A fuel cell failure diagnosis method and apparatus

Abstract

The application discloses a fuel cell fault diagnosis method and device, comprising: S1: constructing and verifying a proton exchange membrane fuel cell (PEMFC) system agent model; S2: analyzing a PEMFC fault feature based on global sensitivity; S3: proposing an online diagnosis method based on sensitivity, namely, constructing a diagnosis method flow based on sensitivity, including a model self-adaption stage, feature processing and screening, constructing a fault representation equation, fault evaluation and analysis, determining a threshold and a sensitivity index, and verifying the proposed method. By constructing the agent model of the PEMFC, the efficiency and accuracy of fault diagnosis are significantly improved. The use of the agent model reduces the demand for computing resources and can quickly identify fault features. The global sensitivity analysis based on the Sobol sequence enhances the identification ability of fault features and ensures that fault signals can be accurately captured under different working conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to a fuel cell fault diagnosis method and device, belonging to the technical field of battery. BACKGROUND

[0002] Overuse of fossil fuels not only causes damage to the environment on which humans depend for survival, but also plunges humans into a serious energy crisis. Fuel cells have the advantages of high efficiency, environmental protection, and flexible use compared to traditional energy conversion methods. Among the many fuel cells, Proton Exchange Membrane Fuel Cell (PEMFC) is attracting attention due to its high energy conversion rate, low operating temperature, and fast start-up. In practical applications, faults such as flooding and membrane drying inevitably affect the stability and durability of the PEMFC system. Therefore, in order to meet the durability requirements of PEMFC, reliable fault diagnosis technology needs to be developed to realize real-time monitoring of the system's operating state and health condition.

[0003] Currently, the methods for PEMFC fault diagnosis mainly include model-based, experimental test-based, and data-driven methods. The model-based diagnosis method compares the residual error between the actual output value and the simulation value of the PEMFC system model to identify different faults. The decision-making process is mostly based on fixed thresholds or rules, which cannot be adaptively adjusted with changes in system characteristics, resulting in inaccurate or unstable diagnosis results. The experimental test-based fault diagnosis method mainly relies on experimental approaches such as electrochemical impedance spectroscopy, visualization technology, fault operation method, and external magnetic field measurement. It usually needs to be performed offline, making it difficult to meet the current demand for rapid diagnosis. The data-driven diagnosis method uses data analysis technology and machine learning algorithms to diagnose and analyze operating data. Compared to other diagnosis methods, this method only needs to use system input and output data to evaluate system faults, making the diagnosis method more convenient and flexible. However, the data-driven diagnosis method requires a large amount of training data, and it is difficult to perform fault experiments on real fuel cell systems to obtain the required experimental data, limiting the generalization ability of the data-driven diagnosis method. Therefore, by establishing a PEMFC model, the lack of real data can be compensated for, various fault types and scenarios can be simulated, and the robustness of pattern recognition can be increased, enabling the diagnosis algorithm to have more comprehensive learning ability.

[0004] However, the current data-driven diagnosis method is basically only concerned with the diagnosis of single fault, and there are relatively few studies on simultaneous faults. When multiple faults exist in the PEMFC system at the same time, the fault characteristics will interact with each other, and even mix or combine into a new fault mode, thereby increasing the complexity of diagnosis. At present, a new type of synchronous fault diagnosis algorithm based on the incremental multi-label classification network of multi-label classification chain and a sparse stacking auto-encoder diagnosis method based on deep learning network are proposed to realize synchronous fault diagnosis. However, these methods still need a large amount of simultaneous fault data for training. For untrained single fault and simultaneous fault mode, the diagnosis accuracy is not reliable enough. When processing single fault data, single-label diagnosis method will assign a unique label for each fault, while simultaneous faults need to be treated as a new independent label for training, and the label complexity will increase exponentially with the increase of concurrent fault combination mode. SUMMARY

[0005] The present application provides a fuel cell fault diagnosis method and device, which realizes model adaptive update through feature cleaning strategy based on sensitivity and correlation analysis and determines normal and fault states of real-time operation data, and constructs a diagnosis method based on sensitivity based on the sensitivity analysis results of the updated RBF proxy model.

[0006] Technical scheme: A fuel cell fault diagnosis method comprises the following steps:

[0007] S1: Proton exchange membrane fuel cell (PEMFC) system proxy model construction and verification;

[0008] S101: PEMFC system theoretical model construction and verification;

[0009] S102: Select system fault type and abnormal feature according to the PEMFC system theoretical model established in S101;

[0010] S103: Based on the results of S101 and S102, construct a proxy model for global sensitivity analysis and verify the accuracy;

[0011] S2: Based on the results of S1, analyze the PEMFC fault features based on global sensitivity;

[0012] S201: Use the proxy model constructed in S103 to analyze the sensitivity of the PEMFC system fault;

[0013] S202: Perform correlation analysis on the sensitive features selected in S201 to obtain the final sensitive feature set;

[0014] S203: Forming evaluation indexes of the feature screening method, including precision dimension evaluation indexes and time dimension evaluation indexes;

[0015] S204: Based on the evaluation indexes of S203, comparing the performance of different feature screening methods under different feature quantities, and selecting the feature screening method based on sensitivity analysis as the diagnosis method;

[0016] S3: Based on the results of S1 and S2, an online diagnosis method based on sensitivity is proposed, that is, a diagnosis method based on sensitivity is constructed, including a model self-adaptation stage, feature processing and screening, construction of a fault representation equation, fault evaluation and analysis, determination of a threshold and a sensitivity index, and verification of the proposed method.

[0017] Preferably, S101 specifically comprises:

[0018] The PEMFC system theoretical model comprises a PEMFC stack model and a PEMFC auxiliary equipment model, which are integrated to obtain the PEMFC system theoretical model; wherein the PEMFC stack model comprises a PEMFC output voltage model, an anode mass flow model, a cathode mass flow model, and a membrane hydration model; the PEMFC auxiliary equipment model comprises an air compressor model, a cooler model, a humidifier model, an intake and exhaust manifold model, and a control system model; and the specific construction method is as follows:

[0019] Construction of the PEMFC stack model:

[0020] Construction of the PEMFC output voltage model:

[0021] The actual output voltage of the PEMFC single cell is obtained by subtracting three voltage losses caused by polarization phenomena from the ideal voltage:

[0022]

[0023] In the formula, represents the thermodynamic theoretical electromotive force; , and respectively represent the activation, concentration and ohmic loss voltages;

[0024] Construction of the anode mass flow model:

[0025] The anode mass flow model aims to describe the mass flow process of hydrogen and water vapor in the anode of the PEMFC and the interaction between them; according to the mass conservation law, the change of each gas is represented as:

[0026]

[0027] In the formula, and These represent the flow rates of hydrogen and water vapor flowing into the anode flow field, respectively. , and These are the flow rates of hydrogen, water vapor, and liquid water flowing out of the anode flow field, respectively. The flow rate of hydrogen gas consumed; The mass of water vapor in the anode flow field; The mass of hydrogen gas in the anode flow field; This indicates the flow rate of water vapor transferred from the proton exchange membrane side;

[0028] Cathode mass flow model construction:

[0029] The changes in each gas are represented as follows:

[0030]

[0031]

[0032]

[0033] In the formula, , and These are the flow fields flowing into the cathode. , And the flow rate of water vapor; , , and These represent the flow rates of each gas and liquid water exiting the cathode flow field, respectively. and These are the resources consumed during PEMFC operation. With the generated water vapor flow rate; Let be the mass of oxygen in the cathode flow field. The mass of water vapor in the cathode flow field; The mass of nitrogen in the cathode flow field;

[0034] Membrane hydration model construction:

[0035] The water transport volume caused by electric drag is affected by the current density and the electric drag coefficient. The water transport volume caused by electric drag is expressed as:

[0036]

[0037] In the formula, This is the electric drag coefficient. This is the electric drag coefficient. Current density

[0038] During PEMFC system operation, the cathode is the main water generation site, resulting in a higher water content on the cathode side and a concentration gradient. Therefore, water will diffuse from the cathode to the anode, alleviating the uneven distribution of some water content. The water amount from the cathode to the anode caused by concentration diffusion is expressed as:

[0039]

[0040] where, is the diffusion coefficient of water in the membrane; is the water concentration; is the vertical distance of the proton exchange membrane;

[0041] PEMFC auxiliary equipment model construction:

[0042] Air compressor model construction:

[0043] The flow rate of the air compressor is obtained:

[0044]

[0045] where, is the air compressor characteristic parameter; is the air compressor blade tip linear velocity, represents the air compressor blade diameter, represents a constant related to the Mach number;

[0046] The temperature change of air after compression by the air compressor is expressed as:

[0047]

[0048] where, is the working efficiency of the air compressor; is the temperature at the inlet of the air compressor; is the pressure at the inlet of the air compressor; is the pressure at the outlet of the air compressor, is the specific heat coefficient of air at a fixed pressure;

[0049] Cooler model construction:

[0050] The relative humidity of air leaving the cooler caused by temperature change is expressed by the following formula:

[0051]

[0052] where, and are the relative humidity and pressure of air at standard atmospheric pressure; is the water vapor pressure at standard atmospheric pressure; is the pressure of cooled air; Tg is the temperature of the gas after air cooling; Pw is the saturated water vapor pressure, T is the temperature at standard atmospheric pressure;

[0053] The flow rate of dry air and water vapor when leaving the cooler is given by:

[0054]

[0055] where, RH is the relative humidity of the air at the cooler outlet; Fda is the flow rate of dry air in the cooler; Fw is the flow rate of water vapor; Ftot is the total flow rate in the cooler; Mw is the molar mass of water vapor, Ma is the molar mass of air; , Pda and Pw are the partial pressures of dry air and water vapor inside the cooler, respectively;

[0056] Humidifier model construction:

