PEMFC residual service life prediction method based on multi-feature fusion

Abstract

The invention discloses a PEMFC residual service life prediction method based on multi-feature fusion, and the method comprises the steps: carrying out the related feature screening of an original data set through employing a Pearson's correlation coefficient method, carrying out the denoising of the screened features through employing adaptive noise complete set empirical mode decomposition, and carrying out the reconstruction of each denoised feature data based on statistical features; introducing Kent chaotic mapping and a linear adjustment strategy to improve a northern eagle algorithm (NGO) to construct an INGO; the weight and the offset of a channel attention mechanism (CAM) are optimized by using INGO, and the optimal weight and the optimal offset are found by using information entropy to perform feature fusion; the method comprises the following steps: introducing Logistic-Tent chaotic mapping and an index adaptive strategy to improve a GKSO (Genetic Kery Shark Algorithm) so as to construct IGKSO; the IGKSO is used to optimize the weight and bias of the bidirectional long short-term memory network (BiLSTM), and the mean square error (MSE) is used to find the optimal weight and bias to predict the remaining service life (RUL).

Description

Technical Field

[0001] The present invention relates to the field of new energy battery detection, and particularly to a method for predicting the remaining service life of PEMFC based on multi-feature fusion. Background Technique

[0002] As an efficient, clean and energy conversion device with broad application prospects, PEMFC (Proton Exchange Membrane Fuel Cell) has attracted wide attention globally in recent years. In the transportation field, PEMFC vehicles are regarded as a key technical approach to achieve carbon neutrality in the transportation field. Compared with traditional fuel vehicles, fuel cell vehicles have a longer driving range and a faster hydrogen refueling speed, which can effectively solve the "range anxiety". In terms of distributed power sources, Proton Exchange Membrane Fuel Cells (PEMFC) can supply power to remote areas, hospitals, data centers and other regions, enhancing the stability and flexibility of power supply. In addition, in specific fields such as aerospace and ocean exploration, Proton Exchange Membrane Fuel Cells have broad application prospects due to their high specific capacity and strong environmental adaptability.

[0003] During the storage and operation of PEMFC, degradation phenomena are likely to occur, resulting in accelerated performance loss and shortened service life. The degradation phenomenon of the battery will lead to a significant decline in the energy supply effect and endurance of the hydrogen fuel cell. If the condition life of PEMFC is not monitored in advance and corresponding measures are not taken, both production and travel may be affected. Existing technologies mostly adopt model-driven methods that rely on electrochemical mechanisms for RUL prediction. However, the complex multi-physical field coupling characteristics inside PEMFC lead to great difficulties in modeling and limited generalization ability. Therefore, there is an urgent need for a data-driven RUL prediction method that integrates multi-feature dynamic correlation analysis, adaptive optimization algorithms and deep learning to provide theoretical support for PEMFC health management. Summary of the Invention

[0004] Objective of the Invention: To solve the problems mentioned in the background technique, the present invention discloses a method for predicting the remaining service life of PEMFC based on multi-feature fusion, which combines an improved Northern Goshawk optimization algorithm, a channel attention mechanism, an improved Genghis Khan optimization algorithm and BiLSTM to perform feature fusion and RUL prediction of PEMFC, significantly improving the prediction accuracy.

[0005] Technical Solution:

[0006] The present invention discloses a method for predicting the remaining service life of PEMFC based on multi-feature fusion, and the method includes the following steps:

[0007] Step 1: Establish an original dataset of PEMFC, and use the Pearson correlation coefficient method to screen and reconstruct relevant features of the original dataset;

[0008] Step 2: Introduce the Kent chaotic map and linear adjustment strategy to improve the Northern Goshawk Optimization (NGO) algorithm to construct the Improved Northern Goshawk Optimization (INGO) algorithm;

[0009] Step 3: Optimize the weights and biases of the Channel Attention Mechanism (CAM) using INGO, and use information entropy to find the optimal weights and biases to perform feature fusion on the preprocessed dataset;

[0010] Step 4: Introduce the Logistics-Tent chaotic map and exponential adaptive strategy to improve the Gengis Khan Shark Optimization (GKSO) algorithm to construct the Improved Gengis Khan Shark Optimization (IGKSO) algorithm;

[0011] Step 5: Optimize the weights and biases of the Bidirectional Long Short-Term Memory (BiLSTM) network using the IGKSO algorithm, use the Mean Squared Error (MSE) to find the optimal weights and biases, and use the data after feature fusion to predict the remaining useful life of the proton exchange membrane fuel cell through the improved BiLSTM.

