A method and system for SOFC system stack performance classification and fault diagnosis

Abstract

The present invention discloses a method and system for grading and diagnosing the performance of a SOFC system stack, and relates to the technical field of fault judgment of the performance degradation level of the stack, including: using the characteristic parameters of the SOFC system such as power, combustion chamber temperature, and stack temperature as data sets to construct a Bayesian artificial neural network structure. When training the model, high-dimensional sample data will lead to a long training time and too complicated calculations. The characteristic variables are sorted by an algorithm to overcome such problems. The number of hidden layer neurons is searched by an algorithm to find the optimal solution for the number of hidden layer neurons. The optimal number of hidden layer neurons is input into the Bayesian artificial neural network model to obtain the best model for fault diagnosis of the performance degradation level of the stack. Predictions are made based on this model to improve the correct diagnosis rate of a single category. The feature selection algorithm is combined with the Bayesian artificial neural network to obtain a better feature combination, effectively improving the overall fault diagnosis efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of fault judgment of battery stack performance degradation level, and more specifically to a method and system for SOFC system battery stack performance grading and fault diagnosis. Background Art

[0002] As people pay more and more attention to the use of clean energy for ships, fuel cells play an increasingly important role in ship propulsion. As a green and clean energy technology, fuel cells are receiving more and more attention. Among them, solid oxide fuel cells (SOFC) show significant environmental advantages with their high efficiency and silent power generation characteristics. Traditional unmanned boat power systems mainly rely on fuel oil, which will cause mechanical noise and emission pollution during the use of fuel oil. SOFC technology can directly convert fuel gas into electrical energy through electrochemical reactions, thereby providing a more efficient and quiet power solution. This can not only reduce carbon emissions during navigation, but also use local water for temperature management of the SOFC system, contributing to environmental protection and the sustainable development of ship energy supply. In addition, some unmanned boats carrying fuel cells have been conducting demonstration voyages, which highlights the widespread demand for fuel cells in unmanned boats. Therefore, in order for SOFC to be effectively applied to unmanned boats, efficient fault diagnosis is crucial.

[0003] Many methods have been applied to fault diagnosis of SOFC systems, such as: data-driven proton exchange membrane fuel cell (PEMFC) diagnostic methods based on deep learning, fault detection methods based on state observers, and diagnostic methods for gas supply fault detection in SOFC systems based on principal component analysis and exponentially weighted moving average control charts. However, fuel cell systems have different fault types and detection methods due to their different stack types. The differences are mainly caused by the significant differences in the operating temperatures of PEMFC and SOFC. From the perspective of fault diagnosis, the diagnostic accuracy of SOFC is lower than that of PEMFC.

[0004] The efficient operation of SOFC systems is essential for energy conversion and environmental protection. However, due to the complexity within the system and the diversity of characteristic variables, such as output power, hydrogen inlet temperature, and combustion chamber temperature, fault diagnosis becomes extremely challenging. In this case, designing an accurate fault diagnosis method is not only crucial to ensure system stability and reliability, but also the key to maximizing energy efficiency and reducing maintenance costs.

[0005] Therefore, how to propose a SOFC system stack performance grading and fault diagnosis method and system, analyze the key features that have an important impact on the fault diagnosis results, improve them using optimization algorithms, improve the accuracy and reliability of faults, and improve the fault diagnosis efficiency of the SOFC system through a graded and quantified approach are issues that technical personnel in this field urgently need to solve. Summary of the invention

[0006] In view of this, the present invention provides a method and system for grading and diagnosing the performance of a SOFC system stack, analyzes the key features that have an important impact on the fault diagnosis results, improves them using an optimization algorithm, improves the accuracy and reliability of faults, and improves the fault diagnosis efficiency of the SOFC system by a graded and quantified manner, thereby promoting the widespread promotion and optimization of the SOFC system in practical applications. In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0007] A method for SOFC system stack performance classification and fault diagnosis, comprising:

[0008] Acquire operating data from the SOFC system, perform data preprocessing, and construct a sample data set;

[0009] Sort the feature variables in the sample data set to obtain the best feature subset;

[0010] A Bayesian artificial neural network is constructed with the best feature subset as input and the fuel cell performance degradation level as output;

[0011] Change the number of input features, record the correct diagnosis rate under different levels of battery stack performance degradation, and determine the best feature combination model;

[0012] Using particle swarm algorithm, the Bayesian artificial neural network of the best feature combination model is optimized to obtain the latest correct diagnosis rate, thus obtaining the best fault diagnosis model;

[0013] Acquire real-time operation data and input it into the optimal fault diagnosis model to obtain fault diagnosis results.

