A method for intelligent detection and repair of faults in methanol fuel cell stacks

Abstract

The present invention relates to the technical field of fuel cell fault detection, and specifically to an intelligent detection and repair method for faults in a methanol fuel cell stack. It synchronously collects the stack voltage, load power, methanol flow rate, and micro water vapor content to generate monitoring data for triggering abnormal troubleshooting. Based on this, it conducts cross-parameter correlation verification to output the root cause of the fault, calls historical data to construct and pre演 a candidate set to determine the target repair strategy, and then starts multi-module linkage repair to generate status data. Finally, it updates the prediction model and synchronizes the cluster to achieve global optimization. In the present invention, by constructing a closed-loop mechanism of pre-monitoring, determination, pre演, repair, and optimization, the fault handling is changed from discrete response to continuous collaborative control, enhancing the stack state convergence ability and repair consistency, forming a differentiated repair path based on the fault scenario, and promoting cluster strategy collaboration and quality convergence through data-driven model and strategy updates, thereby improving the system-level operation and maintenance efficiency and management ability.

Description

Technical Field

[0001] This invention relates to the field of fuel cell fault detection technology, and in particular to an intelligent detection and repair method for methanol fuel cell stack faults. Background Technology

[0002] During the operation of methanol fuel cell stacks, various factors such as load changes, fuel supply fluctuations, accumulation of water vapor generated by the reaction, and stack aging can easily lead to fault phenomena such as overload, abnormal water vapor, and voltage decay. If these faults are not identified and effectively handled in a timely manner, they often result in a decrease in stack output performance, a deterioration in operational stability, and even the risk of continuous shutdown. How to monitor the operating status of methanol fuel cell stacks in real time, provide early warnings, and achieve targeted repairs has become an important research direction in the field of fuel cell fault detection and control.

[0003] In existing technologies, the detection of faults in methanol fuel cell stacks often relies on threshold triggering mechanisms for a single monitoring quantity. For example, an abnormal state is judged based solely on whether one of the voltage, load, or humidity exceeds the limit. Although some solutions introduce time series prediction methods to predict faults, their model outputs usually adopt a linear correlation with fixed weights. This does not fully consider the voltage fluctuation characteristics during stack operation and the dynamic impact of changes in operating status at different time steps on the prediction results. As a result, the ability to adapt to complex dynamic operating conditions is insufficient, and it is difficult to balance the lead time and accuracy of fault prediction. In practical applications, the above methods are prone to problems such as delayed warnings and insufficient identification of potential faults.

[0004] Furthermore, existing technologies for determining water vapor anomalies mostly rely on a simple linear relationship between load power and water vapor content, failing to consider the nonlinear changes in water production behavior under load power overload, or the coupling effect of methanol input flow on reaction water production. This results in limited accuracy of cross-parameter verification, making misjudgments likely. Due to inaccurate fault root cause identification, the matching between subsequent fault classification and repair strategies is often insufficient, potentially triggering unnecessary repair operations. This not only increases system energy consumption and control burden but also hinders the long-term stable and reliable operation of the methanol fuel cell stack. Moreover, existing technologies typically treat the stack as a lumped parameter object, relying solely on a few external quantities such as terminal voltage, total power, or humidity for diagnosis and adjustment. This makes it difficult to describe the continuous evolution of methanol concentration, moisture migration, heat distribution, and local current density in the spatial dimension. Consequently, even if the external output returns to normal temporarily, potential problems such as localized overheating, localized water accumulation, or localized fuel shortages may still exist within the stack, making precise control of the work process and its uniformity difficult. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides an intelligent detection and repair method for methanol fuel cell stack faults. The technical solution is as follows: A method for intelligent detection and repair of faults in a methanol fuel cell stack includes the following steps: S1. Multi-parameter pre-monitoring and anomaly triggering: Simultaneously collect and preprocess four types of parameters, namely stack operating voltage, load power, methanol input flow rate, and micro-moisture content, to generate the first monitoring data; Based on the first monitoring data, an anomaly investigation process is triggered or continuous cyclical monitoring is performed. S2, Layered fault root cause localization: Receive the first monitoring data, reconstruct the methanol concentration field, moisture field, temperature field and local current density field inside the stack based on the infinite-dimensional dynamic system state observer, carry out cross-parameter correlation verification and scenario-based classification, and output the fault root cause determination result; S3. Repair strategy pre-play and feasibility evaluation: Receive the fault root cause determination result, the internal state field and the first monitoring data, call historical repair state data to construct a set of candidate repair strategies, and pre-play the target work trajectory, cumulative work amount and local state convergence trend under the action of each candidate repair strategy based on the reduced-order model of the infinite-dimensional dynamic system, perform pre-play evaluation of each candidate repair strategy, and output the strategy evaluation result and the target repair strategy. S4. Intelligent linkage repair: Receive the target repair strategy and the strategy evaluation result, start the multi-module linkage repair process of the corresponding scenario, and perform closed-loop precise control of methanol input flow, load power, drying intensity, activation intensity and preheating power according to the target work trajectory to generate repair status data; S5. Iterative optimization of model and strategy: Receive the repair status data and historical fault dataset, update the monitoring threshold and prediction model, status observation model and work control parameters, and synchronize them to the same cluster of electric stacks to achieve global strategy optimization.

[0006] As a further aspect of the present invention, the preprocessing in step S1 includes: Outlier removal, data standardization, and timestamp alignment are performed on the four types of parameters collected synchronously to generate the first monitoring data with consistent time sequence. The outlier removal is used to remove abnormal sampling points that deviate from the range of historical normal operation data; the data standardization is used to eliminate the dimensional differences of different parameters; and the timestamp alignment is used to unify parameter data from different sampling periods to the same time interval.

[0007] As a further aspect of the present invention, step S1 further includes S1-1 trend prediction-based anomaly triggering, specifically: Construct an LSTM time series prediction model, input the first monitoring data, and calculate using the following formula. Predicted parameter values ​​at time : ; in, These are the weight coefficients of the LSTM hidden layer; The output of the LSTM hidden layer at time t; The dynamic weighting factor for the time step at time t; This is the correction factor for the stack voltage; Let t be the operating voltage of the fuel cell stack. For the bias term of the LSTM model; When the parameter prediction value When the value exceeds the preset threshold range, the fault root cause location process in step S2 is directly triggered. The LSTM time series prediction model takes the first monitoring data as a time step sequence as input, and adjusts the weight coefficients of the LSTM hidden layer based on the historical fault dataset. , Pile voltage correction factor LSTM model bias term and time step dynamic weighting factor Perform fitting; The time step dynamic weight factor The forecast is dynamically adjusted based on the fluctuation range of the monitoring data to reflect the proportion of the impact of current monitoring data and historical monitoring data on the forecast results.

[0008] As a further aspect of the present invention, step S2 further includes S2-1 cross-parameter correlation verification, specifically including: Construct a correlation model between operating load and water vapor content. Input the load power, trace water vapor content, and methanol input flow rate data from the first monitoring data, and calculate the theoretical water vapor content under the corresponding operating condition using the following formula. : ; in, The correlation coefficient for the quadratic term of load power; Let t be the actual load power of the fuel cell stack. The correlation coefficient for methanol flow rate; Let t be the methanol input flow rate; Basic correction coefficient; Actual micro-water vapor content With the theoretical water vapor content Compare the results and eliminate any incorrect judgments that do not conform to the reaction pattern.

[0009] As a further aspect of the present invention, the correlation model between the operating load and the water vapor content comprehensively considers the nonlinear water production characteristics under load power overload and the coupled influence of methanol input flow rate on the reaction water production. The correlation coefficient of the quadratic term of the load power Methanol flow correlation coefficient and basic correction coefficient The data was obtained by fitting experimental data under different load and flow conditions.