[0012] Furthermore, in Step 1, an original dataset of PEMFC is established, and Pearson correlation coefficient method is used to screen relevant features from the original dataset. The specific implementation is as follows:

[0013] The Pearson correlation coefficient is used to calculate the correlation strength between each feature. The formula is as follows:

[0014]

[0015] where: x i and y i are the i-th values of different features in the data respectively; and are the means of different features in the data respectively; n is the number of samples. Usually, the value range of r is between [-1, 1]. When r = 1, it indicates a perfect positive correlation; when r = -1, it indicates a perfect negative correlation; when r = 0, it indicates no linear relationship; when 0 < r < 1, it indicates a positive correlation; when -1 < r < 0, it indicates a negative correlation;

[0016] The t-distribution test method is used for significance test. According to the calculated t value, the corresponding p value is found through the t-distribution table. The p value between every two features is compared with the set significance level a. If p < a, the two features have a significant correlation, and several features with stronger correlation are selected as the basic features for feature fusion.

[0017] Furthermore, for the selected basic features, the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is used for denoising, and the denoised feature data is reconstructed based on the statistical feature means and variances to obtain new feature signals.

[0018] Furthermore, the specific implementation of Step 2 is as follows:

[0019] For the initial population of the NGO, the Kent chaotic map is introduced to optimize the initial population to improve the diversity and ergodicity of the population. The formula is as follows:

[0020]

[0021] Where: x n is the value of the nth iteration, and β ∈ (0, 1);

[0022] To make the obtained mapping value within the value range of the initial population of the NGO, assuming the value range of the NGO is [L, H], to make the mapping value within this range, it is calculated by the following formula:

[0023] X n = L + (H - L)x n

[0024] Where: X n is the initial population of the nth iteration of the NGO;

[0025] Based on the change rate of the fitness value, a linear adjustment is made to the search radius R of the INGO. When the change rate is large, it indicates that the algorithm is performing a global search. However, when the change rate is small, it indicates that the algorithm may be trapped in a local optimum. At this time, R is increased to enhance the global search ability. When the change rate approaches 0 and the algorithm is close to convergence, R is decreased to improve the local search accuracy. The formula is as follows:

[0026]

[0027] Where: K1 = 0.1, K2 = 0.2, R max = 2, R min = 0, is the set threshold, which is the basis for judging the magnitude of the change rate of the fitness value, and K1 and K2 are adjustment coefficients.

[0028] Furthermore, the specific implementation of step 3 is as follows:

[0029] For the preprocessed dataset where q is the number of samples and z is the sample feature dimension, the following operations are performed on the multi-feature fusion:

[0030] For each feature dimension of, global average pooling is performed to obtain a vector F avg , and global max pooling is performed on each feature dimension of U to obtain the vector F max ;

[0031] Combine F avg and F maxPerform splicing to obtain a fused global vector F with a length of 2z g , learn channel weights, and input F g into the first fully connected layer, and obtain the channel attention weight vector D through the weight W g and the bias b g , and finally obtain the fused feature vector U out , optimize the weights and biases of CAM through the INGO in step 2, and use the information entropy as the fitness function of INGO. The formula is as follows:

[0032]

[0033] In the formula: H(Y) is the entropy of the fused feature Y; y i is the fused feature value, and p(y i ) is the probability of the fused feature value appearing.