[0057] The flow rate of the reaction gas before entering the humidifier is composed of water vapor and dry air, and is given by:

[0058]

[0059] where, , Fda and Fw are the flow rates of dry air and water vapor entering the humidifier, respectively;

[0060] The amount of water added to the dry air by the humidifier is given by:

[0061]

[0062] where, Pw is the water vapor partial pressure, Fw is the flow rate of water vapor out of the humidifier, Pda is the partial pressure of dry air; Fda is the flow rate of dry air at the humidifier inlet;

[0063] The total pressure of the gas at the humidifier outlet is given by:

[0064]

[0065] Intake and exhaust manifold model construction:

[0066] Considering that the pressure difference before and after the air passes through the intake manifold is very small, the flow rate of the gas out is given by:

[0067]

[0068] where, is the supply manifold outlet flow constant; is the intake manifold outlet gas pressure; is the pressure at the intake manifold inlet;

[0069] Due to the high temperature of the air from the air compressor, the temperature and pressure of the air will change as it passes through the intake manifold. The change in the supply manifold is:

[0070]

[0071]

[0072] where, is the supply manifold lumped volume; is the gas temperature in the supply manifold; is the temperature of the air exiting the air compressor; is the gas flow into the supply manifold, is the specific gas constant for air, is the mass of gas in the supply manifold;

[0073] The temperature difference of the air from the stack to the exhaust manifold is small and is neglected. Therefore, we have:

[0074]

[0075] is the pressure in the exhaust manifold; is the gas temperature in the exhaust manifold, which is approximately the temperature of the stack; and are the gas flows into and out of the exhaust manifold, respectively; is the lumped volume of the exhaust manifold, is the gas constant; is the molar mass of air;

[0076] The operating state of the nozzle is represented by the ratio of the upstream pressure to the downstream pressure:

[0077]

[0078] When this ratio is greater than the critical pressure ratio, the nozzle is in a subcritical state; otherwise, when this ratio is less than or equal to the critical pressure ratio, the nozzle is in a supercritical state. The exhaust manifold flow in different states is represented as:

[0079]

[0080] where, is the manifold discharge coefficient;​​​​​​​​​​​​​​​​​​ Indicates the opening degree of the exhaust valve;

[0081] Control system model construction:

[0082] The oxygen ratio is determined based on the ratio between the air flow rate supplied by the air compressor and the air flow rate required by the fuel cell current.

[0083]

[0084] The reference value for the air flow rate of the air compressor is:

[0085]

[0086] In the formula, The molar mass of dry air. This refers to the dry air flow rate at the air compressor inlet.

[0087] The use of a proportional valve to rapidly adjust the hydrogen flow rate and control the pressure across the membrane is expressed as follows:

[0088]

[0089] In the formula, The pressure linear gain of the proportional valve; This is the adjustment coefficient for the pressure drop between the gas supply manifold and the cathode flow field.

[0090] In a preferred embodiment, S102 specifically includes:

[0091] System failure types include f1 contaminants entering the fuel cell stack, f2 increased flow resistance (simulated flooding), f3 exhaust manifold leakage, f4 intake manifold leakage, and f5 cooling system malfunction (simulated membrane drying).

[0092] Abnormal characteristics include air compressor output pressure, air intake flow rate, exhaust manifold outlet pressure, oxygen ratio, air intake pressure, hydrogen intake pressure, air compressor output flow rate, cathode outlet pressure, anode relative humidity, cathode relative humidity, anode outlet pressure, output voltage, and operating current.

[0093] Fault f1 indicates that the system has been contaminated during operation, resulting in a reduction in the catalytic active area and consequently a decrease in current density. Its specific function is:

[0094]

[0095] in, This is the operating current of the fuel cell stack. Indicates the coefficient of variation of fault parameters; effective working area of the PEMFC; since the current density is an important factor affecting the total voltage of the stack, the fault f1 will directly affect the output voltage of the PEMFC system;

[0096] The fault f2 represents the change of the cathode air flow resistance to simulate the flooding of the cathode diffusion layer, and its specific function is:

[0097]

[0098] wherein, is the total flow rate at the cathode outlet; is the cathode outlet aperture constant; and are the cathode pressure and atmospheric pressure, respectively;

[0099] The fault f3 represents the leakage of gas in the exhaust manifold, and the exhaust manifold output flow is multiplied by a coefficient k3 between 0 and 1 to simulate it, and its specific function is:

[0100]

[0101] wherein, represents the gas flow rate at the exhaust manifold outlet; represents the manifold discharge coefficient; represents the opening degree of the exhaust valve; represents the pressure of the exhaust manifold; represents the gas temperature in the exhaust manifold, which is approximately the temperature of the stack; is the specific heat coefficient of air at a fixed pressure;

[0102] The fault f4 represents the leakage of gas in the intake manifold, and its specific function is:

[0103]

[0104] wherein, represents the gas flow rate out of the intake manifold; is the supply manifold outlet aperture constant; and are the pressure of the intake manifold and the cathode side, respectively;

[0105] The fault f5 represents the loss of control of the internal temperature of the PEMFC, and is simulated by changing the set stack working temperature, and its specific function is:

[0106]

[0107] Preferably, the S103 is specifically:

[0108] A radial basis function (RBF) neural network was chosen as the surrogate model. Fault parameters included decreased activation area, decreased water discharge (i.e., increased flow resistance), decreased air discharge, decreased air supply, and increased stack operating temperature. The range of parameters in the dataset was determined based on the maximum and minimum values ​​of each fault parameter, including normal and fault datasets. Initial fault parameter samples were generated using Sobol sequences as the training input set and simulated in the PEMFC system theoretical model to obtain abnormal feature data, which was then used as the training sample output set. These samples were divided into training, validation, and test sets at ratios of 0.70, 0.15, and 0.15. The number of hidden layers, minimum performance gradient, and maximum number of training iterations were set for the RBF surrogate model. Training was performed using the Levenberg-Marquardt method, with a coefficient of determination R0. 2 To evaluate the accuracy of the proxy model during the training, testing, and validation processes;

[0109] The selected initial fault parameter samples were input into the theoretical model and the surrogate model respectively, and the average error of each feature between the surrogate model and the theoretical model was compared under different sample numbers to verify the accuracy of the surrogate model. The surrogate model was selected to replace the theoretical model to obtain sensitivity results. The consistency between the surrogate model and the theoretical model was evaluated by the average error. The smaller the average error, the closer the prediction results of the surrogate model are to the results of the theoretical model, thus verifying the accuracy of the surrogate model.

[0110] The average error is as follows:

[0111]

[0112] In the formula, N is the number of samples; M is the number of indicators; Let i be the output value of the j-th proxy model for index i; Let be the output value of the j-th theoretical model for index i.

[0113] In a preferred embodiment, the radial basis function (RBF) neural network used as a surrogate model specifically comprises:

[0114] According to the RBF neural network structure, the input layer is used to receive input data, represented as follows: The hidden layer uses a set of kernel functions to transform the input layer data. After the transformation, the output of the i-th hidden layer node is represented as:

[0115]

[0116] In the formula, The Euclidean norm in the input space; Center of kernel function; Let be the bandwidth parameter of the i-th kernel function;

[0117] The output of the jth node of the output layer is:

[0118]

[0119] wherein, is the connection weight from the ith node of the hidden layer to the jth node of the output layer; is the bias of the jth node of the output layer; This is the ith base function.

[0120] Preferably, the determination coefficient R 2 Specifically,

[0121]

[0122] wherein, n is the number of samples; is the output value of the ith proxy model; is the output value of the ith theoretical model; is the mean value of the n sample points y.

[0123] Preferably, the S201 and S202 are specifically:

[0124] S201: Analyzing the fault sensitivity of the PEMFC system: Since the main effect sensitivity index is the same as the total effect sensitivity index when analyzing the sensitivity of each abnormal feature under different faults of the PEMFC system, only the main effect sensitivity index is analyzed:

[0125] The sensitivity index under different fault variables is converted into the proportion of each fault variable under different faults, and it is assumed that the features with a proportion greater than 0.1, i.e., 10%, are sensitive features, the parameters with a proportion between 0.05 and 0.1 are relatively sensitive features, and the parameters with a proportion lower than 0.05 are insensitive features;

[0126] S202: Analyzing the correlation between the sensitive features in S201 to obtain related features with strong correlation with the sensitive features, selecting sensitive features with a correlation coefficient higher than 0.6, then retaining the sensitive feature with the highest sensitivity coefficient, and comprehensively obtaining the sensitive feature set A and the sensitive feature set B under the f1-f5 faults;

[0127] Among them, the sensitive feature set A: exhaust manifold outlet pressure, hydrogen inlet pressure, cathode outlet pressure, cathode relative humidity, output voltage;

[0128] The sensitive feature set B: exhaust manifold outlet pressure, cathode outlet pressure, cathode relative humidity, anode outlet pressure, output voltage;

[0129] The sensitive feature set A and the sensitive feature set B are respectively judged by the K-neighbor method to classify the performance under different working conditions, and the sensitive feature set with the highest accuracy and the shortest training time is selected as the final sensitive feature set.

[0130] Preferably, the precision dimension evaluation index and the time dimension evaluation index are specifically:

[0131] The precision dimension evaluation index is specifically:

[0132] The indicators for evaluating the performance of the fault diagnosis model include accuracy, precision and recall rate. In the confusion matrix, the concerned-predicted categories include positive-positive TP, positive-negative FN, negative-positive FP and negative-negative TN.

[0133] The accuracy represents the proportion of correctly predicted samples in the total number of samples, specifically:

[0134]

[0135] The precision is only for samples predicted as positive, and is defined as the proportion of successfully predicted positive samples in all positive samples, specifically:

[0136] The recall rate is based on the actual occurrence of the category, and is defined as the proportion of correctly predicted positive samples

[0137] in all actual positive samples, specifically:

[0138] In the formula, TP represents the number of samples correctly predicted as positive; FN represents the number of samples incorrectly predicted as negative; TN represents the number of samples correctly predicted as negative; and FP represents the number of samples incorrectly predicted as positive.