[0010] As a further aspect of the present invention, step S2 further includes S2-2 AI-based scenario-based fault classification, specifically as follows: A lightweight CNN+SVM fusion model is adopted. The first monitoring data and the data after cross-parameter correlation verification are input to classify the stack faults in a scenario-based manner. The faults are divided into three categories: overload, excessive moisture, and voltage decay. The root cause determination result of the fault is output.

[0011] As a further aspect of the present invention, step S3, which involves performing a preliminary evaluation of each candidate repair strategy, includes: Based on the fault root cause determination results, the first monitoring data, and historical repair status data, the parameter change trend, fault reduction magnitude, operational disturbance degree, and resource call cost after the execution of each candidate repair strategy are evaluated, and the target repair strategy is obtained by screening according to the preset evaluation rules. The candidate repair strategies include at least different module combination repair strategies corresponding to overload, excessive moisture, and voltage decay.

[0012] As a further aspect of the present invention, step S4 further includes S4-1 adaptive intensity repair adjustment, specifically including: Based on the severity of the fault according to the fault root cause determination results, the fault is divided into mild, moderate and severe levels, and the operating power of the repair module is adjusted for different levels. Among them, when the operating condition is overloaded, the power regulation module and the methanol flow regulation module will work together; when the water vapor exceeds the standard, the self-filtration module and the drying module will work together; and when the voltage decays, the stack activation program will be started. During the operation of step S4, the first monitoring data generated in step S1 is collected every 2 seconds. The operating parameters of the repair module are dynamically adjusted according to the real-time changes of the first monitoring data until the parameters corresponding to the fault are restored to the normal range, and the repair status data is generated.

[0013] As a further aspect of the present invention, step S5 further includes: S5-1 Cluster Strategy Synchronization Optimization: The updated monitoring thresholds and prediction model parameters are synchronized to the monitoring modules of all fuel cell stacks in the same cluster through the edge data transmission channel to achieve unified optimization of global strategies; S5-2 Threshold and Model Self-Update: Based on the reinforcement learning framework, the repair state data and historical fault dataset are input, and the weight coefficient wL, voltage correction coefficient kV, bias term bL and time step dynamic weight factor γt of the LSTM time series prediction model are updated, as well as the correlation coefficients kP, kF and basic correction coefficient bF of the load-water content correlation model. The state observer parameters, reduced-order model parameters and work control objective function weights are also updated.

[0014] As a further aspect of the present invention, in the iterative optimization of the model and strategy, the updated monitoring threshold, prediction model parameters, working load and water vapor content correlation model parameters, state observer parameters, reduced-order model parameters, and work control objective function weights are synchronized to all stacks in the same cluster, and then used to uniformly optimize the fault monitoring, fault root cause location, repair strategy simulation, and intelligent linkage repair process of the stacks in the same cluster.

[0015] As a further aspect of the present invention, an infinite-dimensional dynamic system state observer is constructed. Using the spatial coordinate z of the fuel cell flow channel direction or the membrane electrode thickness direction as the distribution dimension, a state vector X(z,t) is established, which includes the methanol concentration field c(z,t), the moisture field w(z,t), the temperature field T(z,t), the membrane water content field λ(z,t), and the local current density field i(z,t). The state vector is then reconstructed online based on the first monitoring data.

[0016] As a further aspect of the present invention, the pre-evaluation in step S3 includes based on the instantaneous power output. and the cumulative work done within the preset evaluation time window H The work deviation, local current density variance, moisture field deviation, and temperature field deviation corresponding to each candidate repair strategy are calculated, and the target repair strategy is obtained by screening based on these parameters.

[0017] As a further aspect of the present invention, the closed-loop precise control in step S4 adopts a model predictive control method after the order reduction of the infinite-dimensional dynamic system, and jointly adjusts the methanol input flow rate, load power, drying intensity, activation intensity and preheating power to make the actual work trajectory track the target work trajectory and suppress local overheating, local water accumulation and local material shortage.

[0018] As a further embodiment of the present invention, step S5 updates the state observer parameters, the reduced-order model parameters, and the weights of the work control objective function while updating the monitoring threshold and the prediction model, and synchronizes them to the same cluster of fuel cells to achieve cluster-level precise work uniform optimization.

[0019] The beneficial effects of the technical solutions provided by the embodiments of the present invention include at least the following: This invention constructs a closed-loop handling mechanism that integrates pre-monitoring, root cause determination, strategy pre-simulation, coordinated repair, and iterative optimization. It incorporates methanol concentration, moisture, temperature, and local current density fields into a unified state observation and predictive control framework. This transforms the fault handling process from discrete response to continuous collaborative control, enhancing the stack's state convergence capability and repair consistency in complex operating scenarios. It achieves precise and coordinated control of the target work trajectory, cumulative work, and local work uniformity. Differentiated repair paths are formed based on the fault scenario and severity, ensuring a higher degree of matching between repair actions and operating states, thus improving the targeting and execution stability of the control process. Furthermore, the repair state data drives continuous updates to monitoring thresholds, predictive models, and strategy rules, promoting strategy coordination and operational quality convergence among stacks in the same cluster, thereby improving system-level operation and maintenance efficiency and cluster-level management capabilities. Attached Figure Description

[0020] Figure 1 This is a flowchart of the intelligent detection and repair process for methanol fuel cell stack faults in this invention. Figure 2 This is a flowchart of the intelligent fault detection and repair process for the vehicle-mounted methanol fuel cell stack of the present invention. Figure 3 This is a flowchart illustrating the intelligent fault detection and repair process for a stationary distributed power generation methanol fuel cell stack cluster according to the present invention. Detailed Implementation

[0021] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0022] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the term "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0023] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0024] Please see Figure 1Example 1: General implementation method for intelligent detection and repair of methanol fuel cell stack faults; This invention provides a technical solution: a method for intelligent detection and repair of faults in a methanol fuel cell stack, comprising the following steps: Step S1: Multi-parameter pre-monitoring and anomaly triggering First, the stack operation status data acquisition system is started to synchronously collect four types of core operating parameters during the operation of the methanol fuel cell stack. The four types of core operating parameters include: stack operating voltage, load power, methanol input flow rate, and micro-water vapor content.

[0025] Specifically, the operating voltage of the fuel cell stack is preferably acquired using a voltage sampling circuit. The voltage sampling circuit can be implemented using an isolated voltage sensor, a Hall voltage sensor, or a voltage acquisition unit composed of a precision voltage divider network, an isolation amplifier, and an analog-to-digital converter. The preferred sampling period is 100ms, and the unit is V. Using the above-mentioned existing technologies, continuous acquisition of the fuel cell stack terminal voltage or single-chip voltage can be achieved. The isolated sampling scheme is beneficial for suppressing common-mode interference and improving system safety.

[0026] The load power is preferably calculated by the controller after synchronous sampling of voltage and current, or it can be directly obtained using existing power analyzers, power transmitters, or electrical parameter acquisition modules. When using the calculation method, the load power can be obtained in real time using the following expression: ; in, for Real-time load power, in kW; for Real-time voltage data, in V; for The actual current is collected at all times, in amperes (A). To avoid confusion with the normalized variables input to subsequent models, this embodiment uses the superscript "A". The variable “” represents the actual measurement in the physical quantity space, and the preferred sampling period is 200ms.

