Intelligent fault diagnosis method for solid oxide fuel cell system

Abstract

This invention relates to an intelligent fault diagnosis method for solid oxide fuel cell systems. By constructing a classifier array consisting of K parallel and independent support vector machine sub-models, the method first trains a classifier array with feature decoupling capabilities using a single fault sample, thus achieving the identification of concurrent faults. Then, by constructing a rule base for the physical logical relationship between faults and symptoms for logical verification, the physical reliability of the results is improved. Ultimately, this method achieves high-precision concurrent fault diagnosis with only single fault data, while meeting online real-time requirements. It solves two major technical challenges: the inability to diagnose concurrent faults in solid oxide fuel cell systems without concurrent fault samples, and the poor real-time performance of diagnosing high-dimensional coupled systems.

Description

Technical Field

[0001] This invention relates to the field of fuel cell monitoring technology, and more specifically, to an intelligent fault diagnosis method for solid oxide fuel cell systems. Background Technology

[0002] Solid oxide fuel cells (SOFCs) have broad prospects in distributed power generation and hydrogen energy utilization due to their advantages such as high energy conversion efficiency, fuel flexibility, and low emissions. However, SOFC systems are complex, high-dimensional, strongly coupled, and nonlinear systems, and their long-term reliability and safety face challenges.

[0003] In the existing technology, most SOFC fault diagnosis research is still at the level of offline identification of single faults, which has three major bottlenecks: 1) It cannot effectively handle concurrent faults without samples; 2) The diagnosis results are only "present or absent", lacking the quantification of "severity" and cannot support maintenance decisions; 3) The model is fixed and cannot adapt to the performance degradation and feature drift in the long-term operation of the system. Summary of the Invention

[0004] The technical problem to be solved by this invention is how to address the issues of existing technologies having a single diagnostic dimension, lacking adaptability, and being unable to provide quantitative evaluation.

[0005] This invention provides an intelligent fault diagnosis method for a solid oxide fuel cell system, comprising: Step 1: Obtain multi-dimensional runtime sequence parameters of the solid oxide fuel cell system under normal operating conditions and K preset single fault modes, and construct a training sample database; Step 2: Preprocess the training samples in the training sample database to obtain the training low-dimensional feature dataset. Step 3: Construct a classifier array consisting of K parallel and independent support vector machine sub-models, and train each support vector machine sub-model using an asymmetric strategy. The asymmetric strategy includes: based on the training low-dimensional feature dataset, constructing a corresponding asymmetric training sample set for each training support vector machine sub-model for training. Step 4: Collect the real-time operating data of the solid oxide fuel cell system and preprocess it to obtain a real-time low-dimensional feature dataset. Input the low-dimensional feature dataset into the trained classifier array. Each support vector machine sub-model outputs a binary diagnostic label and a corresponding confidence score. Step 5: Concatenate the binary diagnostic labels into a K-dimensional fault state vector, and concatenate the confidence scores into a K-dimensional confidence vector to generate a fault diagnosis report.

[0006] Compared with existing technologies, this application has the following advantages: by constructing a classifier array composed of K parallel and independent support vector machine sub-models, a classifier array with feature decoupling capability is first trained using single fault samples, thereby realizing the identification of concurrent faults; it achieves the technical effect of completing high-precision concurrent fault diagnosis with only single fault data and meeting online real-time requirements; it solves the two major technical problems of solid oxide fuel cell systems being unable to diagnose concurrent faults under conditions without concurrent fault samples and the poor real-time performance of high-dimensional coupled systems.

[0007] In one possible implementation, the multidimensional runtime sequence parameters collected in step 1 include at least one or more of the following: voltage and current characteristic parameters, temperature distribution parameters, gas pressure parameters, gas flow rate parameters, and gas composition parameters. Each parameter is collected at a fixed frequency to form an independent time series; and data segments are extracted according to a preset time window as training samples to build a training sample database.

