CN115799580A - OS-ELM fuel cell fault diagnosis method based on optimized FCM training - Google Patents

Abstract

The invention provides an OS-ELM fuel cell fault diagnosis method based on optimized FCM training, which uses OS-ELM as a fault diagnosis classification method, and classifies an obtained sample set by using a classification algorithm in view of the characteristic that OS-ELM can learn online for updating OS-ELM weight. FCM is used as a soft classification method, and PSO and GA are used for carrying out improved optimization on FCM to obtain the optimal classification effect. By using the OS-ELM algorithm based on the optimized FCM training, the effect of fault classification can be continuously optimized by using a small number of initial samples and subsequent continuously obtained samples, and the accuracy and the rapidity of the fault diagnosis of the PEMFC are realized. 1000 groups of samples are classified through the improved FCM, the samples are divided into a training set and a testing set, and the OS-ELM is trained and tested, wherein the training accuracy is 99.33%, the testing accuracy is 99.25%, and the service time is 0.45s.

Description

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

Technical Field

The invention belongs to the field of hydrogen fuel cells, and particularly relates to an OS-ELM fuel cell fault diagnosis method based on optimized FCM training.

Background

Proton Exchange Membrane Fuel Cells (PEMFCs) are favored in the fields of transportation, portable power sources, distributed power generation, etc. because of their advantages of high efficiency, high power density, low operating temperature, zero emission, etc. PEMFCs, as a complex nonlinear system involving coupling of multiple physical domains, have many factors that affect their health status. When in a fault condition, the output power, durability, and lifetime of the PEMFC are affected. Therefore, how to use an effective method to quickly and accurately judge the operating condition of the PEMFC is very important for prolonging the service life of the PEMFC and large-scale application in multiple fields.

The existing fault diagnosis is divided into three methods, namely a model-based method, a data-driven method and an experimental test method. The fault diagnosis based on the model needs to know the structure and the operation mechanism of the pile, a proper model is established according to the requirement, and the fault diagnosis based on the model can be divided into a mechanism model, a semi-empirical model, an empirical model and a data driving model according to different modeling methods.

Literature 1 (von Nelumbo nucifera, zhaoqian, yongjia. Proton exchange membrane fuel cell fault diagnosis based on nonlinear dynamic model [ J/OL ]. Chemical science report:

1-18[2022-10-10]. Http:// kns.cnki.net/kcms/detail/11.1946.TQ.20220630.2217.005. Html) a slip film observer is established on the PEMFC nine-stage state space model to generate a residual error, a fault characteristic 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 carried out through the Euclidean distance.

Document 2 (luzhongchang, liu lotus, yangyang, xie changjun. PEMFC fault classification based on FCM clustering and B0 algorithm [ J/OL ]. Battery:

1-5[2, 2022-10-10]. Http: and// kns.cnki.net/kcms/detail/43.1129.tm.20220322.1833.003.html) using Randles as an equivalent model of the fuel cell, performing equivalent circuit fitting on experimental data by using a least square method, selecting a feature vector to construct a feature set, performing class division on the sample set by using FCM, providing 10 groups of data with unsatisfied membership requirements, and performing fault diagnosis division on the data by using a Bayesian optimization (B0) algorithm.

Document 3 (C.Lv, H.Wang and Y.Tian, "Fuzzy Interval Observer-Based Model Reference Fault Tolerant Control estimating Fault for PEMFC," 2020Chinese Automation Congress (CAC), 2020, pp.5484-5489, doi:

10.1109/CACC51589.2020.9326724) can realize normal working conditions and water fault working conditions by establishing a Takagi-Sugeno (t-s) model of the PEMFC, and realize water fault detection based on a t-s fuzzy interval observer and introduction of regional pole arrangement. A simple PEMFC model, while reducing the computational effort, does not exhibit well the variations inside the stack. While the complicated PEMFC model can well reflect the operation state of the fuel cell, the modeling requirement is high, and the structure and operation mechanism of the PEMFC need to be known in detail. In order to reduce the calculation amount of the system while knowing the state of the inside of the pile as much as possible, it is necessary to select a proper PEMFC model. The fault diagnosis based on data driving uses an intelligent algorithm to perform identification and judgment of faults according to input and output information of the galvanic pile, such as voltage, current, air pressure, air humidity and the like.