[0014] Optionally, the characteristic variables in the sample data set include: output power, maximum temperature gradient of the combustion chamber, maximum temperature of the combustion chamber, fuel cell stack hydrogen outlet temperature, fuel cell stack hydrogen inlet temperature, fuel cell stack air inlet temperature, fuel cell stack air outlet temperature, flue gas inlet-air temperature, flue gas outlet-air temperature, flue gas inlet-hydrogen temperature, flue gas outlet-hydrogen temperature, hydrogen inlet temperature, hydrogen outlet temperature, air inlet temperature, air outlet temperature, fuel cell current, fuel cell voltage, combustion chamber heater temperature increment, combustion chamber intermediate temperature increment, combustion chamber outlet temperature increment, fuel cell stack air side temperature increment, combustion chamber inlet temperature increment, hydrogen pressure, air flow feedback, combustion chamber outlet temperature, combustion chamber intermediate temperature, combustion chamber inlet temperature, fuel cell stack maximum temperature, fuel cell stack maximum temperature gradient, combustion chamber heater, hydrogen flow feedback, hydrogen flow setting, fuel cell stack air temperature increment and air pressure data.

[0015] Optionally, the data preprocessing includes: performing data processing by data normalization:

[0016] ;

[0017] Among them, X and Y are the data before and after normalization, , are the maximum and minimum values ​​of the data before processing, , are the maximum and minimum values ​​of the processed data, , .

[0018] Optionally, the sorting of the feature variables in the sample data set includes: using the ReliefF-mRMR algorithm to calculate the weights of the feature variables and sorting the features to obtain an M-dimensional optimal feature subset.

[0019] Optionally, constructing the Bayesian artificial neural network includes: setting the number of input layer nodes of the Bayesian artificial neural network to the number M of feature variables in the optimal feature subset, where M is 1 to 10, the number of hidden layer nodes is X, where X can be selected as 6, and the number of output layer nodes is Y, where Y can be selected as 3.

[0020] Optionally, the changing of the number of input features, recording the correct diagnosis rates at different levels of battery stack performance degradation, and determining the optimal feature combination model includes: reducing the feature variables with smaller weights one by one, using the optimal feature variable as input to record the correct diagnosis rates at three levels of battery stack performance degradation: slight degradation, moderate degradation, and severe degradation, and determining the optimal feature combination model.

[0021] Optionally, the use of a particle swarm algorithm to optimize the Bayesian artificial neural network of the best feature combination model includes: using a particle swarm algorithm to search for the number of hidden layer neurons in the Bayesian artificial neural network of the best feature combination model to obtain the optimal number of neurons; inputting the optimal number of neurons in the hidden layer into the Bayesian artificial neural network model for optimization to obtain the latest correct diagnosis rate, thereby obtaining the best fault diagnosis model.

[0022] Optionally, the particle swarm algorithm is used to search for the number of hidden layer neurons of the Bayesian artificial neural network of the best feature combination model, and the optimal number of neurons obtained includes:

[0023] S1: Construct a test set to initialize the inertia factor, the maximum number of iterations, and the minimum error parameter for algorithm termination;

[0024] S2: Calculate the first i The fitness value of each individual in the generation population, where the fitness value of the individual is defined as the overall correct diagnosis rate of the test set;

[0025] S3: Iterative optimization, speed update and population update, update the individual optimality to search for the optimal number of hidden layer neurons;

[0026] S4: Determine whether the particle swarm algorithm meets the termination condition. If so, stop the operation and input the optimal number of hidden layer neurons into the Bayesian artificial neural network. Otherwise, return to step S2.