[0139] Since precision and recall rate are mutually restrictive, in order to comprehensively consider, effectively balance between precision and recall rate, and use together with accuracy to evaluate the diagnosis performance, a comprehensive index F1_Score is introduced, specifically:

[0140]

[0141] The time dimension evaluation index is specifically: the execution time of the feature selection method is measured by the timing function to formulate the time dimension evaluation index.

[0142]

[0143] ​​In a preferred embodiment, S3 specifically includes:

[0144] Model Adaptation Phase: Based on real-time operational data, the system performance model and RBF fault proxy model are adaptively updated, thereby generating accurate data that changes with system performance and characteristics based on the adaptive model;

[0145] Feature processing and filtering: Based on the Sobol global sensitivity analysis method, fault features are filtered and cleaned to identify the features with the greatest impact on the fault, and the interaction between the two is evaluated.

[0146] Constructing fault characterization equations: Based on Sobol global sensitivity analysis, the final sensitivity feature set obtained by S202 is decomposed into the relative influence between different input parameters. When there is no interaction in the system, linear equations can be constructed to describe the input parameters. With output variables The relationship between them is shown in the following formula:

[0147]

[0148] In the formula, It is the intercept term; Each input parameter The coefficient represents the degree of influence of this parameter on the output variable Y;

[0149] If the Sobol global sensitivity analysis results show the existence of interactions or nonlinear relationships, these can be expressed by adding interaction terms to the equations, specifically:

[0150]

[0151] In the formula, These are the regression coefficients of the constant term and the first-order term; These are the regression coefficients of the interaction term and higher-order terms;

[0152] Fault Assessment and Analysis: Based on the sensitivity results and changes in sensitive characteristics, the least squares method is used to infer the changes in fault parameters. Then, the absolute values ​​of the residuals of the PEMFC sensitive characteristics are compared. Compared to its normal threshold To determine if the system has malfunctioned, when If it is, it is diagnosed as faulty; otherwise, it is diagnosed as normal.

[0153] Determine the threshold and sensitivity index: Adjust the threshold and sensitivity coefficient based on the changes in sensitive characteristics obtained from the fault assessment to ensure that the system's operating characteristics and fault modes can be captured, thereby improving the accuracy and reliability of fault diagnosis;

[0154] Based on the principle of main effect, each fault input parameter is multiplied by its corresponding Sobol sensitivity coefficient, and is added to construct a linear equation describing the influence degree of different faults on sensitive characteristics, which is specific as follows:

[0155]

[0156] In the formula, is the sensitivity coefficient; is the label of the sensitive characteristic; is the label of the fault; Yi is the output of each sensitive characteristic; is the sensitive characteristic of each system fault;

[0157] Suppose that any performance parameter Yi of PEMFC is affected by m factors, which are respectively denoted as The static mathematical model of PEMFC is as follows:

[0158]

[0159] At this time, the sensitive characteristic of each system fault is The output of each sensitive characteristic is When the influencing factors change within a certain range, the output of each sensitive characteristic becomes The deviation value of the influencing factor is The deviation amount of the output of each sensitive characteristic is The relative deviation between the output of each sensitive characteristic of PEMFC and the relative deviation of the corresponding influencing factor of the fault is defined as the sensitive characteristic coefficient of the output of each sensitive characteristic to the influencing factor of the fault, which is specific as follows:

[0160]

[0161] According to the constructed fault representation equation and the sensitive characteristic coefficient, suppose that the equation group containing m equations and n unknowns is as follows:

[0162]

[0163] In the formula, is an mxn matrix, that is, the fault representation matrix composed of the sensitive characteristic coefficients; x is an n-dimensional vector representing the change of each system fault sensitive characteristic, and b is an m-dimensional vector representing the right side of the equation group, that is, the change of the output of the sensitive characteristic in the system;

[0164] The objective of the least square method is to solve the unknown number vector so that the residual square sum is minimum. Therefore, the residual vector is defined as follows:

[0165]

[0166] The objective is to minimize the sum of squared residuals, i.e., minimize the error function as:

[0167]

[0168] where, denotes the Euclidean norm of the residual vector .

[0169] A fuel cell fault diagnosis method device, comprising a data acquisition module, a model construction module, a feature screening module, an adaptive updating module, a fault diagnosis module;

[0170] The data acquisition module is responsible for real-time monitoring and collecting key operating parameters of the fuel cell system;

[0171] The model construction module is used to construct a PEMFC system theoretical model and design a radial basis neural network proxy model;

[0172] The feature screening module is used to screen and clean the features based on the sensitivity analysis results and correlation analysis, and find out the features that have the greatest impact on the fault;

[0173] The adaptive updating module is used to continuously optimize the model parameters according to new operating data and diagnostic experience, and to perform adaptive updating of the system performance model and the adaptive RBF fault proxy model;

[0174] The fault diagnosis module is used to combine the adaptive model obtained from the real-time data and the screened feature set to establish a fault feature representation equation, and to perform real-time fault diagnosis on the fuel cell system.

[0175] Beneficial effects: By constructing the proxy model of the proton exchange membrane fuel cell (PEMFC), the efficiency and accuracy of fault diagnosis are significantly improved. The use of the proxy model greatly reduces the demand for computing resources, and the proxy model can quickly perform global sensitivity analysis to quickly identify the fault features that have the most significant impact on system performance. In addition, the introduction of global sensitivity analysis based on Sobol sequence enhances the ability to identify fault features, ensuring that key fault signals can be accurately captured under different working conditions. At the same time, the feature cleaning strategy proposed in the present application effectively removes redundant information between sensitive features through correlation analysis, optimizes the feature set, and improves the accuracy and reliability of fault diagnosis. The adaptive updating mechanism enables the diagnostic system to continuously optimize based on real-time data and historical diagnostic experience, improving the adaptability and long-term stability of the system. It meets the demand for fast and accurate fault diagnosis of PEMFC systems, has good generalization ability and robustness, and can provide effective diagnostic support for various fault modes, greatly improving the safety and economy of fuel cell system operation. BRIEF DESCRIPTION OF DRAWINGS

[0176] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only embodiments of the present application, and those skilled in the art can obtain other drawings according to the provided drawings without creating any inventive labor.

[0177] Figure 1 The method flowchart of the present application; Figure 2 Comparison of test polarization curve and model simulation polarization curve; Figure 3 Comparison of simulation and test system output characteristics under transient conditions, where (a) is the system output voltage, and (b) is the error curve; Figure 4 RBF neural network proxy model training results, where (a) is, and (b) is; Figure 5 Proxy model test result graph, where (a-l) are fault features e1, e2, e3, e4, e5, e6, e7, e8, e9, e10, e11, and e12, respectively; Figure 6 Sensitivity of each fault feature under 60A working condition, (a-e) are fault f1, f2, f3, f4, and f5, respectively; Figure 7 Feature cleaning strategy flowchart based on correlation analysis; Figure 8 Correlation analysis scatter plot of each feature under 60A working condition; Figure 9The results of Spearman correlation analysis for each feature under 60 A operating conditions; Figure 10 This is a schematic diagram of feature cleaning; Figure 11 Performance graphs of different feature selection methods under different numbers of features; Figure 12 Here is a flowchart of a sensitivity-based diagnostic method; Figure 13 This is a diagram showing the confusion matrix results for a single fault diagnosis. Detailed Implementation

[0178] 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.

[0179] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0180] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0181] like Figure 1 As shown, S1: Construction and verification of the proxy model for the proton exchange membrane fuel cell (PEMFC) system;

[0182] S101: Construction and Verification of the Theoretical Model of PEMFC System;

[0183] The PEMFC system theoretical model comprises a PEMFC stack model and a PEMFC auxiliary equipment model, and the PEMFC stack model and the PEMFC auxiliary equipment model are integrated to obtain the PEMFC system theoretical model; wherein the PEMFC stack model comprises a PEMFC output voltage model, an anode mass flow model, a cathode mass flow model and a membrane hydration model; and the PEMFC auxiliary equipment model comprises an air compressor model, a cooler model, a humidifier model, an intake and exhaust manifold model and a control system model.

[0184] In order to ensure the accuracy of the system model, firstly, the output characteristics of the PEMFC are evaluated by comparing the polarization curves of the test and simulation under the steady state condition. Secondly, the dynamic response curves of the test and simulation are compared to verify the response ability of the model to transient changes under the transient condition. Figure 2 The polarization curves obtained by the test and simulation are compared. As can be seen from the figure, the polarization curve of the PEMFC stack model is basically consistent with the test result, and the average relative error between the test value and the simulation value is 0.32%, and the maximum error occurs at the current of 200 A, which is 1.28%.

[0185] Figure 3 The comparison of the system output voltage between the simulation and the test under the transient load is shown. As can be seen from the figure, with the increase of the load current, the system output voltage shows a downward trend, which is basically consistent with the trend in the actual operation process of the PEMFC. The maximum relative error of the system output voltage is 3.67%, and the average relative error is only 0.54%. It is considered that the PEMFC system model built can simulate the working state of the actual stack under the steady state and the transient state.

