrolytic cell, forming a temperature field; the thickness sensing module collects the electrode thickness at each measuring point; the pressure sensing module collects the hydrogen outlet pressure and oxygen outlet pressure; and the composition analysis module detects the metal ion concentration of the alkaline solution bypass sample. Extracting real-time edge features: Perform time-frequency analysis on the noise signal collected by the electrochemical noise monitoring module, calculate the noise resistance Rn, and extract transient peak features; perform wavelet packet decomposition on the acoustic emission waveform collected by the acoustic emission sensing module, extract the energy distribution of different frequency bands, and identify the frequency band of the gasket micro-leakage event; calculate the uniformity index of the temperature field and identify the abnormal temperature gradient region; calculate the voltage dispersion of the cell voltage and identify the reverse polarity voltage; Adaptive correction of dynamic operating conditions: Read the current operating condition parameters, call the pre-stored operating condition-feature mapping library, and normalize the collected data and extracted features to eliminate spurious changes caused by load fluctuations; Dual-channel parallel diagnostics: First diagnostic channel: Residual analysis based on a physical model: Constructing a lumped parameter model of the electrolyzer, including an electrochemical model, a thermal equilibrium model, and a fluid model. Comparing the real-time measured values ​​of the electrolyzer with the predicted values ​​of the lumped parameter model to generate a residual sequence; when the residual sequence exceeds a set adaptive threshold, a model mismatch alarm is triggered, indicating a potential fault risk in the electrolyzer; Second diagnostic channel: Data-driven pattern recognition: Inputting extracted real-time edge features into a graph neural network diagnostic model, which outputs the probability distribution and confidence level of each fault type; fusing the diagnostic results from the two channels, and outputting a diagnostic conclusion based on the fusion result.

[0011] Furthermore, it also includes fault warning: faults are divided into multiple levels, which are divided sequentially according to the urgency of the fault. Each level has corresponding triggering conditions and output action instructions.

[0012] Furthermore, it also includes remaining life prediction: For lifetime prediction of electrode corrosion: Based on the accumulation rate of metal ion concentration and combined with the electrochemical noise trend, the Wiener process model is used to predict the remaining time for the effective electrode thickness to reach the threshold. Prediction of gasket aging: Based on the bolt strain relaxation curve and combined with the Arrhenius temperature acceleration model, the time when the gasket compression rebound rate drops to the critical value is predicted. For diaphragm blockage prediction: Based on the increasing trend of voltage dispersion and the growth rate of noise resistance, a long-term memory network is used to predict when the diaphragm differential pressure will reach the alarm value.

[0013] Furthermore, it also includes generating operation and maintenance decision suggestions: based on the diagnostic conclusions and combined with the handling decisions in the fault knowledge base, operation and maintenance suggestions are automatically generated.

[0014] Furthermore, the maintenance recommendations include: Slight corrosion: It is recommended to adjust the operation and maintenance current density and reduce the load operation; For localized blockages: backflushing or chemical cleaning is recommended, and a cleaning solution formulation suggestion will be provided. Gasket leakage: It is recommended to plan a shutdown, retighten the end plate bolts to the target torque value, and if retightening is ineffective, arrange for gasket replacement; The electrode is severely corroded: a major overhaul and replacement are recommended, and a spare parts procurement warning should be provided.

[0015] Furthermore, Stockwell transform is used to perform time-frequency analysis on the noise signal collected by the electrochemical noise monitoring module. The electrochemical noise monitoring module includes a three-electrode system, a zero-resistance galvanometer, and a potentiometer. The three-electrode system is installed in the bypass flow cell of the alkaline solution circulation pipeline of the electrolytic cell. The three-electrode system includes a working electrode, a reference electrode, and an auxiliary electrode. The zero-resistance galvanometer is used to measure the coupling current between the working electrode and the auxiliary electrode, and the potentiometer is used to measure the potential of the working electrode relative to the reference electrode. The noise resistance Rn is the ratio of the standard deviation of the potential noise to the standard deviation of the current noise.

[0016] Furthermore, the fiber Bragg grating strain monitoring module includes a fiber Bragg grating strain sensor, a fiber Bragg grating temperature sensor, a fiber Bragg grating demodulator, and a gasket condition assessment unit. The fiber Bragg grating strain sensor has multiple measuring points arranged along the axial direction of the end plate bolts or multiple measuring points evenly arranged circumferentially around the outer perimeter of the pole frame. The fiber Bragg grating temperature sensor is used for temperature compensation of the fiber Bragg grating strain sensor. The fiber Bragg grating demodulator is used to receive signals read by the fiber Bragg grating strain sensor and the fiber Bragg grating temperature sensor and send the signals to the gasket condition assessment unit. The gasket condition assessment unit calculates the gasket stress relaxation rate based on the strain value and identifies micro-leakage by combining strain fluctuation spectrum analysis.

[0017] Furthermore, the working method of the graph neural network diagnostic model is as follows: Constructing the input layer: The extracted real-time edge features form a multi-dimensional feature vector; the multi-dimensional feature vector has a dimension of M, which, combined with the time window length T, forms temporal feature input data that matches the dimension; Graph construction layer: The electrolyzer is modeled as a graph structure G=(V,E), where the node set V contains the various chambers of the electrolyzer, the alkali inlet, the hydrogen and oxygen outlet, and various sensing modules, and the edge set E contains physical connection relationships and logical connection relationships. The physical connection relationships include pipeline connection relationships and circuit connection relationships, and the logical connection relationships are the association relationships of components in the same area. Based on the fixed physical topology matrix A_physical and the real-time data correlation similarity matrix A_data(t), a dynamic adjacency matrix A(t) is constructed, and the calculation formula is: A(t)=α•A_physical+(1-α)•A_data(t), where α is the weight coefficient, and 0≤α≤1; A_data(t) is calculated according to the correlation of changes in real-time time series data of various sensing modules. Construct a spatiotemporal feature extraction layer: extract time-dimensional and spatial-dimensional features from the input data respectively: use a gated causal convolutional network to process time-dimensional features and extract single-node time-dependent features H_time; use a graph attention network to process spatial-dimensional features, calculate attention coefficients between nodes, aggregate neighbor node information, and extract spatial-dependent features H_space; Constructing a fusion layer: Adaptively fuse the time-dependent feature H_time and the spatial-dependent feature H_space through a gated fusion unit to obtain the fusion feature H_fusion. The calculation formula is: H_fusion=z⊙H_time+(1-z)⊙H_space, where z is the update gate weight calculated by the sigmoid function, and 0≤z≤1, and ⊙ is the element-wise multiplication operation; Constructing the output layer: Input the fused features H_fusion into the fully connected layer, and then output the classification probability of various faults in the electrolytic cell through the Softmax classifier to complete the fault diagnosis.

[0018] Furthermore, the component analysis module includes a bypass sampling pump, a filtration unit, a dilution unit, a component analysis device, and a waste liquid recovery unit. The bypass sampling pump is used to periodically extract the electrolyte. The filtration unit is used to remove air bubbles and solid particles from the electrolyte. The dilution unit is used to dilute the regularly extracted electrolyte to the measurement range of the component analysis device. The component analysis device is used to analyze the concentration of various metal ions in the diluted electrolyte. The waste liquid recovery unit is used to return the analyzed electrolyte to the pipeline of the electrolytic cell.