[0008] Compared with existing technologies, by clearly defining the specific categories and acquisition methods of multi-dimensional runtime sequence parameters, and adopting a comprehensive monitoring scheme covering multiple physical fields such as voltage, current, temperature, pressure, flow rate, and composition, this approach solves the technical problems of missed or misdiagnosed cases caused by single feature dimensions and incomplete information in traditional diagnostics. By systematically collecting key parameters directly related to electrochemical performance, thermal management, and material balance, it provides a complete data foundation for subsequent fault feature extraction and model training, thereby achieving the technical effect of comprehensively improving fault feature characterization capabilities and enhancing model discrimination accuracy.

[0009] In one possible implementation, the voltage and current characteristic parameters include one or more of the following: total stack voltage, stack output current, stack power, individual cell voltage, and stack DC internal resistance. The temperature distribution parameters include one or more of the following: stack anode inlet temperature, stack anode outlet temperature, stack cathode inlet temperature, stack cathode outlet temperature, stack center temperature, and reformer temperature. The gas pressure parameters include one or more of the following: fuel flow anode inlet pressure, fuel flow anode outlet pressure, oxidant flow cathode inlet pressure, oxidant flow cathode outlet pressure, and fuel supply pipeline pressure. The gas flow parameters include one or more of the following: the anode inlet volume or mass flow rate of fuel gas, the anode outlet flow rate, the cathode inlet flow rate of air or oxygen, and the cathode outlet exhaust flow rate. The gas composition parameters include one or more of the following: hydrogen concentration of fuel gas at the anode inlet, hydrogen concentration and carbon monoxide concentration of tail gas at the anode outlet, water vapor concentration at the anode outlet, oxygen concentration at the cathode outlet, fuel utilization rate, and oxidant utilization rate.

[0010] Compared with existing technologies, this approach further refines and enumerates parameters such as voltage and current characteristics, temperature distribution, gas pressure, gas flow rate, and gas composition. By employing specific sensor monitoring points and calculation parameters as feature sources, it solves the problems of vague feature definitions and weak feasibility. By clearly defining key parameters such as the total voltage of the fuel cell stack, the inlet / outlet temperature / pressure / flow rate of each flow channel, gas concentration, and utilization rate, the data acquisition objectives are clear, and the physical meaning of the features is explicit. This achieves the technical effect of guiding engineering implementation and ensuring that the constructed features accurately reflect the essential changes in various system faults.

[0011] In one possible implementation, step 2 specifically includes: Step 201: For each dimension of the running parameter time series of each training sample in the training sample database, calculate its mean, standard deviation, maximum value, minimum value and first-order linear trend slope to form the original static feature vector of the training sample. Step 202: Calculate the mean and standard deviation of the original static feature vector; based on the mean and standard deviation, perform Z-score standardization on the original static feature vector to obtain the standardized feature vector; Step 203: Based on the standardized feature vectors of all training samples, the covariance matrix is ​​calculated and eigenvalue decomposition is performed using principal component analysis; the cumulative variance contribution rate of the first L principal components is calculated, and the first L feature vectors whose cumulative variance contribution rate exceeds a preset threshold are selected as column vectors to form a projection matrix; the standardized feature vectors of all training samples are mapped to the L-dimensional principal component space using the projection matrix to obtain the corresponding training low-dimensional feature dataset.

[0012] Compared with existing technologies, the data preprocessing technique that combines time-series statistical feature extraction (mean, standard deviation, extreme values, trend) with principal component analysis for dimensionality reduction solves the problems of high dimensionality, large redundancy, high noise, and low efficiency of directly using raw runtime time-series data for model training.

[0013] In one possible implementation, step 3 specifically includes: Step 301: For K preset fault modes, initialize K independent support vector machine sub-models. ; Step 302, for the first Support Vector Machine Model An asymmetric training sample set is constructed from the training low-dimensional feature dataset and its corresponding label set; the asymmetric training sample set includes a positive sample set. and negative sample set ; The positive sample set Including the tag as the first Low-dimensional feature vector of a single fault: ; In the formula, In the training low-dimensional feature dataset, the first... A low-dimensional feature vector; Indicates the first Binary labels for low-dimensional feature vectors, The Middle The bit is 1; The negative sample set Low-dimensional feature vectors containing labels for normal state and other fault types: ; Step 303, for the support vector machine sub-model Using the corresponding asymmetric training sample set , among which, if , If not, then ; And using radial basis functions as kernel functions, the soft-margin SVM dual optimization problem of the kernel function is solved, with the expression: ; ; ; In the formula, Represents the support vector machine sub-model Total number of samples in the training set , Let represent the Lagrange multiplier, corresponding to each training sample; Indicates the penalty coefficient. Represents the kernel function. ,in, , The kernel width parameter is represented by a grid search and cross-validation method to determine the optimal parameter pair on the training set. The optimal Lagrange multiplier vector is obtained by solving the dual optimization problem of soft-margin SVM using the sequence minimum optimization algorithm. ; Step 304: Obtain the support vector machine sub-model based on the support vectors and their corresponding Lagrange multiplier vectors. Decision function The calculation formula is: ; In the formula, Represents the support vector machine sub-model index set, Indicates the bias term. ,in, To meet Standard support vectors; like Then it is predicted that there exists a first... Single fault; if Then the prediction does not exist. Single fault.