[0186] S102: According to the PEMFC system theoretical model established in S101, the system fault type and abnormal feature are selected;

[0187] The system fault type and abnormal feature are specifically shown in Table 1:

[0188]

[0189] The fault f1 represents that the system is polluted during the operation process, which reduces the catalytic active area and further reduces the current density. The specific function is:

[0190]

[0191] wherein, is the working current of the stack, represents the change coefficient of the fault parameter; is the effective working area of the PEMFC; since the current density is an important factor affecting the total voltage of the stack, the fault f1 will directly affect the output voltage of the PEMFC system;

[0192] The fault f2 represents the change of the cathode air flow resistance to simulate the flooding of the cathode diffusion layer, and its specific function is:

[0193]

[0194] wherein, is the total flow rate at the cathode outlet; is the cathode outlet aperture constant; and are the cathode pressure and atmospheric pressure, respectively;

[0195] The fault f3 represents the leakage of the exhaust manifold, and the exhaust manifold output flow is multiplied by a coefficient k3 between 0 and 1 to simulate it, and its specific function is:

[0196]

[0197] wherein, represents the gas flow rate at the exhaust manifold outlet; represents the manifold discharge coefficient; represents the opening degree of the exhaust valve; represents the pressure of the exhaust manifold; represents the gas temperature in the exhaust manifold, which is approximately the temperature of the stack; is the specific heat coefficient of air at constant pressure;

[0198] The fault f4 represents the leakage of the intake manifold, and its specific function is:

[0199]

[0200] wherein, represents the gas flow rate out of the intake manifold; is the outlet aperture constant of the supply manifold; and are the pressure of the intake manifold and the cathode side, respectively;

[0201] The fault f5 represents the temperature runaway inside the PEMFC, and it is simulated by changing the set stack operating temperature, and its specific function is:

[0202]

[0203] S103: Based on the results of S101 and S102, a proxy model for global sensitivity analysis is constructed and the accuracy is verified;

[0204] In view of the fact that global sensitivity analysis using a theoretical model of a PEMFC system requires a large amount of computational cost and time, a RBF neural network is selected as a surrogate model to replace the cumbersome theoretical model for global sensitivity analysis, and the fault parameters include a decrease in activation area, a decrease in the amount of water discharged, i.e. an increase in flow resistance, a decrease in the amount of air discharged, a decrease in the amount of air supplied, and an increase in the operating temperature of the stack, as shown in Table 2:

[0205]

[0206] The range of parameters in the data set is determined according to the maximum and minimum values of each fault parameter, including the normal data set and the fault data set. The present application uses Sobol sequences with more uniform distribution and lower diversity, and can achieve higher computational accuracy with a relatively small number of sampling points. 2000 initial fault parameter samples are generated as a training input set, which is input into the theoretical model of the PEMFC system for simulation to obtain 12 abnormal feature data as a training sample output set, wherein the working current is fixed and is not analyzed; and the training set, the validation set and the test set are divided according to the proportions of 0.70, 0.15 and 0.15, the number of hidden layers of the RBF surrogate model is set to 15, the minimum performance gradient is set to 1e-7, the maximum number of training iterations is set to 1000, and the Levenberg-Marquardt method is used for training, and the determination coefficient R 2 is used to evaluate the accuracy of the surrogate model in the training, testing and validation processes; the determination coefficients R 2 are all above 0.95.

[0207] The selected initial fault parameter samples are input into the theoretical model and the surrogate model, respectively, and the average errors of each feature between the surrogate model and the theoretical model under different sample numbers are compared to verify the accuracy of the surrogate model, and the surrogate model is selected to replace the theoretical model to obtain the sensitivity results; the consistency between the surrogate model and the theoretical model is evaluated by the average error, and the smaller the average error, the closer the prediction results of the surrogate model to the results of the theoretical model, thereby verifying the accuracy of the surrogate model; the simulation values of the surrogate model and the theoretical model in the embodiment are basically consistent, showing high accuracy, and therefore the established surrogate model can replace the theoretical model to obtain accurate sensitivity results;

[0208] Figure 4 The results of the final training, testing and validation sets of the RBF neural network surrogate model are shown. As can be seen from the figure, the determination coefficients R 2 are all above 0.95. Further, the selected initial training samples are input into the theoretical model and the surrogate model, respectively, and the average errors of each feature between the surrogate model and the theoretical model under different sample numbers are compared, and the results are as shown in Figure 5The simulation values of the agent model and the theoretical model are basically consistent, and show high precision.

[0209] The average error is specifically:

[0210]

[0211] In the formula, N is the number of samples; M is the number of indexes; is the output value of the jth agent model of the index i; is the output value of the jth theoretical model of the index i.

[0212] The radial basis function (RBF) neural network as the agent model is specifically:

[0213] According to the structure of the RBF neural network, the input layer is used to receive input data, which is represented as The hidden layer changes the input layer data through a set of kernel functions, and the output of the ith hidden layer node after the change is represented as:

[0214]

[0215] In the formula, is the Euclidean norm on the input space; is the kernel function center; is the bandwidth parameter of the ith kernel function;

[0216] The output of the jth node of the output layer is:

[0217]

[0218] In the formula, is the connection weight of the ith node of the hidden layer to the jth node of the output layer; is the bias of the jth node of the output layer; This is the ith basis function.

[0219] The determination coefficient R 2 Specifically:

[0220]

[0221] In the formula, n is the number of samples; is the output value of the ith agent model; is the output value of the ith theoretical model; is the mean value of the n sample points y.

[0222] S2: Based on the results of S1, analyze the PEMFC fault features based on global sensitivity;

[0223] S201: Analyze the PEMFC system fault sensitivity by using the proxy model built in S103;

[0224] The Sobol global sensitivity method is used in this embodiment to analyze the sensitivity of PEMFC system faults under different working conditions. The main effect and total effect sensitivity index of 60 A working condition, Figure 6 The main effect and total effect sensitivity index of this working condition are shown. As can be seen from the figure, the impact degree under different faults is different, and there is some interaction between some features, but the value is very small. Therefore, when analyzing the sensitivity of each abnormal symptom feature under different faults of the PEMFC system, since the main effect sensitivity index is the same as the total effect sensitivity index, only the main effect sensitivity index is analyzed:

[0225] The sensitivity index under different fault variables is converted into the proportion of each fault variable under different faults, and it is assumed that the features with a proportion greater than 0.1, i.e. 10%, are sensitive features, the parameters with a proportion between 0.05 and 0.1 are relatively sensitive features, and the parameters with a proportion less than 0.05 are not sensitive features. In addition, in order to more intuitively understand the sensitivity of different feature parameters to faults, the sensitive features are marked as , the relatively sensitive features are marked as , and the insensitive features are not marked. The sensitive feature set under different working conditions is shown in Tables 3-8:

[0226]

[0227]

[0228] In summary, among the five faults under different operating conditions, fault f1 can only be identified using feature e12, which is consistent with its generation mechanism. Specifically, due to stack contamination, the catalytic active area decreases, directly affecting the output voltage. For fault f2, features e1, e2, e4, e5, e6, e7, e8, and e11 are key features. For fault f2, feature e3 is the most critical, but under certain operating conditions, features e1, e2, e4, e5, e6, e7, e8, and e11 are also not negligible. For fault f4, e1, e2, e4, e5, e6, e7, and e11 are key features, similar to f2 and f3. Faults f2, f3, and f4 all cause changes in flow and pressure in different subsystems of the PEMFC; therefore, their sensitive features are mostly related to flow and pressure, which is consistent with the sensitivity analysis results. Fault f5 is mainly identified using features e9, e10, and e12. This is because fault f5 is temperature-related, thus directly affecting the humidity and output characteristics of the PEMFC system. Furthermore, it can be observed that features e1, e2, e4, e5, e6, e7, e8, and e11 in faults f2, f3, and f4 exhibit similar sensitivity indices and trends. This may be because flow rate and pressure in the PEMFC system are coupled variables, resulting in a strong correlation between these fault variables, which can interfere with the fault identification process. Therefore, it is necessary to analyze the correlation between fault variables to remove redundant information and resolve the interference problem caused by related features.

[0229] S202: Perform correlation analysis on the sensitive features selected in S201 to obtain the final set of sensitive features;

[0230] Sensitivity analysis can identify features sensitive to faults, but it overlooks the correlations between these features. In fault diagnosis practice, information overlap and redundancy can interfere with identification and consequently lead to performance degradation. Therefore, this paper proposes a feature cleaning strategy based on correlation analysis to remove redundancy between sensitive features and supplement more information for performance improvement, such as... Figure 7 As shown, the entire strategy mainly includes a sensitivity analysis phase, a feature cleaning phase, and a feature supplementation phase. In the sensitivity analysis phase, the Sobol global sensitivity analysis method is used to identify features that significantly impact diagnostic performance. In the feature cleaning phase, features that could lead to redundant or misleading diagnoses are removed or adjusted to improve the model's accuracy and reliability. Finally, the feature supplementation phase aims to enhance diagnostic performance by adding additional information.

[0231] Feature cleaning was performed based on correlation analysis, and correlation analysis was conducted on the original 12 sensitive features. Figure 8 The correlation analysis scatter plot of various features under 60 A operating conditions is shown. Figure 9The final Spearman correlation analysis results are given. As can be seen from the figure, there is significant redundancy or correlation between some features.

[0232] Through correlation analysis, the correlation between features is analyzed, and the correlation features with strong correlation with multiple feature parameters are obtained, and the correlation coefficient higher than 0.6 is selected out, as shown in Figure 10 .

[0233] Among e1, e2, e4, e5, e6, e7, e8 and e11, there is a high correlation (correlation coefficient higher than 0.6) between each other. For the coupling of the eight features, e8 is retained due to its high sensitivity coefficient, and the other features are removed. For e9 and e10, e10 is retained due to its high sensitivity coefficient, and e9 is removed. The sensitive feature set of each fault under 60 A working condition is cleaned by using the feature cleaning method described above, wherein the sensitive features under f1 fault are: e12; the sensitive features under f2 fault are: e1, e2, e4, e5, e6, e7, e8, e9, e11, wherein e1, e2, e4, e5, e6, e7, e8, e11 are correlated, and e8 with the highest sensitive feature proportion is taken, and the final sensitive feature is e8; the sensitive features under f3 fault are: e3 and e8, which are not correlated, and the final sensitive features are e3 and e8; the sensitive features under f4 fault are: e1, e2, e4, e5, e6, e7, e9, e11, wherein e1, e2, e4, e5, e6, e7, e11 are correlated, and e6 or e11 (e6 and e11 are equal) with the highest sensitive feature proportion is taken, and the final sensitive feature is e6 or e11; the sensitive features under f5 fault are: e9 and e10. Among them, e9 and e10 are correlated, and e10 with the highest sensitive feature proportion is taken, and the final sensitive feature is e10. In summary, the feature set for fault recognition under 60 A working condition is finally formed, as shown in Table 9.

