OS-ELM fuel cell fault diagnosis method based on optimized FCM training

Abstract

The present invention proposes an OS-ELM fuel cell fault diagnosis method based on optimized FCM training, using OS-ELM as a fault diagnosis classification method. In view of the characteristics of OS-ELM that can learn online, a classification algorithm is used to classify the obtained sample set for updating the OS-ELM weight. FCM is a soft classification method, and PSO and GA are used to improve and optimize it to obtain the best classification effect. The OS-ELM algorithm based on optimized FCM training can continuously optimize the effect of fault classification with a small number of initial samples and subsequent continuously obtained samples, so as to achieve the accuracy and rapidity of PEMFC fault diagnosis. 1000 groups of samples are classified by improved FCM and divided into training sets and test sets, and OS-ELM is trained and tested. The training accuracy is 99.33%, the test accuracy is 99.25%, and the use time is 0.45s.

Description

Technical Field

[0001] The present invention belongs to the field of hydrogen fuel cells, and in particular relates to an OS-ELM fuel cell fault diagnosis method based on optimized FCM training. Background Art

[0002] Proton exchange membrane fuel cells (PEMFCs) are favored in transportation, portable power, distributed power generation and other fields due to their high efficiency, high power density, low operating temperature and zero emissions. As a complex nonlinear system involving the coupling of multiple physical domains, many factors affect the health of PEMFC. When in a fault state, the output power, durability and life of PEMFC will be affected. Therefore, how to use effective methods to quickly and accurately judge the operating conditions of PEMFC is crucial to extending the life of PEMFC and its large-scale application in multiple fields.

[0003] The existing fault diagnosis methods can be divided into three categories: model-based, data-driven and experimental test-based. Model-based fault diagnosis requires some understanding of the battery stack structure and operating mechanism, and the establishment of a suitable model according to the needs. According to different modeling methods, it can be divided into mechanism models, semi-empirical models, empirical models and data-driven models.

[0004] Reference 1 (Feng Nenglian, Zhao Qianqian, Yong Jiawang. Fault diagnosis of proton exchange membrane fuel cells based on nonlinear dynamic model [J / OL]. Journal of Chemical Industry:

[0005] 1-18[2022-10-10].http: / / kns.cnki.net / kcms / detail / 11.1946.TQ.20220630.2217.005.html)A sliding film observer is established on the ninth-order state-space model of PEMFC to generate residuals, a fault feature matrix and a fault sensitivity matrix are established, the Euclidean distance between the fault relative fault sensitivity and the theoretical relative fault sensitivity is calculated, and fault identification is performed through the Euclidean distance.

[0006] Reference 2 (Lu Zhongchang, Liu Furong, Yang Yang, Xie Changjun. PEMFC fault classification based on FCM clustering and B0 algorithm [J / OL]. Battery:

[0007] 1-5[2022-10-10].http: / / kns.cnki.net / kcms / detail / 43.1129.tm.20220322.1833.003.html) Randles is used as the equivalent model of the fuel cell, the least squares method is used to fit the equivalent circuit of the experimental data, the feature vector is selected to construct the feature set, the FCM is used to classify the sample set, and 10 groups of data whose membership does not meet the requirements are proposed, and the Bayesian optimization (B0) algorithm is used to divide the data for fault diagnosis.

[0008] Document 3 (C.Lv, H.Wang and Y.Tian, ​​"Fuzzy Interval Observer-Based ModelReference Fault Tolerant Control against Flooding Fault for PEMFC," 2020Chinese Automation Congress (CAC), 2020, pp.5484-5489, doi:

[0009] 10.1109 / CAC51589.2020.9326724) By establishing the Takagi-Sugeno (ts) model of PEMFC, normal operating conditions and water fault conditions can be realized, and water fault detection is realized based on the ts fuzzy interval observer and the introduction of regional pole arrangement. Although the simple PEMFC model can reduce the amount of calculation, it cannot well show the changes inside the stack. Although the complex PEMFC model can well reflect the operating status of the fuel cell, the modeling requirements are high and a detailed understanding of the structure and operating mechanism of PEMFC is required. In order to understand the internal status of the stack as much as possible while reducing the amount of calculation of the system, it is very necessary to select a suitable PEMFC model. Data-driven fault diagnosis uses intelligent algorithms to identify and judge faults based on the input and output information of the stack such as voltage, current, air pressure, and gas humidity.