[0027] The methanol input flow rate is preferably collected using a liquid flow detection unit. This unit can be a micro-flow Coriolis mass flow meter, a thermal micro-flow meter, a volumetric micro-flow meter, or a flow rate calculated by back-calculating the flow rate using a metering pump pulse frequency. When higher accuracy and stronger media adaptability are required, a Coriolis flow meter or a micro-flow controller from the existing technology is preferred, with a preferred sampling period of 500 ms and the unit expressed as L / min. If a mass flow meter is used at the front end, the controller converts the methanol density into volumetric flow rate before using it in subsequent calculations.

[0028] The moisture content is preferably collected using a humidity and dew point detection unit. The humidity and dew point detection unit can be implemented using an optical cold mirror dew point meter, a heated humidity probe, a thin film capacitive humidity sensor with a temperature compensation algorithm, or a dew point transmitter. In near-condensation, high-humidity environments of fuel cells, heated humidity probes or optical cold mirror dew point meters are preferred to improve continuous measurement stability. A sampling period of 1 second is preferred. If the front-end sensor outputs relative humidity, dew point temperature, or absolute humidity, the controller will convert it to absolute moisture content using a humid air conversion model and use it as "micro-moisture content" for subsequent comparisons.

[0029] The collected raw monitoring data first enters the preprocessing stage, which includes outlier removal, data standardization, and timestamp alignment to generate first monitoring data with consistent time sequence. To avoid inconsistencies in variable space between paragraphs, this embodiment explicitly stipulates that: all quantities involved in physical mechanism comparison, threshold engineering judgment, and execution module control shall use physical quantity space data; all quantities involved in neural network training and inference may use normalized model input data, and the two are mapped to each other through forward normalization and denormalization operations.

[0030] Among them, outlier removal is preferably adopted. The criterion is as follows: when the parameter value corresponding to a certain sampling point exceeds the range of plus or minus three standard deviations of the average value of the parameter's historical normal operation data, the sampling point is determined to be an abnormal sampling point and is removed. To avoid time series breaks after removing outliers, the missing data after removal can be compensated by linear interpolation to ensure data continuity.

[0031] Furthermore, when multiple consecutive sampling points cross the boundary simultaneously, they are not directly identified as single-point noise, but are first labeled as pending verification and then judged jointly by the subsequent trend model and cross-parameter verification, in order to avoid the false deletion of real faults.

[0032] Data standardization preferably employs the min-max normalization method to uniformly map sampled data of different dimensions to a uniform normalization level. The interval, its normalized expression is: ; in, This represents the normalized parameter value. This represents the original parameter value. and These represent the minimum and maximum values ​​of the corresponding parameters in the historical normal operation samples, respectively. Therefore, the normalized values ​​are... Since it is a dimensionless quantity, if it is subsequently fed into the LSTM model, all the calculations in the model will be performed in a dimensionless space, which can avoid the problem of inconsistent dimensions.

[0033] Timestamp alignment is used to unify parameter data from different sampling periods to the same time interval. Preferably, linear interpolation is used to align four types of parameters—pile operating voltage, load power, methanol input flow rate, and trace water vapor content—to a 1-second time interval, thus forming time-consistent first monitoring data. The first monitoring data is preferably stored in a two-dimensional array, where each row corresponds to the four types of parameter values ​​under a unified timestamp, and each column corresponds to continuous time-series data for one type of parameter. To ensure clear data retrieval relationships between segments, the first monitoring data includes at least: timestamp, actual voltage, and so on. Actual current (If the power is calculated from voltage and current, retain it) Actual load power Actual methanol input flow rate Actual micro-water vapor content , and the corresponding normalized data.

[0034] In a preferred embodiment, after generating the first monitoring data, an internal state field observer for the fuel cell stack is further constructed based on the theory of infinite-dimensional dynamical systems. The state field observer establishes a state vector with the spatial coordinate z along the flow channel direction or the film electrode thickness direction of the fuel cell stack as the distribution dimension. Where c(z,t) is the methanol concentration field, w(z,t) is the moisture field, T(z,t) is the temperature field, λ(z,t) is the membrane water content field, and i(z,t) is the local current density field; the observer reconstructs the state vector online based on the first monitoring data, boundary inputs, and historical calibration parameters to obtain the continuous distribution state inside the stack that is difficult for external sensors to measure directly, providing a state basis for subsequent fault location, strategy simulation, and precise control of work.

[0035] After obtaining the first monitoring data, execute S1-1 trend prediction anomaly trigger.

[0036] Specifically, an LSTM time series prediction model is constructed by serializing the first monitoring data into a time step sequence. Predict parameters at any given time.

[0037] Preferably, the LSTM time series prediction model has 4 neurons in the input layer, corresponding to the four types of monitoring parameters; 32 neurons in the hidden layer; and 4 neurons in the output layer, corresponding to... Predicted values ​​of four types of parameters at time.

[0038] Furthermore, a sliding time window approach can be used to construct the model input samples, for example, by taking data from the past 5 consecutive timestamps. The feature matrix is ​​used as an input sample.

[0039] The LSTM time series prediction model is calculated using the following formula. Predicted parameter values ​​at time : ; Among them, the parameter prediction value This includes predicted values ​​for stack operating voltage, load power, methanol input flow rate, and micro-moisture content; These are the weight coefficients of the LSTM hidden layer; for LSTM hidden layer output at time; for The dynamic weighting factor of the time step at any given moment; This is the correction factor for the stack voltage; for The input characteristic value corresponding to the stack operating voltage at any given time; This is the bias term for the LSTM model.

[0040] The engineering significance of the above expression lies in: outputting from the hidden layer. Characterizing multi-parameter historical coupling features, in order to The bias term characterizes the impact of the current voltage state on the correction results of future parameter predictions. The overall output of the model is calibrated. Since all input terms in the formula are taken from a unified input feature space, therefore... The model outputs predicted values, which can be recovered to the corresponding physical quantity predicted values ​​through inverse normalization.

[0041] During the model training phase, historical failure datasets and historical normal operation datasets are used as training samples to adjust the parameters. , , as well as Perform fitting.

[0042] Preferably, 1200 sets of historical fault data and 800 sets of historical normal operation data can be used for training, and the model parameters can be updated through the backpropagation algorithm.

[0043] Time step dynamic weighting factor The system is dynamically adjusted based on the fluctuation range of monitoring data to reflect the relative impact of current and historical monitoring data on the prediction results. For example, it can be adjusted based on the mean time-series standard deviation of each monitoring parameter over the past 10 seconds: when the mean standard deviation is greater than 0.1, the system will adjust accordingly. Adjusted to 0.8 to enhance the sensitivity of the current perturbation data to the prediction results; when the mean standard deviation is less than 0.05, Adjusted to 1.2 to improve the contribution of historical conditions to forecast stability, when the mean standard deviation is between 0.05 and 0.1. Keep it at 1.0.

[0044] After the model prediction is completed, the predicted values ​​of the parameters are compared with the corresponding preset threshold ranges. The preset threshold ranges are preferably determined based on the 95% confidence interval of historical normal operation data. When the predicted value of any parameter exceeds the preset threshold range, the system directly triggers the fault root cause location process in step S2. When the predicted values ​​of all parameters are within the preset threshold range, the continuous cyclic monitoring state is maintained.

[0045] Step S2: Root cause localization of layered faults After an abnormal trigger signal is detected in step S1, the fault location module receives the first monitoring data and performs a hierarchical fault root cause location process. The hierarchical fault root cause location includes: S2-1 cross-parameter correlation verification and S2-2 AI-based scenario-based fault classification. To avoid confusion between theoretical water vapor content and actual micro-water vapor content, this embodiment explicitly uses different symbols in this step: Indicates the theoretical water vapor content. This indicates the actual micro-water vapor content.