[0034] Furthermore, the specific implementation of step 4 is as follows:

[0035] In the hunting stage, GKSO searches for the optimal position of the target in the search space. In this stage, an exponential adaptive strategy is introduced. In the early stage of iteration, the step size is relatively large, making the optimization range wider. As the number of iterations increases, the step size decreases exponentially. At this time, a more precise search for the optimal solution can be carried out in the better area. The formula is as follows:

[0036]

[0037] In the formula: φ is the initial step size, φ = 1; θ is the attenuation coefficient, θ = 2;

[0038] Introduce the Logistics-Tent chaotic mapping to optimize the initial population of GKSO, traverse the entire solution space and the Tent chaotic mapping to make the initial population more evenly distributed in the solution space, improve the quality of the initial solution, and enhance the advantages of the global search ability. The formula is as follows:

[0039]

[0040] In the formula: γ ∈ (0, 4); (·) mod1 is the modulo operation, and x i is the position of the i-th population.

[0041] Furthermore, the specific implementation of step 5 is as follows:

[0042] IGKSO is used to optimize the weights and biases of BiLSTM, with MSE as the fitness function of IGKSO. That is, IGKSO is used to find a set of weights and bias vectors, which are substituted into BiLSTM for prediction. When the error between the predicted value and the true value is minimized, that is, when MSE is minimized, this set of weights and bias vectors is the optimal solution found by IGKSO optimization. The formula is as follows:

[0043]

[0044] Where: y i are input features, is the prediction result.

[0045] Beneficial effects:

[0046] 1. This paper uses the Pearson correlation coefficient to screen highly correlated features, denoises them using the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), and reconstructs the denoised feature data based on statistical features (mean and variance). The filtered dataset can further help subsequent models effectively capture the core laws of battery performance degradation, improving the stability of long-term predictions and the ability to generalize across scenarios.

[0047] 2. The present invention introduces Kent chaotic mapping and linear adjustment strategy to improve the Northern Goshawk algorithm (NGO) to construct INGO, avoid the algorithm from falling into local optimality, improve search accuracy, and use information entropy to find the optimal weight and bias of CAM to perform feature fusion to obtain more comprehensive and more relevant features, thereby enhancing the accuracy and robustness of subsequent model predictions.

[0048] 3. The present invention introduces the Logistics-Tent chaos mapping and exponential adaptive strategy to improve the Genghis Khan Shark Algorithm (GKSO) to construct IGKSO, so that the initial population is more evenly distributed in the solution space, the quality of the initial solution is improved, and the optimal solution of the URL is more accurately sought in a better area. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is the overall flow chart of the present invention;

[0050] Figure 2 This is the INGO flow chart of the present invention;

[0051] Figure 3 IGKSO flow chart of the present invention;

[0052] Figure 4 This is a characteristic signal diagram after INGO-CAM fusion of the present invention;

[0053] Figure 5 2 is a comparison chart of the prediction results of the embodiment of the present invention and various models. Detailed implementation mode

[0054] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0055] As Figure 1 shown, the present invention discloses a method for predicting the remaining service life of PEMFC based on multi-feature fusion, and the steps are as follows:

[0056] Step 1: Establish a PEMFC original data set, and use the Pearson correlation coefficient method to screen and reconstruct relevant features of the original data set;

[0057] The Pearson correlation coefficient is used to calculate the correlation strength between each feature, and the formula is as follows:

[0058]

[0059] In the formula: x i and y i are respectively the i-th values of different features in the data; and are respectively the means of different features in the data; n is the number of samples. Usually, the value range of r is between [-1, 1]. When r = 1, it means a perfect positive correlation; when r = -1, it means a perfect negative correlation; when r = 0, it means no linear relationship; when 0 < r < 1, it means a positive correlation; when -1 < r < 0, it means a negative correlation.

[0060] After calculating the Pearson correlation coefficient, a significance test is also required. The significance test measures the possibility that the observed correlation coefficient actually exists in the population by calculating the p-value. The test method used is the t-distribution test, and the formula is as follows:

[0061]

[0062] According to the calculated t-value, the corresponding p-value is found through the t-distribution table. In this embodiment, the p-values between each two features are directly obtained through SPSS software. The p-value between each two features is compared with the set significance level a. In this embodiment, a = 0.05. If p < a, then the two features have a significant correlation.

[0063] From the above operations, several features with strong correlations screened in the original data set in this embodiment are used as the basic features of the fusion features.