[0010] Reference 4 (Yuan Tiejiang, Guo Zelin, Fang Tong. PEMFC fault diagnosis based on spatiotemporal characteristics of operation data and stacking ensemble learning [J / OL]. Proceedings of the CSEE:

[0011] 1-10[2022-10-10].http: / / kns.cnki.net / kcms / detail / 11.2107.tm.20220707.1344.003.html) Firstly, the voltage, current, temperature and other sensor data of PEMFC are weighted, and the spatial and temporal features are obtained by using kernel principal component analysis (KPCA) and long short-term memory neural network (LSTM). Then, a spatiotemporal feature set is constructed, and a Stacking ensemble learning framework with base classifier and meta-classifier is established to classify the operating status of PEMFC, which can realize fast and accurate diagnosis of PEMFC system faults.

[0012] Document 5 (R.Ma, H.Dang, R.Xie, L.Xu and D.Zhao, "Online Fault Diagnosis for Open-Cathode PEMFC Systems Based on Output Voltage Measurements and Data-Driven Method," in IEEE Transactions on Transportation Electrification, vol.8, no.2, pp.2050-2061, June 2022, doi:

[0013] 10.1109 / TTE.2021.3114194.) used a PEMFC open cathode fault diagnosis method based only on output voltage measurement. The sample was constructed by extracting feature vectors from the output voltage of each cell using t-distributed random embedding (t-SNE), and the sample was fault diagnosed using extreme gradient boosting (XGBoost). Experiments showed that five health states and fault levels can be effectively identified.

[0014] Document 6 (J.Lu, Y.Gao, L.Zhang, K.Li and C.Yin, "PEMFC water management faultdiagnosis method based on principal component analysis and support vectordata description," IECON 2021-47th Annual Conference of the IEEE Industrial Electronics Society, 2021, pp.1-8, doi:

[0015] 10.1109 / IECON48115.2021.9589931.) A single cell voltage is used as a diagnostic variable, principal component analysis (PCA) is used to extract feature information, and support vector data description (SVDD) is used to construct a hypersphere. Each hypersphere contains feature information, and fault detection is achieved by considering the size of the hypersphere and the distance from the sample to the center of the hypersphere. Data-driven fault diagnosis does not require understanding of the internal structure of the system. Fault diagnosis can be performed well by training the intelligent algorithm using sample data obtained by sensors. However, for the commercialization of PEMFC, an increase in the number of sensors will increase costs, and a large number of training data samples will increase training time and data acquisition issues, which are all unfavorable conditions. Fault diagnosis based on experimental tests uses special methods such as the magnetic field generated by the stack current to perform fault diagnosis. This type of method is currently less used. Summary of the invention

[0016] To solve the above technical problems, the present invention provides an OS-ELM fuel cell fault diagnosis method based on optimized FCM training, which proposes an online sequential extreme learning machine (OS-ELM) algorithm based on optimized fuzzy C-mean clustering algorithm (FCM) training, collects battery stack voltage and current sensor data, performs fast Fourier transform (EFT) analysis, obtains electrochemical impedance spectroscopy (EIS) data, fits it using an improved equivalent circuit, and selects a suitable feature vector to construct a sample set. The sample set is classified using the FCM optimized by particle swarm algorithm (PSO) and genetic algorithm (GA), and OS-ELM is trained using the divided sample set, and its fault diagnosis effect is tested. In order to more intuitively understand how its diagnostic effect is, the same sample is used to train and test other intelligent algorithms, and a comparison is made.

[0017] In order to achieve the above object, the present invention adopts the following technical scheme:

[0018] An OS-ELM fuel cell fault diagnosis method based on optimized FCM training includes the following steps:

[0019] Step 1: Optimize FCM training based on particle swarm algorithm and genetic algorithm; use FCM to divide the existing sample set to obtain the initial optimal cluster center and objective function; initialize two populations, and use particle swarm algorithm and genetic algorithm to optimize the objective function respectively; when the objective function found by particle swarm algorithm and genetic algorithm is better than the initial optimal objective function, update the objective function and continue to find the optimal objective function until the iteration condition is met; at this time, the objective function and cluster center are globally optimal;