[0234]

[0235] The sensitive feature set under 60 A working condition can be selected as e3, e6, e8, e10, e12 or e3, e8, e10, e11, e12 by comprehensively screening the sensitive features under five faults. Table 10 shows the sensitive features after screening under different working conditions.

[0236]

[0237] To verify the performance of the sensitive feature set in different working conditions, it needs to be brought into the evaluation model to study the performance of each feature selection method under different feature quantities. In terms of evaluation model, the K-neighbor algorithm is selected, which has good generalization ability and flexibility, and is suitable for multi-classification problems. Run the K-neighbor algorithm on the training set, select a suitable k value (neighbor number). For each sample in the test set, find the k most similar samples in the training set. According to the known categories of these most similar samples, determine the category of the test sample by majority voting. Calculate the accuracy or other performance indicators of the model. Compare the performance of the model using different feature sets to find the best combination. According to the evaluation results, select the feature set with the best performance. Finally, the performance of the sensitive feature set in different working conditions is shown in Table 11. As can be seen from the table, the sensitive feature set selected by the sensitivity feature selection algorithm performs slightly better than the original data set, and the training time is also significantly reduced.

[0238]

[0239] As can be seen from Table 10, the sensitive feature set in different working conditions all contains e3, e8, e12. Features e9 and e10 are related features, and in different working conditions, the sensitivity index is large or small. In addition to features e3, e8 and e12 that are sensitive in all working conditions, features e2 and e11 appear frequently. Therefore, in the case of considering all working conditions, features e3, e8 and e12 can be selected, and one of features e2 and e11 is selected, and one of features e9 and e10 is selected. In order to simplify the feature selection process, it is necessary to fuse the features of each working condition. Table 12 shows the performance of different sensitive feature sets in all working conditions. As can be seen from the table, in all working conditions, the sensitive feature set e2, e3, e8, e9, e12 shows the best performance and has a shorter training time. Although this combination is slightly lower in accuracy than training with different sensitive data sets for different working conditions, it has reduced training time. This indicates that the selected feature set has excellent generalization performance.

[0240] Table 12 Performance comparison of different sensitive feature sets in all working conditions

[0241]

[0242] S203: Formulate evaluation indicators of feature selection method, including precision dimension evaluation indicators and time dimension evaluation indicators formulation;

[0243] Precision dimension evaluation index formulation: the indexes for evaluating the performance of the fault diagnosis model include accuracy, precision and recall rate, in the confusion matrix as shown in Table 13, the attention-predicted class includes positive-positive TP, positive-negative FN, negative-positive FP and negative-negative TN;

[0244]

[0245] Accuracy represents the correctly predicted samples The proportion in the total number of samples, specifically:

[0246] Precision is only for samples predicted as positive, defined as the proportion of all positive samples

[0247] Successfully predicted as positive, specifically:

[0248] Recall rate is based on the true occurrence of the class, defined as the proportion of all actual positive samples Correctly predicted as positive, specifically:

[0249]

[0250] In the formula, TP represents the number of samples correctly predicted as positive; FN represents the number of samples incorrectly predicted as negative for the attention class; TN represents the number of samples correctly predicted as negative; FP represents the number of samples incorrectly predicted as positive for the attention class;

[0251] Accuracy is the most commonly used index to represent the overall performance of the diagnosis model, and high precision indicates that the model correctly identifies more samples. However, in the case of class imbalance, accuracy can be misleading. Since precision and recall are mutually restrictive, in order to comprehensively consider, effectively trade off between precision and recall, and use together with accuracy to evaluate the diagnosis performance, a comprehensive index F1_Score is introduced, specifically:

[0252]

[0253] Time dimension evaluation index formulation: the time dimension evaluation index is formulated by measuring the execution time of the feature selection method through the timing function. The main concern of the present application is various feature selection methods, so the focus is on the execution time cost of the feature selection method, and the time required for model training is not involved, mainly to compare the time of the feature selection method.

[0254]

[0255] ​S204: Based on the evaluation index of S203, compare the performance of different feature screening methods under different feature quantities, and select the feature screening method based on sensitivity analysis as the diagnostic method;

[0256] To compare different feature screening methods, a variety of common feature screening methods were constructed, including filter methods such as Relief Feature Selection, Relieff, Max-Relevance and Min-Redundancy, MRMR, and Neighborhood Components Analysis, NCA, wrapper feature selection methods such as Recursive Feature Elimination, RFE, embedded feature selection methods such as Random Forest, RF, eXtreme Gradient Boosting, XGBoost, and the sensitivity feature screening method proposed in the present application. The performance of different feature screening methods under different feature quantities was compared. Table 14 shows the feature importance ranking of various feature screening methods.

[0257]

[0258] According to the feature importance ranking of different feature selection methods in Table 14, the performance of different feature selection methods under different feature quantities was obtained by bringing them into the evaluation model, as shown in Table 15. Figure 11

[0259] As can be seen from the figure, as the number of features used for diagnosis increases, the performance of the classifier model shows a continuous upward trend until it reaches redundancy. When the number of features is small, the sensitivity analysis, MRMR, NCA and RFE methods perform well, among which the sensitivity analysis, NCA and PFE exhibit the best stability, all reaching stability after five features. These methods can effectively select features with high information content from the data and consider factors such as correlation and importance between features during the selection process. Therefore, the feature sets selected by them have better discrimination ability.

[0260] ​As can be seen from Table 15, the feature selection methods with good effects have highly similar importance rankings. In contrast, the methods with poor effects such as MRMR, Relieff and RF classify e4 and e7 (with lower importance in other methods) as the top five important features, which is the main reason for their poor performance. Overall, the filter methods perform poorly except for NCA. This is because they are only based on the statistics or correlation of a single feature for selection without considering the complex relationship between features. Compared with the filter methods, the embedded methods have a slight advantage, and the performance of the wrapper method is obviously better than that of other methods. Compared with the three traditional feature selection methods, the feature selection method based on sensitivity analysis combines the correlation analysis and sensitivity analysis strategies, can retain important features while removing redundant information, and make the final selected feature set have better discrimination ability. Therefore, the selection method based on sensitivity performs best under all feature numbers.

[0261] The time taken by different feature selection methods is also given, as shown in Table 15. The filter methods are independent of the model and select a feature subset according to the relationship between features, so the computational cost is small. Among them, the mutual information method takes the shortest time of 1.46 s, but the performance is poor; the performance of the PFE method is significantly better than that of other methods, but since it adopts a strategy similar to exhaustive search, it takes the longest time of 5369.53 s; the feature cleaning strategy based on correlation analysis and sensitivity analysis performs best and the computational cost is only 10.14 s, which is slightly higher than that of some filter methods with poor performance.

[0262]

[0263] As shown in Figure 12 , S3: based on the results of S1 and S2, an online diagnosis method based on sensitivity is proposed, i.e., a diagnosis method based on sensitivity is constructed, including a model adaptive stage, feature processing and screening, construction of a fault representation equation, fault evaluation and analysis, determination of a threshold and a sensitivity index, and verification of the proposed method.

[0264] Model adaptive stage: based on real-time running data, the system performance model and the RBF fault proxy model are adaptively updated, so as to generate accurate data that changes with system performance and characteristics based on the adaptive model;

[0265] Feature processing and screening: based on the Sobol global sensitivity analysis method, the fault features are screened and cleaned, the features with the greatest impact on the fault are found, and the interaction between them is evaluated;

[0266] Building the fault characterization equation: According to Sobol global sensitivity analysis, the final sensitive feature set obtained in S202 is split into the relative influence between different input parameters, when there is no interaction in the system, a linear equation can be built to describe the relationship between the input parameters and the output variable , as shown in the following formula:

[0267]

[0268] In the formula, is the intercept term; is the coefficient of each input parameter , which represents the degree of influence of the parameter on the output variable Y;

[0269] If the Sobol global sensitivity analysis result shows that there is interaction or nonlinear relationship, it can be expressed by adding interaction terms to the equation, specifically:

[0270]

[0271] In the formula, is the regression coefficient of the constant term and the first-order term; is the regression coefficient of the interaction term and the high-order term;

[0272] Fault evaluation and analysis: According to the sensitivity results and the change of sensitive features, the least squares method is used to back-propagate the fault parameter change, and then by comparing the absolute value of the residual of the PEMFC sensitive feature with its threshold value when the system is normal, if is diagnosed as having a fault, otherwise, it is diagnosed as normal;

[0273] Determine the threshold and sensitivity index: According to the change of sensitive features obtained by fault evaluation, adjust the threshold and sensitivity coefficient to ensure that the running characteristics and fault modes of the system can be captured, so as to improve the accuracy and reliability of fault diagnosis;

[0274] According to the sensitivity diagnosis method flow, first, the data set needs to be processed through the feature screening strategy based on sensitivity and correlation, and the sensitive feature set under different working conditions is obtained, see Table 10 for the selected sensitive features under different working conditions. In addition, the global sensitivity analysis result of PEMFC system shows that the main effect and total effect values of each sensitive feature under different faults are basically consistent. This means that under the sensitive features, the interaction between these fault parameters is very small, and each fault is independent of each other. Based on the principle of main effect, multiply each fault input parameter by its corresponding Sobol sensitivity coefficient, and add them to build a linear equation to describe the influence degree of different faults on sensitive features, specifically:

[0275]

[0276] In the formula, Sensitivity coefficient; These are labels for sensitive features; The fault number is represented by Yi; Yi represents the output of each sensitive feature. These are the fault-sensitive characteristics of each system;

[0277] Furthermore, to simplify the global sensitivity analysis process, each fault parameter is analyzed individually, without considering combinations of all parameters. A specific sensitivity index is provided for each fault parameter, indicating the degree to which a change in that parameter affects the abnormal symptoms. Assume that any performance parameter Yi of PEMFC is affected by m factors, denoted as follows: Then the static mathematical model of PEMFC is:

[0278]

[0279] At this point, the fault sensitivity characteristics of each system are: The output of each corresponding sensitive feature is When influencing factors When changes occur within a certain range, the outputs of each sensitive feature become The deviation value of the influencing factors The deviation of the output of each sensitive feature is The proportional relationship between the relative deviation of the output of each sensitive feature of PEMFC and the relative deviation of the corresponding influencing factor causing the fault is defined as the sensitivity characteristic coefficient of the output of each sensitive feature to the influencing factor of the fault, specifically:

[0280]

[0281] Under fixed operating conditions (e.g., 60 A), the characteristic coefficients of the sensitive features selected after global variance sensitivity analysis and data cleaning were constructed, and the results are shown in Table 16:

[0282]

[0283] Based on the constructed fault characterization equations and sensitivity coefficients, assume a system of equations containing m equations and n unknowns, specifically:

[0284]

[0285] In the formula, is an mxn matrix, i.e. the fault characterization matrix composed of sensitivity feature coefficients; x is an n-dimensional vector representing the changes in the individual system fault sensitive features, b is an m-dimensional vector representing the right-hand side terms of the equation system, i.e. the changes in the outputs of the sensitive features in the system;

[0286] The goal of the least squares method is to solve the unknown vector x such that the residual sum of squares is minimized, thus the residual vector is defined as:

[0287]

[0288] The goal is to minimize the residual sum of squares, i.e. minimize the error function as:

[0289]

[0290] where, represents the Euclidean norm of the residual vector .

[0291] This means that an optimal fault parameter change is needed to be found such that the solution of the sensitivity standard equation can best approximate the changes in the sensitive features collected from the PEMFC. In addition, in order to confirm whether it is caused by the PEMFC system fault, a decision threshold is applied to the changes in the sensitive features. Specifically, for a certain fault, the part of the change rate of which exceeds the threshold is set to 1 to form a binary matrix. For example, the part of the change rate of which exceeds the decision threshold in the f1 fault parameter is set to 1; the part of the change rate of which exceeds the decision threshold in the f2 fault parameter is set to 1; and so on. Finally, the obtained binary matrix is compared with the real fault label information to infer the specific fault type occurred.

[0292] Verification of the sensitivity diagnosis method:

[0293] where the fault diagnosis data set is constructed;

[0294] In order to collect sufficient information about the normal behavior of the PEMFC system, the PEMFC system proxy model is run and the variable data is obtained. In addition, in order to simulate the real sensor sampling environment, Gaussian noise with a variance of 1 is added to the current input to simulate the fault noise of the fuel cell under actual working conditions. Finally, 18000 sets of experimental data are derived with a sampling time of 1s. 70% of the 18000 samples are selected as the training data set, and the test data is selected from the remaining samples.

[0295] For the diagnosis of synchronous faults, considering that in actual engineering scenarios, systems usually work in normal state and faults rarely occur, let alone simultaneous faults. Although the state monitoring system continuously collects system operation data, due to the fact that normal data is much more than fault data, the working condition data used for diagnosis presents a serious unbalanced characteristic. In order to be closer to the actual situation and cover various possible fault types and simultaneous occurrence, five different types of faults f1 to f5 are introduced, and the fault degree of each type varies between its specified range. Then 12600 groups of single fault data and 180 groups of simultaneous fault data are selected as the training set to reflect the scarcity of simultaneous fault data in actual situations, in addition, 5400 groups of single fault and 5820 groups of simultaneous fault test data are selected as test data. The two data sets do not overlap each other to simulate the serious imbalance of actual working condition data. Through such design, the generalization ability and robustness of the model can be better evaluated, so as to optimize and improve the performance of the model more effectively.

[0296] Among them, the performance evaluation index:

[0297] In the comparison of different feature selection methods, F1_Score is introduced to evaluate the effect of feature selection. For single fault diagnosis performance, the comprehensive evaluation index can still be used to effectively compromise between precision and recall rate. But when evaluating the performance of synchronous faults (multi-label classifier), in order to comprehensively consider the case of partial correct prediction, the macro F1 value (Macro-F1) is introduced as the evaluation index. Macro F1 value calculates the F1 value of each class and takes the average of the F1 values of all classes, which more accurately reflects the performance of the model on each class, rewards the model even if it is not completely correct to predict all labels, and can get a certain degree of recognition, so as to more accurately evaluate the overall performance.

[0298] For the multi-fault data test set , the prediction label data of the classifier prediction output is expressed as:

[0299]

[0300] Among them, the number of faults in the data set is , and the number of test set samples is ns.

[0301] The correct label data of the test set can be expressed as:

[0302]

[0303] ​By comparing the prediction matrix and the label matrix of the classifier, the difference between the two can be quantified to evaluate the classification performance of the multi-label classifier. By comparing the prediction matrix and the real label matrix , the precision of the classifier can be calculated as:

[0304]

[0305] Similarly, the recall and under multi-label are obtained, which are:

[0306]

[0307]

[0308] Among them, the fault diagnosis result analysis:

[0309] In order to verify the feasibility of the sensitivity-based diagnosis method, the classification effect is tested by using single fault and simultaneous fault data sets respectively. The confusion matrix of single fault diagnosis result is shown in Figure 13 , and the macro-average F1 evaluation index is introduced to measure the diagnosis performance, as shown in Table 17. This method achieves more than 90% accuracy for single fault and simultaneous fault, and the diagnosis time is also within an acceptable range. Therefore, the sensitivity-based diagnosis method can effectively diagnose the fault state of the PEMFC system.

[0310]

[0311] Among them, the diagnosis performance evaluation and comparative study includes:

[0312] In the process of single fault data, single-label diagnosis method assigns a unique label to each fault, while concurrent faults are treated as a new independent label for training, and the label complexity increases exponentially with the increase of concurrent fault combination. In contrast, multi-label diagnosis method allows a sample to have multiple labels, usually using one-hot encoding to handle labels, which can more effectively capture and represent the combination relationship between different faults and understand the complex situation of system failure more comprehensively, especially when dealing with concurrent faults of PEMFC system. In order to verify the superiority of sensitivity-based diagnosis method, different diagnosis models are trained using the same training data set, including traditional single-label learning algorithms such as decision tree (DT), linear discriminant analysis (LDA), support vector machine (SVM) and K-nearest neighbor algorithm (KNN). In addition, multi-label learning algorithms such as multi-label decision tree (ML-DT), multi-label random forest (ML-RF) and multi-label K-nearest neighbor algorithm (ML-KNN) are also included.

[0313] Table 18 shows the performance of the sensitivity-based fault diagnosis method proposed in the present application and several common machine learning classification algorithms in the test data set. As can be seen from Table 18, except for LDA, other diagnosis methods all obtain F1 accuracy of more than 99% in single fault diagnosis. The sensitivity-based diagnosis method diagnoses 5400 groups of data in 0.264 s, which is more than 5 times faster than the KNN method with the highest precision. Its diagnosis process is more efficient. In summary, the sensitivity-based diagnosis algorithm not only has lower computational burden, but also has better diagnosis performance.

[0314]

[0315] Table 19 shows the performance comparison of single-label and multi-label diagnosis methods trained using the same test and training sets in the case of simultaneous fault data scarcity. The results show that both single-label and multi-label diagnosis methods can achieve an F1 accuracy of more than 90% in the presence of single fault data. However, the multi-label fault method outperforms the single-label classification method in terms of diagnosis accuracy and time consumption in identifying simultaneous faults. This is due to the multi-label method's ability to better capture the correlation and common features between different faults, thereby improving diagnosis accuracy, which corresponds to the analysis above. In contrast, the single-label diagnosis algorithm performs poorly in this regard, as it can only treat simultaneous faults as a new class for training, resulting in poor generalization ability and robustness.

[0316]

[0317] In Table 19, the performance of the fault diagnosis algorithm based on sensitivity analysis and several common multi-label machine learning classification algorithms in handling simultaneous faults is also compared. As can be seen from the table, among the multi-label methods, ML-DT has the highest macro-average F1 accuracy in handling simultaneous faults. However, its single fault F1 accuracy is 97.0%, lower than the sensitivity-based 99.1%. This is because the sensitivity-based algorithm uses a specific diagnosis mechanism for simultaneous faults, which decomposes and characterizes simultaneous faults. Therefore, for simultaneous faults, the sensitivity-based diagnosis algorithm shows the highest macro-average F1 accuracy (92.5%), which is 7.5% higher than the highest multi-label method, and it can efficiently diagnose simultaneous faults with different distribution characteristics using only single fault data, with a diagnosis time of only 0.308 s.

[0318] A fuel cell fault diagnosis method device, comprising a data acquisition module, a model construction module, a feature screening module, an adaptive updating module, and a fault diagnosis module. The data acquisition module is responsible for real-time monitoring and collecting key operating parameters of the fuel cell system. The model construction module is used to construct a PEMFC system theoretical model and design a radial basis function neural network proxy model. The feature screening module is used to screen and clean features based on sensitivity analysis results and correlation analysis, and to find the features that have the greatest impact on faults. The adaptive updating module is used to continuously optimize model parameters based on new operating data and diagnosis experience, and to perform adaptive updating of the system performance model and the adaptive RBF fault proxy model. The fault diagnosis module is used to combine the adaptive model obtained from real-time data and the screened feature set to establish a fault feature characterization equation, and to perform real-time fault diagnosis of the fuel cell system.