Literature 4 (yuanjianjiang, guolin, fangtong. PEMFC fault diagnosis based on run data space-time features and Stacking ensemble learning [ J/OL ]. Chinese electro-mechanical engineering report:

1-10[ 2-10-10] http: // kns.cnki.net/kcms/detail/11.2107.tm.20220707.1344.003.html) firstly weights sensor data of the PEMFC such as voltage, current, temperature and the like, obtains space and time characteristics by using Kernel Principal Component Analysis (KPCA) and long-and-short time memory neural network (LSTM), constructs a space-time characteristic set, and establishes a Stacking integrated learning framework with a base classifier and a meta classifier to classify the operating state of the PEMFC, so that the quick and accurate diagnosis of the system fault of the PEMFC can be realized.

Document 5 (R.Ma, H.Dang, R.Xie, L.xu and D.ZHao, "on line Fault Diagnosis for Open-Catode PEMFC Systems Based on Output Voltage Measurements and Data-Driven Method," in IEEE Transactions on transfer electronics, vol.8, no.2, pp.2050-2061, june 2022, doi:

10.1109/tte.2021.3114194.) a PEMFC on-cathode fault diagnostic method based only on output voltage measurement was used. And extracting characteristic vectors from the output voltage of each battery by using t-distributed random embedding (t-SNE) to form a sample, and performing fault diagnosis on the sample by adopting extreme gradient boosting (XGboost). Experiments show that five health states and fault degrees can be effectively identified.

Document 6 (J.Lu, Y.Gao, L.Zhang, K.Li and C.Yin, "PEMFC water management fault diagnosis method based on a primary component analysis and sub-vector data description," IECON 2021-47th annular Conference of the IEEE Industrial Electronics society,2021, pp.1-8, doi:

10.1109/iecon48115.2021.9589931.) single cell voltage is used as a diagnostic variable, principal Component Analysis (PCA) is used to extract feature information, support Vector Data Description (SVDD) is used to construct hyper-spheres, each hyper-sphere contains feature information, and fault detection is achieved by considering the size of the hyper-sphere and the distance from the sample to the center of the hyper-sphere. The data-driven fault diagnosis does not need to know the internal structure of the system, and the intelligent algorithm can be well trained by using sample data obtained by the sensor. However, for the commercialization progress of PEMFCs, the cost increases due to the increase in the number of sensors, and the problem of how to acquire data due to the increase in the training time of more training data samples is a disadvantage. The fault diagnosis based on the experimental test is carried out by using special methods such as a magnetic field generated by the current of the galvanic pile, and the like, and the methods are less applied at present.

Disclosure of Invention

In order to solve the technical problems, the invention provides a 0S-ELM fuel cell fault diagnosis method based on optimized fuzzy C-means clustering algorithm (FCM) training, which provides an online sequence overrun learning machine (0S-ELM) algorithm based on optimized fuzzy C-means clustering algorithm (FCM) training, collects stack voltage and current sensor data, performs fast Fourier transform (EFT) analysis to obtain Electrochemical Impedance Spectroscopy (EIS) data, uses an improved equivalent circuit to fit the EIS data, and selects a proper feature vector to construct a sample set. And classifying the sample set by using the FCM optimized by a Particle Swarm Optimization (PSO) and a Genetic Algorithm (GA), training the OS-ELM by using the classified sample set, and testing the fault diagnosis effect. In order to more intuitively understand how the diagnosis effect is, the same sample is used for training and testing other intelligent algorithms and comparing the intelligent algorithms.

In order to achieve the purpose, the invention adopts the following technical scheme:

an OS-ELM fuel cell fault diagnosis method based on optimized FCM training comprises the following steps:

step 1, optimizing FCM training based on a particle swarm algorithm and a genetic algorithm; dividing an existing sample set by using FCM to obtain an initial optimal clustering center and an initial optimal target function; initializing two populations, and optimizing a target function by respectively using a particle swarm algorithm and a genetic algorithm; when the target function searched by the particle swarm algorithm and the genetic algorithm is superior to the initial optimal target function, updating the target function, and continuously searching the optimal target function until an iteration condition is met; at this time, the objective function and the clustering center are globally optimal;