[0019] An electrolytic cell fault prediction device includes a sensing unit, an edge computing unit, and a cloud-based intelligent unit. The edge computing unit is connected to both the sensing unit and the cloud-based intelligent unit. The sensing unit sends real-time data to the edge computing unit. The edge computing unit sends feature data to the cloud-based intelligent unit, which updates the model. The sensing unit includes a monitoring system and a clock synchronization system. The edge computing unit includes a multi-channel synchronous data acquisition system, a preprocessing system, and an embedded industrial control computer. The cloud-based intelligent unit includes a data storage and management system, a model training and optimization system, a digital twin system, and an application service system. The monitoring system includes a cell voltage monitoring module, an electrochemical noise monitoring module, a fiber optic strain monitoring module, an acoustic emission sensing module, a temperature sensing module, a thickness sensing module, a pressure sensing module, and a component analysis module. The cell voltage monitoring module collects the voltage of all individual cells and generates a single-cell voltage distribution cloud map. The electrochemical noise monitoring module collects the potential fluctuations between the anode and cathode, as well as the inter-electrode current fluctuations. The fiber optic strain monitoring module collects the strain values ​​of the endplate bolts. The acoustic emission sensing module collects the structural wave signals on the outer side of the endplate. The temperature sensing module collects the temperature on the side of the electrolytic cell, forming a temperature field. The thickness sensing module collects the electrode thickness at each measuring point. The pressure sensing module collects the hydrogen outlet pressure and oxygen outlet pressure. The component analysis module detects the concentration of metal ions in the alkali bypass sample. The clock synchronization system is used to distribute the data synchronization pulses collected by the monitoring system to the multi-channel synchronous data acquisition system; The preprocessing system is used to perform real-time filtering, feature extraction, and Stockwell transform on the transmitted data. The embedded industrial control computer is used to realize adaptive correction of dynamic working conditions, dual-channel parallel diagnosis, and shutdown logic judgment; The data storage and management system is used to store historical databases and fault knowledge bases; The model training and optimization system is used to train and optimize the model; The digital twin system is used to achieve 3D visualization, real-time mapping, and simulation. The application service system is used for fault diagnosis, lifespan prediction, operation and maintenance decision-making, report generation, and alarm push.

[0020] The beneficial effects of this technical solution are as follows: This invention forms an eight-dimensional sensing system. The selection and combination logic of the sensing modules are designed based on the fault propagation mechanism and multi-physics coupling characteristics of the electrolytic cell: micro-leakage of the gasket will simultaneously cause stress changes and fluid disturbances, so a fiber optic strain sensing module and an acoustic emission sensing module are combined; electrode corrosion will simultaneously generate electrochemical signals and metal ion dissolution, so an electrochemical noise sensing module and a component analysis module are paired; flow channel blockage will lead to abnormal temperature distribution and pressure fluctuations, so a temperature sensing module and a pressure sensing module are integrated to realize the monitoring logic of "one fault, multiple sources of verification", which greatly improves the fault accuracy.

[0021] Traditional electrochemical noise analysis often employs statistical methods (standard deviation, skewness) or FFT transform. However, FFT cannot simultaneously guarantee low-frequency resolution and high-frequency temporal resolution, resulting in poor identification of short-term non-stationary events such as pitting corrosion initiation. Addressing the composite characteristics of electrolytic cell electrode corrosion signals—combining slow low-frequency changes (uniform corrosion) with high-frequency instantaneous pulses (pitting corrosion initiation)—Stockwell transform utilizes a frequency-adaptive resolution design. It maintains high frequency resolution in the low-frequency range (<1Hz) to capture uniform corrosion trends, and high temporal resolution in the high-frequency range (100-500Hz) to capture instantaneous pitting corrosion signals, perfectly matching the multi-scale characteristics of corrosion signals.

[0022] Existing technologies rely entirely on experience or shutdown inspections to assess gasket condition. This invention, for the first time, achieves non-invasive online monitoring of electrolytic cell sealing gaskets. It captures the slow changes in gasket stress relaxation using a fiber optic strain sensor module, combined with an acoustic emission sensor module to capture instantaneous signals generated by micro-leakage, forming a dual monitoring mechanism of "slow change + transient," solving the problem that a single sensor cannot fully cover the gasket fault evolution process. Through multi-source data fusion, the accuracy of gasket fault identification is improved and the false alarm rate is reduced. Furthermore, micro-leakage identification is achieved through strain fluctuation spectrum analysis, which is more direct than traditional acoustic emission sensors (due to the bolt serving as the force transmission path, signal attenuation is less).

[0023] Traditional corrosion monitoring relies on electrochemical methods or gravimetric methods; the former is difficult to quantify, while the latter requires stopping the process for sampling. This invention achieves direct quantitative measurement of corrosion rate by monitoring the dissolution concentration of corrosion products in the electrolyte online. It can also distinguish the corrosion behavior of different metal components (e.g., Fe dissolution corresponds to matrix corrosion, Ni dissolution corresponds to coating detachment), providing data support for electrode failure mechanism analysis. When combined with electrochemical noise monitoring, it can simultaneously acquire the corrosion "process signal" (noise) and "result signal" (ion concentration), achieving dual verification of the corrosion state and further improving monitoring accuracy.

[0024] Existing fault diagnosis models either ignore spatial structure (such as MLP) or use fixed graph structures (such as CNN), which cannot adapt to the spatiotemporal propagation characteristics of electrolytic cell faults. This invention introduces the dynamic spatiotemporal coupling characteristics of electrolytic cells into the diagnostic model for the first time. The dynamic adjacency matrix can capture changes in fault propagation paths caused by changes in operating conditions in real time (such as changes in the impact range of flow channel blockage when the load fluctuates). The graph attention mechanism can automatically learn the fault influence weights between different components and accurately identify the coupling between primary and secondary faults.

[0025] This invention can predict the remaining life of electrodes, the aging of gaskets, and the blockage of diaphragms, providing a basis for the replacement and maintenance of key components. Attached Figure Description

[0026] Figure 1 This is a framework diagram of an electrolytic cell fault prediction device according to the present invention. Detailed Implementation

[0027] The following detailed description illustrates the specific implementation method: The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Example 1 An electrolytic cell failure prediction method includes the following steps: 1. Multi-source synchronous monitoring: The cell voltage monitoring module scans all single-cell voltages at a frequency of 10Hz and forms a single-cell voltage distribution cloud map. The electrochemical noise monitoring module samples at a high speed of 1000Hz, recording the potential fluctuations and inter-electrode current fluctuations between the anode and cathode relative to the reference electrode. The fiber optic grating strain monitoring module continuously reads the strain values ​​of the endplate bolts (accuracy ±1με), reflecting the gasket compression state. The acoustic emission sensing module acquires structural wave signals from the outer side of the endplate in the range of 20-200kHz; when a micron-level leak occurs in the gasket, fluid turbulence excites characteristic frequency signals. The temperature sensing module (specifically an infrared thermal imager) scans the side of the electrolytic cell every 30 seconds, forming a temperature field (temperature distribution matrix). The thickness sensing module (specifically an ultrasonic thickness probe) acquires the electrode thickness at each measuring point, obtaining electrode thickness data. The pressure sensing module (specifically a pressure pulsation sensor) acquires the hydrogen outlet pressure and oxygen outlet pressure, performing spectral analysis. The component analysis module (including the online electrolyte component analyzer) takes samples from the alkali bypass every hour to detect the concentration of metal ions in the alkali bypass samples. The installation location, monitoring data, and corresponding faults for each monitoring module are shown in the table below.