[0020] Step 2, fault classification based on OS-ELM: the fault classification is divided into two stages: initialization training and online learning; in the initialization training stage, the sample set is clustered and divided into a test set and a training set using the optimal cluster center obtained in step 1; an initial fault classification identification model is obtained using the training set, and the test set is used to verify the model classification effect; in the online learning stage, external data is collected for diagnosing the current working condition of the proton exchange membrane fuel cell, and the sample data is saved. When the saved sample data meets the requirements, it is classified according to the optimal cluster center and used to update the OS-ELM model;

[0021] Step 3, obtain the OS-ELM fault classification results: use multiple samples for verification, and evenly distribute the samples into four working conditions: normal, membrane dry, flooded and oxygen starvation; divide the samples obtained after clustering into training sample set, online learning sample set and test sample set in a ratio of 3:3:4. The training sample set is used for initial training of the OS-ELM model, the online learning sample set is used to continuously update the model, and the test sample set is used to test the fault diagnosis effect of the model, and the fault classification results of the training set and the test set are obtained.

[0022] Furthermore, the acquisition of the initial optimal cluster center and the objective function in step 1 includes the following steps:

[0023] Step 1.1: Divide the sample set into c classes according to the number of fault types that need to be judged. ij Belonging to each category, μ ij represents the membership between the i-th sample and the j-th class center; where the membership μ ij The following conditions must be met:

[0024] μ ij ∈

[01]

[0025] The objective function of FCM is:

[0026]

[0027] Among them, d ij is the Euclidean distance, which is used to measure the i-th sample x i The distance between the jth distance center and m is the weighting parameter; when the change of the objective function before and after is less than the set value, the iteration of FCM is terminated; at this time, the objective function and cluster center obtained are optimal;

[0028] Step 1.2: Assume that the population number of the particle swarm algorithm and the genetic algorithm is N, and the number of iterations is Iter; set the replication, crossover, and mutation probabilities of the genetic algorithm, determine the variable range, use binary coding, roulette as the selection method, randomly select the initial value of the population, and record the optimal fitness value during the iteration process; set the speed range [V1, V2], position range [X1, X2], learning factors C1, C2 and inertia weight w of the particle swarm algorithm, randomly select the initial speed and initial position of the population, and record the optimal position pbest of each particle and the optimal position gbest of the group during the iteration process; the particle iteration formula is as follows:

[0029]

[0030]

[0031]

[0032] Among them, ω ini is the initial inertia weight, ω end is the inertia weight of the maximum number of iterations, G k is the maximum number of iterations, i.e. Iter, and g is the current number of iterations; and are the velocities of the i-th particle at the k+1 and k iterations, respectively. C1, C2 and w are the learning factor and inertia weight, respectively. r1 and r2 are random values ​​in [0, 1]. and are the individual optimal position and the group optimal position of the i-th particle at the k-th iteration, and are the positions of the i-th particle at k+1 and k iterations respectively.

[0033] During the iteration process, the FCM objective function value is the fitness value and position value of GA and PSO. When the Jm obtained by GA or PSO is better than the current global optimal value, it will be replaced and used as the optimal fitness and optimal position of GA and PSO. Figure 8a and Figure 8b It can be seen that the unoptimized FCM may fall into the local optimum, resulting in some samples not belonging to a certain class well. Using GA and PSO to optimize FCM can effectively avoid FCM from falling into the local optimum, so that each sample belongs to a certain class well.

[0034] Furthermore, the specific steps of the two stages of OS-ELM initialization training and online learning in step 2 are as follows:

[0035] Step 2.1: Initialize the training phase, use sigmoid as the activation function of OS-ELM, and use the training set to initialize the output weight β0 of the single hidden layer feedforward neural network;

[0036] Step 2.2: In the online learning phase, for the externally collected voltage and current data, the impedance spectrum obtained by fast Fourier transform is fitted using an equivalent circuit to obtain the characteristic vectors Rw, Rm and Rp of a sample, R m Used to characterize the water content of proton exchange membrane, R p Used to characterize the charge transfer process, R w Used to characterize the material transfer process; the feature vector is used as the input of OS-ELM to identify whether a fault has occurred and the type of fault; the collected samples are saved, and when the number of samples is certain, FCM based on PSO and GA optimization is used for classification, which is used to update the OS-ELM output weight β; by continuously adjusting the value of the output weight β, the identification effect of the OS-ELM fault classification model is optimized.