step 2, fault classification is carried out based on OS-ELM: the fault classification comprises two stages of initialization training and online learning; in the initial training stage, clustering and dividing a sample set by using the optimal clustering center obtained in the step 1, and dividing the sample set into a test set and a training set; obtaining an initial fault classification identification model by using the training set, and verifying the classification effect of the model by using the test set; in the on-line learning stage, external data is collected for diagnosing the current working condition of the proton exchange membrane fuel cell, sample data is stored, and when the stored sample data meets the requirement, classification is carried out according to the optimal clustering center and the stored sample data is used for updating the 0S-ELM model;

step 3, obtaining an OS-ELM fault classification result: a plurality of samples are used for verification, and the samples are evenly distributed to four working conditions of normal, dry membrane, water logging and oxygen starvation; and (3) clustering the obtained samples according to the weight ratio of 3:3: and 4, dividing the model into a training sample set, an online learning sample set and a test sample set in proportion, wherein the training sample set is used for initially training the OS-ELM model, the online learning sample set is used for continuously updating the model, the test sample set is used for testing the fault diagnosis effect of the model, and the obtained training and test set fault classification results.

Further, the obtaining of the initial optimal cluster center and the objective function in step 1 includes the following steps:

step 1.1: dividing the sample set into c classes according to the number of fault types to be judged, wherein the samples have a certain membership degree mu _ij Belonging to each class, μ _ij Representing the degree of membership between the ith sample and the jth class center; wherein the degree of membership mu _ij The following conditions are satisfied:

μ _ij ∈[01]

the objective function of FCM is:

wherein d is _ij Is the Euclidean distance, which is used to measure the ith sample x _i And a distance between the jth class of distance centers, m being a weighting parameter; when the change of the objective function is smaller than a set value, terminating the iteration of the FCM; at this time, the obtained objective function and the clustering center are optimal;

step 1.2: assuming that the population numbers of the particle swarm algorithm and the genetic algorithm are both N, and the iteration number is Iter; setting the copy, cross and variation probabilities of a genetic algorithm, determining a variable range, using binary coding, roulette as a selection method, randomly selecting an initial value of a population, and recording an optimal fitness value in an iteration process; setting a speed range [ V1, V2], a position range [ X1, X2], learning factors C1, C2 and an inertia weight w of a particle swarm algorithm, randomly selecting an initial speed and an initial position of a population, and recording an optimal position pbest of each particle and an optimal position gbest of the population in an iteration process; the particle iteration formula is as follows:

wherein, ω is _ini Is the initial inertial weight, ω _end Is the inertial weight iterated to the maximum number, G _k Is the maximum number of iterations, i.e. Iter, g is the current number of iterations;

and

the velocities at the i-th particle iterations k +1 and k, respectively, C1, C2 and w are the learning factor and inertial weight, respectively, r ₁ And r ₂ Is [0,1 ]]Is determined by the random value of (a),

and

the individual optimal position and the population optimal position in the ith particle k iteration are respectively,

and

the position at which the ith particle, k +1 and k, iterates, respectively.

In the iteration process, the FCM objective function value is a fitness value and a position value of the GA and the PSO, and when the Jm obtained by the GA or the PSO is superior to a current global optimal value, the Jm is replaced and used as the optimal fitness and the optimal position of the GA and the PSO. As can be seen from fig. 8a and 8b, there may be cases where the unoptimized FCM falls into local optima, resulting in some samples not belonging to a certain class well. The optimization of the FCM by using the GA and the PSO can well avoid the FCM from being trapped in a locally optimal condition, so that each sample well belongs to a certain class.

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

step 2.1: initializing a training stage, using sigmoid for a 0S-ELM activation function, and initializing output weight beta of a single hidden layer feedforward neural network by using a training set ₀ ；

Step 2.2: in the online learning stage, for voltage and current data acquired externally, an impedance spectrum obtained by fast Fourier transform is used, and an equivalent circuit is used for fitting to obtain characteristic vectors Rw, rm, rp and R of a sample _m For characterizing protonsWater content of exchange membrane, R _p For characterizing the charge transport process, R _w For characterizing the substance transport process; the feature vector is used as the input of the OS-ELM, and whether a fault occurs at present or not and the fault type are identified; saving the collected samples, and when the number of the samples is certain, classifying the samples by using FCM after optimization based on PSO and GA for updating the output weight beta of the OS-ELM; and optimizing the identification effect of the OS-ELM fault classification model by continuously adjusting the value of the output weight beta.