[0014] Compared with existing technologies, the use of positive sample sets and negative sample set The core operation of using an asymmetric training sample set to train each SVM sub-model solves the key technical problem of how to enable each SVM sub-model to learn the ability to decouple concurrent faults from single-fault data. Through the asymmetric strategy, each SVM sub-model is forced to learn to distinguish a specific fault from a mixed negative class (normal + all other faults), thereby enabling its decision boundary to distinguish mixed features in concurrent faults. Combined with radial basis function kernel function optimization, the technical effect of enabling the classifier array to obtain accurate and independent concurrent fault component identification capability by training only with single-fault samples is finally achieved.

[0015] In one possible implementation, step 4 specifically includes: Step 401: By continuously collecting real-time operating data of the solid oxide fuel cell at a fixed frequency, a real-time low-dimensional feature dataset is obtained through preprocessing. Step 402: Simultaneously input the real-time low-dimensional feature dataset into the classifier array that has already loaded K support vector machine sub-models, and each support vector machine sub-model independently calculates the decision function. Output K decision functions ; Step 403: Calculate a binary diagnostic label for determining whether a fault exists using the decision function value; ; Where 1 indicates that the support vector machine sub-model determines the existence of the first... A single fault is defined as 0, where 0 indicates that no fault exists. Step 404: Based on the decision function, calculate the confidence score using the Sigmoid function. The calculation formula is as follows: ; In the formula, This is a scaling parameter used to control the sensitivity of the confidence score as the decision function grows. .

[0016] Compared with existing technologies, the Sigmoid function is used to map the decision function of the support vector machine sub-model to a confidence score. The confidence score quantifies the degree of certainty for each fault diagnosis report, thereby ensuring the stability and reliability of online diagnosis.

[0017] In one possible implementation, step 5 specifically includes: Step 501: Combine the K binary diagnostic labels to generate a K-dimensional fault state vector: ; At the same time, the corresponding confidence scores are combined to form a K-dimensional confidence vector: ; Step 502, if the fault state vector If the number of elements with a median value of 1 is greater than 1, it is determined to be a concurrent fault, based on the fault state vector. and confidence vector Generate a fault diagnosis report.

[0018] Compared with existing technologies, this invention introduces a rule-based concurrent fault diagnosis result verification mechanism and calculates a logical consistency score. It adopts a dual-guarantee technology of data-driven preliminary diagnosis and active verification based on mechanistic knowledge, which solves the technical problems of logical contradictions and high false alarm rates that may occur in pure data-driven models under complex concurrent scenarios. By comparing and quantifying the diagnostic assumptions with physical common sense, it effectively filters out false diagnostic results that are physically impossible or invalid, and significantly improves the reliability and interpretability of the final output of the diagnostic system.

[0019] In one possible implementation, after determining a concurrent failure in step 502, a consistency check is further performed based on a preset rule base; specifically including: Based on the statistical analysis of the mechanism and multidimensional operating sequence parameters of solid oxide fuel cell systems, a rule base is established to describe the logical relationship between fault modes and changes in operating parameters. For fault state vector For each fault mode marked as 1, a logical consistency score is calculated based on a preset rule base and real-time operating data, and the logical consistency score is added to the fault diagnosis report.

[0020] In one possible implementation, step 502 further includes: for the fault state vector that has passed the consistency check Based on the decision function value of its corresponding support vector machine sub-model, the fault state vector is quantified through a pre-set regression model mapping. The severity level of each failure mode.

