Fuel cell fault detection method based on data driving

Abstract

The invention discloses a fuel cell fault detection method based on data driving, and the method comprises the steps: 1, collecting the voltage data of a single fuel cell, and constructing a training set and a test set, the training set comprising a first high-dimensional vector, and the test set comprising a second high-dimensional vector; 2, performing dimension reduction processing on the first high-dimensional vector and the second high-dimensional vector by adopting a Fisher discriminant analysis method to obtain a first low-dimensional feature vector and a second low-dimensional feature vector; 3, performing classifier training on the first low-dimensional feature vector by adopting a spherical multi-classification support vector machine, independently constructing a minimum bounding ball for each fault category, and setting a decision threshold for each fault category; and 4, carrying out dimension reduction processing on the vector to be measured to obtain a third low-dimensional feature vector, calculating the distance from the third low-dimensional feature vector to the center of each minimum surrounding ball and the membership score, comparing the maximum value of the membership score with the decision threshold, and judging the fault category of the vector to be measured. According to the invention, early warning and preventive control of faults can be realized, the reliability of the fuel cell is improved, and the service life of the fuel cell is prolonged.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fuel cell, in particular to a data-driven fuel cell fault detection method. BACKGROUND

[0002] Proton exchange membrane fuel cell (PEMFC) system is a complex electrochemical device, its efficient and stable operation depends on the accurate coordination of reaction gas, water, heat management. In actual operation, fuel cell system is prone to various faults, including proton exchange membrane dehydration (membrane dry), cathode waterlogging, reaction gas metering ratio too low, cooling flow channel blockage, etc. If these faults cannot be diagnosed in time and accurately, it will cause the continuous attenuation or irreversible damage of the battery performance.

[0003] At present, fuel cell fault diagnosis mainly adopts the method based on sensor data, which seriously depends on the direct measurement and data of specific physical sensors, by monitoring voltage, temperature, pressure, flow and other parameters, and setting corresponding threshold value for fault judgment. However, this method has obvious limitations: first, its diagnosis ability is limited by the pre-set sensor type and layout, which is difficult to fully capture the complex state of the system; second, the threshold value of single parameter is difficult to accurately distinguish different fault types with similar external performance; in addition, the precision drift and fault of sensor will directly affect the reliability of diagnosis result. Most importantly, the method based on sensor is essentially a post-diagnosis, only when the fault has occurred and significantly affects the sensor measurement value, the alarm will be triggered, which cannot identify and warn in the early stage of fault.

[0004] In contrast, data-driven method provides a new solution for fuel cell fault diagnosis. This kind of method does not depend on accurate physical model or specific sensor configuration, but constructs diagnosis model by mining deep features and rules in historical operation data. Especially in dealing with high-dimensional, nonlinear data, data-driven method shows significant advantages. What's more important is that data-driven method can identify patterns through subtle changes in system running state, and give early warning when fault has not occurred but system has shown abnormal trend, so that the system has the opportunity to avoid the actual occurrence of fault by adjusting the actuator operating conditions, realizing the change from post-processing to pre-prevention.

[0005] However, existing data-driven methods still face some challenges when applied to fuel cell systems. The multi-cell voltage monitoring of fuel cell systems produces high-dimensional data samples, and directly diagnosing in this high-dimensional space can cause a dimension disaster, making it difficult to train a diagnostic model and reducing its generalization ability. In addition, different faults exhibit complex and overlapping distributions in the original high-dimensional voltage space, which is a typical nonlinear separable pattern. Although nonlinear classifiers such as support vector machines can handle such problems, their direct application to high-dimensional raw data can dramatically increase their computational complexity, making it difficult to meet the real-time requirements of controllers.