[0046] In this step, in addition to using the first monitoring data for discrimination, the X(z,t) output by the state field observer is also called as the internal state basis for locating the root cause of the hierarchical fault. The local current density peak, temperature gradient, moisture gradient and methanol concentration gradient are used to characterize the uniformity of internal work distribution and potential instability risk.

[0047] S2-1 Cross-parameter correlation verification First, a correlation model between operating load and water vapor content is constructed to derive the theoretical water vapor content based on load power and methanol input flow rate, thereby eliminating misjudgments inconsistent with the reaction mechanism. The expression for the correlation model between operating load and water vapor content is as follows: ; in, for The theoretical water vapor content under the corresponding operating conditions at any given time, in units of ; The correlation coefficient for the quadratic term of load power; for Real-time load power of the fuel cell stack; The correlation coefficient for methanol flow rate; for Methanol input flow rate at all times; The basic correction factor.

[0048] The engineering significance of the above formula is as follows: The term is used to reflect the nonlinear enhancement of water production behavior, especially in the high-load range, when the load increases. This item is used to reflect the impact of changes in methanol supply on the total amount of water vapor produced in the reaction. This expression is used to absorb baseline offset, environmental factors, and system constant deviations. It belongs to the empirical correlation model and is used for engineering discrimination and misjudgment elimination.

[0049] The load-water content correlation model comprehensively considers the nonlinear water production characteristics under load power overload, as well as the coupled effect of methanol input flow rate on the water production rate. Preferably, the parameters... , as well as It can be obtained by fitting test data under different loads and flow conditions. For example, 100 sets of test data under different conditions can be used to complete the parameter fitting by the least squares method.

[0050] After obtaining the theoretical water vapor content Then, the actual micro-water vapor content will be... Compare the results with the theoretical water vapor content. If the deviation exceeds a preset allowable deviation range, for example, if the following conditions are met: ; in, The allowable deviation threshold is preferably set to 0.5. If the current data does not conform to the reaction pattern, it is considered that there is a risk of misjudgment, and the process returns to step S1 to continue monitoring; if the following conditions are met: ; If the current data is deemed valid, proceed to the S2-2AI scenario-based fault classification stage.

[0051] Furthermore, when deviations exceed the limit twice or more consecutively, the data is not discarded directly, but the batch is marked as a sensor abnormality or a state to be calibrated, triggering sensor self-test, zero-point verification, or maintenance alarm.

[0052] S2-2AI scenario-based fault classification, after completing cross-parameter correlation verification, uses a lightweight CNN+SVM fusion model to classify fuel cell stack faults in a scenario-based manner and outputs the fault root cause determination results.

[0053] Preferably, the model input is the first monitoring data after cross-parameter correlation verification, and can further be composed of four types of parameter data from the past 30 unified timestamps. Feature matrix. The lightweight CNN part preferably adopts the MobileNetV2 structure to extract local temporal features; the SVM part preferably adopts the RBF kernel function to achieve the final classification decision. The input of the fusion model is the preprocessed temporal feature data, and the output is discrete fault type labels and corresponding confidence scores, which can be called by the subsequent policy pre-playback module.

[0054] The fusion model classifies faults into three categories: overload, excessive moisture, and voltage decay.

[0055] Preferably, each fault category can be defined according to the following conditions: overload refers to the load power exceeding the rated value by more than 10%; excessive water vapor refers to the actual micro-water vapor content exceeding the theoretical value by more than 10%; and voltage decay refers to the stack operating voltage being lower than the rated value by more than 5%.

[0056] After classification, the fault root cause determination result is output, and the fault root cause determination result and the current first monitoring data are transmitted together to the repair strategy pre-play and executability evaluation module in step S3.

[0057] Step S3: Remediation Strategy Preview and Feasibility Assessment In this step, the repair strategy pre-simulation module receives the root cause determination results and the first monitoring data, and uses historical repair status data to construct a set of candidate repair strategies. Based on this, a pre-simulation evaluation is performed on each candidate repair strategy, outputting the strategy evaluation results and the target repair strategy. To ensure the full disclosure of this step, the historical repair status data includes: historical fault type, fault severity, the first monitoring data at that time, the actual repair strategy executed, repair duration, parameter changes before and after repair, resource call records, and repair success / failure tags.

[0058] To achieve precise control of work output, this step further constructs a reduced-order pre-simulation model of the infinite-dimensional dynamic system based on the state vector X(z,t). Preferably, modal truncation, Galerkin projection, or POD principal mode extraction can be used to map the high-dimensional continuous distribution state into a low-dimensional modal state vector, which is used to quickly predict the impact of each candidate repair strategy on the internal state field and external work output within a preset evaluation window.

[0059] The candidate repair strategies should include at least different module combination repair strategies corresponding to overload, excessive moisture, and voltage decay. Preferably, a candidate strategy library can be established according to the fault type.

[0060] For example, for overload faults, the following candidate repair strategies can be set: Strategy Only the power regulation module is activated to adjust the load power accordingly; Strategy The power regulation module and the methanol flow regulation module work together to regulate the flow. Strategy In strategy Based on this, a heat dissipation module is further activated to improve thermal stability.

[0061] For faults caused by excessive moisture, the following candidate repair strategies can be set: Strategy : Activate the self-filtering module; Strategy Start the self-filtration module and drying module; Strategy In strategy Based on this, a methanol flow regulation module is used to reduce the rate of water production in the reaction.

[0062] For voltage attenuation faults, the following candidate repair strategies can be set: Strategy Perform a low-intensity activation procedure; Strategy : Perform a medium-intensity activation procedure; Strategy : Execute a high-intensity activation procedure and activate the preheating module.

[0063] The acquisition methods and execution methods of the above-mentioned repair modules can all be implemented based on existing technologies: the power regulation module can be implemented by a DC / DC converter control unit, load manager, or power distribution controller; the methanol flow regulation module can be implemented by a proportional valve, a stepper motor-driven metering pump, or a mass flow controller; the heat dissipation module can be implemented by a fan, a liquid cooling pump, a heat dissipation valve group, and a thermal management controller; the self-filtration module can be implemented by a replaceable adsorption unit or an online filtration channel; the drying module can be implemented by a dryer, a regenerated adsorption bed, or a heating and dehumidification unit; the activation program can be implemented by the fuel cell stack controller outputting a preset current curve; and the preheating module can be executed by a PTC heater, a liquid circuit heater, or a heat exchanger.

[0064] Furthermore, the self-filtration module refers to an online water vapor filtration and droplet separation unit installed in the exhaust channel, circulating gas path, or water vapor collection channel of a methanol fuel cell stack. It is used to separate, trap, and discharge condensate, micro-droplets, methanol mist droplets, and particulate impurities entrained in the airflow under conditions of excessive water vapor. The self-filtration module may include one or more combinations of a gas-liquid separation chamber, a hydrophobic microporous membrane, a demister filter, an adsorption filter layer, a liquid collection chamber, a drain valve, a bypass valve, and a differential pressure detection unit. The gas-liquid separation chamber causes a change in the flow rate or swirling separation of the mixed airflow carrying water vapor; the hydrophobic microporous membrane and demister filter intercept micro-droplets and mist-like moisture; the adsorption filter layer adsorbs residual water vapor or methanol mist droplets; and the liquid collection chamber and drain valve collect and discharge the separated liquid water or condensate.

[0065] When a water vapor exceeding the standard fault occurs, the controller controls the self-filtering module to enter the working state based on the actual micro water vapor content, theoretical water vapor content, differential pressure before and after filtration, and the operating status of the fuel cell stack.