[0027] Optionally, the changing the number of input features, recording the correct diagnosis rates at different battery stack performance degradation levels, and determining the best feature combination model further includes:

[0028] In order to avoid the contingency caused by a single set of data sets, two other sets of data sets are selected to verify the above-mentioned model to verify that the feature combination has a high accuracy and reliability for fault diagnosis, or multiple sets of data sets are selected to verify the optimal feature combination model;

[0029] The best feature combination model is evaluated, and the evaluation indicators include: overall accuracy and single-unit accuracy.

[0030] Optionally, the iterative optimization, speed update and population update include:

[0031] No. ( k+1 The iterative optimization process of the position of the first generation of particles is:

[0032] ;

[0033] in, k is the number of iterations; λis the inertia weight, and its value decreases linearly from 0.9 to 0.4 as the number of iterations increases; is the position of the (k+1)th generation particle; is the velocity of the (k+1)th generation particle; is the position of the optimal particle of the kth generation individual; is the position of the global optimal particle of the kth generation; and is a random number between 0 and 1, and is the learning factor, and its value is 2.

[0034] Optionally, the best fault diagnosis model is analyzed in terms of running time and number of parameters to analyze the performance of the best fault diagnosis model.

[0035] Optionally, a test data set collected by the SOFC system is obtained, and the characteristic variables in the test data set are used as input parameters and input into an optimal fault diagnosis model for SOFC system stack performance grading and fault diagnosis optimization based on Bayesian artificial neural network and feature selection to obtain the correct diagnosis rate results under different degradation levels of stack performance.

[0036] Optionally, a SOFC system stack performance grading and fault diagnosis system comprises:

[0037] Acquisition module: used to obtain operating data from the SOFC system, perform data preprocessing, and construct a sample data set;

[0038] Feature selection module: used to sort the feature variables in the sample data set to obtain the best feature subset;

[0039] Model building module: used to build a Bayesian artificial neural network with the best feature subset as input and the fuel cell performance degradation level as output;

[0040] Arrangement module: used to change the number of input features, record the correct diagnosis rate under different levels of battery stack performance degradation, and determine the best feature combination model;

[0041] Optimization module: used to use particle swarm algorithm to optimize the Bayesian artificial neural network of the best feature combination model to obtain the latest correct diagnosis rate, thereby obtaining the best fault diagnosis model;

[0042] Diagnostic module: used to obtain real-time operating data and input it into the optimal fault diagnosis model to obtain fault diagnosis results.

[0043] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a method and system for SOFC system stack performance classification and fault diagnosis, which has the following beneficial effects:

[0044] The present invention provides a method and system for grading and diagnosing the performance of a SOFC system stack. The characteristic parameters of the SOFC system such as power, combustion chamber temperature, stack temperature, current, and voltage data are used as data sets to construct a Bayesian artificial neural network structure. When training the model, high-dimensional sample data will lead to a long training time and too complicated calculation. The ReliefF-mRMR algorithm, as a data dimension reduction method, can sort the characteristic variables to overcome such problems. The present invention also takes into account the problem that the correct diagnosis rate of a single category is relatively low. The number of hidden layer neurons is searched through the PSO algorithm to find the optimal solution for the number of hidden layer neurons. The optimal number of hidden layer neurons is input into the Bayesian artificial neural network model, thereby obtaining the best model for SOFC stack performance degradation level fault diagnosis. According to this model, the stack performance degradation level can be predicted. The present invention combines the feature selection algorithm with the Bayesian artificial neural network to find a better feature combination and effectively improve the overall fault diagnosis efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0046] Figure 1 A schematic flow chart of a method for SOFC system stack performance classification and fault diagnosis provided by the present invention;

[0047] Figure 2 A logic flow chart of a fault diagnosis system for a SOFC system provided by the present invention;

[0048] Figure 3 This is a graph of the ReliefF feature sorting results provided by the present invention;

[0049] Figure 4 The mRMR feature ranking result diagram provided by the present invention;

[0050] Figure 5 A curve diagram showing the correct diagnosis rate of the SOFC stack provided by the present invention when the performance is slightly degraded;

[0051] Figure 6 A curve diagram showing the correct diagnosis rate of the SOFC stack under moderate degradation of performance provided by the present invention;

[0052] Figure 7 A curve diagram showing the correct diagnosis rate of the SOFC stack under severe performance degradation provided by the present invention;