[0006] The statements herein merely provide background technology related to the present application and do not necessarily constitute the prior art. SUMMARY

[0007] The purpose of the present application is to provide an innovative data-driven fault detection method that can automatically extract the most discriminative features from high-dimensional operating data and establish an accurate and efficient multi-fault classification model, effectively solving the problems of high-dimensional data processing and multi-fault pattern recognition, while maintaining the advantages of data-driven methods and meeting real-time requirements, thereby achieving early warning and preventive control of faults and improving the reliability and service life of fuel cell systems.

[0008] To achieve the above purpose, the present application provides a data-driven fuel cell fault detection method, comprising at least the following steps: Step 1, collecting cell voltage data of a fuel cell under different operating conditions, constructing a training set and a test set, the training set containing a first high-dimensional vector, and the test set containing a second high-dimensional vector; Step 2, using Fisher discriminant analysis to reduce the dimensions of the first and second high-dimensional vectors to first and second low-dimensional feature vectors; Step 3, using a spherical multi-class support vector machine to train a classifier for the first low-dimensional feature vector, independently constructing a minimum enclosing sphere for each fault category in a high-dimensional feature space mapped by a kernel function, and setting a decision threshold for each fault category using the second low-dimensional feature vector; Step 4, using Fisher discriminant analysis to reduce the dimensions of the to-be-tested vector to a third low-dimensional feature vector, calculating the distance from the third low-dimensional feature vector to the center of each minimum enclosing sphere and the membership score, comparing the maximum value of the membership score with the decision threshold, and determining the fault category of the to-be-tested vector.

[0009] Optionally, in step 2, the Fisher discriminant analysis method extracts the most discriminative low-dimensional features by maximizing the inter-class scatter and minimizing the intra-class scatter.

[0010] Optionally, in step 2, the specific steps of the dimension reduction processing include: 1) calculating the mean vector of each category sample and the overall mean vector; 2) constructing the intra-class scatter matrix S w and the inter-class scatter matrix S b ; 3) solving the generalized eigenvalue to obtain the projection vector w i ; wherein λ i is the corresponding eigenvalue of w i ; 4) selecting the projection vectors corresponding to the first L largest generalized eigenvalues to form a projection matrix, wherein L≤C-1, C is the total number of categories; 5) calculating the first low-dimensional feature vector and the second low-dimensional feature vector by the formulas and respectively; wherein Z m is the first low-dimensional feature vector, Z n is the second low-dimensional feature vector, V m is the first high-dimensional vector, V n is the second high-dimensional vector, W is the projection matrix, and T is the transpose of the projection matrix.

[0011] Optionally, in step 3, the specific steps of the classifier training include: 1) for each of the fault categories, obtaining an optimal Lagrange multiplier vector by solving a convex quadratic programming problem with specific constraints; 2) calculating a constant term based on the optimal Lagrange multiplier vector and a kernel function.

[0012] Optionally, in step 3, the calculation formula of the kernel function is , wherein K n,m is the kernel function, Z m is the first low-dimensional feature vector, and Z n is the second low-dimensional feature vector.

[0013] Optionally, in step 3, the setting method of the decision threshold is: inputting the second low-dimensional feature vector into the trained classification model and calculating the membership score for each fault category, arranging the membership scores in each fault category in ascending order, and selecting the arithmetic mean of the first five smallest membership scores from the sequence as the decision threshold of each fault category.

[0014] Optionally, in step 4, the calculation formula of the membership score is ; wherein F i (d i2 ) represents the membership score, d i 2 The distance from the third low-dimensional feature vector to the center of each of the minimum bounding spheres is given.

[0015] Optionally, in step 4, the method for determining the fault category of the vector to be tested is as follows: when the maximum value of the membership score is greater than the decision threshold of a certain fault category, it is classified as that fault category; when the maximum value of the membership score is less than or equal to the decision threshold of all known fault categories, it is classified as a new fault category.

[0016] Optionally, in step 1, the different operating states include normal state, flooded state, low temperature state, and low gas state.

[0017] Optionally, in step 1, the individual unit voltage data comes from measured data of the fuel cell stack or simulation data obtained from running a high-fidelity simulation model.