[0319] The various embodiments described in this specification are implemented in the context of a progressive manner, each embodiment focusing on the differences from other embodiments, and the same or similar parts between embodiments can be mutually referred to. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part description. The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fuel cell failure diagnosis method characterized by comprising: Comprise the following steps: S1: Proton exchange membrane fuel cell PEMFC system agent model construction and verification; S101: PEMFC system theoretical model construction and verification; S102: According to the PEMFC system theoretical model established in S101, select the system fault type and abnormal characteristics; S103: Based on the results of S101 and S102, build an agent model for global sensitivity analysis and verify the accuracy; S2: Based on the results of S1, analyze the PEMFC fault characteristics based on global sensitivity; S201: Use the agent model built in S103 to analyze the sensitivity of the PEMFC system failure; S202: Correlation analysis of the sensitive features selected in S201 to obtain the final sensitive feature set; S203: Develop evaluation indicators for feature selection methods, including precision dimension evaluation indicators and time dimension evaluation indicators; S204: Based on the evaluation indicators in S203, compare the performance of different feature selection methods under different feature quantities, and select the feature selection method based on sensitivity analysis as the diagnosis method; S3: Based on the results of S1 and S2, propose an online diagnosis method based on sensitivity, that is, build a diagnosis method based on sensitivity, including model adaptation stage, feature processing and screening, build fault representation equation, fault evaluation and analysis, determine threshold and sensitivity index, and verify the proposed method.

2. The fuel cell failure diagnosis method according to claim 1, characterized by, S101 is specifically: The PEMFC system theoretical model includes building a PEMFC stack model and a PEMFC auxiliary equipment model, which are integrated to obtain a PEMFC system theoretical model; wherein the PEMFC stack model includes a PEMFC output voltage model, an anode mass flow model, a cathode mass flow model, and a membrane hydration model; the PEMFC auxiliary equipment model includes an air compressor model, a cooler model, a humidifier model, an intake and exhaust manifold model, and a control system model; the specific construction method is: PEMFC stack model construction: PEMFC output voltage model construction: The actual output voltage of a PEMFC cell is obtained by subtracting the three voltage losses caused by polarization phenomena from the ideal voltage: ； wherein represents the thermodynamic theoretical electromotive force; , and respectively represent the activation, concentration difference and ohmic loss voltage; Anode mass flow model construction: The anode mass flow model aims to describe the mass flow process of hydrogen and water vapor in the PEMFC anode and their interaction; according to the law of conservation of mass, the change of each gas is represented as: ； wherein, and are the flow rates of hydrogen and water vapor flowing into the anode flow field, respectively; , and are the flow rates of hydrogen, water vapor and liquid water flowing out of the anode flow field, respectively; is the consumed hydrogen flow rate; is the water vapor mass in the anode flow field; is the hydrogen mass in the anode flow field; represents the water vapor flow rate transferred from the proton exchange membrane side; Cathode mass flow model construction: The change of each gas is represented as: ； ； ； wherein, , and are the flow rates of the , and water vapor flowing into the cathode flow field, respectively; , , and are the flow rates of the respective gases and liquid water flowing out of the cathode flow field, respectively; and are the flow rates of the consumed and water vapor generated by the PEMFC, respectively; is the mass of oxygen in the cathode flow field, is the mass of water vapor in the cathode flow field; is the mass of nitrogen in the cathode flow field; Membrane hydration model construction: The water transport amount caused by electrical drag is affected by current density and electrical drag coefficient, and the water transport amount caused by electrical drag is represented as: ； wherein is the electrical drag coefficient, is the electrical drag coefficient, is the current density In the operation of the PEMFC system, the cathode is the main water generation site, resulting in a high water content on the cathode side, forming a concentration gradient; therefore, water will diffuse from the cathode to the anode, alleviating the uneven distribution of part of the water content; the water amount from the cathode to the anode caused by concentration diffusion is represented as: ； wherein D is the diffusion coefficient of water in the membrane; C is the water concentration; L is the vertical distance of the proton exchange membrane; PEMFC auxiliary equipment model construction: Air compressor model construction: Get the flow of the air compressor: ； wherein is the air compressor characteristic parameter; is the air compressor tip line speed, is the air compressor blade diameter, is a constant related to the Mach number; After the air is compressed by the air compressor, the temperature change is represented as: ； wherein for the working efficiency of the air compressor; for the temperature at the inlet of the air compressor; Table for the pressure at the inlet of the air compressor; for the pressure at the outlet of the air compressor, is the specific heat coefficient of the gas at constant pressure; Cooler model construction: The relative humidity of the air leaving the cooler due to the temperature change is given by: ； wherein and is the relative humidity and pressure of air at standard atmospheric pressure; is the water vapor pressure at standard atmospheric pressure; is the pressure of the cooled air; is the gas temperature of the cooled air; denotes the saturated water vapor pressure, denotes the temperature at standard atmospheric pressure; The flow rate of dry air and water vapor leaving the cooler is given by: ； wherein represents the air humidity ratio at the cooler outlet; represents the flow of dry air in the cooler; represents the flow of water vapor; represents the total flow in the cooler; is the molar mass of water vapor, represents the molar mass of air; , are the partial pressures of dry water vapor and dry air in the cooler, respectively; Humidifier model construction: The flow rate of the reactant gas before entering the humidifier is mainly composed of water vapor and dry air, which is given by: ； wherein , Qd and Qw represent the flow rates of dry air and water vapor into the humidifier, respectively; The water injection amount of the humidifier to the dry air is given by: ； wherein Pw is the water vapor partial pressure, Qw is the water vapor flow rate out of the humidifier, Pda is the dry air partial pressure; Qda is the dry air flow rate at the humidifier inlet; The total pressure of the gas at the outlet of the humidifier is given by: ； Intake and exhaust manifold model construction: Considering that the pressure difference before and after the air passes through the intake manifold is very small, the flow rate of the gas flowing out is given by: ； wherein is the supply manifold outlet flow constant; is the intake manifold outlet gas pressure; denotes the pressure at the intake manifold inlet; Due to the high temperature of the air from the air compressor, the temperature and pressure of the air will change when it passes through the intake manifold, and the change in the supply manifold is: ； ； wherein Vsup is the aggregate volume of the supply manifold; Tsup is the temperature of the gas in the supply manifold; Tout is the temperature of the air flowing out of the air compressor; Qsup is the flow rate of the gas into the supply manifold, Cp is the specific gas constant for air, Msup is the mass of the gas in the supply manifold; The temperature difference of the air after leaving the stack and the exhaust manifold is small and can be ignored; therefore, there is: ； wherein P represents the pressure of the exhaust manifold; T represents the temperature of the gas in the exhaust manifold, approximated by the temperature of the stack; and Gin and Gout represent the gas flow at the inlet and outlet of the exhaust manifold, respectively; V represents the lumped volume of the exhaust manifold, R represents the gas constant; M represents the molar mass of air. The working state of the nozzle is represented by the ratio of the upstream pressure to the downstream pressure: ； When the ratio is greater than the critical pressure ratio, the nozzle is in a subcritical state; otherwise, when the ratio is less than or equal to the critical pressure ratio, the nozzle is in a supercritical state; the exhaust manifold flow rate in different states is given by: ； In the formula, represents the manifold discharge coefficient; represents the opening degree of the exhaust valve; Control system model construction: The peroxide ratio is determined according to the ratio between the air flow rate supplied by the air compressor and the air flow rate required by the stack current: ； The air flow rate reference value of the air compressor is: ； wherein Molar mass of dry air, P = pressure of the compressed air at the compressor outlet, The proportional valve is used to quickly adjust the hydrogen flow rate and control the pressure on both sides of the membrane, which is given by: ； wherein is the pressure linear gain of the proportional valve; is the regulation coefficient of the pressure drop between the gas supply manifold and the cathode flow field.

3. The fuel cell failure diagnosis method according to claim 1, characterized by, The S102 specifically is: System fault types include f1, f2, f3, f4, and f5, which simulate the entry of pollutants into the stack, the increase in flow resistance, the leakage of the exhaust manifold, the leakage of the intake manifold, and the loss of control of the cooling system, respectively; Abnormal features include air compressor output pressure, air intake flow rate, exhaust manifold outlet pressure, peroxide ratio, air intake pressure, hydrogen intake pressure, air compressor output flow rate, cathode outlet pressure, anode relative humidity, cathode relative humidity, anode outlet pressure, output voltage, and working current; Fault f1 represents that the system is contaminated during operation, resulting in a decrease in catalytic active area and a decrease in current density, and its specific function is: ； wherein, is the operating current of the stack, represents the coefficient of variation of the fault parameter; is the effective working area of the PEMFC; since the current density is an important factor affecting the total voltage of the stack, the fault f1 will directly affect the output voltage of the PEMFC system; Fault f2 represents a change in the cathode air flow resistance to simulate water flooding in the cathode diffusion layer, and its specific function is: ； wherein, is the total cathode outlet flow rate; is the cathode outlet orifice constant; and are the cathode pressure and atmospheric pressure, respectively. Fault f3 represents gas leakage in the exhaust manifold, which is simulated by multiplying the exhaust manifold output flow rate by a coefficient k3 between 0 and 1, and its specific function is: ； wherein, represents the gas flow rate at the exhaust manifold outlet; represents the manifold discharge coefficient; represents the opening degree of the exhaust valve; represents the pressure of the exhaust manifold; represents the gas temperature in the exhaust manifold, which is approximated to the temperature of the stack; is the specific heat coefficient of air at a fixed pressure; Fault f4 represents gas leakage in the intake manifold, and its specific function is: ； wherein, represents the flow rate of the gas flowing out of the intake manifold; is the exit hole constant of the supply manifold; and are the pressures of the intake manifold and the cathode side, respectively. Fault f5 represents the loss of control of the internal temperature of the PEMFC, which is simulated by changing the set stack operating temperature, and its specific function is: 。 4. The fuel cell failure diagnosis method according to claim 1, characterized by The S103 specifically is: The radial basis function (RBF) neural network is selected as the surrogate model. The fault parameters include the decrease of activation area, the decrease of water discharge (i.e. the increase of flow resistance), the decrease of air discharge, the decrease of air supply, and the increase of stack operating temperature. The ranges of the parameters in the data set, including the normal data set and the fault data set, are determined according to the maximum and minimum values of each fault parameter. The initial fault parameter samples are generated as the training input set using Sobol sequence, and are input into the PEMFC system theoretical model for simulation to obtain abnormal feature data as the training sample output set. The abnormal feature data is divided into the training set, the validation set and the test set according to the proportions of 0.70, 0.15 and 0.