[0029] Table 1: Installation location, monitoring data, and corresponding faults for each monitoring module

[0030] 2. Extract real-time edge features: (1) Electrochemical noise feature extraction: The noise signal collected by the electrochemical noise monitoring module was analyzed in time and frequency using Stockwell transform (S-transform) to calculate the noise resistance Rn (the ratio of the standard deviation of potential noise to the standard deviation of current noise). A decrease in Rn indicates increased corrosion activity. Transient peak features (peak height, half width at half maximum, and frequency) were extracted to distinguish between uniform corrosion and local pitting corrosion. S-transform can capture the peaks in the signal. The peak height, half width at half maximum, and frequency are determined based on the peaks. If there are no peaks or very few peaks, and they are not high, it is determined to be uniform corrosion. If there are many, high, and rapid peaks, it indicates local pitting corrosion.

[0031] (2) Acoustic emission feature extraction: Wavelet packet decomposition is performed on the acoustic emission waveform collected by the acoustic emission sensing module to extract the energy distribution of different frequency bands and identify the frequency band of the gasket micro-leakage event (usually the gasket micro-leakage is concentrated in 50-80kHz).

[0032] (3) Thermal imaging feature extraction: Calculate the uniformity index of the temperature field (the sum of squares of the deviations between the temperature at each point and the average temperature) and identify abnormal temperature gradient areas (which may be local dry areas caused by flow channel blockage).

[0033] (4) Voltage feature extraction: Calculate the voltage dispersion of the cell voltage (standard deviation / average value of each cell voltage) and identify the reverse polarity voltage (abnormally low or negative voltage, indicating electrode contamination or short circuit).

[0034] 3. Adaptive Correction of Dynamic Operating Conditions: The system reads current operating condition parameters (current, temperature, pressure, alkali flow rate), calls the pre-stored operating condition-feature mapping library, and normalizes the collected data and extracted features to eliminate spurious changes caused by load fluctuations. For example, when the current decreases, the background increase in gas purity is a normal phenomenon. The system adjusts through a dynamic baseline to avoid false alarms.

[0035] 4. Dual-channel parallel diagnosis: First diagnostic channel: Residual analysis based on physical model: Construct a lumped parameter model of the electrolyzer, which includes an electrochemical model, a thermal equilibrium model, and a fluid model. Compare the real-time measured values ​​of the electrolyzer with the predicted values ​​of the lumped parameter model to generate a residual sequence. When the residual sequence exceeds a set adaptive threshold, a model mismatch alarm is triggered, indicating that the electrolyzer has a risk of failure.

[0036] Second diagnostic channel: Data-driven pattern recognition: The extracted real-time edge features are input into the graph neural network diagnostic model, which outputs the probability distribution and confidence level of each fault type; the diagnostic results of the two channels are fused, and a diagnostic conclusion is output based on the fusion result.

[0037] The working method of the graph neural network diagnostic model is as follows: ① Constructing the input layer: The extracted real-time edge features form a multi-dimensional feature vector; the multi-dimensional feature vector has a dimension of M, which, combined with the time window length T, forms temporal feature input data that matches the dimension.

[0038] ② Constructing the Graph Layer: The electrolyzer is modeled as a graph structure G=(V,E), where the node set V contains the various chambers of the electrolyzer, the alkali inlet, the hydrogen and oxygen outlet, and various sensing modules. The edge set E contains physical and logical connections. Physical connections include pipeline connections and circuit connections, while logical connections are the relationships between components within the same region. Based on the fixed physical topology matrix A_physical and the real-time data correlation similarity matrix A_data(t), a dynamic adjacency matrix A(t) is constructed, calculated as: A(t)=α•A_physical+(1-α)•A_data(t), where α is the weight coefficient, and 0≤α≤1. A_data(t) is calculated based on the correlation of real-time time-series data changes of various sensing modules, and can adaptively capture changes in fault propagation paths caused by changes in operating conditions, such as changes in the impact range of flow channel blockage during load fluctuations. The higher the correlation of the voltage change trends of the sensors corresponding to two nodes, the larger the weight value of the corresponding edge.

[0039] ③ Constructing a spatiotemporal feature extraction layer: Temporal and spatial features are extracted from the input data separately. A gated causal convolutional network is used to process the temporal dimension features, filtering out future information leakage and extracting the single-node time dependency feature H_time. A graph attention network is used to process the spatial dimension features, calculating the attention coefficients between nodes, aggregating neighbor node information, and extracting the spatial dependency feature H_space. The gated causal convolutional network only extracts the time dependency features of historical time-series data, eliminating interference from future time-series data in feature extraction.

[0040] ④ Constructing the fusion layer: Adaptively fuse the time-dependent feature H_time and the spatial-dependent feature H_space through a gated fusion unit to obtain the fusion feature H_fusion. The calculation formula is: H_fusion=z⊙H_time+(1-z)⊙H_space, where z is the update gate weight calculated by the sigmoid function, and 0≤z≤1, and ⊙ is the element-wise multiplication operation; ⑤ Construct the output layer: Input the fused features H_fusion into the fully connected layer, and then output the classification probability of various faults in the electrolytic cell through the Softmax classifier to complete the fault diagnosis.

[0041] The advantage of this model lies in its ability to capture the propagation path of faults within the electrolyzer—for example, when the inlet is blocked, multiple downstream chambers experience a sequential voltage increase, a spatiotemporal propagation pattern that only graph neural networks can effectively learn. Comparative experiments show that, in complex fault scenarios, the dynamic GNN model of this invention achieves a diagnostic accuracy of 92%, while the traditional CNN model achieves 68%, and the MLP model only 55%, fully demonstrating its adaptability to the complex fault propagation characteristics of electrolyzers.

[0042] 5. Fault Warning: Faults are divided into multiple levels, which are further divided according to the urgency of the fault. Each level has corresponding triggering conditions and output action commands. In this embodiment, there are three levels. The specific judgment conditions and output actions are shown in the table below.