[0037] Beneficial effects:

[0038] The present invention focuses on fuel cell fault diagnosis technology and conducts research from three aspects: how to build a suitable PEMFC equivalent model, how to classify faults in sample data, and how to perform fault diagnosis. The present invention uses R m , R p and R w As the characteristic quantity of the sample, it can be used as a good standard for fault classification and diagnosis. Using PSO and GA to optimize the FCM algorithm, the optimal cluster center can be obtained, and the membership matrix value of the sample also has obvious discrimination, which greatly improves the accuracy of sample classification and has a great impact on the quality of the training of the OS-ELM algorithm. When using the OS-ELM algorithm to diagnose the sample fault, a small amount of samples are used for training at the beginning, and new samples are continuously input during the diagnosis process to update the weights to continuously optimize the performance of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is the working principle diagram of PEMFC;

[0040] Figure 2 is the Randles equivalent circuit;

[0041] Figure 3 is the Fouquet equivalent circuit;

[0042] Figure 4 is an improved equivalent circuit model;

[0043] Figure 5 This is the EIS schematic diagram;

[0044] Figure 6 EIS diagrams under different working conditions;

[0045] Figure 7a The clustering diagram is not optimized for FCM;

[0046] Figure 7b Optimize clustering graph for FCM;

[0047] Figure 8a This is the unoptimized membership matrix diagram for FCM;

[0048] Figure 8b This is the membership matrix diagram after FCM optimization;

[0049] Fig. 9 Optimize the flow chart for FCM;

[0050] Fig.10a This is the overall flow chart of fault diagnosis;

[0051] Fig.10b is the classification diagram of the training set;

[0052] Fig.10c Classify the confusion matrix for the training set;

[0053] Fig.11a This is the classification diagram of the test set;

[0054] Fig.11b Test set classification confusion matrix. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0056] like Figure 1 As shown in the figure, the working principle of the proton exchange membrane fuel cell (PEMFC) is the reverse reaction of water electrolysis. When hydrogen reaches the anode catalyst layer, it releases positively charged hydrogen ions and negatively charged electrons. The hydrogen ions pass through the electrolyte to the cathode, and the electrons return to the cathode through the external circuit, generating water at the cathode.

[0057] The electrochemical reaction equation is as follows:

[0058] Anode: H2→2H + +2e -

[0059] cathode:

[0060] Overall reaction:

[0061] Common adverse conditions of PEMFC systems include water flooding and membrane drying caused by improper heat and humidity control of the water thermal management system, and oxygen deficiency caused by improper control of the oxygen (air) supply system. If the above-mentioned failures occur in PEMFC, the performance of the power generation system will deteriorate, the efficiency will decrease, and the reliability and life of the battery will be affected. Therefore, the common failures of the proton exchange membrane fuel cell (PEMFC) include membrane drying failure, water flooding failure and oxygen starvation failure.

[0062] The membrane dry failure is: when the internal temperature of the stack is too high, the drainage speed is too fast, the air intake humidity is too low, etc., the PEMFC is prone to membrane dryness. When the membrane dryness occurs, the proton membrane will become dry, resulting in a decrease in proton conductivity, the conductivity of the membrane will be significantly weakened, and the output voltage will drop. The rate of decrease is related to the severity of the membrane dry failure. When the proton membrane is in a serious dry state for a long time, it may cause irreversible damage to the membrane.

[0063] The flooding fault is: when the temperature inside the stack is too low, the stack current density is too high, the drainage link fails, the air humidity is too high, etc., liquid water is easy to accumulate in the gas channel or electrode, resulting in flooding. Excessive liquid water will block the reactant channel, thereby affecting the normal electrochemical reaction. Since the location where PEMFC generates water is at the cathode, the cathode is more prone to flooding.

[0064] The oxygen starvation fault is: when the air supply system of the fuel cell fails, the fuel cell is prone to oxygen starvation. When the fuel cell is oxygen starved, the reaction of the internal gas cannot meet the load demand, resulting in a significant drop in the cell voltage.

[0065] Since the above-mentioned faults occur inside the PEMFC stack, the only feedback obtained is the reduction of output power and voltage. However, it is impossible to know what kind of fault has occurred in the stack just by the change of voltage, which makes it very difficult to solve the specific fault. Therefore, the research on effective PEMFC fault diagnosis technology is very important.