Has the advantages that:

the invention relates to a fuel cell fault diagnosis technology, and researches on three aspects of how to build a proper PEMFC equivalent model, how to carry out fault division on sample data and how to carry out fault diagnosis are carried out _m 、R _p And R _w The characteristic quantity of the sample can be used as a fault dividing and diagnosing standard. The FCM algorithm is optimized by using the PSO and the GA, an optimal clustering center can be obtained, the membership matrix value of the sample also has obvious discrimination, the accuracy of sample classification is greatly improved, and the training quality of the 0S-ELM algorithm is greatly influenced. And (3) carrying out fault diagnosis on the samples by using a 0S-ELM algorithm, training by using a small amount of samples at the beginning, and continuously inputting new samples to update the weight in the diagnosis process so as to continuously optimize the performance of the algorithm.

Drawings

FIG. 1 is a schematic diagram of the operation of a PEMFC;

FIG. 2 is a Randles equivalent circuit;

FIG. 3 is a Fouquet equivalent circuit;

FIG. 4 is a modified equivalent circuit model;

FIG. 5 is a schematic diagram of EIS;

FIG. 6 is an EIS diagram under different working conditions;

FIG. 7a is a FCM unoptimized cluster map;

FIG. 7b is a FCM optimized cluster map;

FIG. 8a is a diagram of FCM unoptimized membership matrices;

FIG. 8b is a diagram of the FCM optimized membership matrix;

FIG. 9 is a FCM optimization flow diagram;

FIG. 10a is an overall flow chart for fault diagnosis;

FIG. 10b is a training set classification diagram;

FIG. 10c is a training set classification confusion matrix;

FIG. 11a is a test set classification diagram;

FIG. 11b test set classification confusion matrix.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

As shown in fig. 1, the operating principle of Proton Exchange Membrane Fuel Cells (PEMFC) is the reverse reaction of the electrolyzed water, hydrogen reaches the anode catalytic layer, which 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 an external circuit, where water is produced.

The electrochemical reaction equation is as follows:

anode: h ₂ →2H ⁺ +2e ^-

Cathode:

and (3) total reaction:

common adverse conditions of the PEMFC system include stack electrode flooding and membrane drying due to improper control of heat and humidity of the water heat management system, and oxygen deficiency due to improper control of the oxygen (air) supply system. If the PEMFC has the above-mentioned failure, the performance of the power generation system is deteriorated, the efficiency is decreased, and the reliability and the life of the battery are affected. Thus, common failures of the Proton Exchange Membrane Fuel Cell (PEMFC) include dry membrane failures, flooding failures, and oxygen starvation failures.

The membrane stem failure is: when the electric pile has the conditions of overhigh internal temperature, overhigh drainage speed, overlow air inlet humidity and the like, the PEMFC is easy to generate a film drying phenomenon. When the membrane is dry, the proton membrane becomes dry, so that the proton conductivity is reduced, the conductivity of the membrane is obviously weakened, the output voltage is reduced, and the reduction rate is related to the severity of the membrane dry failure. When the proton membrane is in a severely dry state for a long period of time, irreversible damage to the membrane may result.

The flooding fault is as follows: when the temperature in the galvanic pile is too low, the current density of the galvanic pile is too high, the drainage link is broken down, the humidity of the air is too high and the like, liquid water is easy to gather in the gas channel or the electrode to cause flooding. The excessive liquid water may block the reactant passage, thereby affecting the malfunction of the electrochemical reaction in normal operation. Because the PEMFC produces water at the cathode, flooding occurs more easily at the cathode.

The oxygen starvation fault is as follows: when the air supply system of the electric pile is in failure, the electric pile is easy to have oxygen starvation failure. When the fuel cell is starved by oxygen, the reaction of internal gas cannot meet the requirement of a load, so that the voltage of a cell node is greatly reduced.

Since several of the above faults occur inside the PEMFC stack, the resulting feedback is only a reduction in output power and a reduction in voltage. However, it is not known what kind of fault occurs in the stack just by the voltage change, which causes great difficulty in specifically solving the specific fault. Therefore, the research of effective PEMFC fault diagnosis technology is crucial.