[0053] Figure 8 A curve diagram showing the correct diagnosis rate of the overall performance of the SOFC stack provided by the present invention;

[0054] Fig. 9 The confusion matrix diagram of the three best feature combinations provided by the present invention;

[0055] Fig.10 The following are the verification result diagrams of the three best feature combinations under the data set A provided by the present invention;

[0056] Fig.11 The three best feature combination verification result diagrams under the data set B provided by the present invention;

[0057] Fig.12 The minimum error variation curve of the test set provided by the present invention;

[0058] Fig.13 A curve diagram showing the change in the number of neurons in the hidden layer provided by the present invention;

[0059] Fig.14 A curve diagram of the correct diagnosis rate for different numbers of neurons under the optimal number of features provided by the present invention;

[0060] Fig.15 Bar graphs of parameters and operation time under different optimal feature numbers provided by the present invention;

[0061] Fig.16 A structural framework diagram of a SOFC system stack performance grading and fault diagnosis system provided by the present invention. DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0063] The embodiment of the present invention discloses a method for SOFC system stack performance classification and fault diagnosis, such as Figure 1 As shown, including:

[0064] Acquire operating data from the SOFC system, perform data preprocessing, and construct a sample data set;

[0065] Sort the feature variables in the sample data set to obtain the best feature subset;

[0066] A Bayesian artificial neural network is constructed with the best feature subset as input and the fuel cell performance degradation level as output;

[0067] Change the number of input features, record the correct diagnosis rate under different levels of battery stack performance degradation, and determine the best feature combination model;

[0068] Using particle swarm algorithm, the Bayesian artificial neural network of the best feature combination model is optimized to obtain the latest correct diagnosis rate, thus obtaining the best fault diagnosis model;

[0069] Acquire real-time operation data and input it into the optimal fault diagnosis model to obtain fault diagnosis results.

[0070] In a specific implementation, the embodiment of the present invention discloses a method for SOFC system stack performance classification and fault diagnosis, comprising the following steps:

[0071] Step 1, set sensors, collect SOFC prototype operation data from the SOFC system, and perform data preprocessing to form a SOFC prototype operation data set; the SOFC prototype operation data includes: output power, maximum temperature gradient of the combustion chamber, maximum temperature of the combustion chamber, stack hydrogen outlet temperature, stack hydrogen inlet temperature, stack air inlet temperature, stack air outlet temperature, flue gas inlet-air temperature, flue gas outlet-air temperature, flue gas inlet-hydrogen temperature, flue gas outlet-hydrogen temperature, hydrogen inlet temperature, hydrogen outlet temperature, air inlet temperature, air outlet temperature, stack current, stack voltage, combustion chamber heater temperature increment, combustion chamber intermediate temperature increment, combustion chamber outlet temperature increment, stack air side temperature increment, combustion chamber inlet temperature increment, hydrogen pressure, air flow feedback, combustion chamber outlet temperature, combustion chamber intermediate temperature, combustion chamber inlet temperature, stack maximum temperature, stack maximum temperature gradient, combustion chamber heater, hydrogen flow feedback, hydrogen flow setting, stack air temperature increment and air pressure data and other SOFC system characteristic variables.

[0072] In the specific implementation process, the SOFC prototype operation data sampling time interval can be set to 1 minute.

[0073] Furthermore, in step 1, the data preprocessing process includes data normalization, and the formula is as follows:

[0074] ;

[0075] Among them, X and Y are the data before and after normalization, , are the maximum and minimum values ​​of the data before processing, , are the maximum and minimum values ​​of the processed data, , .

[0076] Step 2: According to multiple SOFC system characteristic variables such as output power, stack current and stack voltage of the SOFC prototype operation data set, the ReliefF-mRMR algorithm is used to perform feature sorting to obtain the best feature subset.

[0077] Among them, the ReliefF-mRMR algorithm is used for dimensionality reduction. The specific method is to first use the ReliefF algorithm to select feature variables with weights Q(L) greater than 0.16 to obtain a candidate feature subset, such as Figure 3 As shown in Figure 2, the mRMR algorithm is then used to obtain the 10-dimensional optimal feature subset for fault classification. The specific feature ranking is as follows: Figure 4 shown.