[0018] Compared with the prior art, the technical solution of the present invention has at least the following beneficial effects: 1) Traditional data-driven methods can only distinguish existing fault types provided by the training set itself, but cannot distinguish new fault types that have not been classified. The fault detection method proposed in this invention constructs a minimum bounding ball model independently for each known fault category during offline training and sets a decision threshold for each category using the test set. During online diagnosis, by calculating the membership score of the test data to each minimum bounding ball and comparing it with the decision threshold, the fault type can be effectively distinguished: when the maximum value of the membership score is greater than the decision threshold of a certain fault category, it is classified into that fault category; when the maximum value of the membership score is less than or equal to the decision threshold of all known fault categories, it is classified into a new fault category. This ensures that new fault categories are not forcibly classified into the existing fault types provided by the training set, thus improving the accuracy and reliability of diagnosis.

[0019] 2) Furthermore, this detection method combines Fisher Discriminant Analysis (FDA) for dimensionality reduction and Spherical Multi-Class Support Vector Machine (SSM-SVM) classification to perform highly sensitive pattern recognition of subtle changes in the system's operating state. FDA dimensionality reduction maximizes inter-class differences and minimizes intra-class differences, making features of different states easier to distinguish. SSM-SVM constructs a minimum bounding sphere by mapping to a high-dimensional space using kernel functions, which can capture difficult-to-identify nonlinear anomaly patterns. It provides early warning when the system shows abnormal trends before a fault occurs, giving the system the opportunity to avoid the actual occurrence of the fault by adjusting actuator operating conditions, thus realizing a shift from post-processing to pre-event prevention.

[0020] 3) Furthermore, this detection method can determine the fault type by relying solely on the individual voltage data collected by the voltage monitoring instrument, reducing the use of multiple dedicated sensors, lowering system complexity and cost, and making it easy to use.

[0021] 4) Furthermore, in the online diagnosis stage, only the distance from the data to be tested to the center of the smallest enclosing sphere and the membership score need to be calculated. The calculation efficiency is high, which can meet the requirements of real-time diagnosis of the controller and solve the contradiction between high-dimensional data processing and real-time performance. Attached Figure Description

[0022] Figure 1 This is a flowchart of the data-driven fuel cell fault detection method of the present invention.

[0023] Figure 2 This is a schematic diagram of the single-cell hydrogen-air fuel cell system of the present invention.

[0024] Attached image labels: 1. Hydrogen inlet; 2. Fuel cell stack; 3. Hydrogen circulation pump; 4. Anode tail gas valve; 5. Blower; 6. Humidifier; 7. Cooling water tank; 8. Cooling fan; 9. Circulation pump; 10. Individual cell voltage monitoring instrument. Detailed Implementation

[0025] The following detailed description, in conjunction with the accompanying drawings and specific embodiments, provides a further detailed explanation of the data-driven fuel cell fault detection method proposed in this invention. The advantages and features of this invention will become clearer from the following description. It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions, used only to facilitate and clarify the illustration of the embodiments of this invention. Please refer to the accompanying drawings to make the objectives, features, and advantages of this invention more apparent and understandable. It should be understood that the structures, proportions, sizes, etc., depicted in the accompanying drawings are only for illustrative purposes to aid those skilled in the art and are not intended to limit the implementation conditions of this invention. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to the size, without affecting the effects and objectives achieved by this invention, should still fall within the scope of the technical content disclosed in this invention.

[0026] Monitoring the voltage of multiple individual cells in a fuel cell system generates high-dimensional data samples. Existing data-driven methods that directly perform diagnosis within this high-dimensional space suffer from the curse of dimensionality, making diagnostic model training difficult and reducing generalization ability. Furthermore, different faults exhibit complex and overlapping distributions in the original high-dimensional voltage space, representing a typical nonlinear separable pattern. Directly applying support vector machines to the high-dimensional raw data drastically increases its computational complexity, making it difficult to meet the real-time requirements of the controller.