[0066] When the actual moisture content is only slightly above the standard, the self-filtration module is activated first to reduce free water and entrained droplets in the gas path through online separation, interception and drainage. When the actual moisture content continues to rise or exceeds the moderate or severe threshold, the drying module is activated in addition to the self-filtration module. The self-filtration module first removes droplet-like and mist-like moisture, and then the drying module reduces the residual gaseous moisture content, thereby achieving graded dehumidification and synergistic repair.

[0067] The self-filtering module and the drying module have different functions: the self-filtering module is mainly used to remove condensate, droplets, mist, and particulate impurities entrained in the air path, and belongs to the pre-stage online filtration and droplet separation mechanism; the drying module is mainly used to reduce the gaseous water vapor content in the airflow, and belongs to the post-stage dehumidification mechanism. By setting the self-filtering module and the drying module in tandem, it is possible to avoid a large number of droplets directly entering the drying module, which could cause the drying material to become saturated too quickly or reduce the dehumidification efficiency, thereby improving the stability and continuity of the fault repair process for excessive water vapor.

[0068] When conducting a preliminary evaluation of each candidate repair strategy, based on the fault root cause determination results, the first monitoring data, and historical repair status data, the following aspects after the execution of each candidate strategy are comprehensively evaluated: Parameter change trend; The extent of fault reduction; Degree of operational disturbance; Cost of resource allocation.

[0069] Target work deviation and local work uniformity indicators.

[0070] Among them, the parameter change trend is used to describe the direction and convergence speed of the changes in the stack operating voltage, load power, methanol input flow rate, and micro-moisture content within the preset evaluation time window after executing a certain candidate repair strategy; the fault reduction magnitude is used to characterize the degree to which abnormal indicators recover to the normal range; the degree of operational disturbance is used to describe the impact of the repair action on the continuity and stability of the current system output; the resource call cost is used to reflect the power consumed, module occupancy time, methanol adjustment amount, and auxiliary resource call cost of executing the candidate repair strategy. The target work deviation is used to characterize the degree of deviation between the actual output power trajectory and the target work trajectory under the action of the candidate repair strategy, and the local work uniformity index is used to characterize the comprehensive fluctuation level of local current density variance, temperature gradient, and moisture gradient. Regarding the continuity of the change direction, this embodiment further clarifies that: if the first-order difference sign of a parameter remains consistent for multiple consecutive moments within the pre-test window, it is considered that the change direction is continuous; if the first-order difference sign changes repeatedly in adjacent moments, it is considered that the change direction is discontinuous, indicating that the candidate strategy may cause oscillation or control instability, and points will be deducted in the comprehensive score, rather than being directly judged as unexecutable.

[0071] Furthermore, instantaneous power is defined as The cumulative work done within the preset evaluation time window H is defined as The target work deviation is preferably represented by a weighted sum of the cumulative work deviation, instantaneous power tracking error, and local current density variance.

[0072] Preferably, a comprehensive evaluation scoring function can be constructed to weight and integrate the above four indicators. For example: in, The overall evaluation score for candidate repair strategies; Score the trend of parameter changes; Scoring is given based on the degree of fault reduction; Score the degree of operational disturbance; Score the cost of resource retrieval; Scoring is given for deviations in work done towards the target and unevenness in work done in certain areas; These are the corresponding weighting coefficients.

[0073] Step S4: Intelligent Linkage Repair After receiving the target repair strategy and strategy evaluation results, the intelligent linkage repair module initiates the multi-module linkage repair process for the corresponding scenario and generates repair status data.

[0074] In a preferred embodiment, the intelligent linkage repair module further invokes the reduced-order pre-simulation model and target work trajectory output in step S3 to construct a closed-loop controller with the goal of precise work control. The closed-loop controller is based on the current state vector... Based on the first monitoring data, the coordinated adjustment of methanol input flow rate, load power, drying intensity, activation intensity and preheating power is calculated in real time.

[0075] In this step, the S4-1 adaptive strength repair adjustment is performed first.

[0076] Specifically, based on the severity of the fault according to the root cause determination results, the fault is divided into three levels: mild, moderate and severe, and the operating power and linkage intensity of the repair module are adjusted for different levels.

[0077] Preferably, the fault level classification standard can be set as follows: For overload faults: mild is when the load power exceeds the rated value by 10% to 20%; moderate is when the load power exceeds the rated value by 20% to 30%; severe is when the load power exceeds the rated value by more than 30%.

[0078] For water vapor exceeding the standard: mild is when the actual water vapor content exceeds the theoretical value by 10% to 20%; moderate is when the actual water vapor content exceeds the theoretical value by 20% to 30%; severe is when the actual water vapor content exceeds the theoretical value by more than 30%.

[0079] For voltage decay faults: mild is when the fuel cell stack operating voltage is 5% to 10% lower than the rated value; moderate is when the fuel cell stack operating voltage is 10% to 15% lower than the rated value; severe is when the fuel cell stack operating voltage is more than 15% lower than the rated value.

[0080] After determining the fault level, a multi-module linkage repair process for the corresponding scenario is initiated according to the target repair strategy.

[0081] When the fault type is overload, the power regulation module and the methanol flow regulation module work together. For example, in the case of mild overload, the power regulation module reduces the load power to the rated range, while the methanol flow regulation module reduces the methanol input flow proportionally. In the case of moderate overload, the power regulation module, the methanol flow regulation module and the heat dissipation module work together. In the case of severe overload, the power supply to unnecessary auxiliary loads can be further suspended to ensure the recovery of the fuel cell stack.

[0082] When the fault type is excessive moisture, the self-filtration module and the drying module are activated to work together. For example, for mild moisture exceedance, only the self-filtration module can be activated; for moderate moisture exceedance, both the self-filtration module and the drying module are activated simultaneously; for severe moisture exceedance, in addition to the self-filtration module and the drying module, the methanol flow regulation module can also be activated to reduce the rate of water production in the reaction.

[0083] When the fault type is voltage decay, the fuel cell stack activation procedure is initiated. For example, for mild voltage decay, a low-intensity constant current activation method is used; for moderate voltage decay, a medium-intensity constant current activation method is used; for severe voltage decay, the preheating module is activated while increasing the activation intensity to accelerate fuel cell stack recovery. If a multiplier expression method is used, the rated current of the fuel cell stack is used. Using this as a benchmark, for example, 0.1C, 0.2C, and 0.3C correspond to... , , .

[0084] During the repair process, the first monitoring data generated in step S1 is collected every 2 seconds, and the operating parameters of the repair module are dynamically adjusted based on the real-time changes of the first monitoring data. When the parameters corresponding to the fault return to the normal range, the operating power of the repair module is gradually reduced until it stops operating, thereby completing the repair process and generating repair status data.

[0085] Preferably, the closed-loop controller adopts a model predictive control method after order reduction of the infinite-dimensional dynamic system. It solves the control sequence in a rolling manner in each control cycle so that the actual work trajectory tracks the target work trajectory, while constraining the local current density peak, temperature gradient, moisture gradient and methanol concentration gradient to be within a safe range. If the prediction results show that there is a trend of local overheating, local water accumulation or local material shortage, the corresponding control quantity is adjusted first to suppress the internal imbalance.

[0086] Furthermore, when the target parameters are continuous The repair is considered successful only when all sampling periods remain within the normal range. If the parameters exceed the limit again during the rollback process, the previous stable control setting will be restored to continue the repair.

[0087] Preferably, the objective function of the model predictive control simultaneously includes a cumulative work error term, a local current density variance term, a moisture field deviation term, a temperature field deviation term, and a control increment term, thereby taking into account the output work accuracy, internal state uniformity, and control smoothness.