[0064] Add N Gaussian white noise to the original signal y(t) to construct N noise signal sequences δ i (t) Generate N noise signals y i (t)(i=1,2,...,N), the formula is as follows:

[0065] y i (t)=y(t)+δ i (t) (3)

[0066] Where: δ i (t) is the added Gaussian white noise.

[0067] y i (t) Perform EMD decomposition to obtain the first IMF component c of each 1i (t), and take its average as the first IMF component c1(t) and the first residual r1(t) obtained by CEEMDAN decomposition. The formula is as follows:

[0068]

[0069] Starting from the second IMF, adaptive white noise is added each time. Let the jth IMF be c j (t), the jth residual is r j (t). Calculate the jth adaptive white noise δ j (t), the formula is as follows:

[0070] δ ji (t)=σ j c j (t)δ i (t) (5)

[0071] Among them, σ j is a scaling factor that controls the intensity of white noise.

[0072] Add white noise to the residual r j (t) to obtain N noise residuals r ji (t)(i=1,2,...,N), the formula is as follows:

[0073] r ji (t) = r j (t)+δ ji (t) (6)

[0074] to r ji (t) Perform EMD decomposition to obtain the j+1th IMF component c (j+1)i (t) and average it to get c (j+1) (t), the formula is as follows:

[0075]

[0076] Update the residual, the formula is as follows:

[0077] r j+1 (t) = r j (t) - c j+1 (t) (8)

[0078] Repeat formulas (5) to (8) until r j (t) meets the stopping condition. Finally, the representation formula of y(t) is as follows:

[0079]

[0080] For the IMF component c i (t), consider the IMF component with a mean close to 0 and a variance close to 1 as a signal with more noise and remove it, and reconstruct to obtain a new feature signal.

[0081] Step 2: As Figure 2 shown, introduce the Kent chaotic map and the linear adjustment strategy to improve the Northern Goshawk Optimization (NGO) algorithm to construct the Improved Northern Goshawk Optimization (INGO) algorithm;

[0082] The northern goshawk randomly selects a prey and quickly attacks during the exploration process. The formula is as follows:

[0083]

[0084] In the formula: x i,j is the j - th dimension position of the i - th northern goshawk; P i,j is the j - th dimension position of the i - th prey; is the objective function of the position of the i - th prey; F i is the fitness value of the position of the i - th northern goshawk; X i is the position vector of the i - th northern goshawk; is the new position vector of the i - th northern goshawk during the global search process; is the new position of the i - th northern goshawk in the j - th dimension during the global search process; F i new,P1 is the fitness value of the i - th northern goshawk after update during the global search process; ψ is a random number within [0, 1]; I is 0 or 1.

[0085] In this stage, introduce the Kent chaotic map to optimize the initial population of NGO to improve the diversity and ergodicity of the population. The formula is as follows:

[0086]

[0087] In the formula: xn is the value of the nth iteration, and β ∈ (0, 1).

[0088] To make the obtained mapping value within the value range of the initial population of NGO, assuming the value range of NGO is [L, H], to make the mapping value within this range, it is calculated by the following formula:

[0089] X n = L + (H - L)x n (13)

[0090] In the formula: X n is the initial population of the nth iteration of NGO.

[0091] In the development stage, after the northern goshawk attacks its prey, the prey tries to escape, and the northern goshawk continues to chase the prey. This behavior improves the local search ability of the algorithm for the search space. The formula is as follows:

[0092]

[0093] In the formula: t is the current iteration number; T is the maximum iteration number; is the new position of the ith northern goshawk in the jth dimension during the local search process; R is the search radius; is the new position vector of the ith northern goshawk during the local search process; F i new,P2 is the fitness value of the ith northern goshawk after update during the local search process.

[0094] For some highly complex optimization problems with a large number of local optimal solutions, the NGO algorithm may fall into a local optimum and thus cannot obtain the final optimal solution. To address this problem, the present invention makes a linear adjustment to the search radius R based on the change rate of the fitness value. When the change rate is large, it indicates that the algorithm is performing a global search. However, when the change rate is small, it indicates that the algorithm may have fallen into a local optimum. At this time, R is increased to enhance the global search ability. When the change rate approaches 0 and the algorithm is close to convergence, R is decreased to improve the local search accuracy. The formula is as follows:

[0095]

[0096] In the formula: K1 = 0.1, K2 = 0.2, R max = 2, R min = 0, is the set threshold, which is the basis for judging the magnitude of the change rate of the fitness value. K1 and K2 are adjustment coefficients.