[0043] Table 2: Judgment Criteria and Output Actions for Each Level

[0044] 6. Remaining life prediction: (1) Lifetime prediction for electrode corrosion: Based on the cumulative rate of metal ion concentration and combined with the electrochemical noise trend, the Wiener process model is used to predict the remaining time for the effective electrode thickness to reach the threshold. The Wiener process model integrates direct measurement of ion concentration (reflecting the total corrosion) and electrochemical noise characteristics (reflecting the corrosion rate), and introduces a covariate correction term μ(t) (that is, the drift coefficient, a dynamic quantity that changes with time), μ(t) = μ0 + β•f_ECN(t), where f_ECN(t) is the reciprocal of the noise resistance, which can dynamically track the corrosion rate change and solve the problem that traditional models are difficult to adapt to the corrosion acceleration / deceleration process. μ0 is the baseline corrosion coefficient, which is the inherent corrosion baseline value under normal operating conditions, no abnormal corrosion, and no acceleration / deceleration trend, and is the "initial base" of the model. β is the corrosion sensitivity weighting coefficient, which is a proportional coefficient that measures how much the corrosion rate change can be driven by the change of noise signal. It is a constant obtained by fitting historical corrosion data and experimental tests.

[0045] Training method of Wiener process model: Collect historical data of metal ion concentration, electrochemical noise and corresponding actual corrosion thickness of multiple historical electrodes within a preset historical time period, and use these data as training set to train Wiener process model.

[0046] (2) Prediction of gasket aging: Based on the bolt strain relaxation curve and combined with the Arrhenius temperature acceleration model, the time for the gasket compression rebound rate to drop to the critical value is predicted. ① Establish the mapping relationship between bolt strain and gasket rebound rate: Through standard compression rebound test, the rebound rate of the target gasket under different clamping stresses is tested to obtain the stress-rebound rate benchmark curve. Combined with the bolt elasticity relationship, the real-time collected bolt strain value is converted into the current clamping stress of the gasket, and a quantitative conversion formula of bolt strain → gasket stress → compression rebound rate is established. Based on the real-time strain data, the curve of gasket rebound rate decaying with time is fitted, that is, the rebound rate degradation curve. ② Construct a time decay model of strain relaxation: The collected bolt strain relaxation data is fitted, and the exponential decay model is commonly used: ,in Let be the real-time strain of the bolt at time t; Let be the initial preload strain, k be the strain relaxation rate constant, and t be the service time. Converting strain decay to springback decay yields the springback degradation model: , Let be the shim rebound rate at time t. ③ Introducing the Arrhenius temperature-accelerated model to correct the rate: Through multiple sets of high-temperature accelerated aging tests, the activation energy Ea and constant A in the Arrhenius formula were calibrated to establish the temperature-relaxation rate correspondence. Substituting the actual operating temperature, the relaxation rate k under real service conditions was calculated to eliminate the interference of temperature fluctuations on the prediction results. ④ Calculating the critical failure time: The critical rebound rate of the gasket was set. Substitute into the rebound rate degradation model Deformation solution failure time: This refers to the time it takes for the gasket's compression resilience to drop to a critical value, which is also the remaining lifespan of the seal.

[0047] (3) Prediction of diaphragm blockage: Based on the rising trend of voltage dispersion and the growth rate of noise resistance, a long-term memory network (LSTM) is used to predict the time when the diaphragm pressure difference reaches the alarm value. The real-time collected voltage dispersion and noise resistance data are input into the trained LSTM model, and the trend curve of the diaphragm pressure difference in the future period is output. By curve fitting, the time when the pressure difference rises to the preset alarm value is solved, and the remaining alarm time is calculated.

[0048] 7. Generate Operation and Maintenance Decision Recommendations: Based on the diagnostic conclusions and the handling decisions in the fault knowledge base, automatically generate operation and maintenance recommendations. These recommendations include: Slight corrosion: It is recommended to adjust the operation and maintenance current density, reduce the load and operate for 1 month, and observe the trend; For localized blockages: backflushing or chemical cleaning is recommended, and a cleaning solution formulation suggestion will be provided. Gasket leakage: It is recommended to stop the machine in a planned manner and retighten the end plate bolts to the target torque value (specific value). If retightening is ineffective, arrange for gasket replacement. The electrode is severely corroded: a major overhaul and replacement are recommended, and a spare parts procurement warning should be provided (XX days in advance).

[0049] Stockwell transform was used to perform time-frequency analysis on the noise signal acquired by the electrochemical noise monitoring module. The electrochemical noise monitoring module includes a three-electrode system, a zero-resistance galvanometer, and a potentiometer. The three-electrode system is installed in the bypass flow cell of the alkaline solution circulation pipeline of the electrolyzer. The three-electrode system includes a working electrode (made of the same material as the electrode plate), a reference electrode (Hg / HgO), and an auxiliary electrode (platinum sheet). The zero-resistance galvanometer (ZRA) is used to measure the coupling current between the working electrode and the auxiliary electrode, and the potentiometer (input impedance > 10 Ω) is used to measure the coupling current between the working electrode and the auxiliary electrode. 12 Ω) is used to measure the potential of the working electrode relative to the reference electrode; the noise resistance Rn is the ratio of the standard deviation of the potential noise to the standard deviation of the current noise.

[0050] When the electrode corrodes, the metal dissolution and passivation film rupture / repair process will produce characteristic transient fluctuations (electrochemical noise) in the potential and current signals.

[0051] Traditional FFT analysis decomposes the entire signal into sine waves of different frequencies. However, for short-term non-stationary events such as pitting initiation, FFT will "smear" its energy across the entire frequency domain, resulting in blurred features.

[0052] This invention employs Stockwell transform, which combines the advantages of short-time Fourier transform and wavelet transform. It can achieve high frequency resolution in the low-frequency band (<1Hz) and high time resolution in the high-frequency band (100-500Hz), perfectly matching the characteristics of corrosion noise signals.

[0053] Specific algorithm: Calculate the Stockwell transform of the acquired potential / current time series x(t): S(τ,f) = ∫ x(t) · [|f| / √(2π)] · e^(-(τ-t)²f² / 2) · e^(-i2πft) dt The time-frequency spectrum matrix is ​​obtained, and the time-frequency energy peak and the main frequency migration trajectory are extracted from it as corrosion features; where f is the frequency, T is the time, and e^(-(τ-t)²f² / 2) is the frequency-dependent Gaussian window.

[0054] Comparative verification with traditional methods: Through accelerated corrosion experiments, corrosion monitoring was performed on the same electrode samples using the Stockwell transform method of this invention, the traditional FFT method, and the statistical method. The results showed that the Stockwell transform method could identify the characteristic spectral peak at the pitting corrosion initiation stage (corrosion pit depth < 5 μm), while the FFT method required a corrosion pit depth of more than 15 μm to capture a significant signal. The statistical method could not effectively distinguish between pitting corrosion and uniform corrosion. Experiments demonstrated that the sensitivity of this method for identifying pitting corrosion initiation is more than 3 times higher than that of the FFT method, and the corrosion rate measurement error is reduced from ±15% of the traditional method to ±3%.

[0055] The fiber optic strain monitoring module includes a fiber optic strain sensor, a fiber optic temperature sensor, a fiber optic demodulator (wavelength demodulation accuracy 1 pm, corresponding to a strain resolution of approximately 1 με), and a gasket condition assessment unit. The fiber optic strain sensor has multiple measuring points arranged along the axial direction of the end plate bolts or evenly distributed around the perimeter of the pole frame, for example, three measuring points spaced 120° apart around the perimeter of the pole frame. The fiber optic temperature sensor is used for temperature compensation of the fiber optic strain sensor. The fiber optic demodulator receives signals read by the fiber optic strain sensor and the fiber optic temperature sensor and sends these signals to the gasket condition assessment unit. The gasket condition assessment unit calculates the gasket stress relaxation rate based on the strain value and identifies micro-leakage by combining strain fluctuation spectrum analysis.