[0088] The repair status data includes: repair duration, parameter changes before and after repair, repair module operating parameters, fault level, target repair strategy number, repair success or failure flag, and disturbance information during the repair process.

[0089] Further optimization includes: repair start time, end time, duty cycle or output settings of each module, peak resource consumption, rollback count, and fault recurrence flag.

[0090] Step S5: Iterative optimization of model and policy The model and strategy iterative optimization module receives the repair status data and historical fault dataset generated in step S4, updates the monitoring threshold, prediction model parameters, associated model parameters, state field observer parameters, and work control parameters, and synchronizes the update results to the same cluster of electric stacks to achieve global strategy optimization.

[0091] This step includes S5-1 cluster strategy synchronization optimization and S5-2 threshold and model self-update. First, a model self-update system is built based on the reinforcement learning framework.

[0092] Preferably, a DQN reinforcement learning framework is used, and the state space includes first monitoring data, fault root cause determination results, repair state data, and the state vector. The action space includes adjustments to the LSTM model parameters, the parameters of the load-water content correlation model, the monitoring threshold, the state observer parameters, the reduced-order model parameters, and the weights of the work control objective function. DQN is suitable for discrete action spaces, so in this embodiment, the parameter adjustment step size is set to a discrete level to adapt to DQN.

[0093] Preferably, the reward function is set as follows: ; in, This is the reward value; For fault warning accuracy; To improve the success rate of repair; The average repair time is in minutes.

[0094] In the above reward function, and Represented in decimal form between 0 and 1. This is used to scale the time term to an approximately dimensionless quantity, so the overall expression is self-consistent in engineering evaluation.

[0095] After accumulating a certain amount of repair status data, a reinforcement learning iteration is triggered. Preferably, every 100 sets of repair status data are accumulated, a model training is initiated, and samples are extracted from the experience replay pool for learning, updating the monitoring threshold and the following model parameters: wL, kV, bL, γt in the LSTM time series prediction model; kP, kF, bF in the load-water vapor content correlation model; boundary correction parameters, modal weight parameters, and weight coefficients in the work control objective function of the state field observer.

[0096] Preferably, discrete adjustment step sizes can be set for each parameter, for example: The adjustment step size is ; The adjustment step size is ; The adjustment step size is ; The adjustment step size is ; The adjustment step size is ; The adjustment step size is .

[0097] Furthermore, to avoid the reinforcement learning directly affecting the underlying actuators and causing control instability, the reinforcement learning framework is preferably used to update the state observer parameters, reduced-order model parameters, objective function weights, and thresholds, while the real-time adjustment of methanol input flow rate, load power, drying intensity, activation intensity, and preheating power at the underlying level is still performed by the model predictive controller.

[0098] S5-1 Cluster Policy Synchronization Optimization After the parameter update is completed, the updated monitoring thresholds, prediction model parameters, load-moisture content correlation model parameters, state observer parameters, reduced-order model parameters, and work control objective function weights are synchronized to the monitoring modules of all fuel cells in the same cluster through the edge data transmission channel to achieve unified optimization of the global strategy.

[0099] Preferably, the update parameters can be encrypted using the TLS 1.3 encryption protocol before transmission. After receiving the data, each fuel cell stack monitoring module performs decryption and integrity verification. After the verification is successful, the update result is written to the local monitoring module, so that each fuel cell stack in the same cluster can use the unified and optimized strategy parameters during fault monitoring, fault root cause location, repair strategy rehearsal, and intelligent linkage repair.

[0100] Therefore, after the updated monitoring thresholds, prediction model parameters, and parameters of the correlation model between operating load and water vapor content are synchronized to all fuel cell stacks in the same cluster, they can be used to uniformly optimize the fault monitoring, fault root cause location, repair strategy simulation, and intelligent linkage repair process of fuel cell stacks in the same cluster.

[0101] This embodiment achieves early warning, precise location, efficient repair, and continuous optimization of methanol fuel cell stack faults through a complete data acquisition, prediction triggering, hierarchical localization, strategy pre-playing, coordinated repair, and reinforcement learning update process. In a set of preferred application scenarios, the fault warning accuracy rate can reach 96.8%, the repair success rate can reach 95.2%, and the average repair time is reduced by 62% compared to traditional methods. The above performance data are test results of the embodiment and are used to illustrate the technical effects that this solution can achieve, and are not intended to limit the scope of protection of this invention.

[0102] Please see Figure 2 Example 2: Implementation method for intelligent detection and repair of faults in vehicle-mounted methanol fuel cell stacks; This embodiment describes the application scenario of a vehicle-mounted methanol fuel cell stack. In a vehicle-mounted scenario, the stack's operating conditions change frequently, often involving complex conditions such as acceleration, deceleration, idling, and hill climbing. It is also affected by factors such as vehicle vibration, electromagnetic interference, and ambient temperature fluctuations. Therefore, this embodiment adaptively extends the monitoring, positioning, strategy pre-simulation, intelligent repair, and iterative optimization processes based on Embodiment 1, but the overall step division and core formulas remain consistent with the claims. Unless otherwise specified, the definitions of symbols, data space definitions, and the calling relationships between paragraphs are the same as in Embodiment 1.

[0103] Step S1: Multi-parameter pre-monitoring and anomaly triggering In vehicle-mounted scenarios, the fuel cell stack operating voltage sensor preferably adopts a double-layer electromagnetic shielding structure to reduce the impact of vehicle motors and electronic control systems on monitoring accuracy; the methanol flow sensor preferably adopts a vibration-resistant packaging structure; and the micro-moisture content sensor preferably adopts a heated optical dew point meter to reduce the impact of ambient temperature changes.

[0104] In the preprocessing stage, in addition to outlier removal, data standardization, and timestamp alignment as described in Example 1, a Kalman filtering step can be added to suppress vibration noise and electromagnetic interference noise. The input to the Kalman filter is the measured sequence of each sensor, and the output is the filtered measured parameter sequence. Then, it enters the normalization and time alignment steps, thus avoiding inconsistencies between data segments.

[0105] In S1-1 trend-based anomaly triggering, the LSTM time series prediction model can be retrained using historical samples from vehicle-mounted scenarios.

[0106] Preferably, training is performed using 1500 sets of historical fault data from vehicle scenarios and 1000 sets of historical normal operation data. In the vehicle scenario, the voltage correction coefficient... It can be individually calibrated according to the characteristics of the vehicle platform, for example, set to 0.25. Simultaneously, the time step dynamic weighting factor... It can dynamically change according to the vehicle's operating conditions. For example, when the vehicle is idling, it can increase... To enhance the weighting of historical data; when the vehicle is in a rapid acceleration condition, reduce... To enhance the impact of current data on prediction results.

[0107] Step S2: Root cause localization of layered faults In the S2-1 cross-parameter correlation verification, the methanol input flow rate changes more rapidly in the vehicle scenario, therefore the methanol flow rate correlation coefficient... A calibration range adapted to in-vehicle scenarios can be adopted. In addition, a mechanism for comparing historical data of the same vehicle model and under the same operating conditions can be introduced to further eliminate misjudgments caused by sudden changes in driving conditions.

[0108] In S2-2AI scenario-based fault classification, a lightweight CNN+SVM fusion model can be trained using federated learning. Each vehicle terminal trains the model locally, uploading only the model parameters instead of the raw data. The cloud aggregates the parameters and then distributes the updated model, thereby improving the privacy and security of vehicle user data while ensuring classification accuracy. The classification results still output three core fault categories: overload, excessive moisture, and voltage decay, to ensure consistency with the claims.