[0066] Impedance includes resistance, capacitance and inductance. For the PEMFC system, its internal characteristics are mainly impedance and capacitance. Therefore, the series-parallel circuit of resistance and capacitance is often used as its equivalent circuit. Since the reaction of the PEMFC system is mainly concentrated in the cathode, its equivalent circuit model is mainly based on the cathode equivalent circuit. The Randles model is the most practical and typical electrochemical equivalent circuit model, which is often used in the basic analysis of fuel cells. Figure 2As shown. m is the ohmic impedance, R ct is the transfer resistance, C dl is a double layer capacitor.

[0067] Faced with more and more different PEMFC research objects, scholars have proposed more complex equivalent circuit models and used electrochemical characterization elements such as constant phase distribution element (CPE) and Warburg. Warburg is used to characterize the loss associated with diffusion phenomena and the loss associated with mass transfer, and is represented by Zw in the circuit. CPE is used to simulate the capacitance behavior of the electrolyte or electrode interface and the impedance arc deformation caused by the porosity of the electrode.

[0068] Fouquet et al. used CPE to replace the double-layer capacitor and added Warburg to the circuit. The model is the Fouquet model. The structure of the Fouquet model is as follows: Figure 3 shown.

[0069] In order to obtain more information about the internal parameters of the battery and simplify the complexity of the equivalent circuit, use Figure 4 The equivalent circuit is shown.

[0070] The model consists of basic components R, C and CPE. m Used to characterize the water content of proton exchange membrane, R p and CPE are used to characterize the charge transfer process, R w and C w Used to characterize material transport processes.

[0071] The PEMFC system is a complex electrochemical reaction system, so it is difficult to observe the internal operating state of the battery intuitively. Electrochemical impedance spectroscopy (EIS) is an effective tool for studying electrochemical systems and electrochemical reaction mechanisms. It can be used to study and explain electrochemical phenomena based on the equivalent circuit model. The basic principles of EIS are as follows: Figure 5 As shown, an AC current signal is applied to the PEMFC through a DC-DC circuit. By measuring the voltage signal under the current signal, the PEMFC impedance information under the frequency signal can be obtained. The impedance value can be obtained by the following formula:

[0072] I(t)=I m sin(2πf0t+θ)

[0073]

[0074] Among them, I m is the amplitude of the current sine wave, θ is the phase of the current sine wave, f0 is the frequency of the sine wave, U m is the amplitude of the voltage sine wave, is the phase of the voltage sine wave.

[0075] Among them, the above formula is the current signal of the PEMFC system, and the following formula is the PEMFC voltage signal under the frequency current signal. By changing the expressions of the two from time domain to frequency domain, the impedance expression can be obtained as shown in the formula:

[0076] |Z(jw)|=U m / I m

[0077] Z(jw)=|Z(jw)|e j(ψ-θ)

[0078] Among them, Z(jw) is the impedance value of the battery at the current frequency, which is a complex number, Z(jw)| is the impedance amplitude, and ψ-θ is the angle between the impedance and the real axis.

[0079] When the experiment is performed with an excitation signal containing a rich frequency band, comprehensive impedance spectrum information can be obtained, an EIS diagram can be drawn, and the parameters of the components in the equivalent circuit can be obtained through parameter identification.

[0080] Figure 6 The EIS curves under fault and normal conditions under the same conditions show that the EIS curves under different operating conditions of PEMFC are quite different. Compared with normal operating conditions, when a flooding fault occurs, the impedance at high frequency is slightly smaller than that at normal conditions, and the impedance at low frequency is significantly increased compared with normal operating conditions, and the diffusion arc will become larger. The reason why the high-frequency resistance becomes smaller is that the water content of the membrane increases, the membrane impedance decreases, and the reason why the low-frequency resistance increases is that the water content of the membrane increases, resulting in the obstruction of the reactant transmission. When a membrane dry fault occurs, the EIS curve moves to the right as a whole, and both the low-frequency and high-frequency impedances increase. The reason for the increase in high-frequency resistance is that the membrane lacks water, resulting in an increase in membrane impedance. The reason for the increase in low-frequency resistance is that the membrane impedance increases, causing the reaction rate of the stack to decrease. When an oxygen starvation fault occurs, the most obvious feature is that the low-frequency impedance is much larger than the impedance under normal operating conditions, because the material transmission is limited due to insufficient supply of reaction gas.