[0021] Compared with existing technologies, this method, based on fault identification and verification, further solves the technical limitation of traditional diagnostic methods that can only determine whether a fault exists but cannot assess the severity of the fault by introducing a pre-set regression model and using a technique that maps the decision function of the support vector machine sub-model to the fault severity level.

[0022] In one possible implementation, the method for constructing the regression model specifically includes: For the The system identifies different failure modes, obtains sample data under varying severity parameters, processes the data in step 2, and then inputs it into the trained support vector machine sub-model. The decision function is obtained; the absolute value of the decision function is used as input, and the severity level is used as output to fit the regression model. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating an embodiment of this application. Detailed Implementation

[0024] First, those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the embodiments of this application and are not intended to limit the scope of protection of the embodiments of this application. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0025] In the description of the embodiments of this application, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of this application based on the specific circumstances.

[0026] In the embodiments of this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0027] The present application will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0028] See Figure 1As shown in the embodiments, this application discloses an intelligent fault diagnosis method for a solid oxide fuel cell system, aiming to solve the problem of concurrent fault diagnosis caused by the lack of concurrent fault samples and the complexity of high-dimensional data calculation in the prior art. The method of this application realizes high-precision and high-reliability online diagnosis of complex concurrent faults by constructing an intelligent diagnosis framework that integrates data-driven and mechanism knowledge, using only single fault data.

[0029] The implementation of this application will now be described in detail with reference to a specific embodiment; in this embodiment, K=3, that is, three typical single faults are preset: fuel leakage F1, air leakage F2, and electrode delamination F3; The intelligent fault diagnosis method for solid oxide fuel cell systems in this application includes: Step 1: Obtain multi-dimensional runtime sequence parameters of the solid oxide fuel cell system under normal operating conditions and K preset single fault modes, and construct a training sample database; In this embodiment, a sensor network deployed on the solid oxide fuel cell system or a high-fidelity inverse model is used to systematically collect multidimensional operational sequence parameters of the solid oxide fuel cell system under four states: normal operation, fuel leakage only (F1), air leakage only (F2), and electrode delamination only (F3). The label for the normal operation state is... The label for F1 vehicles that only experience fuel leaks is: The label for F2 is only for air leaks. The label for F3, which only involves electrode delamination, is... .

[0030] The collected multidimensional time-series parameters need to comprehensively cover the multi-physics state of the solid oxide fuel cell system. Specifically, the multidimensional operating time-series parameters should include at least voltage and current characteristic parameters, temperature distribution parameters, gas pressure parameters, gas flow rate parameters, and gas composition parameters. The voltage and current characteristic parameters include the total voltage of the fuel cell stack, the output current of the fuel cell stack, the power of the fuel cell stack, the voltage of each individual cell, and the DC internal resistance of the fuel cell stack. The temperature distribution parameters include the anode inlet temperature, anode outlet temperature, cathode inlet temperature, cathode outlet temperature, center temperature of the fuel cell stack, and reformer temperature. The gas pressure parameters include the fuel flow anode inlet pressure, the fuel flow anode outlet pressure, the oxidant flow cathode inlet pressure, the oxidant flow cathode outlet pressure, and the fuel supply pipeline pressure. The gas flow parameters include the anode inlet volume or mass flow rate of fuel gas, the anode outlet flow rate, the cathode inlet flow rate of air or oxygen, and the cathode outlet exhaust flow rate. The gas composition parameters include the hydrogen concentration of the fuel gas at the anode inlet, the hydrogen and carbon monoxide concentrations of the tail gas at the anode outlet, the water vapor concentration at the anode outlet, the oxygen concentration at the cathode outlet, the fuel utilization rate, and the oxidant utilization rate.

[0031] Each parameter is collected at a fixed frequency to form a time series; data segments after steady state or fault stabilization are extracted according to a preset time window and used as training samples to build a training sample database.