[0097] Step 3: Optimize the weights and biases of the channel attention mechanism CAM through INGO, and use information entropy to find the optimal weights and biases to perform feature fusion on the preprocessed dataset;

[0098] For the preprocessed dataset where q is the number of samples and z is the sample feature dimension (number of channels), the following operations are performed on multi-feature fusion:

[0099] For each feature dimension of, perform global average pooling, and the formula is as follows:

[0100]

[0101] In the formula: F avg is a vector of length z obtained through global average pooling.

[0102] For each feature dimension of U, perform global max pooling, and the formula is as follows:

[0103]

[0104] In the formula: F max is a vector of length z obtained through global max pooling.

[0105] Concatenate F avg and F max to obtain a fused global vector F g , and the formula is as follows:

[0106] F g = [F avg , F max (20)

[0107] Learn the channel weights, input F g into the first fully connected layer, and obtain the channel attention weight vector D through the weight W g and the bias b g , and finally obtain the fused feature vector U out , and the formula is as follows:

[0108]

[0109] In the formula: σ(·) is the Sigmoid function.

[0110] Optimize the weights and biases of CAM through INGO in Step 2, and use information entropy as the fitness function of INGO, and the formula is as follows:

[0111]

[0112] Where: H(Y) is the entropy of the fused feature Y; y i is the fused feature value, and p(y i ) is the probability of the occurrence of the fused feature value. For the information entropy of the fused feature, the larger the entropy value, the stronger the correlation of the fused feature and the more comprehensive the information. The feature signal fused by INGO-CAM is as shown in Figure 4 .

[0113] Step 4: As shown in Figure 3 , introduce the Logistics-Tent chaotic map and the exponential self-adaptive strategy to improve the Genghis Khan shark algorithm GKSO to construct the IGKSO algorithm;

[0114] In the hunting stage, GKSO searches for the optimal position of the target in the search space. The formula is as follows:

[0115]

[0116] Where: is the position of the i-th population at the j-th dimension at time t; ub and lb are the upper and lower bounds of the search space respectively; r1 is a random number within [0,1]; t is the current iteration number.

[0117] In this stage, introduce the exponential self-adaptive strategy. In the early stage of iteration, the step size is relatively large, making the optimization range wider. As the number of iterations increases, the step size decreases exponentially. At this time, a more precise search for the optimal solution can be carried out in the better area. The formula is as follows:

[0118]

[0119] Where: φ is the initial step size, φ = 1; θ is the attenuation coefficient, θ = 2.

[0120] In the movement stage, in order to obtain better prey, the Genghis Khan shark continuously approaches the best prey position by smell. The formula is as follows:

[0121]

[0122] Where: is the position of the i-th population at the j-th dimension at time t + 1 in the movement stage; is the optimal population position at the j-th dimension at time t; s is the smell intensity when the Genghis Khan shark moves towards the best prey. This factor depends on the smell concentration of the prey. The formula is as follows:

[0123] s = mε μ (26)

[0124] Where: m = 1.5; ε = 1; μ is a random number within [0,1].

[0125] GKSO retains the first two optimal individual positions and updates the positions of other individuals to make them close to the optimal position. This behavior is called elite leadership and the formula is as follows:

[0126]

[0127] During the foraging phase, the Genghis Khan shark approaches its prey rapidly in a parabolic shape, as shown in the following formula:

[0128]

[0129] Where r2 is a random number in [0,1]; λ is 0 or 1; J is a parameter that controls the moving step size, and its formula is as follows:

[0130]

[0131] Where: |ω(t+1)|=1-2ω 4 (t), w(0) = 0.1.