15. The number of hidden layers, the minimum performance gradient and the maximum number of training iterations of the RBF surrogate model are set, and the Levenberg-Marquardt method is used for training. The determination coefficient R 2 is used to evaluate the accuracy of the surrogate model in the training, testing and validation processes. The selected initial fault parameter samples are input into the theoretical model and the surrogate model, and the average errors of each feature between the surrogate model and the theoretical model under different sample sizes are compared to verify the accuracy of the surrogate model, and the surrogate model is selected to replace the theoretical model to obtain the sensitivity results; The consistency between the surrogate model and the theoretical model is evaluated by the average error, and the smaller the average error, the closer the prediction results of the surrogate model to the results of the theoretical model, thereby verifying the accuracy of the surrogate model; The average error specifically is: ； In the formula, N is the number of samples; M is the number of indexes; is the output value of the jth agent model of the index i; is the output value of the jth theoretical model of the index i.

5. The fuel cell failure diagnosis method according to claim 4, characterized by, The radial basis function (RBF) neural network as a surrogate model specifically is: According to the RBF neural network structure, the input layer is used to receive input data, denoted as The hidden layer changes the input layer data through a set of kernel functions, and the output of the i-th hidden layer node after the change is denoted as: ； wherein is the Euclidean norm on the input space; is the kernel function center; is the bandwidth parameter of the ith kernel function; The output of the jth node of the output layer is: ； In the formula, is the connection weight from the i-th node of the hidden layer to the j-th node of the output layer; is the bias of the j-th node of the output layer.

6. The fuel cell failure diagnosis method according to claim 4, characterized by The determination coefficient R 2 Specifically: ； In the formula, n is the number of samples; is the output value of the ith agent model; is the output value of the ith theoretical model; is the mean value of n sample points y.

7. The fuel cell failure diagnosis method according to claim 1, characterized by, The S201 and S202 are specifically: S201: Analyzing the fault sensitivity of the PEMFC system: Since the main effect sensitivity index and the total effect sensitivity index are the same when analyzing the sensitivity of each abnormal feature under different faults of the PEMFC system, only the main effect sensitivity index is analyzed: Convert the sensitivity index under different fault variables into the proportion of each fault variable under different faults, and assume that the features with a proportion greater than 0.1, i.e., 10%, are sensitive features, the parameters with a proportion between 0.05 and 0.1 are relatively sensitive features, and the parameters with a proportion less than 0.05 are insensitive features; S202: Analyzing the correlation between the sensitive features in S201 to obtain related features with strong correlation with the sensitive features, selecting sensitive features with a correlation coefficient higher than 0.6, then retaining the sensitive feature with the highest sensitivity coefficient, and combining the sensitive features under f1-f5 faults to obtain sensitive feature set A and sensitive feature set B; Among them, the sensitive feature set A: exhaust manifold outlet pressure, hydrogen inlet pressure, cathode outlet pressure, cathode relative humidity, output voltage; The sensitive feature set B: exhaust manifold outlet pressure, cathode outlet pressure, cathode relative humidity, anode outlet pressure, output voltage; The sensitive feature set A and the sensitive feature set B are respectively classified by K-neighbor method to judge their performance under different working conditions, and the sensitive feature set with the highest accuracy and the shortest training time is selected as the final sensitive feature set.

8. The fuel cell failure diagnosis method according to claim 1, characterized by, The precision dimension evaluation index and the time dimension evaluation index are specifically: Precision dimension evaluation index: The indicators for evaluating the performance of the fault diagnosis model include accuracy, precision, and recall rate. In the confusion matrix, the concerned-predicted class includes positive-positive TP, positive-negative FN, negative-positive FP, and negative-negative TN. Accuracy represents the proportion of correctly predicted samples out of the total number of samples specifically: ； Accuracy was defined as the proportion of all positive samples that were successfully predicted as positive specifically:​ ； Recall is defined as the proportion of all actual positive samples that are correctly predicted as positive, specifically: true positives, i.e., the number of correctly predicted positive samples divided by the total number of actual positive samples. ； In the formula, TP represents the number of samples correctly predicted as positive; FN represents the number of samples incorrectly predicted as negative; TN represents the number of samples correctly predicted as negative; and FP represents the number of samples incorrectly predicted as positive. Since precision and recall rate are mutually restrictive, in order to comprehensively consider, effectively balance between precision and recall rate, and use together with accuracy to evaluate the diagnosis performance, a comprehensive index F1_Score is introduced, which is specifically: ； Time dimension evaluation index: The time dimension evaluation index is established by measuring the execution time of the feature selection method through the timing function.

9. The fuel cell failure diagnosis method according to claim 1, characterized by, The S3 is specifically: Model adaptive stage: Based on real-time running data, the system performance model and the RBF fault proxy model are updated adaptively, so as to generate accurate data that changes with system performance and characteristics based on the adaptive model; Feature processing and screening: Based on the Sobol global sensitivity analysis method, the fault features are screened and cleaned to find the features that have the greatest impact on the fault and evaluate the interaction between them; Building fault characterization equation: according to Sobol global sensitivity analysis, the final sensitive feature set obtained in S202 is split into the relative influence between different input parameters, when there is no interaction in the system, a linear equation can be built to describe the relationship between the input parameters and the output variable , as shown in the following formula: ； wherein is an intercept term; is a coefficient of each input parameter representing the degree of influence of that parameter on the output variable Y; If Sobol global sensitivity analysis results show that there is interaction or nonlinear relationship, it can be expressed by adding interaction terms to the equation, which is: ； wherein is the regression coefficient of the constant term and the first order term; is the regression coefficient of the interaction term and the higher order term; Fault evaluation and analysis: according to the sensitivity results and the change of sensitive features, the least square method is used to deduce the change of fault parameters, and then the absolute value of the residual of the PEMFC sensitive features is compared with the threshold value in the normal state to determine whether the system is faulty . When , it is diagnosed as faulty, otherwise, it is diagnosed as normal . Determine the threshold and sensitivity index: According to the change of sensitive characteristics obtained by fault evaluation, adjust the threshold and sensitivity coefficient to ensure that the running characteristics and fault modes of the system can be captured, so as to improve the accuracy and reliability of fault diagnosis; Based on the principle of main effect, multiply each fault input parameter by its corresponding Sobol sensitivity coefficient, and add them to construct a linear equation describing the influence of different faults on sensitive characteristics, which is: ； wherein is a sensitivity coefficient; is a label of a sensitive feature; is a label of a fault; Yi is an output of each sensitive feature; is each system fault sensitive feature; Assume that any performance parameter Yi of the PEMFC is affected by m factors, denoted as The static mathematical model of the PEMFC is: ； At this time, the sensitive features of each system fault are The outputs of the corresponding sensitive features are When the influencing factor Changes within a certain range, the output of each sensitive feature becomes The deviation value of the influencing factor is The deviation amount of the output of each sensitive feature is The proportional relationship between the relative deviation of the output of each sensitive feature of the PEMFC and the relative deviation of the corresponding influencing factor causing the fault is defined as the sensitive characteristic coefficient of the output of each sensitive feature to the fault influencing factor, specifically: ； According to the constructed fault characteristic equation and sensitive characteristic coefficient, assume that the equation group contains m equations and n unknowns, which is: ； wherein is an mxn matrix, i.e. a fault characterization matrix composed of sensitivity feature coefficients; x is an n-dimensional vector representing changes in the individual system fault sensitivity features, and b is an m-dimensional vector representing the right-hand side terms of the equation system, i.e. changes in the output of the sensitivity features in the system; The goal of the least squares method is to solve the unknown vector such that the sum of the squares of the residuals is minimized, and thus the residual vector is defined as: ； The objective is to minimize the sum of squared residuals, that is, to minimize the error function: ； wherein denotes the Euclidean norm of the residual vector .

10. An apparatus for implementing the fuel cell failure diagnosis method according to any one of claims 1 to 9, characterized by, It includes data acquisition module, model construction module, feature selection module, adaptive updating module and fault diagnosis module. The data acquisition module is responsible for real-time monitoring and collecting key operating parameters of the fuel cell system; The model construction module is used to construct the theoretical model of PEMFC system and design the radial basis function neural network proxy model; The feature selection module is used to select and clean the features based on the sensitivity analysis results and correlation analysis, and find out the features that have the greatest impact on the fault; The adaptive updating module is used to optimize the model parameters according to the new operating data and diagnostic experience, and to update the system performance model and adaptive RBF fault proxy model; The fault diagnosis module is used to combine the adaptive model obtained from real-time data and the filtered feature set to establish a fault feature representation equation, and to perform real-time fault diagnosis on the fuel cell system.

Images