The impedance includes resistance, capacitance and inductance. For PEMFC systems, their internal characteristics are dominated by impedance and capacitance, and therefore, often only a series-parallel circuit of a resistor and a capacitor is used as its equivalent circuit. Since the PEMFC system reaction is mainly focused on the cathode, the equivalent circuit model thereof is mainly based on the cathode equivalent circuit. Randles model is the most practical and typical electrochemical equivalent circuit modelIt is commonly used for basic analysis of fuel cells. The Randles model is shown in figure 2. Wherein R is _m Is ohmic impedance, R _ct To transfer resistance, C _dl Is a double-layer capacitor.

In the face of more diverse PEMFC research subjects, scholars propose more complex equivalent circuit models and use electrochemical characterization elements such as constant phase distribution elements (CPE), walburg (Warburg), etc. Warburg is used to characterize the losses associated with the diffusion phenomenon and the losses associated with mass transport, represented in the circuit by Zw. CPE was used to simulate the resistive arc deformation caused by the capacitive behavior of the electrolyte or electrode interface and the porosity of the electrode.

Fouquet et al used CPE instead of double layer capacitors and added Warburg to the circuit, which is a Fouquet model whose structure is shown in FIG. 3.

In order to obtain more abundant information of the internal parameters of the battery and simplify the complexity of the equivalent circuit, the equivalent circuit shown in fig. 4 is used.

The model consists of the basic elements R, C and CPE. R _m Is used for characterizing the water content, R, of the proton exchange membrane _p And CPE is used to characterize the charge transport process, R _w And C _w For characterizing the substance transport process.

The PEMFC system is a complicated electrochemical reaction system, so that it is difficult to visually observe the operation state of the inside of the cell. An Electrochemical Impedance Spectroscopy (EIS) is an effective tool for researching an electrochemical system and an electrochemical reaction mechanism, and an electrochemical phenomenon can be well researched and explained by means of an equivalent circuit model. The basic principle of EIS is shown in fig. 5, an alternating current signal is applied to the PEMFC through the DC-DC circuit, and impedance information of the PEMFC at the frequency signal can be obtained by measuring a voltage signal under the current signal, and the impedance value can be obtained by the following formula:

I(t)＝I _m sin(2πf ₀ t+θ)

wherein, I _m Is the amplitude of the current sine wave, θ is the phase of the current sine wave, f ₀ Frequency of sine wave, U _m Is the amplitude of the voltage sine wave,

is the phase of the voltage sine wave.

The above formula is the current signal of the PEMFC system, the following formula is the PEMFC voltage signal under the frequency current signal, the expression of the two is changed from time domain to frequency domain, and the expression of the obtained impedance is shown as the following formula:

|Z(jw)|＝U _m /I _m

Z(jw)＝|Z(jw)|e ^j(ψ-θ)

wherein Z (jw) is an impedance value of the battery at the current frequency, and is a complex number, Z (jw) | is an impedance amplitude, and ψ - θ is an angle between the impedance and the real axis.

When the experiment is performed through the action of excitation signals containing rich frequency bands, comprehensive impedance spectrum information can be obtained, an EIS (impedance enhanced spectroscopy) graph is drawn, and parameters of elements in the equivalent circuit can be obtained through parameter identification.

FIG. 6 is an EIS curve chart under the same condition, and it can be seen that the EIS curves of the PEMFC under different conditions are greatly different. Compared with the normal working condition, when the water flooding fault occurs, the impedance at high frequency is slightly smaller than that at normal frequency, the impedance at low frequency is obviously increased compared with the normal working condition, and the diffusion arc is enlarged. The reason that the high-frequency resistance becomes small is that the water content of the membrane is increased, the impedance of the membrane is reduced, and the transmission of reactants is blocked due to flooding because of the large low-frequency resistance. When the dry membrane fault occurs, the EIS curve integrally moves to the right, and the low-frequency impedance and the high-frequency impedance become larger. The reason why the high frequency resistance becomes large is that the film lacks water, resulting in an increase in film resistance. The reason why the low frequency resistance becomes large is that the reactor reaction rate decreases due to the increase in the membrane impedance. In the event of an oxygen starvation fault, the most obvious feature is that the low frequency impedance is much greater than the impedance under normal operating conditions due to the limited mass transport caused by the insufficient supply of reactant gas.