[0081] The EIS curves of PEMFC under different working conditions are quite different. There are 5 parameters that can be obtained from the equivalent circuit fitted by EIS. Selecting appropriate characteristic quantities can reduce the amount of calculation on the basis of accurately reflecting the operating state of the system. CPE can be considered as a constant value and cannot be used as a sample feature. The material transfer process of PEMFC mainly affects the low-frequency impedance area. On the basis of reducing complexity, Rw is better than Cw. Therefore, Rw, Rm and Rp are selected. R m Used to characterize the water content of proton exchange membrane, R p Used to characterize the charge transfer process, R wIt is used to characterize the material transfer process as the characteristic quantity of the sample.

[0082] The present invention performs fuel cell fault diagnosis based on FCM optimized by PSO and GA, specifically including:

[0083] (1) Fuzzy C-means clustering algorithm (FCM)

[0084] FCM is a popular clustering method that distributes the data to different clusters and then determines the distance between these clusters to divide them. It is a soft classification method.

[0085] FCM divides the given sample set X into c (2≤c≤n) categories, and the cluster center of each category is {v1, v2, ..., v c}, the sample has a certain membership degree μ ij Belong to various categories, among which μ ij The following conditions must be met:

[0086] μ ij ∈

[01]

[0087] The objective function of FCM is:

[0088]

[0089] Among them, d ij is the Euclidean distance, which is used to measure the j-th sample x j The distance between the i-th distance center and m is the weighting parameter.

[0090] (2) Particle Swarm Optimization (PSO) and Genetic Algorithm (GA)

[0091] Both GA and PSO algorithms are algorithms for finding the optimal solution. GA finds the optimal solution by iteratively searching multiple individuals, while PSO finds the optimal solution by iteratively updating the optimal position of each individual and the optimal position of the group.

[0092] The GA algorithm first randomly generates a population, calculates the fitness of the individuals, and then performs replication, crossover, and mutation operations according to probability. The best individual and the optimal solution are output, and the cycle repeats until the iteration ends.

[0093] PSO first initializes the initial velocity and initial position of each particle, calculates the fitness of the particle, obtains the current optimal position, and updates the optimal position of each particle and the optimal position of the group until the iteration ends.

[0094] (3) Comparison of experimental results before and after FCM optimization

[0095] Since FCM is very sensitive to the initial value of clustering and is prone to falling into local minima, PSO and GA are used to optimize FCM to obtain the global minimum. However, when PSO or GA is used alone, premature convergence is prone to occur. Therefore, PSO and GA are combined, and the optimal solution of the two is used as the optimal solution for the entire group to avoid falling into local optimality. In order to verify the feasibility of this method, the experimental results before and after FCM optimization are compared. The results are shown in the figure. Figure 7a , Figure 7b and Figure 8a , Figure 8b shown.

[0096] from Figure 7a , Figure 7b and Figure 8a , Figure 8b It can be seen that when FCM is used alone to cluster samples, some cluster centers are too far away from the optimal cluster center. As a result, the membership matrix values ​​of the samples have no obvious discrimination. Using PSO and GA to optimize FCM, it can be seen that the membership matrix values ​​under the optimized cluster centers have obvious discrimination and can be used to classify fault categories well.

[0097] The flowchart of PSO and GA optimization FCM is as follows: Fig. 9 As shown in the figure: Use FCM to divide the used sample set, obtain the initial optimal cluster center and objective function J, and use them as the initial optimal values ​​of PSO and GA. Initialize two populations X1 and X2, initialize the parameters of PSO and GA; perform GA operation on X1, perform PSO operation on X2, and calculate the objective function J; when the objective function found by GA and PSO is better than the initial optimal objective function, update the objective function and continue to find the optimal objective function until the iteration condition is met. At this time, the objective function and cluster center are globally optimal.

[0098] The following describes in detail the fault classification based on OS-ELM.