[0032] Step 2 involves preprocessing the training samples in the training sample database to obtain a low-dimensional training feature dataset; specifically, this includes: Step 201: For each dimension of the running parameter time series of each training sample in the training sample database, calculate its mean, standard deviation, maximum value, minimum value and first-order linear trend slope to form the original static feature vector of the training sample. Step 202: Calculate the mean and standard deviation of the original static feature vector; based on the mean and standard deviation, perform Z-score standardization on the original static feature vector to obtain the standardized feature vector; Step 203: Based on the standardized feature vectors of all training samples, the covariance matrix is ​​calculated and eigenvalue decomposition is performed using principal component analysis; the cumulative variance contribution rate of the first L principal components is calculated, and the first L feature vectors whose cumulative variance contribution rate exceeds a preset threshold are selected as column vectors to form a projection matrix; the standardized feature vectors of all training samples are mapped to the L-dimensional principal component space using the projection matrix to obtain the corresponding training low-dimensional feature dataset.

[0033] Step 3: Construct a classifier array consisting of K parallel and independent support vector machine (SVM) sub-models, and train each SVM sub-model using an asymmetric strategy. The asymmetric strategy includes constructing an asymmetric training sample set to train the SVM models; specifically: Step 301: Initialize three independent support vector machine sub-models for three preset fault modes. .

[0034] Step 302, for the first Support Vector Machine Model An asymmetric training sample set is constructed from the training low-dimensional feature dataset and its corresponding label set; the asymmetric training sample set includes a positive sample set. and negative sample set This application's embodiment uses the identification of fuel leak only (F1) as an example: The positive sample set Including tags Low-dimensional feature vectors: ; In the formula, Represents the first in a low-dimensional feature dataset A low-dimensional feature vector; Indicates the first Binary labels for a low-dimensional feature vector; The negative sample set Includes tags (Normal state) and tagged as (Only air leak F2 occurred) and Low-dimensional eigenvectors of (only electrode delamination F3 occurs): ; This is used to construct a support vector machine sub-model. Used to learn to distinguish between fuel leaks and other non-fuel leak conditions; Similarly, construct a support vector machine sub-model. and support vector machine sub-model The asymmetric training sample set.

[0035] Step 303, for the support vector machine sub-model Using the corresponding asymmetric training sample set , among which, if , If not, then ; And using radial basis functions as kernel functions, the soft-margin SVM dual optimization problem of the kernel function is solved, with the expression: ; ; ; In the formula, Represents the support vector machine sub-model Total number of samples in the training set , Let represent the Lagrange multiplier, corresponding to each training sample; Indicates the penalty coefficient. Represents the kernel function. ,in, , The kernel width parameter is represented by a grid search and cross-validation method to determine the optimal parameter pair on the training set. The optimal Lagrange multiplier vector is obtained by solving the dual optimization problem of soft-margin SVM using the sequence minimum optimization algorithm. ; Step 304: Obtain the support vector machine sub-model based on the support vectors and their corresponding Lagrange multiplier vectors. Decision function The calculation formula is: ; In the formula, Represents the support vector machine sub-model index set, Indicates the bias term. ,in, To meet Standard support vectors; like Then it is predicted that there exists a first... Single fault; if Then the prediction does not exist. Single fault.

[0036] Step 4 involves collecting real-time operational data from the solid oxide fuel cell system, preprocessing it to obtain a real-time low-dimensional feature dataset, and inputting this dataset into a trained classifier array. Each support vector machine sub-model outputs a binary diagnostic label and a corresponding confidence score. Specifically, this includes: Step 401: Real-time operating data of solid oxide fuel cells are continuously collected at a fixed frequency, and data segments are extracted using a preset time window. The data is then preprocessed to obtain a low-dimensional feature dataset. In this embodiment, the preprocessing in step 401 is performed according to steps 201-203, including normalization and feature dimensionality reduction. Step 402: Simultaneously input the real-time low-dimensional feature dataset into the three loaded support vector machine sub-models. In the classifier array, each support vector machine sub-model independently calculates the decision function. Output K decision functions ; Step 403: Calculate a binary diagnostic label for determining whether a fault exists using the decision function value; ; Where 1 indicates that the support vector machine sub-model determines the existence of the first... A single fault is defined as 0, where 0 indicates that no fault exists. Step 404: Based on the decision function, calculate the confidence score using the Sigmoid function. The calculation formula is as follows: ; In the formula, This is a scaling parameter used to control the sensitivity of the confidence score as the decision function grows. .