[0056] When the gasket is installed, it is compressed and generates a rebound force. This force is transmitted to the end plate bolts through the pole frame, causing the bolts to undergo tensile strain.

[0057] When the gasket experiences stress relaxation, creep, or aging shrinkage, its reaction force on the pole frame decreases, and the tensile strain of the bolt decreases accordingly.

[0058] The system monitors the bolt strain value ε(t) in real time and compares it with the initial installation strain ε0 to calculate the gasket stress relaxation rate R=[ε0-ε(t)] / ε0.

[0059] When R exceeds the set threshold (e.g., 15%), it is determined that the gasket's sealing ability has decreased, and the bolts need to be tightened again.

[0060] If the strain recovers rapidly after tightening but loosens again in a short period of time, it indicates that the gasket has undergone plastic deformation or broken and needs to be replaced.

[0061] Meanwhile, the fluctuation characteristics of the strain signal can reflect the pulsating transmission of fluid pressure. When a micro-leak occurs in the gasket, the pressure fluctuation caused by the leakage channel will be superimposed on the strain signal, which can be identified through spectrum analysis.

[0062] Synergistic Effect of Sensor Module Combination: Dual verification of gasket condition is achieved through the collaborative monitoring of a fiber Bragg grating strain sensor module and an acoustic emission sensor module. The fiber Bragg grating strain sensor module primarily captures the slow changes in gasket stress relaxation (timescale: days-months), while the acoustic emission sensor module primarily captures the instantaneous signals generated by micro-leakage (timescale: milliseconds-seconds). After data fusion, the gasket fault identification accuracy increases from 78% with a single sensor to 95%, and the false alarm rate decreases from 12% to below 2%.

[0063] The component analysis module includes a bypass sampling pump, a filtration unit, a dilution unit, a component analysis device, and a waste liquid recovery unit. The bypass sampling pump is used to periodically extract the electrolyte; the filtration unit is used to remove air bubbles and solid particles from the electrolyte; the dilution unit is used to dilute the regular electrolyte to the measurement range of the component analysis device; the component analysis device is used to analyze the concentration of various metal ions in the diluted electrolyte; and the waste liquid recovery unit is used to return the analyzed electrolyte to the pipeline of the electrolytic cell.

[0064] During corrosion, electrodes (especially anodes) release metal ions (Fe2+) into the electrolyte. 2+ / 3+ Ni 2+ Measuring the concentration and rate of change of metal ions in the electrolyte can directly quantify the corrosion rate.

[0065] The system automatically samples and analyzes once per hour, recording the concentrations of elements such as Fe, Ni, and Cr.

[0066] By applying Faraday's law, the cumulative metal loss is calculated from the ion concentration, and then the remaining electrode thickness is estimated.

[0067] When a sudden increase in the concentration of a certain element is detected (such as Fe concentration rising from ppb level to ppm level), it indicates that an accelerated corrosion event has occurred (such as passivation film rupture or pitting corrosion outbreak).

[0068] Accelerated aging tests can be performed by automatically injecting tracers (such as phosphoric acid) into the electrolyte and observing the corrosion response, which can quickly assess the electrode's tolerance to specific operating conditions. Example

[0069] like Figure 1As shown, an electrolytic cell fault prediction device includes a sensing unit, an edge computing unit, and a cloud intelligent unit. The edge computing unit is connected to both the sensing unit and the cloud intelligent unit. The sensing unit sends real-time data to the edge computing unit. The edge computing unit sends feature data to the cloud intelligent unit, which can update the model. The sensing unit includes a monitoring system and a clock synchronization system. The edge computing unit includes a multi-channel synchronous data acquisition system, a preprocessing system, and an embedded industrial control computer. The cloud intelligent unit includes a data storage and management system, a model training and optimization system, a digital twin system, and an application service system.

[0070] The monitoring system includes a cell voltage monitoring module, an electrochemical noise monitoring module, a fiber optic grating strain monitoring module, an acoustic emission sensing module, a temperature sensing module, a thickness sensing module, a pressure sensing module, and a composition analysis module. The cell voltage monitoring module collects the voltage of all individual cells and generates a single-cell voltage distribution cloud map. The electrochemical noise monitoring module collects the potential fluctuations between the anode and cathode, as well as the inter-electrode current fluctuations. The fiber optic grating strain monitoring module collects the strain values ​​of the endplate bolts. The acoustic emission sensing module collects the structural wave signals on the outer side of the endplate. The temperature sensing module collects the temperature on the side of the electrolytic cell and generates a temperature field. The thickness sensing module collects the electrode thickness at each measuring point. The pressure sensing module collects the hydrogen outlet pressure and oxygen outlet pressure. The composition analysis module detects the concentration of metal ions in the alkaline bypass sample.

[0071] The clock synchronization system is used to distribute the data synchronization pulses collected by the monitoring system to the multi-channel synchronous data acquisition system.

[0072] The preprocessing system is used to perform real-time filtering, feature extraction, and Stockwell transform on the transmitted data.

[0073] Embedded industrial control computers are used to achieve adaptive correction of dynamic operating conditions, dual-channel parallel diagnosis, and shutdown logic judgment.

[0074] The data storage and management system is used to store historical databases and fault knowledge bases.

[0075] The model training and optimization system is used to train and optimize models.

[0076] Digital twin systems are used to achieve 3D visualization, real-time mapping, and simulation.

[0077] The application service system is used for fault diagnosis, life prediction, operation and maintenance decision-making, report generation, and alarm push.

[0078] The following examples illustrate several specific implementation methods: I. Early Warning of Gasket Micro-Leakage After six months of operation, the following anomaly was detected in the system of an alkaline electrolytic cell: Fiber Bragg grating strain sensing module: The strain value of the end plate bolts slowly decreased from the initial 850με to 735με (relaxation rate 13.5%), close to the 15% threshold.

[0079] Acoustic emission sensing module: detects intermittent burst signals in the 50-80kHz frequency band, with an amplitude approximately three times that of the background noise.

[0080] Infrared thermal imager: The temperature of the outer wall at the corresponding location dropped locally (about 0.5°C), suspected to be due to the absorption of heat by the evaporation of alkali solution through micro-leakage.

[0081] Cell voltage: The voltage fluctuations of the two nearby cells increase, and the dispersion increases.

[0082] Diagnostic process: Acoustic emission event features were extracted from the edge side, and the matching degree with the gasket micro-leakage template in the fault knowledge base reached 87%.

[0083] The graph neural network diagnostic model integrates the above features and outputs a "gasket seal failure" probability of 92%, with high confidence.

[0084] The system triggered a level 2 warning, recommending "planned shutdown to check the gaskets, and recommending a retightening torque of 650 N·m".

[0085] Processing result: During a planned maintenance window, the operator shut down the machine and discovered that the gasket had been slightly extruded and deformed.