[0132] During the foraging phase, Genghis Khan sharks also face external factors (such as factors that threaten their own safety or competing predators), so they have a set of self-protection mechanisms to deter predators and quickly escape. The formula is as follows:

[0133]

[0134] Where: k1 is a random number in the range [-1, 1]; k2 is a normally distributed random number with a mean of 0 and a standard deviation of 1; ρ is an adaptive random parameter; a1, a2, and a3 are three random numbers calculated using the following formula:

[0135]

[0136] Where: rand is a random number in [0,1]; l1 is 0 or 1; Ω min =0.2,Ω max =1.2.

[0137] In addition, the randomly generated X1 i (t) and X2 i (t) can be expressed as:

[0138]

[0139] X k The calculation formula of (t) is as follows:

[0140] X k (t) = l2*(X p (t)-X r (t))+X r(t) (34)

[0141] Where: X p (t), X r (t) are two position solutions randomly selected at time t; l2 is 0 or 1.

[0142] Due to the limitation of the search mechanism of the GKSO algorithm, in the early iteration, the shark individuals can quickly search in the solution space. However, as the number of iterations increases, the individuals are easily attracted by the current local optimal solution, and their optimization range is prematurely reduced, resulting in falling into the local optimum. To address this problem, the Logistics-Tent chaotic mapping is introduced to optimize the initial population. Its advantage lies in inheriting the advantages of the Logistics chaotic mapping to traverse the entire solution space and the Tent chaotic mapping to make the initial population more evenly distributed in the solution space, improving the quality of the initial solution and enhancing the global search ability. The formula is as follows:

[0143]

[0144] Where: γ ∈ (0, 4); (·)mod1 is the modulo operation, x i is the position of the i-th population. To make the mapping value within the range, it is the same as formula (13) above.

[0145] Step 5: Use the IGKSO algorithm to optimize the weights and biases of the bidirectional long short-term memory network BiLSTM, find the optimal weights and biases using the mean square error MSE, and predict the remaining useful life of the proton exchange membrane fuel cell through the improved BiLSTM using the data after feature fusion.

[0146] BiLSTM considers both past and future information, enabling the model to better capture the relationships existing in the sequence data, and effectively solving the problems of gradient vanishing and gradient explosion of the traditional recurrent neural network by introducing a gating mechanism. The specific description is as follows:

[0147] Determine which information to discard from the cell state through the forget gate. It receives the hidden state h t-1 at the previous moment and the input x t at the current moment, then concatenates them, and uses the sigmoid activation function to obtain a vector f t between 0 and 1. Each element of this vector corresponds to each element of the cell state C t-1 . If a certain element in f t is close to 0, it means that most of the information in the corresponding cell state will be forgotten. If a certain element in f t is close to 1, it means that most of the information in the corresponding cell state will be retained. The formula is as follows:

[0148] ft = σ(W f · [h t-1 , x t + b f ) (36)

[0149] Where: W f is the weight matrix of the forget gate; b f is the bias term; σ is the sigmoid function.

[0150] Update the cell state through the input gate, determining which new information is added to the cell state. It consists of two parts. The first part determines which information to update through the sigmoid function, and the formula is as follows:

[0151] i t = σ(W i · [h t-1 , x t + b i ) (37)

[0152] Where: W i is the weight matrix of the input gate; b i is the bias term.

[0153] The second part creates a new candidate value vector through the tanh function which takes h t-1 and x t and then passes through a fully connected layer and the tanh function. The formula is as follows:

[0154]

[0155] Where: W c is the candidate weight matrix; b c is the bias term.

[0156] The updated cell state c t The formula is as follows:

[0157]

[0158] Control the final output through the output gate. First, it takes h t-1 and x t passes through a fully connected layer and the sigmoid function to input a value between 0 and 1 to determine which parts of the cell state will be output. The formula is as follows:

[0159] o t = σ(W o · [h t-1 , x t + b0) (40)

[0160] Where: W o is the weight matrix of the output gate; b o is the bias term.

[0161] Finally, the updated cell state c t is multiplied by the output of the output gate through the tanh function to obtain the final hidden state h t as the current output and also as part of the input for the next time step. The formula is as follows:

[0162] h t = o t * tanh(c t ) (41)

[0163] Through the operations of these gates, the LSTM can effectively remember key information, flexibly update and output information when processing sequential data. For the BiLSTM, its output is determined by the output of the forward LSTM and the output of the backward LSTM jointly, and the output y t of the BiLSTM is obtained by passing through the hidden layer. The expression is as follows:

[0164]

[0165] Where: b y are the output weights of the forward LSTM, the output weights of the backward LSTM, and the output bias term of the BiLSTM, respectively.