[0109] Step S3: Remediation Strategy Preview and Feasibility Assessment In automotive scenarios, the candidate repair strategy set considers not only the battery stack repair module but also the synergistic effects of the vehicle's power control unit, thermal management unit, and auxiliary battery system. For example, for overload faults, strategies can be preset to reduce the power of auxiliary loads such as air conditioning and entertainment systems, or to start the auxiliary battery to bear part of the vehicle's output load. In the preliminary evaluation of each candidate repair strategy, in addition to examining the fault reduction capability, the impact on vehicle driving smoothness, power response, and range can be further examined, and the target repair strategy with the optimal overall benefit is selected. Vehicle-side data can be provided by the vehicle controller, power battery management system, and thermal management controller via the vehicle bus. Simultaneously, the vehicle's driving demand can be mapped to a target work trajectory to balance power response and the uniformity of the battery stack's internal state.

[0110] Step S4: Intelligent Linkage Repair In vehicle-mounted scenarios, the intelligent linkage repair module forms a collaborative repair system with the vehicle's power control unit and thermal management unit. For example, when an overload fault occurs, linkage repair can be achieved by reducing auxiliary electrical load, limiting peak power output, and utilizing auxiliary battery power. When a moisture exceeding standard fault occurs, the working intensity of the drying module can be dynamically adjusted according to the vehicle's speed. When a voltage decay fault occurs and the ambient temperature is low, preheating can be performed before activation to reduce the risk of damage caused by high-current activation under low-temperature conditions. Relevant data such as vehicle speed, ambient temperature, and auxiliary battery state of charge can be obtained from existing vehicle-mounted sensors and controllers. In vehicle-mounted scenarios, the target work trajectory can be issued in real time by the vehicle controller according to states such as acceleration, cruising, and regenerative braking.

[0111] Step S5: Iterative optimization of model and policy In automotive scenarios, reinforcement learning reward functions can be further used to introduce vehicle range improvement metrics, for example: ; in, This improves vehicle range. Simultaneously, cluster strategy synchronization optimization leverages edge computing gateways to achieve low-latency synchronization and aggregation optimization of vehicle model parameters within a given area.

[0112] Under optimized testing conditions, the fault warning accuracy rate of this embodiment can reach 97.2%, the repair success rate can reach 95.8%, the average repair time is shortened by 68% compared with traditional vehicle repair methods, and the vehicle range is increased by 4.2%. The above data are the effect data of the embodiment.

[0113] Please see Figure 3 Example 3: Implementation method for intelligent fault detection and repair of stationary distributed power generation methanol fuel cell stack clusters; This embodiment describes a stationary distributed power generation methanol fuel cell stack cluster scenario. In this scenario, multiple methanol fuel cell stacks with the same or similar power ratings form a power generation cluster for supplying power to industrial parks or commercial buildings. Therefore, this embodiment, based on Embodiment 1, further emphasizes cluster collaborative monitoring, cluster-level fault location, cluster-level repair strategy simulation, and cluster-level coordinated repair and optimization. Unless otherwise specified, the symbol definitions, dimensional processing, and data flow logic are the same as in Embodiment 1.

[0114] Step S1: Multi-parameter pre-monitoring and anomaly triggering Cluster-level multi-parameter pre-monitoring preferably adopts a unified acquisition architecture. A data acquisition server is set up in the central control room to synchronously acquire the operating voltage, load power, methanol input flow rate, and micro-moisture content of each fuel cell stack via industrial Ethernet. In addition to outlier removal, standardization, and timestamp alignment, the raw data is also aligned to the cluster-level time to form the first monitoring dataset of the cluster. The first monitoring dataset of the cluster is preferably stored in the form of a three-dimensional array, corresponding to the timestamp, fuel cell stack number, and parameter type. The sensors of each individual stack in the cluster still use the existing acquisition devices in Example 1. The central acquisition server is responsible for unifying the time reference and protocol conversion.

[0115] In the S1-1 trend-based anomaly triggering, a cluster-level LSTM time series prediction model can be constructed to jointly predict the future state of multiple fuel cell stacks. When the predicted future value of any fuel cell stack exceeds the threshold range, the cluster-level fault root cause localization process is triggered.

[0116] Step S2: Root cause localization of layered faults In the S2-1 cross-parameter correlation verification, a cluster-based comparison mechanism can be further introduced. That is, for abnormal fuel cell stacks, other fuel cell stacks within the cluster whose load power deviation is within a defined range are selected as a reference group. The actual micro-moisture content of the abnormal fuel cell stack is compared with the theoretical average moisture content of the reference group to improve the ability to eliminate false positives. Here, the theoretical average moisture content of the reference group is still determined by the values ​​of each reference fuel cell stack. and The result was obtained by substituting into the same correlation model.

[0117] In S2-2AI scenario-based fault classification, each fuel cell stack can use a federated learning approach of local training and central aggregation to update the parameters of a lightweight CNN+SVM model, so as to balance cluster data security and model accuracy.

[0118] Step S3: Remediation Strategy Preview and Feasibility Assessment In cluster scenarios, the candidate repair strategy set considers not only the repair modules within a single fuel cell stack, but also the cluster load scheduling module and shared auxiliary modules. For example, for overload faults, a candidate repair strategy can be preset to transfer part of the load from the abnormal stack to other idle stacks; for excessive moisture faults, a candidate repair strategy can be preset to invoke the shared drying module; for voltage decay faults, a candidate repair strategy can be preset to invoke the activation resources of other idle stacks and the shared preheating module. During the pre-evaluation, in addition to evaluating the repair effect of a single fuel cell stack, the overall power generation efficiency, load balance, and system energy consumption changes of the cluster are further evaluated to determine the target repair strategy. Load scheduling information and shared module occupancy information can be provided by the energy management system and the central controller. Simultaneously, the target work distribution deviations of each fuel cell stack are considered to achieve cluster-level work balance.

[0119] Step S4: Intelligent Linkage Repair In cluster scenarios, the intelligent linkage repair module can work collaboratively with the cluster load scheduling module and the shared auxiliary module. For example, in overload scenarios, it can transfer part or all of the load from the abnormal fuel cell stack to other fuel cell stacks; in scenarios with excessive moisture, it can schedule the shared drying module to support the abnormal fuel cell stack; in scenarios with voltage decay, it can call upon idle fuel cell stack activation resources and cooperate with the cluster shared preheating module for recovery. During the repair process, it continuously collects the cluster's first monitoring data and dynamically adjusts the repair parameters and load scheduling parameters until the abnormal fuel cell stack returns to normal. Furthermore, the central controller can also perform rolling optimization of the load transfer amount based on the total target work of the cluster and the internal state field distribution of each fuel cell stack.

[0120] Step S5: Iterative optimization of model and policy In cluster scenarios, a multi-agent reinforcement learning framework is preferred for threshold and model self-updating. Each stack can be considered an independent agent, and the model parameters, load transfer amount, and shared module power are jointly optimized by comprehensively considering its own operating data, the operating data of other stacks in the cluster, the cluster load scheduling data, and the target work allocation deviation of each stack. Preferably, the reward function can be set as follows: ; in, For the overall power generation efficiency of the cluster; To reduce the overall energy consumption of the cluster; This represents the percentage of average downtime due to cluster failures.

[0121] After the parameter update is completed, the central control server will synchronize the updated monitoring thresholds, prediction model parameters, as well as the associated model parameters, state observer parameters, work allocation parameters, and work control objective function weights to all stack monitoring modules in the cluster, thereby achieving unified optimization of the cluster. If necessary, the optimized model parameters can also be synchronized to other similar distributed generation clusters to achieve cross-cluster strategy sharing.