[0037] Step 5: Concatenate the binary diagnostic labels into a K-dimensional fault state vector, and concatenate the confidence scores into a K-dimensional confidence vector to generate a fault diagnosis report, which specifically includes: Step 501: Combine the three binary diagnostic labels to generate a 3D fault state vector: ; At the same time, the corresponding confidence scores are combined to form a 3-dimensional confidence vector: ; Step 502, if the fault state vector If the number of elements with a median value of 1 is greater than 1, it is determined to be a concurrent fault, based on the fault state vector. and confidence vector Generate a fault diagnosis report.

[0038] In addition, after determining concurrent faults in step 502 and before generating a fault diagnosis report, the method also includes analyzing the fault state vector based on a preset rule base. Perform consistency checks; specifically including: Based on the statistical analysis of the mechanism and multidimensional operating sequence parameters of solid oxide fuel cell systems, a rule base is established to describe the logical relationship between fault modes and changes in operating parameters. For fault state vector For each fault mode marked as 1, a logical consistency score is calculated based on a preset rule base and real-time operating data, and the logical consistency score is added to the fault diagnosis report.

[0039] This application embodiment constructs a rule base describing the logical relationship between fault modes and changes in operating parameters. Specifically, it includes: based on the electrochemical, thermodynamic, and fluid dynamic mechanisms of solid oxide fuel cell systems, and combined with statistical analysis of the above training samples, summarizing the logical relationships describing strong correlations between faults and symptoms. This application embodiment establishes rules for fault modes (F2) where only air leakage occurs and fault modes (F3) where only electrode delamination occurs, respectively: Rule 1 (for fuel leak only F1): If a fuel leak exists, the anode inlet pressure should show a significant downward trend, i.e. And the stack voltage has a corresponding decrease, that is ; Rule 2 (Air Leakage Only F2): If an air leak exists, the cathode inlet pressure exhibits a significant decreasing trend, i.e. ; Rule 3 (only electrode delamination F3 occurs): If electrode delamination exists, the stack voltage will decrease significantly, i.e. And the pressure change is not significant; that is... ; in, , , , , All are preset thresholds.

[0040] Taking electrode delamination as an example, when a solid oxide fuel cell experiences electrode delamination, fractures perpendicular to the main current path are generated inside the originally smooth electrodes. This causes an insulating barrier layer to form at the fracture site, preventing effective charge transfer. Consequently, the effective active sites for electrochemical reactions are destroyed, and the electrochemical performance of the fuel cell stack significantly decreases. Notably, this failure process directly affects the effective contact area between the anode and cathode of the fuel cell stack for electrochemical reactions, resulting in a substantial reduction in this area.

[0041] Therefore, the simulation of electrode stratification faults in this application is achieved by reducing the effective reaction area between individual stacks. This method of simulating electrode stratification will directly lead to a synchronous increase in stack series connection and polarization resistance.

[0042] Next, for the fault state vector that has passed the consistency check... Based on the decision function value of its corresponding support vector machine sub-model, the fault state vector is quantified through a pre-set regression model mapping. The severity level of each failure mode.

[0043] First, construct a regression model for the th... The system identifies different failure modes, obtains sample data under varying severity parameters, processes the data in step 2, and then inputs it into the trained support vector machine sub-model. The decision function is obtained; the absolute value of the decision function is used as input, and the severity level is used as output to fit the regression model. Taking a fuel leak F1 as an example, different leak severity parameters (such as crack coefficient ω) are set in the simulation to generate corresponding data. After processing in the second step, the data is input into the pre-trained support vector machine sub-model. A series of decision functions are obtained. , fitting Regression relationship with crack coefficient ω This is the severity regression model for F1 when only a fuel leak occurs.

[0044] In addition, new operational data is collected periodically or when the overall confidence level of the diagnostic results is consistently below a preset threshold, and incremental learning algorithms are used to update the principal component analysis projection matrix and / or the support vector machine sub-model.

[0045] In the description of the embodiments of this application, it should be noted that the terms "inner" and "outer" and other terms indicating direction or positional relationship are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and does not indicate or imply that the device or component must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this application.