EIS curve of PEMFC under different working conditionsThe difference is large, 5 parameters can be obtained by an equivalent circuit fitted by the EIS, and the calculation amount can be reduced by selecting proper characteristic amount on the basis of accurately reflecting the operation state of the system. CPE can be considered a constant value and cannot be characterized as a sample. The PEMFC mass transfer process mainly influences a low-frequency impedance area, and Rw is better than Cw on the basis of reducing complexity. Therefore, rw, rm and Rp are selected for use. R _m Is used for characterizing the water content, R, of the proton exchange membrane _p For characterizing the charge transport process, R _w Is used for characterizing the substance transport process as a characteristic quantity of the sample.

The invention carries out fuel cell fault diagnosis based on FCM optimized by PSO and GA, and specifically comprises the following steps:

(1) Fuzzy C mean value clustering algorithm (FCM)

FCM is a popular clustering method, which assigns the data to different clusters, and then determines the distance between the clusters to classify the data, and belongs to soft classification.

FCM divides a given sample set X into classes c (2 ≦ c ≦ n) with the cluster center for each class { v ≦ n ≦ v ≦ m ₁ ，v ₂ ，...，v _c With a certain membership degree mu of the sample _ij Belonging to various classes, wherein _ij The following conditions are satisfied:

μ _ij ∈[01]

the objective function of FCM is:

wherein, d _ij Is the Euclidean distance, which is used to measure the jth sample x _j And the distance between the ith class and the center, m being a weighting parameter.

(2) Particle Swarm (PSO) and Genetic (GA) algorithms

The GA and PSO algorithms are both algorithms for finding the optimal solution, the GA is found by iterative search of a plurality of individuals, and the PSO is found by iteratively updating the optimal positions of the individuals and the optimal positions of the population.

The GA algorithm firstly generates a population at random, calculates the fitness of individuals, and then carries out copying, crossing and mutation operations according to the probability. And outputting the optimal individual and the optimal solution, and repeating the steps until the iteration is finished.

The PSO firstly initializes the initial speed and the initial position of each particle, calculates the fitness of the particles to obtain the current optimal position, and updates the optimal position of each particle and the optimal position of the group until the iteration is finished.

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

Because the FCM is very sensitive to the initial clustering value and is easy to fall into a local minimum value, the FCM is optimized by using PSO and GA to obtain a global minimum value. However, when the PSO or the GA is used alone, the premature phenomenon is likely to occur, so that the PSO and the GA are combined, and the optimal solution of the PSO and the GA is used as the optimal solution of the whole population, so that the PSO and the GA are prevented from falling into local optimal solution. To verify the feasibility of this method, the results of the experiment before and after FCM optimization were compared and shown in fig. 7a, 7b and 8a, 8 b.

As can be seen from fig. 7a, 7b and 8a, 8b, when clustering samples using FCM alone, it may happen that some cluster centers are too far from their optimal cluster center positions. Resulting in no significant discrimination of the membership matrix values of the samples. The FCM is optimized by using the PSO and the GA, and it can be seen that the membership matrix value under the clustering center obtained after optimization has obvious discrimination, and fault categories can be well divided.

The flow chart of PSO and GA for optimizing FCM is shown in FIG. 9: and (4) dividing the used sample set by using FCM to obtain an initial optimal clustering center and an objective function J, and taking the initial optimal clustering center and the objective function J as initial optimal values of PSO and GA. Initializing two populations X1 and X2, and initializing parameters of PSO and GA; performing GA operation on the X1, performing PSO operation on the X2, and calculating an objective function J; and when the target function searched by the GA and the PSO is superior to the initial optimal target function, updating the target function, and continuously searching the optimal target function until an iteration condition is met. At this time, the objective function and the cluster center are globally optimal.

The OS-ELM based fault classification is described in detail below.

(1) Introduction of OS-ELM Algorithm

An Extreme Learning Machine (ELM) is an algorithm applied to training a Single hidden Layer Feedforward neural Network (SLFN), and the traditional SLFN has many application achievements in the fields of pattern recognition, signal processing, short-term prediction and the like due to the characteristics of simple structure, high training speed, higher generalization capability and the like [23]. Since ELM is a batch based algorithm, this means that in the training phase, it needs to obtain all training data, rather than updating online with the arrival of new data, but in the training process of the model, it often happens that training samples gradually increase with the accumulation of experimental data. In order to meet the requirement of incremental learning, an online sequential ultralimit learning machine (0S-ELM) algorithm is proposed on the basis of ELM in the prior art. The OS-ELM is divided into two parts, the first part is to calculate the initial output weight through a small number of training samples, and the second part is to learn online and update the weight through a new sample.