[0086] After tightening to the recommended torque, the strain returned to 820με, the acoustic emission signal disappeared, and the fault was resolved.

[0087] II. Early Warning of Accelerated Electrode Corrosion Rate An electrolyzer has been operating under a fluctuating renewable energy power supply mode for a long time, experiencing frequent start-ups and shutdowns. The system monitoring detected: Electrochemical noise: Stockwell transform spectrum shows a sustained increase in energy in the low frequency range (<1Hz), with the noise resistance Rn decreasing from 1200Ω·cm² to 350Ω·cm².

[0088] Electrolyte composition: Fe ion concentration gradually increased from 5 ppb to 28 ppb, and the rate of increase accelerated (from 0.1 ppb / day to 0.5 ppb / day).

[0089] Cell voltage: The dispersion of the voltage of the three cells in the corresponding region increased from 2% to 5%, and showed an upward trend.

[0090] Diagnostic process: Stockwell transform identifies typical pitting corrosion characteristics: high-frequency transient spikes appear in the time spectrum (corresponding to pitting initiation), while low-frequency energy continues to increase (corresponding to uniform corrosion acceleration).

[0091] Based on the Fe ion dissolution rate, the Wiener process model predicts that if the current operating mode is maintained, the remaining electrode lifespan will be approximately 8 months (less than the designed lifespan of 24 months).

[0092] The system triggered a Level 1 alert and generated the following recommendations: "Optimize the start-stop strategy to reduce the number of daily start-stop cycles; adjust the current rise rate from 100A / s to 50A / s; consider adding a corrosion inhibitor to the electrolyte."

[0093] Processing result: After the operators adopted the suggestions and optimized the control strategy, subsequent monitoring showed that the Fe ion dissolution rate dropped to 0.2 ppb / day, and the corrosion trend was brought under control.

[0094] To visually demonstrate the technological advancements of this invention, the system is compared with existing typical monitoring technologies (traditional DCS monitoring + regular maintenance mode), and the experimental results are shown in Table 3.

[0095] Table 3 Comparison of core performance indicators of the present invention and existing technologies.

[0096] The above effects are explained in detail below: (i) Significantly improve the ability to identify early and subtle faults. 1. Gasket micro-leakage early warning: Detected 15-30 days in advance. Comparative tests conducted on a commercially operating 1000 Nm³ / h alkaline electrolyzer showed that: Existing technology: Traditional pressure monitoring and gas purity analysis only trigger an alarm when there is a significant leak in the gasket (leaking amount of about 5-10 mL / min). At this time, the oxygen concentration in hydrogen has already risen from 0.2% to 0.6% (close to the 0.8% alarm threshold), which is only one step away from the risk of hydrogen-oxygen cross-contamination and explosion.

[0097] This invention utilizes a fiber optic strain sensor module to continuously monitor bolt strain changes, issuing an early warning when the gasket stress relaxation rate reaches 12% (leakage <1mL / min). Simultaneously, an acoustic emission sensor module captures characteristic leakage signals in the 50-80kHz frequency band, corroborating the strain data. Operators can then retighten the bolts during planned maintenance windows, preventing unplanned downtime.

[0098] Performance data: During 12 months of field testing, the system identified 17 early gasket anomalies, 15 of which were resolved by tightening the bolts, and only 2 required gasket replacement. Compared to the control group (which did not use this system), the number of unplanned downtimes due to gasket leakage decreased from 5 times / year to 1 time / year.

[0099] 2. Electrode corrosion detection sensitivity increased by 10 times. Traditional voltage monitoring only shows a significant change in chamber voltage (voltage increase of 20-50mV) when the electrode active area loss exceeds 10%. This invention, however, employs a combined monitoring method of electrochemical noise and electrolyte ion concentration. Electrochemical noise: The Stockwell transform was used to extract corrosion features, which can identify the characteristic spectrum at the stage of pitting corrosion initiation (corrosion pit depth <10μm, corrosion loss <1%). Experimental data show that when the noise resistance Rn decreases from 1200Ω·cm² to 800Ω·cm² (a decrease of 33%), the corresponding electrode mass loss is only 0.8%.

[0100] Ion concentration monitoring: When the Fe³⁺ concentration rises from the background of 5 ppb to 15 ppb (an increase of 10 ppb) by the ICP-MS online analyzer, the corresponding electrode corrosion rate is about 0.5 μm / month. At this time, the voltage change is less than 5 mV, which is completely undetectable by traditional monitoring.

[0101] Results data: At a hydrogen production station powered by a fluctuating power supply, the system provided an early warning three months in advance of the accelerated corrosion trend of electrodes in a certain area. After guiding the operators to optimize the start-up and shutdown strategy, the corrosion rate dropped from 0.5 μm / month to 0.15 μm / month, and the electrode life was extended by about 40%.

[0102] Traditional fixed-threshold alarm mechanisms frequently generate false alarms due to frequent load fluctuations (adjustment range 20%-100%, ramp rate up to 100% / s) caused by renewable energy grid connection. Statistics from a wind power hydrogen production project show that traditional systems generate an average of 3-5 false alarms per day when wind speed fluctuations are severe, leading to "alarm fatigue" among operators.

[0103] This invention employs dynamic operating condition adaptive correction technology: Establish a working condition-feature mapping library covering current densities of 0.2-1.0 A / cm². 2 A complete working environment with a temperature of 60-90℃ and a pressure of 0.1-1.0MPa.

[0104] The extracted features are normalized according to operating conditions. For example, when the current drops from the rated value to 30%, the increase in the background purity of the gas is a normal phenomenon, and the system automatically adjusts the alarm baseline.

[0105] Verification data: During a three-month field test, the system processed 2,386 changes in operating conditions, triggering 37 valid alarms, with only one false alarm (due to a temporary sensor malfunction). The false alarm rate was 2.7%, a 90% reduction compared to the over 30% of traditional systems. Operators' responsiveness to alarms significantly improved.

[0106] In actual operation of an electrolytic cell, multiple faults often occur simultaneously, making it difficult to distinguish them using traditional methods. For example, an increase in the voltage dispersion in a small chamber could be caused by electrode corrosion, flow channel blockage, or even flow deviation due to internal leakage in the gasket.

[0107] This invention employs a graph neural network diagnostic model, demonstrating advantages in the following cases: Test Case: After 18 months of operation, the following phenomenon occurred in an electrolytic cell: Area A (compartments 20-25): Voltage rise 50-80mV Area B (cells 40-45): Voltage drop of 20-30mV The overall oxygen concentration in hydrogen increased slightly (0.3% → 0.45%). Traditional diagnostic challenges: Increased voltage dispersion may indicate multiple faults, making it difficult to determine the primary cause.