[0046] In the description of this application, the references to terms such as "an embodiment," "some embodiments," "in this embodiment," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0047] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A smart fault diagnosis method for a solid oxide fuel cell system, characterized in that, include: Step 1: Obtain multi-dimensional runtime sequence parameters of the solid oxide fuel cell system under normal operating conditions and K preset single fault modes, and construct a training sample database; Step 2: Preprocess the training samples in the training sample database to obtain the training low-dimensional feature dataset. Step 3: Construct a classifier array consisting of K parallel and independent support vector machine sub-models, and train each support vector machine sub-model using an asymmetric strategy. The asymmetric strategy includes: based on the training low-dimensional feature dataset, constructing a corresponding asymmetric training sample set for each training support vector machine sub-model for training. Step 4: Collect the real-time operating data of the solid oxide fuel cell system and preprocess it to obtain a real-time low-dimensional feature dataset. Input the real-time low-dimensional feature dataset into the trained classifier array. Each support vector machine sub-model outputs a binary diagnostic label and a corresponding confidence score. Step 5: Concatenate the binary diagnostic labels into a K-dimensional fault state vector, and concatenate the confidence scores into a K-dimensional confidence vector to generate a fault diagnosis report.

2. The intelligent fault diagnosis method for a solid oxide fuel cell system according to claim 1, characterized in that, The multidimensional runtime sequence parameters collected in step 1 include at least one or more of the following: voltage and current characteristic parameters, temperature distribution parameters, gas pressure parameters, gas flow rate parameters, and gas composition parameters. Each parameter is collected at a preset frequency to form an independent time series; and data segments are extracted according to a preset time window as training samples to build a training sample database.

3. The intelligent fault diagnosis method for a solid oxide fuel cell system according to claim 2, characterized in that, The voltage and current characteristic parameters include one or more of the following: total stack voltage, stack output current, stack power, individual cell voltage, and stack DC internal resistance. The temperature distribution parameters include one or more of the following: stack anode inlet temperature, stack anode outlet temperature, stack cathode inlet temperature, stack cathode outlet temperature, stack center temperature, and reformer temperature. The gas pressure parameters include one or more of the following: fuel flow anode inlet pressure, fuel flow anode outlet pressure, oxidant flow cathode inlet pressure, oxidant flow cathode outlet pressure, and fuel supply pipeline pressure. The gas flow parameters include one or more of the following: the anode inlet volume or mass flow rate of fuel gas, the anode outlet flow rate, the cathode inlet flow rate of air or oxygen, and the cathode outlet exhaust flow rate. The gas composition parameters include one or more of the following: hydrogen concentration of fuel gas at the anode inlet, hydrogen concentration and carbon monoxide concentration of tail gas at the anode outlet, water vapor concentration at the anode outlet, oxygen concentration at the cathode outlet, fuel utilization rate, and oxidant utilization rate.

4. The intelligent fault diagnosis method for a solid oxide fuel cell system according to claim 3, characterized in that, Step 2 specifically includes: Step 201: For each dimension of the running parameter time series of each training sample in the training sample database, calculate its mean, standard deviation, maximum value, minimum value and first-order linear trend slope to form the original static feature vector of the training sample. Step 202: Calculate the mean and standard deviation of the original static feature vector; based on the mean and standard deviation, perform Z-score standardization on the original static feature vector to obtain the standardized feature vector; Step 203: Based on the standardized feature vectors of all training samples, the covariance matrix is ​​calculated and eigenvalue decomposition is performed using principal component analysis; the cumulative variance contribution rate of the first L principal components is calculated, and the first L feature vectors whose cumulative variance contribution rate exceeds a preset threshold are selected as column vectors to form a projection matrix; the standardized feature vectors of all training samples are mapped to the L-dimensional principal component space using the projection matrix to obtain the corresponding training low-dimensional feature dataset.