(2) OS-ELM fault classification results

The whole flow chart of the PEMFC fault diagnosis algorithm is shown in FIG. 10 a: the fault classification based on the OS-ELM is divided into two stages of initial training and online learning. In the initial training stage, the optimal clustering center obtained by the improved FCM is used for carrying out clustering division on the sample set, and the sample set is divided into a test set and a training set. And obtaining an initial fault classification identification model by using the training set, and verifying the classification effect of the model by using the test set. And in the process of the on-line learning stage, acquiring external data for diagnosing the current working condition of the PEMFC, storing the sample data, and when the stored data meet the requirements, performing class division according to the optimal clustering center and updating the 0S-ELM model. First, a sample set is divided through the optimized FCM for training and testing the OS-ELM model. And in the process of the on-line learning stage, acquiring external data for diagnosing the current working condition of the PEMFC, storing the sample data, and when the stored data meet the requirements, performing class division according to the optimal clustering center and updating the OS-ELM model.

To test the feasibility of the algorithm, 1000 samples were used for validation, and the samples were equally assigned to four conditions, normal, dry membrane, flooded and oxygen starved. And (3) clustering the obtained samples according to the weight ratio of 3:3: the 4-bit column is divided into a training sample set, an online learning sample set and a testing sample set, the training sample set is used for initially training the OS-ELM model, the online learning sample set is used for continuously updating the model, the testing sample set is used for testing the fault diagnosis effect of the model, and the obtained training set and testing set fault classification results are shown in fig. 10b and fig. 10 c.

The categories 1, 2, 3 and 4 represent normal, dry membrane, flooding and oxygen starvation working conditions respectively, and as can be seen from fig. 11a and 11b, the training set and test set centralized diagnosis algorithm diagnoses the normal and dry membrane working conditions completely and correctly, diagnoses the flooding and oxygen starvation working conditions slightly deviate, the overall accuracy of the training set is 99.33%, the overall accuracy of the test set is 99.25%, and the feasibility of the algorithm is verified.

(3) Comparison of different algorithms

In order to verify the diagnostic effect of the algorithm, the same sample data is divided into a training sample set and a testing machine according to the ratio of 6: 4, a Support Vector Machine (SVM) algorithm, a nearest neighbor classification (KNN) algorithm and a BP neural network are used for diagnosis and classification, the accuracy of fault classification is compared, and the comparison result is shown in the following table.

As can be seen from the table, the accuracy of the OS-ELM algorithm trained on the basis of the optimized FFCM to the fault classification of the fuel cell is higher than that of the other three algorithms, so that the algorithm has certain feasibility, and a new thought is provided for fault diagnosis of the fuel cell.

It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims (3)

1. An OS-ELM fuel cell fault diagnosis method based on optimized FCM training is characterized by comprising the following steps:

step 1, optimizing FCM training based on a particle swarm algorithm and a genetic algorithm; dividing an existing sample set by using FCM to obtain an initial optimal clustering center and an initial optimal target function; initializing two populations, and optimizing a target function by respectively using a particle swarm algorithm and a genetic algorithm; when the target function searched by the particle swarm algorithm and the genetic algorithm is superior to the initial optimal target function, updating the target function, and continuously searching the optimal target function until an iteration condition is met; at the moment, the objective function and the clustering center are globally optimal;

step 2, fault classification is carried out based on OS-ELM: the fault classification comprises two stages of initialization training and online learning; in the initial training stage, clustering and dividing a sample set by using the optimal clustering center obtained in the step 1, and dividing the sample set into a test set and a training set; obtaining an initial fault classification identification model by using the training set, and verifying the classification effect of the model by using the test set; in the on-line learning stage, external data is collected for diagnosing the current working condition of the proton exchange membrane fuel cell, sample data is stored, and when the stored sample data meets the requirement, classification is carried out according to the optimal clustering center and the stored sample data is used for updating the OS-ELM model;

step 3, obtaining an OS-ELM fault classification result: a plurality of samples are used for verification, and the samples are evenly distributed to four working conditions of normal, dry membrane, water logging and oxygen starvation; and (3) clustering the obtained samples according to the weight ratio of 3:3: and 4, dividing the model into a training sample set, an online learning sample set and a test sample set in proportion, wherein the training sample set is used for initially training the OS-ELM model, the online learning sample set is used for continuously updating the model, the test sample set is used for testing the fault diagnosis effect of the model, and the obtained training and test set fault classification results.