[0108] Diagnostic process of this invention: The graph neural network diagnostic model constructs a dynamic adjacency matrix to analyze the spatiotemporal propagation pattern of faults. The anomaly was first detected near the inlet (Area A) and then spread downstream over time (Area B). Thermal imaging shows that area A has a higher temperature (dry area due to lack of liquid), while area B has a normal temperature but a lower voltage (possibly due to dilution by alkali solution). Acoustic emission detected a 50kHz leakage characteristic near region A. Model output fusion diagnostics: Main fault: Blockage of the liquid inlet channel in area A (94% probability), resulting in localized liquid shortage and voltage increase. Secondary failure: Due to reduced alkali distribution, gas purity in region B is affected (probability 87%). Potential hazard: The gasket near area A may experience a micro-leakage due to thermal stress (76% probability). Verification results: A shutdown inspection confirmed that crystals were indeed blocking the flow channels in sections 20-25 of the liquid inlet distribution plate, and the nearby gaskets showed slight thermal deformation. After cleaning and tightening the flow channels as recommended, all parameters returned to normal. The diagnostic accuracy far exceeds that of traditional methods.

[0109] To address the progressive failure of electrode corrosion, this invention establishes a mechanism-data hybrid-driven lifetime prediction model, achieving online assessment of the remaining lifetime of core components of an electrolytic cell for the first time.

[0110] 1. Electrode corrosion remaining life prediction Based on Fe in the electrolyte 3+ / Ni 2+ The concentration accumulation rate, combined with the electrochemical noise trend, is used to predict the remaining time for the remaining effective thickness of the electrode to reach the threshold using the Wiener process model.

[0111] Verification data: Accelerated aging tests were conducted in an electrolytic cell, operating continuously for 2000 hours at 1.2 times the rated current density. The system predicts the remaining lifespan in real time and compares it with the actual failure time. In the middle of the lifespan (50% remaining), the prediction error is approximately ±5%. In the later stages of the lifetime (with 20% remaining), the prediction error narrows to ±3%. The overall mean absolute percentage error (MAPE) is 7.2%, which is better than the traditional single model's over 15%. 2. Gasket relaxation life prediction Based on the bolt strain relaxation curve (cumulative monitoring data for 12 months) and combined with the Arrhenius temperature acceleration model, the time when the gasket compression rebound rate drops to the critical value is predicted.

[0112] Validation data: The model predicts that the relaxation life of a certain type of gasket at the rated temperature (85℃) is 5.8 years. After 5.5 years of actual operation, the strain relaxation rate exceeded the 25% threshold, requiring re-tightening. With a prediction error of 5.2%, it provides an accurate basis for planned overhauls.

[0113] 1. Enhanced security Hydrogen-oxygen cross-contamination risk prevention: Early warning is issued at the gasket micro-leak stage (oxygen concentration in hydrogen <0.5%) to prevent the leak from escalating to a dangerous concentration (>1.0%). During field testing, two potential hydrogen-oxygen cross-contamination incidents were successfully prevented.

[0114] Hotspot warning: Infrared thermal imaging identifies local temperature anomalies (temperature difference > 5℃), detecting 3 potential "dry zone" hazards in advance and preventing the diaphragm from burning through.

[0115] 2. Improved economic efficiency Reduced energy consumption: Timely warnings of electrode corrosion and flow channel blockage, guiding cleaning and maintenance, ensuring the electrolytic cell always operates within its high-efficiency range. After implementing this system in a certain project, the average annual DC power consumption decreased by 4.2% (approximately 1.5 million kWh / year, equivalent to approximately 900,000 RMB in electricity costs).

[0116] Unplanned downtime reduced: The number of unplanned downtime events decreased from an average of 4.2 times per year to 0.8 times per year. With each downtime loss estimated at 200,000 yuan, the annual savings amounted to 680,000 yuan.

[0117] Extended overhaul cycle: Precision condition inspection replaces periodic overhaul, extending the overhaul cycle of electrolytic cells from 5 years to 7 years, reducing the average annual overhaul cost by 28%.

[0118] Precise spare parts management: Life prediction guides spare parts procurement, reducing spare parts inventory capital tied up by approximately 35%.

[0119] To facilitate patent examination and technical evaluation, the core technical indicators of this invention are compared with the best levels reported in existing published literature:

[0120] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0121] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for predicting electrolytic cell failures, characterized in that: Includes the following steps: Multi-source data synchronous monitoring: The cell voltage inspection module collects the voltage of all individual cells; The electrochemical noise monitoring module collects potential fluctuations and inter-electrode current fluctuations between the anode and cathode; the fiber optic grating strain monitoring module collects strain values ​​of the endplate bolts; the acoustic emission sensing module collects structural wave signals from the outer side of the endplate; the temperature sensing module collects the temperature on the side of the electrolytic cell to form a temperature field; the thickness sensing module collects the electrode thickness at each measuring point; the pressure sensing module collects the hydrogen outlet pressure and oxygen outlet pressure; and the composition analysis module detects the concentration of metal ions in the alkaline bypass sample. Extracting real-time edge features: Perform time-frequency analysis on the noise signal collected by the electrochemical noise monitoring module, calculate the noise resistance Rn, and extract transient peak features; perform wavelet packet decomposition on the acoustic emission waveform collected by the acoustic emission sensing module, extract the energy distribution of different frequency bands, and identify the frequency band of the gasket micro-leakage event; calculate the uniformity index of the temperature field and identify the abnormal temperature gradient region; calculate the voltage dispersion of the cell voltage and identify the reverse polarity voltage; Adaptive correction of dynamic operating conditions: Read the current operating condition parameters, call the pre-stored operating condition-feature mapping library, and normalize the collected data and extracted features to eliminate spurious changes caused by load fluctuations; Dual-channel parallel diagnostics: First diagnostic channel: Residual analysis based on a physical model: Constructing a lumped parameter model of the electrolyzer, including an electrochemical model, a thermal equilibrium model, and a fluid model. Comparing the real-time measured values ​​of the electrolyzer with the predicted values ​​of the lumped parameter model to generate a residual sequence; when the residual sequence exceeds a set adaptive threshold, a model mismatch alarm is triggered, indicating a potential fault risk in the electrolyzer; Second diagnostic channel: Data-driven pattern recognition: Inputting extracted real-time edge features into a graph neural network diagnostic model, which outputs the probability distribution and confidence level of each fault type; fusing the diagnostic results from the two channels, and outputting a diagnostic conclusion based on the fusion result.

2. The method for predicting electrolytic cell failures according to claim 1, characterized in that: It also includes fault warning: faults are divided into multiple levels, which are divided in order of the urgency of the fault. Each level has a corresponding triggering condition and output action command.

3. The method for predicting electrolytic cell failures according to claim 1, characterized in that: It also includes remaining life prediction: For lifetime prediction of electrode corrosion: Based on the accumulation rate of metal ion concentration and combined with the electrochemical noise trend, the Wiener process model is used to predict the remaining time for the effective electrode thickness to reach the threshold. Prediction of gasket aging: Based on the bolt strain relaxation curve and combined with the Arrhenius temperature acceleration model, the time when the gasket compression rebound rate drops to the critical value is predicted. For diaphragm blockage prediction: Based on the increasing trend of voltage dispersion and the growth rate of noise resistance, a long-term memory network is used to predict when the diaphragm differential pressure will reach the alarm value.

4. The method for predicting electrolytic cell failures according to claim 1, characterized in that: It also includes generating operation and maintenance decision suggestions: based on the diagnostic conclusions and combined with the handling decisions in the fault knowledge base, operation and maintenance suggestions are automatically generated.