[0099] (1) Introduction to OS-ELM algorithm

[0100] Extreme learning machine (ELM) is an algorithm used to train single hidden layer feedforward neural network (SLFN). Traditional SLFN has many application results in pattern recognition, signal processing, short-term prediction and other fields due to its simple structure, fast training speed and high generalization ability

[23] . Because ELM is a batch-based algorithm, it means that in the training stage, it needs to obtain all training data instead of updating online with the arrival of new data. However, in the training process of the model, the training samples often increase gradually with the accumulation of experimental data. In order to meet the requirements of incremental learning, the online sequential extreme learning machine (OS-ELM) algorithm is proposed in the prior art based on ELM. OS-ELM is divided into two parts. The first part is to calculate the initial output weights through a small number of training samples, and the second part is online learning to update the weights through new samples.

[0101] (2) OS-ELM fault classification results

[0102] The overall flow chart of PEMFC fault diagnosis algorithm is as follows Fig.10a As shown: The fault classification based on OS-ELM is divided into two stages: initialization training and online learning. In the initialization training stage, the sample set is clustered and divided into a test set and a training set using the optimal cluster center obtained by the improved FCM. The training set is used to obtain an initial fault classification identification model, and the test set is used to verify the model classification effect. During the online learning stage, external data is collected to diagnose the current working condition of PEMFC, and the sample data is saved. When the saved data meets the requirements, the categories are divided according to the optimal cluster center and used to update the OS-ELM model. First, the sample set is divided by the optimized FCM for training and testing the OS-ELM model. During the online learning stage, external data is collected to diagnose the current working condition of PEMFC, and the sample data is saved. When the saved data meets the requirements, the categories are divided according to the optimal cluster center and used to update the OS-ELM model.

[0103] In order to test the feasibility of the algorithm, 1000 samples were used for verification, and the samples were evenly distributed into four working conditions: normal, membrane dry, flooded, and oxygen starvation. The samples obtained after clustering were divided into training sample sets, online learning sample sets, and test sample sets in a ratio of 3:3:4. The training sample set was used for initial training of the OS-ELM model, the online learning sample set was used to continuously update the model, and the test sample set was used to test the fault diagnosis effect of the model. The fault classification results of the training set and the test set are shown in the figure below. Fig.10b , Fig.10c shown.

[0104] Category 1, 2, 3, and 4 represent normal, membrane dry, flooded, and oxygen starved conditions, respectively. Fig.11a , Fig.11b It can be seen that the diagnostic algorithm in the training set and the test set is completely correct in diagnosing normal and membrane-dry conditions, but has some deviations in diagnosing flooding and oxygen starvation conditions. The overall accuracy of the training set is 99.33%, and the overall accuracy of the test set is 99.25%. The feasibility of the algorithm has been verified.

[0105] (3) Comparison of different algorithms

[0106] In order to verify the diagnostic effect of the algorithm, the same sample data is divided into a training sample set and a test set according to a 6:4 ratio. The support vector machine (SVM) algorithm, the nearest neighbor classification (KNN) algorithm and the BP neural network are used for diagnostic classification. Their fault classification accuracy is compared. The comparison results are shown in the following table.

[0107]

[0108] It can be seen from the table that the accuracy of the OS-ELM algorithm based on optimized FFCM training for fuel cell fault classification is higher than that of the other three algorithms, indicating that the algorithm has certain feasibility and provides a new idea for fuel cell fault diagnosis.

[0109] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An OS-ELM fuel cell fault diagnosis method based on optimized FCM training, characterized in that: The steps include: Step 1: Optimize FCM training based on particle swarm algorithm and genetic algorithm; use FCM to divide the existing sample set to obtain the initial optimal cluster center and objective function; initialize two populations, and use particle swarm algorithm and genetic algorithm to optimize the objective function respectively; when the objective function found by particle swarm algorithm and genetic algorithm is better than the initial optimal objective function, update the objective function and continue to find the optimal objective function until the iteration condition is met; at this time, the objective function and cluster center are globally optimal; Step 2, fault classification based on OS-ELM: the fault classification is divided into two stages: initialization training and online learning; in the initialization training stage, the sample set is clustered and divided into a test set and a training set using the optimal cluster center obtained in step 1; an initial fault classification identification model is obtained using the training set, and the test set is used to verify the model classification effect; in the online learning stage, external data is collected to diagnose the current working condition of the proton exchange membrane fuel cell, and the sample data is saved. When the saved sample data meets the requirements, it is classified according to the optimal cluster center and used to update the OS-ELM model; Step 3, obtain the OS-ELM fault classification results: use multiple samples for verification, and evenly distribute the samples into four working conditions: normal, membrane dry, flooded and oxygen starvation; divide the samples obtained after clustering into training sample set, online learning sample set and test sample set in a ratio of 3:3:

4. The training sample set is used for initial training of the OS-ELM model, the online learning sample set is used to continuously update the model, and the test sample set is used to test the fault diagnosis effect of the model, and the fault classification results of the training set and the test set are obtained.