[0027] To address the shortcomings of existing technologies, this invention provides an innovative data-driven fault detection method that can automatically extract the most discriminative features from high-dimensional operational data and establish an accurate and efficient multi-fault classification model. This effectively solves the challenges of high-dimensional data processing and multi-fault mode recognition, while maintaining the advantages of data-driven methods and meeting real-time requirements. This enables early warning and preventive control of faults, thereby improving the reliability and service life of fuel cell systems.

[0028] Specifically, such as Figure 1 As shown, the present invention provides a data-driven method for detecting fuel cell faults, comprising at least the following steps: Step 1: Collect single-cell voltage data of fuel cells under different operating conditions, and construct training set and test set. The training set contains a first high-dimensional vector, and the test set contains a second high-dimensional vector.

[0029] In some embodiments, different operating states include normal state, flooded state, cryogenic state, and low-gas state. Individual cell voltage data are derived from measured data of the fuel cell stack or simulation data obtained from running a high-fidelity simulation model.

[0030] In this embodiment, a single-cell hydrogen-air fuel cell system consisting of 10 fuel cell cells connected in series is used (e.g., Figure 2 (As shown). The system operates as follows: Hydrogen enters the anode channel of fuel cell stack 2 through hydrogen inlet 1. Unreacted hydrogen is recovered from the outlet of fuel cell stack 2 and returned to the inlet of fuel cell stack 2 via hydrogen circulation pump 3. Anode gas is discharged from the anode tail gas valve 4. Air is humidified by blower 5 and humidifier 6 before entering the cathode channel of fuel cell stack 2. Hydrogen and oxygen in the air undergo an electrochemical reaction inside fuel cell stack 2 to generate electricity and heat. Cooling water stored in cooling water tank 7 enters the cooling channel inside fuel cell stack 2 through cooling fan 8 and circulation pump 9 to cool fuel cell stack 2. Individual cell voltage monitor 10 is used to monitor the individual cell voltage of fuel cell stack 2 in real time and upload the records.

[0031] During implementation, the system operated for 30 minutes under various conditions: normal state, flooded state (increasing the efficiency of humidifier 6 to increase the moisture content of the intake air), low temperature state (increasing the speed of cooling fan 8 to lower the operating temperature of stack 2 below the calibration temperature), and low air state (reducing the speed of blower 5). In the last 15 minutes of operation under each condition, single-cell voltage data were recorded every 9 seconds, resulting in 100 sets of fuel cell single-cell voltage data for each condition. Fifty sets were randomly selected as the training set, and the remaining 50 sets were used as the test set. The original vectors and FDA projected vectors in the training set are shown in Table 1. Table 1. Original vectors and FDA projection vectors of the training set Step 2: Use Fisher discriminant analysis to reduce the dimensionality of the first high-dimensional vector and the second high-dimensional vector into a first low-dimensional feature vector and a second low-dimensional feature vector.

[0032] Fisher discriminant analysis (FDA) extracts the most discriminative low-dimensional features by maximizing between-class divergence and minimizing within-class divergence, thereby reducing data dimensionality and saving computational resources. The specific steps of dimensionality reduction are as follows: 1) Calculate the mean vector of each category of samples and the overall mean vector. The specific calculation method is as follows: 2) Construct the intraclass scatter matrix S w and the inter-class scatter matrix S b The specific calculation method is as follows: Where T is the transpose matrix.

[0033] 3) Solve for the generalized eigenvalues. The specific calculation method is as follows: in, w i Let λ be the projection vector of the high-dimensional to low-dimensional mapping to be solved. i for w i The corresponding eigenvalues.

[0034] 4) Select the projection vectors corresponding to the first L largest generalized eigenvalues ​​to form a projection matrix, where L≤C-1 and C is the number of categories.