[0166] The weights and biases of the BiLSTM are optimized by IGKSO. The MSE is used as the fitness function of IGKSO. That is, a set of weight and bias vectors are found by IGKSO, and these weight and bias vectors are substituted into the BiLSTM for prediction. When the error between the predicted value and the true value is minimized, that is, when the MSE is minimized, this set of weight and bias vectors is the optimal solution found by IGKSO optimization. The formula is as follows:

[0167]

[0168] Where: y i is the input feature, is the prediction result.

[0169] To better reflect the prediction results of the model, the feasibility of the experiment is demonstrated through a verification scheme in this embodiment. The dataset used comes from the publicly available dataset of the French FCLAB Research Consortium. Since the original dataset is too large, this embodiment modifies the dataset and experiments with the first 200 features sampled. By comparing with existing prediction models, it is found that the training accuracy of the test set of IGKSO-BiLSTM is improved the most compared with other methods, where the accuracy is represented by the coefficient of determination (R 2 ) and the formula is as follows:

[0170]

[0171] In the formula: is the average value of the input features. For this prediction accuracy, the closer it is to 100%, the better the prediction result.

[0172] In addition, the experiment also compares from the root mean square error (RMSE), mean absolute error (MAE), etc. to prove the superiority of this model. The experimental results are shown in Table 1, and the comparison chart of the prediction results of each model is as shown in Figure 5 .

[0173] Table 1 Comparison results of prediction of each model

[0174]

[0175] The above description of the embodiments enables those skilled in the art to implement or use the present invention. Various modifications to the embodiments will be obvious to those skilled in the art. The general principles of the present invention can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention should not be limited to the embodiments shown herein, but should cover the widest scope consistent with the principles and novel features disclosed in the present invention.

Claims

1. A method for predicting the remaining service life of PEMFC based on multi-feature fusion, characterized in that The method includes the following steps: Step 1: Establish the original PEMFC dataset, and use the Pearson correlation coefficient method to screen and reconstruct the relevant features of the original dataset; Step 2: Introduce the Kent chaotic map and linear adjustment strategy to improve the Northern Goshawk Optimization (NGO) algorithm to construct the Improved Northern Goshawk Optimization (INGO) algorithm; Step 3: Use INGO to optimize the weights and biases of the Channel Attention Mechanism (CAM), and use information entropy to find the optimal weights and biases to perform feature fusion on the preprocessed dataset; Step 4: Introduce the Logistics-Tent chaotic map and exponential adaptive strategy to improve the Gengis Khan Sharks Optimization (GKSO) algorithm to construct the Improved Gengis Khan Sharks Optimization (IGKSO) algorithm; Step 5: Use the IGKSO algorithm to optimize the weights and biases of the Bidirectional Long Short-Term Memory (BiLSTM) network, use the Mean Squared Error (MSE) to find the optimal weights and biases, and use the data after feature fusion to predict the remaining useful life of the proton exchange membrane fuel cell through the improved BiLSTM.

2. The method for predicting the remaining service life of PEMFC based on multi-feature fusion according to claim 1, characterized in that, In Step 1, the original dataset of PEMFC is established, and the Pearson correlation coefficient method is used to screen the relevant features of the original dataset. The specific implementation is as follows: The Pearson correlation coefficient is used to calculate the correlation strength between each feature. The formula is as follows: Where: x i and y i are respectively the i-th values of different features in the data; and are respectively the means of different features in the data; n is the number of samples. Usually, the value range of r is between [-1, 1]. When r = 1, it indicates a perfect positive correlation; when r = -1, it indicates a perfect negative correlation; when r = 0, it indicates no linear relationship; when 0 < r < 1, it indicates a positive correlation; when -1 < r < 0, it indicates a negative correlation; The t-distribution test method is used for significance testing. According to the calculated t-value, the corresponding p-value is found through the t-distribution table. The p-value between every two features is compared with the set significance level α. If p < α, the two features have a significant correlation, and several features with stronger correlation are selected as the basic features for feature fusion.