2. The OS-ELM fuel cell fault diagnosis method based on optimized FCM training according to claim 1 is characterized in that: The acquisition of the initial optimal cluster center and the objective function in step 1 includes the following steps: Step 1.1: Divide the sample set n into c classes according to the number of fault types that need to be judged, and the samples are classified into c classes with a certain membership degree μ. ij Belonging to each category, μ ij represents the membership between the jth sample and the i-th class center; where the membership μ ij The following conditions must be met: The objective function of FCM is: Among them, d ij is the Euclidean distance, which is used to measure the j-th sample x j The distance between the i-th distance center and m is the weighted parameter; iterate continuously, and when the change of the objective function is less than the set value, the iteration of FCM is terminated; at this time, the objective function and cluster center obtained are optimal, and the objective function value is used as the optimal value of the initial population of the particle swarm algorithm and the genetic algorithm; Step 1.2: Assume that the population size of both the particle swarm algorithm and the genetic algorithm is N = 200, and the number of iterations is Iter = 100; set the crossover probability of the genetic algorithm to 0.6, the mutation probability to 0.1, determine the variable range to [0,1], use binary coding, roulette as the selection method, the initial value of the population is randomly selected, and the optimal objective function value and the cluster center value corresponding to the objective function value are recorded during the iteration process; set the speed range of the particle swarm algorithm to [-0.0001, 0.0001], the position range to [0,1], the learning factors c1 and c2 are both 0.2, and the inertia weight w, the initial speed and initial position of the population are randomly selected. During the iteration, the optimal position pbest of each particle and the optimal position gbest of the population are recorded. The optimal position of the population is the optimal objective function value in the current particle swarm algorithm iteration; similarly, the cluster center value corresponding to the objective function value is recorded; the minimum value of the optimal objective function values ​​of the particle swarm algorithm and the genetic algorithm is taken, and compared with the current global optimal objective function value; if it is less than, the objective function is updated, and the optimal objective function is continuously searched until the iteration conditions are met; the objective function and cluster center at the end of the iteration are the global optimal; The particle iteration formula is as follows: Among them, ω ini is the initial inertia weight, with a value of 0.5; ω end is the inertia weight of the maximum number of iterations, with a value of 0.1; G k is the maximum number of iterations, i.e. Iter, and g is the current number of iterations; and are the velocities of the i-th particle at the k+1 and k iterations, c1, c2, and w are the learning factor and inertia weight, r1 and r2 are random values ​​in [0,1], and are the individual optimal position and the group optimal position of the i-th particle at the k-th iteration, and are the positions of the i-th particle at k+1 and k iterations respectively.

3. The OS-ELM fuel cell fault diagnosis method based on optimized FCM training according to claim 2 is characterized in that: The specific steps of the two stages of OS-ELM initialization training and online learning in step 2 are as follows: Step 2.1: Initialize the training phase. The OS-ELM activation function uses sigmoid and the training set is used to initialize the output weights of the single hidden layer feedforward neural network to obtain the fault classification and identification model. Step 2.2: In the online learning phase, for the externally collected voltage and current data, the impedance spectrum obtained by fast Fourier transform is fitted using an equivalent circuit to obtain a sample feature vector R w , R m and R p , R m Used to characterize the water content of proton exchange membrane, R p Used to characterize the charge transfer process, R w Used to characterize the material transfer process; the feature vector is used as the input of OS-ELM to identify whether a fault has occurred and the type of fault; the collected samples are saved, and when the number of samples reaches the quantity requirement, FCM based on PSO and GA optimization is used for classification, which is used to update the output weight of OS-ELM; by continuously adjusting the value of the output weight, the fault classification and identification model is updated to optimize the identification effect of the model.

Images