[0078] Step 2.1, using the ReliefF algorithm, assign different weights to features based on the correlation between each feature and the degradation level of the stack performance. The weight of a feature is proportional to its classification contribution. Assuming L is a feature in the data set, its updated weight can be expressed as:

[0079] ;

[0080] in, is the updated weight value, is the weight value before updating, L is a feature in the data set, n is the number of samplings, t represents the tth nearest neighbor sample, and the function Represents the sample U and In Features L The difference between them, p(C) is the proportion of the category, p(class(U)) is the category proportion of a randomly selected sample, express The i-th nearest neighbor sample in , express The numerical features of the i-th nearest neighbor sample in are expressed as follows:

[0081] ;

[0082] in, Representation sample and The difference between the features L, Represents feature L and sample The difference between Represents feature L and sample The difference between them, max(L) is the maximum value of feature L, and min(L) is the minimum value of feature L.

[0083] Step 2.2, using the mRMR algorithm, by calculating the size of the mutual information between the features in the SOFC system and between the features and the fault categories, retain an optimal feature subset for fault diagnosis in the SOFC system, so as to maximize the correlation between the features and the fault categories and minimize the redundancy between the selected features. If α and β are two continuous random variables, and represents the probability corresponding to α and β, and the joint probability is , then the formula for the mutual information (MI) between them is:

[0084] ;

[0085] The formula of mRMR principle is as follows:

[0086] ;

[0087] Where, T(W, C) is the information correlation between feature space W and category C. The maximum mutual information between feature space W and category C is sought to find the feature combination that can best distinguish different categories. R(W) is the difference between features in feature space W. R(W) is minimized to make the features as independent as possible and avoid redundant information. mRMR is the maximum relevance minimum redundancy algorithm. d is the integral operator symbol. C is the target category. W is the feature set. is the number of characteristic quantities; , represents the i-th and j-th features.

[0088] The mRMR principle is denoted as , by using the Mutual Information Quotient (MIQ) standard , the forward sequence search method is used to find the approximate optimal fault feature. Assuming Ω is the original feature set, there are Feature subset , in the remaining features Select the The following characteristics must be met:

[0089] ;

[0090] Step 3, taking the best feature subset of the SOFC data set as input and the fuel cell performance degradation level as output, designing the structure of the Bayesian artificial neural network; setting the number of input layer nodes of the Bayesian artificial neural network to the number of feature variables M in the best feature subset, where M is 1 to 10, the number of hidden layer nodes is X, where X can be selected as 6, and the number of output layer nodes is Y, where Y can be selected as 3.

[0091] Step 4: Change the number of input features, record the correct diagnosis rate under different degradation levels of the battery stack performance, and determine the optimal feature combination model. In order to avoid the randomness brought by a single set of data sets, select the other two sets of data sets to verify the above-obtained optimal feature combination model.

[0092] Step 5: Use the particle swarm algorithm to search for the number of hidden layer neurons of the Bayesian artificial neural network to obtain the optimal number of neurons; input the optimal number of hidden layer neurons into the Bayesian artificial neural network model for optimization to obtain the latest correct diagnosis rate, thereby determining the optimal fault diagnosis model;

[0093] Among them, the PSO optimization algorithm is used, such as Figure 2 As shown in the figure, the specific steps for searching the number of hidden layer neurons of the Bayesian artificial neural network are:

[0094] Step 5.1: Construct a test set to initialize parameters such as the inertia factor, the maximum number of iterations, and the minimum error at algorithm termination, such as Fig.12 As shown;

[0095] Step 5.2: Calculate the first i The fitness value of each individual in the generation population;

[0096] Step 5.3, iterative optimization, continuously updating the speed and population, and continuously updating the individual optimum to search for the optimal number of hidden layer neurons;

[0097] Step 5.4: Determine whether the PSO algorithm meets the termination conditions. If so, stop the operation and input the optimal number of hidden layer neurons into the Bayesian artificial neural network; otherwise, return to step 5.2.

[0098] By constructing a model to identify the three levels simultaneously, the effect after iteration is to satisfy the diagnosis of the three levels simultaneously.

[0099] Furthermore, in step 5.2, the fitness value of the individual is defined as the overall correct diagnosis rate of the SOFC system test set.