5. The method for predicting electrolytic cell failures according to claim 4, characterized in that: The maintenance recommendations include: Slight corrosion: It is recommended to adjust the operation and maintenance current density and reduce the load operation; For localized blockages: backflushing or chemical cleaning is recommended, and a cleaning solution formulation suggestion will be provided. Gasket leakage: It is recommended to plan a shutdown, retighten the end plate bolts to the target torque value, and if retightening is ineffective, arrange for gasket replacement; The electrode is severely corroded: a major overhaul and replacement are recommended, and a spare parts procurement warning should be provided.

6. The method for predicting electrolytic cell failures according to claim 1, characterized in that: Stockwell transform was used to perform time-frequency analysis on the noise signal collected by the electrochemical noise monitoring module. The electrochemical noise monitoring module includes a three-electrode system, a zero-resistance galvanometer, and a potentiometer. The three-electrode system is installed in the bypass flow cell of the alkaline solution circulation pipeline of the electrolytic cell. The three-electrode system includes a working electrode, a reference electrode, and an auxiliary electrode. The zero-resistance galvanometer is used to measure the coupling current between the working electrode and the auxiliary electrode, and the potentiometer is used to measure the potential of the working electrode relative to the reference electrode. The noise resistance Rn is the ratio of the standard deviation of potential noise to the standard deviation of current noise.

7. The method for predicting electrolytic cell failures according to claim 1, characterized in that: The fiber Bragg grating strain monitoring module includes a fiber Bragg grating strain sensor, a fiber Bragg grating temperature sensor, a fiber Bragg grating demodulator, and a gasket condition assessment unit. The fiber Bragg grating strain sensor has multiple measuring points arranged along the axial direction of the end plate bolts or multiple measuring points evenly arranged circumferentially around the outer perimeter of the pole frame. The fiber Bragg grating temperature sensor is used for temperature compensation of the fiber Bragg grating strain sensor. The fiber Bragg grating demodulator receives signals read by the fiber Bragg grating strain sensor and the fiber Bragg grating temperature sensor and sends the signals to the gasket condition assessment unit. The gasket condition assessment unit calculates the gasket stress relaxation rate based on the strain value and identifies micro-leakage by combining strain fluctuation spectrum analysis.

8. The method for predicting electrolytic cell failures according to claim 1, characterized in that: The working method of the graph neural network diagnostic model is as follows: Constructing the input layer: The extracted real-time edge features form a multi-dimensional feature vector; the multi-dimensional feature vector has a dimension of M, which, combined with the time window length T, forms temporal feature input data that matches the dimension; Graph construction layer: The electrolyzer is modeled as a graph structure G=(V,E), where the node set V contains the various chambers of the electrolyzer, the alkali inlet, the hydrogen and oxygen outlet, and various sensing modules, and the edge set E contains physical connection relationships and logical connection relationships. The physical connection relationships include pipeline connection relationships and circuit connection relationships, and the logical connection relationships are the association relationships of components in the same area. Based on the fixed physical topology matrix A_physical and the real-time data correlation similarity matrix A_data(t), a dynamic adjacency matrix A(t) is constructed, and the calculation formula is: A(t)=α•A_physical+(1-α)•A_data(t), where α is the weight coefficient, and 0≤α≤1; A_data(t) is calculated according to the correlation of changes in real-time time series data of various sensing modules. Construct a spatiotemporal feature extraction layer: extract time-dimensional and spatial-dimensional features from the input data respectively: use a gated causal convolutional network to process time-dimensional features and extract single-node time-dependent features H_time; use a graph attention network to process spatial-dimensional features, calculate attention coefficients between nodes, aggregate neighbor node information, and extract spatial-dependent features H_space; Constructing a fusion layer: Adaptively fuse the time-dependent feature H_time and the spatial-dependent feature H_space through a gated fusion unit to obtain the fusion feature H_fusion. The calculation formula is: H_fusion=z⊙H_time+(1-z)⊙H_space, where z is the update gate weight calculated by the sigmoid function, and 0≤z≤1, and ⊙ is the element-wise multiplication operation; Constructing the output layer: Input the fused features H_fusion into the fully connected layer, and then output the classification probability of various faults in the electrolytic cell through the Softmax classifier to complete the fault diagnosis.

9. The method for predicting electrolytic cell failures according to claim 1, characterized in that: The component analysis module includes a bypass sampling pump, a filtration unit, a dilution unit, a component analysis device, and a waste liquid recovery unit. The bypass sampling pump is used to periodically extract electrolyte. The filtration unit is used to remove air bubbles and solid particles from the electrolyte. The dilution unit is used to dilute the regular electrolyte to the measurement range of the component analysis device. The component analysis device is used to analyze the concentration of various metal ions in the diluted electrolyte. The waste liquid recovery unit is used to return the analyzed electrolyte to the pipeline of the electrolytic cell.

10. An electrolytic cell fault prediction device, characterized in that: The system includes a sensing unit, an edge computing unit, and a cloud-based intelligent unit. The edge computing unit is connected to both the sensing unit and the cloud-based intelligent unit. The sensing unit sends real-time data to the edge computing unit. The edge computing unit sends feature data to the cloud-based intelligent unit, which updates the model. The sensing unit includes a monitoring system and a clock synchronization system. The edge computing unit includes a multi-channel synchronous data acquisition system, a preprocessing system, and an embedded industrial control computer. The cloud-based intelligent unit includes a data storage and management system, a model training and optimization system, a digital twin system, and an application service system. The monitoring system includes a cell voltage monitoring module, an electrochemical noise monitoring module, a fiber optic strain monitoring module, an acoustic emission sensing module, a temperature sensing module, a thickness sensing module, a pressure sensing module, and a component analysis module. The cell voltage monitoring module collects the voltage of all individual cells and generates a single-cell voltage distribution cloud map. The electrochemical noise monitoring module collects the potential fluctuations between the anode and cathode, as well as the inter-electrode current fluctuations. The fiber optic strain monitoring module collects the strain values ​​of the endplate bolts. The acoustic emission sensing module collects the structural wave signals on the outer side of the endplate. The temperature sensing module collects the temperature on the side of the electrolytic cell, forming a temperature field. The thickness sensing module collects the electrode thickness at each measuring point. The pressure sensing module collects the hydrogen outlet pressure and oxygen outlet pressure. The component analysis module detects the concentration of metal ions in the alkali bypass sample. The clock synchronization system is used to distribute the data synchronization pulses collected by the monitoring system to the multi-channel synchronous data acquisition system; The preprocessing system is used to perform real-time filtering, feature extraction, and Stockwell transform on the transmitted data. The embedded industrial control computer is used to realize adaptive correction of dynamic working conditions, dual-channel parallel diagnosis, and shutdown logic judgment; The data storage and management system is used to store historical databases and fault knowledge bases; The model training and optimization system is used to train and optimize the model; The digital twin system is used to achieve 3D visualization, real-time mapping, and simulation. The application service system is used for fault diagnosis, lifespan prediction, operation and maintenance decision-making, report generation, and alarm push.

Images