2. The OS-ELM fuel cell fault diagnosis method based on optimized FCM training of claim 1, wherein the obtaining of the initial optimal cluster center and objective function in step 1 comprises the following steps:

step 1.1: dividing a sample set n into c classes according to the number of fault types to be judged, wherein the samples have a certain membership degree mu _ij Belonging to each class, μ _ij Representing the degree of membership between the jth sample and the ith class center; wherein the degree of membership mu _ij The following conditions are satisfied:

the objective function of FCM is:

wherein, d _ij Is the Euclidean distance, which is used to measure the jth sample x _j And the distance between the class i distance centers, m being a weighting parameter; continuously iterating, and when the change of the target function is smaller than a set value, terminating the iteration of the FCM; at the moment, the obtained objective function and the clustering center are optimal, and the objective function value is used as an initial population optimal value of the particle swarm algorithm and the genetic algorithm;

step 1.2: assuming that the population numbers of the particle swarm algorithm and the genetic algorithm are both N =200, and the iteration frequency is Iter =100; setting the crossover probability of the genetic algorithm to be 0.6, the mutation probability to be 0.1, and determining the variable range to be [0,1]Using binary coding and roulette as a selection method, randomly selecting an initial value of a population, and recording an optimal objective function value and a cluster center value corresponding to the objective function value in an iteration process; setting the speed range of the particle swarm algorithm to [ -0.0001,0.0001]In the range of [0,1 ]]Learning factor c ₁ ，c ₂ All of which are 0.2, and the inertial weight w, the initial velocity and initial position of the population are randomly selected, iteratedIn the generation process, recording the optimal position pbest of each particle and the optimal position gbest of the group, wherein the optimal position of the group is the optimal objective function value in the iteration of the current particle swarm algorithm; similarly, recording the cluster center value corresponding to the objective function value; taking the minimum value of the optimal objective function values of the particle swarm algorithm and the genetic algorithm, and comparing the minimum value with the current global optimal objective function value; if the target function is smaller than the iteration condition, updating the target function, and continuously searching for the optimal target function until the iteration condition is met; the objective function and the clustering center at the end of iteration are globally optimal;

the particle iteration formula is as follows:

wherein, ω is _ini Is the initial inertial weight, value 0.5; omega _end Is the inertial weight iterated to the maximum number of times, with a value of 0.1; g _k Is the maximum number of iterations, i.e. Iter, g is the current number of iterations;

and

the velocities at the i-th particle iterations k +1 and k, respectively, c ₁ 、c ₂ And w are the learning factor and inertial weight, r, respectively ₁ And r ₂ Is [0,1 ]]Is determined by the random value of (a),

and

the individual optimal position and the population optimal position in the ith particle k iteration are respectively,

and

the position at which the ith particle, k +1 and k, iterates, respectively.

3. The method for diagnosing the faults of the OS-ELM fuel cell based on the optimized FCM training as claimed in claim 2, wherein the specific steps of the OS-ELM initialization training and the online learning in the step 2 are as follows:

step 2.1: in the initial training stage, the OS-ELM activation function uses sigmoid and a training set to initialize the output weight of a single hidden layer feedforward neural network to obtain a fault classification identification model;

step 2.2: in the on-line learning stage, for externally acquired voltage and current data, an impedance spectrum obtained by fast Fourier transform is used, and an equivalent circuit is used for fitting to obtain a characteristic vector R of a sample _w 、R _m And R _p ，R _m Is used for characterizing the water content, R, of the proton exchange membrane _p For characterizing the charge transport process, R _w Is used to characterize the substance transport process. The feature vector is used as an input of the OS-ELM, and identifies whether a fault occurs currently and the type of the fault. Storing the collected samples, and classifying by using FCM after optimization based on PSO and GA when the number of the samples meets the quantity requirement, wherein the FCM is used for updating the output weight of the OS-ELM; and updating the fault classification identification model by continuously adjusting the value of the output weight, and optimizing the identification effect of the model.

Images