[0100] Further, in step 5.3, ( k+1 The iterative optimization process of the position of the first generation of particles is:

[0101] ;

[0102] in, k is the number of iterations; λ is the inertia weight, and its value decreases linearly from 0.9 to 0.4 as the number of iterations increases; is the position of the (k+1)th generation particle; is the velocity of the (k+1)th generation particle; is the position of the optimal particle of the kth generation individual; is the position of the global optimal particle of the kth generation; and is a random number between 0 and 1, and is the learning factor, and its value is 2.

[0103] In a specific embodiment, in step 1, there are 2206 sets of SOFC prototype operation data collected, including variables such as power, hydrogen inlet and outlet temperature, hydrogen pressure, hydrogen flue gas inlet and outlet temperature, etc. The 2206 sets of data are divided into a training set and a test set in a ratio of 3:1;

[0104] The training set data is used to train the Bayesian artificial neural network in the above step 3. The input value during training is the best feature variable of 1544 random data sets, including power, hydrogen inlet and outlet temperature, hydrogen pressure, hydrogen flue gas inlet and outlet temperature and other characteristic parameters. The output value is the fault type corresponding to the 1544 random data sets, that is, the performance degradation level of the fuel cell stack.

[0105] The test set data is used to test and evaluate the generated SOFC system stack performance fault diagnosis model based on Bayesian artificial neural network and feature selection. The input value during the test is the best feature variable of another 662 sets of random data, including output power, hydrogen inlet and outlet temperature, hydrogen pressure, hydrogen flue gas inlet and outlet temperature and other characteristic parameters. The output value is the fault type corresponding to the random 662 sets of data, that is, the performance degradation level of the stack. The fault diagnosis results of the test set are shown in Figure 1. Figure 5 , Figure 6 , Figure 7 and Figure 8 As shown. The model is evaluated based on the test results. The evaluation indicators use two different standard performance evaluation criteria: overall correct diagnosis rate (OCDR) and single correct diagnosis rate (CDR). The higher the value, the higher the correct diagnosis rate, indicating that the model's diagnostic performance is more accurate. The overall correct diagnosis rate and single correct diagnosis rate are defined as:

[0106] ;

[0107] Where TP 11 TP 22 TP 33 Respectively, they represent the number of samples that the model correctly predicts as slightly degraded, moderately degraded, and severely degraded. ij It indicates the number of samples that the model mistakenly predicts to be in state j when they are actually in state i, and TD indicates the number of samples in the overall test set.

[0108] In a specific embodiment, a SOFC system stack performance grading and fault diagnosis system, such as Fig.16 As shown, including:

[0109] Acquisition module: used to obtain operating data from the SOFC system, perform data preprocessing, and construct a sample data set;

[0110] Feature selection module: used to sort the feature variables in the sample data set to obtain the best feature subset;

[0111] Model building module: used to build a Bayesian artificial neural network with the best feature subset as input and the fuel cell performance degradation level as output;

[0112] Arrangement module: used to change the number of input features, record the correct diagnosis rate under different levels of battery stack performance degradation, and determine the best feature combination model;

[0113] Optimization module: used to use particle swarm algorithm to optimize the Bayesian artificial neural network of the best feature combination model to obtain the latest correct diagnosis rate, thereby obtaining the best fault diagnosis model;

[0114] Diagnostic module: used to obtain real-time operation data and input it into the optimal fault diagnosis model to obtain fault diagnosis results. Specifically, the characteristic variables in the real-time operation data are used as input parameters and input into the optimal fault diagnosis model for SOFC system stack performance grading and fault diagnosis optimization based on Bayesian artificial neural network and feature selection to obtain the correct diagnosis rate results under different degradation levels of stack performance.

[0115] In a specific embodiment, the input layer nodes of the Bayesian artificial neural network at the beginning are the best feature number (1 to 10) in the input best feature subset, the hidden layer nodes are 6, and the output layer nodes are 3, and the neural network is trained. When the input features are 3 best features, under 3 different degradation levels, the correct diagnosis rate is low in the state of slight degradation of the battery stack, and it is verified by multiple data sets, such as Fig. 9 , Fig.10 and Fig.11 The present invention uses particle swarm to search the number of hidden layer neurons in the Bayesian artificial neural network to optimize the model. The model effect when the number of neurons is 7 to 16 is recorded as follows: Fig.13 As shown in Figure 2, when the three best feature variables are determined, the correct diagnosis rate of the model constructed by 8 hidden layer neurons is relatively high, as shown in Figure 2. Fig.14 shown.