[0122] Under optimized testing conditions, this embodiment can reduce the overall failure rate of the power cluster by 62% compared to traditional methods, increase power generation efficiency by 8.5%, and reduce overall energy consumption of the power cluster by 7.2%. The above data are the effect data of the embodiment.

[0123] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent detection and repair of faults in a methanol fuel cell stack, characterized in that, Includes the following steps: S1. Multi-parameter pre-monitoring and anomaly triggering: Simultaneously collect and preprocess four types of parameters, namely stack operating voltage, load power, methanol input flow rate, and micro-moisture content, to generate the first monitoring data; Based on the first monitoring data, an anomaly investigation process is triggered or continuous cyclical monitoring is performed. S2, Layered fault root cause localization: Receive the first monitoring data, reconstruct the methanol concentration field, moisture field, temperature field and local current density field inside the stack based on the infinite-dimensional dynamic system state observer, carry out cross-parameter correlation verification and scenario-based classification, and output the fault root cause determination result; S3. Repair strategy pre-play and feasibility evaluation: Receive the fault root cause determination result, the internal state field and the first monitoring data, call historical repair state data to construct a set of candidate repair strategies, and pre-play the target work trajectory, cumulative work amount and local state convergence trend under the action of each candidate repair strategy based on the reduced-order model of the infinite-dimensional dynamic system, perform pre-play evaluation of each candidate repair strategy, and output the strategy evaluation result and the target repair strategy. S4. Intelligent linkage repair: Receive the target repair strategy and the strategy evaluation result, start the multi-module linkage repair process of the corresponding scenario, and perform closed-loop precise control of methanol input flow, load power, drying intensity, activation intensity and preheating power according to the target work trajectory to generate repair status data; S5. Iterative optimization of model and strategy: Receive the repair status data and historical fault dataset, update the monitoring threshold and prediction model, status observation model and work control parameters, and synchronize them to the same cluster of fuel cells to achieve global optimization.

2. The intelligent detection and repair method for methanol fuel cell stack faults according to claim 1, characterized in that: Step S1 further includes S1-1 trend prediction-based anomaly triggering, specifically: Construct an LSTM time series prediction model, input the first monitoring data, and calculate using the following formula. Predicted parameter values ​​at time : ； in, These are the weight coefficients of the LSTM hidden layer; The output of the LSTM hidden layer at time t; The dynamic weighting factor for the time step at time t; This is the correction factor for the stack voltage; Let t be the operating voltage of the fuel cell stack. For the bias term of the LSTM model; When the parameter prediction value When the value exceeds the preset threshold range, the fault root cause location process in step S2 is directly triggered. The LSTM time series prediction model takes the first monitoring data as a time step sequence as input, and adjusts the weight coefficients of the LSTM hidden layer based on the historical fault dataset. , Pile voltage correction factor LSTM model bias term and time step dynamic weighting factor Perform fitting; The time step dynamic weight factor The forecast is dynamically adjusted based on the fluctuation range of the monitoring data to reflect the proportion of the impact of current monitoring data and historical monitoring data on the forecast results.

3. The intelligent detection and repair method for methanol fuel cell stack faults according to claim 2, characterized in that: Step S2 further includes S2-1 cross-parameter correlation verification, specifically including: Construct a correlation model between operating load and water vapor content. Input the load power, trace water vapor content, and methanol input flow rate data from the first monitoring data, and calculate the theoretical water vapor content under the corresponding operating condition using the following formula. : ； in, The correlation coefficient for the quadratic term of load power; Let t be the actual load power of the fuel cell stack. The correlation coefficient for methanol flow rate; Let t be the methanol input flow rate; Basic correction coefficient; Actual micro-water vapor content With the theoretical water vapor content Compare the results and eliminate any incorrect judgments that do not conform to the reaction pattern.

4. The intelligent detection and repair method for methanol fuel cell stack faults according to claim 3, characterized in that: Step S2 further includes S2-2 AI-based scenario-based fault classification, specifically: A lightweight CNN+SVM fusion model is adopted. The first monitoring data and the data after cross-parameter correlation verification are input to classify the stack faults in a scenario-based manner. The faults are divided into three categories: overload, excessive moisture, and voltage decay. The root cause determination result of the fault is output.

5. The intelligent detection and repair method for methanol fuel cell stack faults according to claim 1, characterized in that: In step S3, the pre-evaluation of each candidate repair strategy includes: Based on the fault root cause determination results, the first monitoring data, and historical repair status data, the parameter change trend, fault reduction magnitude, operational disturbance degree, and resource call cost after the execution of each candidate repair strategy are evaluated, and the target repair strategy is obtained by screening according to the preset evaluation rules. The candidate repair strategies include at least different module combination repair strategies corresponding to overload, excessive moisture, and voltage decay.

6. The intelligent detection and repair method for methanol fuel cell stack faults according to claim 1, characterized in that: Step S4 further includes S4-1 adaptive intensity repair adjustment, specifically including: Based on the severity of the fault according to the fault root cause determination results, the fault is divided into mild, moderate and severe levels, and the operating power of the repair module is adjusted for different levels. Among them, when the working condition is overloaded, the power regulation module and the methanol flow regulation module work together; when the water vapor exceeds the standard, the self-filtration module [1][A42] and the drying module work together; when the voltage decays, the stack activation program is started. During the operation of step S4, the first monitoring data generated in step S1 is collected every 2 seconds. The operating parameters of the repair module are dynamically adjusted according to the real-time changes of the first monitoring data until the parameters corresponding to the fault are restored to the normal range, and the repair status data is generated.

7. The intelligent detection and repair method for methanol fuel cell stack faults according to claim 3, characterized in that: Step S5 further includes: [A43] S5-1 Cluster Strategy Synchronization Optimization: The updated monitoring thresholds and prediction model parameters are synchronized to the monitoring modules of all fuel cell stacks in the same cluster through the edge data transmission channel to achieve unified optimization of global strategies; S5-2 Threshold and Model Self-Update: Based on the reinforcement learning framework, the repair state data and historical fault dataset are input, and the weight coefficient wL, voltage correction coefficient kV, bias term bL and time step dynamic weight factor γt of the LSTM time series prediction model are updated, as well as the correlation coefficients kP, kF and basic correction coefficient bF of the load-water content correlation model. The state observer parameters, reduced-order model parameters and work control objective function weights are also updated.

8. The intelligent detection and repair method for methanol fuel cell stack faults according to claim 1, characterized in that: The infinite-dimensional dynamic system state observer uses the spatial coordinate z of the fuel cell flow channel direction or the membrane electrode thickness direction as the distribution dimension to construct a state vector X(z,t) containing the methanol concentration field c(z,t), moisture field w(z,t), temperature field T(z,t), membrane water content field λ(z,t), and local current density field i(z,t), and reconstructs the state vector online based on the first monitoring data.

9. The intelligent detection and repair method for methanol fuel cell stack faults according to claim 8, characterized in that: The pre-performance evaluation in step S3 includes based on instantaneous power output. and the cumulative work done within the preset evaluation time window H The work deviation, local current density variance, moisture field deviation, and temperature field deviation corresponding to each candidate repair strategy are calculated, and the target repair strategy is obtained by screening based on these parameters; where Ut is the stack operating voltage at time t, and It is the stack operating current at time t.

10. The intelligent detection and repair method for methanol fuel cell stack faults according to claim 9, characterized in that: The closed-loop precise control in step S4 adopts the model predictive control method after the infinite-dimensional dynamic system is reduced in order. It jointly adjusts the methanol input flow rate, load power, drying intensity, activation intensity and preheating power to make the actual work trajectory track the target work trajectory and suppress local overheating, local water accumulation and local material shortage.

Images