[0035] 5) Through the formula and Calculate the first low-dimensional eigenvector and the second low-dimensional eigenvector respectively; where Z m Z is the first low-dimensional eigenvector. n V is the second low-dimensional feature vector. m V is the first high-dimensional vector. n Let W be the second high-dimensional vector, W be the projection matrix, and T be the transpose of the projection matrix. The first and second low-dimensional feature vectors will be used as inputs to the subsequent classifier.

[0036] Step 3: Use a spherical multi-class support vector machine to train the classifier on the first low-dimensional feature vector. Construct a minimum bounding sphere for each fault category in the high-dimensional feature space of the kernel function mapping, and use the second low-dimensional feature vector to set a decision threshold for each fault category.

[0037] The specific steps for classifier training are as follows: 1) Construct and solve the dual optimization problem, transforming the SSM-SVM process into an optimization problem, using the following method: For category Ω i The classification model consists of Lagrange multiplier vectors α i =[α1 i , α2 i ,…, α n i ] T The vector is represented by the following convex quadratic programming problem: Its constraints are: For z n ∈Ω i c n i = +1, for z n ∉Ω i c n i = -1.

[0038] Q n i = c n i c m i k n , m k n , m This is the kernel function. In this embodiment, a Gaussian kernel is used, and the calculation formula is: 2) Solve the optimization problem in 1). In this embodiment, the quadprog function in MATLAB is used to solve the problem. For each fault category i, a set of optimal Lagrange multiplier vectors α are obtained. i ; 3) Based on the results in 2), calculate the center and radius of each class in the high-dimensional feature space. Explicit calculation of the center and radius for each class is to determine the distance of the data to be tested from the center during real-time operation and compare it with the radius of that class, thereby determining the degree to which the data belongs to that class. This method is computationally intensive. In this embodiment, a kernel function and Lagrange multipliers are used to directly calculate the degree to which the data belongs. To facilitate subsequent calculations, the constant C is calculated in advance in this step. i value: 4) The method for setting the decision threshold is as follows: input the second low-dimensional feature vector into the trained classification model and calculate the membership score for each fault category. Sort the membership scores in each fault category in ascending order, select the first 5 smallest membership scores from the sequence, calculate the arithmetic mean as the decision threshold for each fault category.

[0039] Step 4: Use Fisher discriminant analysis to reduce the dimension of the vector to be tested to a third low-dimensional feature vector. Calculate the distance and membership score of the third low-dimensional feature vector to the center of each minimum bounding sphere. Compare the maximum value of the membership score with the decision threshold to determine the fault category of the vector to be tested.

[0040] The test data is input into the constructed model to determine the fault type. When a new set of individual cell voltage data is collected, it forms the test vector v. test To determine the vector to be tested, v test In this embodiment, the specific steps for determining which state to belong to are as follows: 1) Through the formula Calculate the third low-dimensional eigenvector; where Z test Let V be the vector to be measured, W be the projection matrix, and V be the vector to be measured. test This is the third low-dimensional feature vector.

[0041] 2) Through the formula Calculate the distance from the third low-dimensional eigenvector to the center of each minimum bounding sphere; where d i 2 is the distance from the third low-dimensional eigenvector to the center of each minimum bounding sphere.

[0042] 3) Through the formula Calculate the membership score; where F i (d) i 2 ) represents the membership score.

[0043] The method for determining the fault category of the vector under test is as follows: when the maximum value of the membership score is greater than the decision threshold of a certain fault category, it is classified into that fault category; when the maximum value of the membership score is less than or equal to the decision threshold of all known fault categories, it is classified into a new fault category.

[0044] In summary, the fault detection method provided by this invention can automatically extract the most discriminative features from high-dimensional operational data and establish an accurate and efficient multi-fault classification model, effectively solving the problems of high-dimensional data processing and multi-fault mode recognition. While maintaining the advantages of data-driven methods, it meets real-time requirements, thereby achieving early warning and preventive control of faults and improving the reliability and service life of fuel cell systems.