[0116] In order to better illustrate the performance superiority of the above model, we choose to analyze it in terms of running time and number of parameters. Fig.15The bar graph shows the number of parameters and running time under different feature numbers. At this time, the input layer nodes of the optimized Bayesian artificial neural network are the 3 optimal input feature numbers, the hidden layer nodes are 8, and the output layer nodes are 3.

[0117] The above analysis shows that the present invention proposes a method for SOFC system stack performance classification and fault diagnosis. The experimental results show that compared with the model using all 34 features, the fault diagnosis model using only 3 best features can shorten the training time by 16 seconds while maintaining a high correct diagnosis rate. In terms of model complexity, the number of parameters during model operation is significantly reduced by 67.2%. This finding not only emphasizes the importance of feature selection in fault diagnosis, but also proves the feasibility and effectiveness of the fault diagnosis method developed by the present invention.

[0118] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0119] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for SOFC system stack performance classification and fault diagnosis, characterized in that: include: Acquire operating data from the SOFC system, perform data preprocessing, and construct a sample data set; Sort the feature variables in the sample data set to obtain the best feature subset; A Bayesian artificial neural network is constructed with the best feature subset as input and the fuel cell performance degradation level as output; Change the number of input features, record the correct diagnosis rate under different levels of battery stack performance degradation, and determine the best feature combination model; Using particle swarm algorithm, the Bayesian artificial neural network of the best feature combination model is optimized to obtain the latest correct diagnosis rate, thus obtaining the best fault diagnosis model; The method of using a particle swarm algorithm to optimize the Bayesian artificial neural network of the best feature combination model includes: using a particle swarm algorithm to search for the number of neurons in the hidden layer of the Bayesian artificial neural network of the best feature combination model to obtain the best number of neurons; inputting the best number of neurons in the hidden layer into the Bayesian artificial neural network model for optimization to obtain the latest correct diagnosis rate, thereby obtaining the best fault diagnosis model; The particle swarm algorithm is used to search for the number of hidden layer neurons of the Bayesian artificial neural network of the best feature combination model, and the best number of neurons is obtained, including: S1: Construct a test set to initialize the inertia factor, the maximum number of iterations, and the minimum error parameter for algorithm termination; S2: Calculate the fitness value of each individual in the i-th generation population according to the fitness function, where the fitness value of the individual is defined as the overall correct diagnosis rate of the test set; S3: Iterative optimization, speed update and population update, update the individual optimality to search for the optimal number of hidden layer neurons; S4: Determine whether the particle swarm algorithm meets the termination condition. If so, stop the operation and input the optimal number of hidden layer neurons into the Bayesian artificial neural network. Otherwise, return to step S2. Acquire real-time operation data and input it into the optimal fault diagnosis model to obtain fault diagnosis results.

2. A method for SOFC system stack performance classification and fault diagnosis according to claim 1, characterized in that: The characteristic variables in the sample data set include: output power, maximum temperature gradient of the combustion chamber, maximum temperature of the combustion chamber, fuel cell stack hydrogen outlet temperature, fuel cell stack hydrogen inlet temperature, fuel cell stack air inlet temperature, fuel cell stack air outlet temperature, flue gas inlet-air temperature, flue gas outlet-air temperature, flue gas inlet-hydrogen temperature, flue gas outlet-hydrogen temperature, hydrogen inlet temperature, hydrogen outlet temperature, air inlet temperature, air outlet temperature, fuel cell current, fuel cell voltage, combustion chamber heater temperature increment, combustion chamber intermediate temperature increment, combustion chamber outlet temperature increment, fuel cell stack air side temperature increment, combustion chamber inlet temperature increment, hydrogen pressure, air flow feedback, combustion chamber outlet temperature, combustion chamber intermediate temperature, combustion chamber inlet temperature, fuel cell stack maximum temperature, fuel cell stack maximum temperature gradient, combustion chamber heater, hydrogen flow feedback, hydrogen flow setting, fuel cell stack air temperature increment and air pressure data.