3. The method for predicting the remaining service life of PEMFC based on multi-feature fusion according to claim 2, wherein, For the screened basic features, the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is used for denoising, and the denoised feature data is reconstructed based on the statistical feature mean and variance to obtain a new feature signal.

4. The method for predicting the remaining service life of PEMFC based on multi-feature fusion according to claim 1 or 3, characterized in that, The specific implementation of Step 2 is as follows: For the initial population of NGO, the Kent chaotic map is introduced to optimize the initial population to improve the diversity and ergodicity of the population. The formula is as follows: where: x n is the value of the nth iteration, β ∈ (0, 1); To make the obtained mapping value within the value range of the NGO initial population, assuming the value range of NGO is [L, H], in order to make the mapping value within this range, it is calculated through the following formula: X n = L + (H - L)x n where: X n is the initial population of the n-th iteration of the NGO; Based on the change rate of the fitness value, a linear adjustment is made to the search radius R of INGO. When the change rate is large, it indicates that the algorithm is performing global search. However, when the change rate is small, it indicates that the algorithm may fall into a local optimum. At this time, R is increased to enhance the global search ability. When the change rate approaches 0 and the algorithm is close to convergence, R is decreased to improve the local search accuracy. The formula is as follows: In the formula: K1 = 0.1, K2 = 0.2, R max = 2, R min = 0, is the set threshold value and is the basis for judging the magnitude of the change rate of the fitness value. K1 and K2 are adjustment coefficients.

5. The method for predicting the remaining service life of PEMFC based on multi-feature fusion according to claim 4, wherein The specific implementation of Step 3 is as follows: For the preprocessed dataset where q is the number of samples and z is the sample feature dimension, the following operations are performed on multi-feature fusion: For each feature dimension of perform global average pooling to obtain a vector F of length z avg and perform global max pooling on each feature dimension of U to obtain vector F max ; Concatenate F avg and F max to obtain a fused global vector F of length 2z g , learn the channel weights, and input F g into the first fully-connected layer to obtain the channel attention weight vector D through the weight W g and bias b g , and finally obtain the fused feature vector U out , optimize the weights and biases of CAM through INGO in step 2, and use the information entropy as the fitness function of INGO. The formula is as follows: Where: H(Y) is the entropy of the fused feature Y; y i is the fused feature value, and p(y i ) is the probability of the occurrence of the fused feature value.

6. The method for predicting the remaining service life of PEMFC based on multi-feature fusion according to claim 1, characterized in that The specific implementation of Step 4 is as follows: In the hunting stage, GKSO searches for the optimal position where the target is located in the search space. In this stage, an exponential adaptive strategy is introduced. In the early stage of iteration, the step size is relatively large, making the optimization range wider. As the number of iterations increases, the step size decreases exponentially. At this time, a more precise search for the optimal solution can be carried out in the better region. The formula is as follows: In the formula: φ is the initial step size, φ = 1; θ is the attenuation coefficient, θ = 2; Introduce the Logistics-Tent chaotic map to optimize the initial population of GKSO, traverse the entire solution space and the Tent chaotic map to make the initial population more evenly distributed in the solution space, improve the quality of the initial solution, and enhance the advantages of the global search ability. The formula is as follows: where: γ ∈ (0, 4); (·) mod 1 is the modulo operation, and x i is the position of the i-th population.

7. The method for predicting the remaining service life of PEMFC based on multi-feature fusion according to claim 6, characterized in that, The specific implementation of step 5 is as follows: Optimize the weights and biases of BiLSTM through IGKSO, and use MSE as the fitness function of IGKSO. That is, find a set of weight and bias vectors through IGKSO, substitute this set of weight and bias vectors into BiLSTM for prediction. When the error between the predicted value and the true value is the smallest, that is, when MSE is the smallest, this set of weight and bias vectors is the optimal solution found by IGKSO for optimization. The formula is as follows: where: y i is the input feature, is the prediction result.

Images