[0045] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0046] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A data-driven method for detecting fuel cell faults, characterized in that, Include at least the following steps: Step 1: Collect single-cell voltage data of fuel cells under different operating conditions, and construct training set and test set. The training set contains a first high-dimensional vector, and the test set contains a second high-dimensional vector. Step 2: Use Fisher discriminant analysis to reduce the dimensionality of the first high-dimensional vector and the second high-dimensional vector into a first low-dimensional feature vector and a second low-dimensional feature vector. Step 3: Use a spherical multi-class support vector machine to train the classifier on the first low-dimensional feature vector, independently construct a minimum bounding sphere for each fault category in the high-dimensional feature space of the kernel function mapping, and use the second low-dimensional feature vector to set a decision threshold for each fault category; Step 4: Use Fisher discriminant analysis to reduce the dimension of the vector to be tested to a third low-dimensional feature vector. Calculate the distance and membership score of the third low-dimensional feature vector to the center of each minimum bounding sphere. Compare the maximum value of the membership score with the decision threshold to determine the fault category of the vector to be tested.

2. The method as described in claim 1, characterized in that, In step 2, the Fisher discriminant analysis method extracts the most discriminative low-dimensional features by maximizing inter-class divergence and minimizing intra-class divergence.

3. The method as described in claim 2, characterized in that, Step 2, the specific steps of the dimensionality reduction process include: 1) Calculate the mean vector of each category of samples and the overall mean vector; 2) Construct the intraclass scatter matrix S w and the inter-class scatter matrix S b ; 3) Solving for generalized eigenvalues To obtain the projection vector w i ; where λ i For w i The corresponding eigenvalues; 4) Select the projection vectors corresponding to the first L largest generalized eigenvalues ​​to form a projection matrix, where L≤C-1 and C is the total number of categories; 5) Through the formula and Calculate the first low-dimensional feature vector and the second low-dimensional feature vector respectively; where Z m Z is the first low-dimensional feature vector. n V is the second low-dimensional feature vector. m Let V be the first high-dimensional vector. n Let W be the second high-dimensional vector, W be the projection matrix, and T be the transpose of the projection matrix.

4. The method as described in claim 1, characterized in that, Step 3, the specific steps for training the classifier include: 1) For each of the aforementioned fault categories, the optimal Lagrange multiplier vector is obtained by solving a convex quadratic programming problem with specific constraints; 2) Calculate the constant term based on the optimal Lagrange multiplier vector and kernel function.

5. The method as described in claim 1, characterized in that, In step 3, the formula for calculating the kernel function is as follows: , where K n,m Z is the kernel function. m Z is the first low-dimensional feature vector. n This is the second low-dimensional feature vector.

6. The method as described in claim 1, characterized in that, In step 3, the method for setting the decision threshold is as follows: input the second low-dimensional feature vector into the trained classification model and calculate the membership score for each fault category. Sort the membership scores of each fault category in ascending order, and select the first 5 smallest membership scores from the sequence to calculate the arithmetic mean as the decision threshold for each fault category.

7. The method as described in claim 1, characterized in that, In step 4, the formula for calculating the membership score is as follows: Among them, F i (d) i 2 ) represents the membership score, d i 2 The distance from the third low-dimensional feature vector to the center of each of the minimum bounding spheres is given.

8. The method as described in claim 1, characterized in that, In step 4, the method for determining the fault category of the vector to be tested is as follows: when the maximum value of the membership score is greater than the decision threshold of a certain fault category, it is classified into that fault category; when the maximum value of the membership score is less than or equal to the decision threshold of all known fault categories, it is classified into a new fault category.

9. The method as described in claim 1, characterized in that, In step 1, the different operating states include normal state, flooded state, low temperature state, and low gas state.

10. The method as described in claim 1, characterized in that, In step 1, the individual unit voltage data comes from measured data of the fuel cell stack or simulation data obtained from running a high-fidelity simulation model.

Images