3. A method for SOFC system stack performance classification and fault diagnosis according to claim 1, characterized in that: The data preprocessing includes: performing data processing by data normalization: Among them, X and Y are the data before and after normalization, X max , X min are the maximum and minimum values ​​of the data before processing, Y max , T min is the maximum and minimum value of the processed data, Y max =1,Y min =0.

4. A method for SOFC system stack performance classification and fault diagnosis according to claim 1, characterized in that: The sorting of the characteristic variables in the sample data set includes: using the ReliefF-mRMR algorithm to calculate the weights of the characteristic variables and sorting the characteristics to obtain an M-dimensional optimal characteristic subset.

5. A method for SOFC system stack performance classification and fault diagnosis according to claim 1, characterized in that: The Bayesian artificial neural network construction includes: setting the number of input layer nodes of the Bayesian artificial neural network to the number M of feature variables in the optimal feature subset, the number of hidden layer nodes to X, and the number of output layer nodes to Y.

6. A method for SOFC system stack performance classification and fault diagnosis according to claim 1, characterized in that: The changing the number of input features, recording the correct diagnosis rate under different battery stack performance degradation levels, and determining the best feature combination model also includes: Selecting multiple sets of data sets to verify the optimal feature combination model; The best feature combination model is evaluated, and the evaluation indicators include: overall accuracy and single-unit accuracy.

7. A method for SOFC system stack performance classification and fault diagnosis according to claim 1, characterized in that: The iterative optimization, speed update and population update include: The iterative optimization process for the position of the (k+1) generation particles is: Where k is the number of iterations; λ is the inertia weight, whose value decreases linearly from 0.9 to 0.4 as the number of iterations increases; X k +1 is the position of the (k+1)th generation particle; V k+1 is the velocity of the (k+1)th generation particle; P k is the position of the optimal particle of the kth generation individual; P best k is the position of the global optimal particle of the kth generation; γ1 and γ2 are random numbers between 0 and 1, and ζ1 and ζ2 are learning factors.

8. A SOFC system stack performance grading and fault diagnosis system, using the performance grading and fault diagnosis method of claims 1-7, characterized in that: include: Acquisition module: used to obtain operating data from the SOFC system, perform data preprocessing, and construct a sample data set; Feature selection module: used to sort the feature variables in the sample data set to obtain the best feature subset; Model building module: used to build a Bayesian artificial neural network with the best feature subset as input and the fuel cell performance degradation level as output; Arrangement module: used to change the number of input features, record the correct diagnosis rate under different levels of battery stack performance degradation, and determine the best feature combination model; Optimization module: used to use particle swarm algorithm to optimize the Bayesian artificial neural network of the best feature combination model to obtain the latest correct diagnosis rate, thereby obtaining the best fault diagnosis model; The method of using a particle swarm algorithm to optimize the Bayesian artificial neural network of the best feature combination model includes: using a particle swarm algorithm to search for the number of neurons in the hidden layer of the Bayesian artificial neural network of the best feature combination model to obtain the best number of neurons; inputting the best number of neurons in the hidden layer into the Bayesian artificial neural network model for optimization to obtain the latest correct diagnosis rate, thereby obtaining the best fault diagnosis model; The particle swarm algorithm is used to search for the number of hidden layer neurons of the Bayesian artificial neural network of the best feature combination model, and the best number of neurons is obtained, including: S1: Construct a test set to initialize the inertia factor, the maximum number of iterations, and the minimum error parameter for algorithm termination; S2: Calculate the fitness value of each individual in the i-th generation population according to the fitness function, where the fitness value of the individual is defined as the overall correct diagnosis rate of the test set; S3: Iterative optimization, speed update and population update, update the individual optimality to search for the optimal number of hidden layer neurons; S4: Determine whether the particle swarm algorithm meets the termination condition. If so, stop the operation and input the optimal number of hidden layer neurons into the Bayesian artificial neural network. Otherwise, return to step S2. Diagnostic module: used to obtain real-time operating data and input it into the optimal fault diagnosis model to obtain fault diagnosis results.

Images