5. The intelligent fault diagnosis method for a solid oxide fuel cell system according to claim 4, characterized in that, Step 3 specifically includes: Step 301: For K preset fault modes, initialize K independent support vector machine sub-models. ; Step 302, for the first Support Vector Machine Model An asymmetric training sample set is constructed from the training low-dimensional feature dataset and its corresponding label set; the asymmetric training sample set includes a positive sample set. and negative sample set; The positive sample set Including the tag as the first Low-dimensional feature vector of a single fault: ； In the formula, In the training low-dimensional feature dataset, the first... A low-dimensional feature vector; Indicates the first Binary labels for low-dimensional feature vectors, The Middle The bit is 1; The negative sample set Low-dimensional feature vectors containing labels for normal state and other fault types: ； Step 303, for the support vector machine sub-model Using the corresponding asymmetric training sample set , among which, if , If not, then ; And using radial basis functions as kernel functions, the soft-margin SVM dual optimization problem of the kernel function is solved, with the expression: ； ； ； In the formula, Represents the support vector machine sub-model Total number of samples in the training set , Let represent the Lagrange multiplier, corresponding to each training sample; Indicates the penalty coefficient. Represents the kernel function. ,in, , The kernel width parameter is represented by a grid search and cross-validation method to determine the optimal parameter pair on the training set. The optimal Lagrange multiplier vector is obtained by solving the dual optimization problem of soft-margin SVM using the sequence minimum optimization algorithm. ; Step 304: Obtain the support vector machine sub-model based on the support vectors and their corresponding Lagrange multiplier vectors. Decision function The calculation formula is: ； In the formula, Represents the support vector machine sub-model index set, Indicates the bias term. ,in, To meet Standard support vectors; like Then it is predicted that there exists a first... Single fault; if Then the prediction does not exist. Single fault.

6. The intelligent fault diagnosis method for a solid oxide fuel cell system according to claim 5, characterized in that, Step 4 specifically includes: Step 401: Real-time low-dimensional feature dataset is obtained by continuously collecting real-time operating data of solid oxide fuel cells at a fixed frequency, followed by normalization and feature dimensionality reduction processing. Step 402: Simultaneously input the real-time low-dimensional feature dataset into the classifier array that has already loaded K support vector machine sub-models, and each support vector machine sub-model independently calculates the decision function. Output K decision functions ; Step 403: Calculate a binary diagnostic label for determining whether a fault exists using the decision function value; ； Where 1 indicates that the support vector machine sub-model determines the existence of the first... A single fault is defined as 0, where 0 indicates that no fault exists. Step 404: Based on the decision function, calculate the confidence score using the Sigmoid function. The calculation formula is as follows: ； In the formula, This is a scaling parameter used to control the sensitivity of the confidence score as the decision function grows. .

7. The intelligent fault diagnosis method for a solid oxide fuel cell system according to claim 6, characterized in that, Step 5 specifically includes: Step 501: Combine the K binary diagnostic labels to generate a K-dimensional fault state vector: ； At the same time, the corresponding confidence scores are combined to form a K-dimensional confidence vector: ； Step 502, if the fault state vector If the number of elements with a median value of 1 is greater than 1, it is determined to be a concurrent fault, based on the fault state vector. and confidence vector Generate a fault diagnosis report.

8. The intelligent fault diagnosis method for a solid oxide fuel cell system according to claim 7, characterized in that, After determining a concurrent fault in step 502, the method further includes analyzing the fault state vector based on a preset rule base. Perform consistency checks; specifically including: Based on the statistical analysis of the mechanism and multidimensional operating sequence parameters of solid oxide fuel cell systems, a rule base is established to describe the logical relationship between fault modes and changes in operating parameters. For fault state vector For each fault mode marked as 1, a logical consistency score is calculated based on the rule base and real-time operational data, and the logical consistency score is added to the fault diagnosis report.

9. The intelligent fault diagnosis method for a solid oxide fuel cell system according to claim 8, characterized in that, Step 502 further includes: for the fault state vector that has passed the consistency check Based on the decision function value of its corresponding support vector machine sub-model, the fault state vector is quantified through a pre-set regression model mapping. The severity level of each failure mode.

10. The intelligent fault diagnosis method for a solid oxide fuel cell system according to claim 9, characterized in that, The specific methods for constructing the regression model include: For the The system identifies different failure modes, obtains sample data under varying severity parameters, processes the data in step 2, and then inputs it into the trained support vector machine sub-model. The decision function is obtained; the absolute value of the decision function is used as input, and the severity level is used as output to fit the regression model.

Images