Method for predicting residual service life of proton exchange membrane fuel cell based on VMD-IASO-BiLSTM

Abstract

The invention discloses a VMD-IASO-BiLSTM-based method for predicting the remaining service life of a PEMFC (Proton Exchange Membrane Fuel Cell), and the method comprises the following steps: 1) building an original data set of the PEMFC, and carrying out the wavelet hard threshold (WHT) of the original data to remove a noise signal; 2) performing modal decomposition on the de-noised data by using variational modal decomposition (VMD), screening clean signals by using normal distribution, and reconstructing new voltage signals; 3) respectively introducing Tent chaotic mapping, a Gaussian mutation strategy and a hyperbolic tangent strategy to improve an atomic search optimization algorithm (ASO) to construct an IASO; 4) carrying out optimization iteration on the weight and bias of the bidirectional long short-term memory network (BiLSTM) by using an IASO, and finding out an optimal weight and bias parameter by using a mean square error (MSE) to carry out prediction; the feasibility and superiority of the method are proved, the iteration speed of the model can be greatly improved, the training time is shortened, and the accuracy is improved.

Description

Technical Field

[0001] The invention relates to the field of proton exchange membrane fuel cell quality prediction, and in particular to a method for predicting the remaining service life of a proton exchange membrane fuel cell based on VMD-IASO-BiLSTM. Background Art

[0002] As an important part of clean energy technology, the remaining useful life (RUL) prediction of proton exchange membrane fuel cells (PEMFC) is of great significance to improving system reliability and reducing maintenance costs. The current mainstream prediction methods include model-based, data-driven and hybrid methods, but there are still the following technical bottlenecks:

[0003] Limitations of signal processing and feature extraction: Existing technologies mostly use a single signal decomposition method combined with traditional denoising methods (such as wavelet denoising), but lack a screening mechanism, which can easily introduce noise interference or lose key aging features. When adapting to the RUL prediction needs of fuel cells of different working conditions and models, traditional optimization algorithms (such as particle swarm optimization algorithm PSO, sparrow optimization algorithm) are prone to fall into local optimality and slow convergence when optimizing parameters of the prediction model due to the lack of targeted tuning of PEMFC life prediction scenarios. Therefore, there is an urgent need for a model and optimization algorithm that is more suitable for the prediction of the remaining service life of proton exchange membrane fuel cells, which can provide users with necessary information more accurately and quickly. Summary of the invention

[0004] Purpose of the invention: To solve the problems mentioned in the background technology, the present invention discloses a method for predicting the remaining service life of a proton exchange membrane fuel cell based on VMD-IASO-BiLSTM. The denoised data is subjected to modal decomposition and normal distribution by VMD to further screen the data. The Tent chaos mapping, Gaussian mutation strategy and hyperbolic tangent strategy are introduced to improve the atomic search optimization algorithm (ASO) to construct IASO to optimize and predict BiLSTM. The prediction is performed through the model, which significantly improves the prediction accuracy in this field.

[0005] Technical solution:

[0006] The present invention discloses a method for predicting the remaining service life of a proton exchange membrane fuel cell based on VMD-IASO-BiLSTM, the method comprising the following steps:

[0007] S1 establishes the original PEMFC data set and removes the noise signal of the original data through WHT;

[0008] S2 uses VMD to perform modal decomposition on the denoised data, uses normal distribution to filter clean signals, and reconstructs new voltage signals;

[0009] S3 introduces Tent chaos mapping, Gaussian mutation strategy and hyperbolic tangent strategy to improve ASO and construct IASO algorithm;

[0010] S4 uses the IASO algorithm to iteratively optimize the weights and biases of the BiLSTM model, uses MSE to find the optimal weight and bias parameters, and trains and predicts through the improved BiLSTM model.

[0011] Furthermore, the denoised data described in S2 is subjected to modal decomposition using VMD, and clean signals are screened using normal distribution to reconstruct new voltage signals, which is specifically implemented as follows:

[0012] VMD redefines the intrinsic mode function of limited bandwidth with stricter constraints. The intrinsic mode component is defined as the component mode function of amplitude-frequency modulation. The mathematical expression is as follows:

[0013] u k (t) = A k (t)cos(φ k (t))

[0014] Where: A k (t) is the signal u k (t) envelope amplitude; φ k (t) is the instantaneous phase;

[0015] Assume that the original signal f(t) is decomposed into k components. To ensure that the original signal is decomposed into multiple narrowband modal components with center frequencies, and the sum of the estimated bandwidths of each mode is minimized, the constraint condition is that the sum of all modes is equal to the original signal. The constrained variational model formula of VMD is as follows:

[0016]

[0017] Where: u k is each mode function; ω k is the center frequency of each mode; δ(t) is the Dirac function.

[0018] Furthermore, in order to solve the optimal problem of construction constraints, the constrained variational problem is transformed into an unconstrained variational problem. By taking advantage of the quadratic penalty term and the Lagrange multiplier method, the augmented Lagrangian function is introduced, and the formula is as follows:

[0019]

[0020] Where: α is the penalty parameter; λ is the Lagrange multiplier;

[0021] For all ω ≥ 0, the update formulas for each component and the corresponding center frequency are as follows:

[0022]

[0023] Where: are the Fourier transforms of f(t), u(t), and λ(t), respectively, ω is the frequency, n is the number of iterations, and γ represents the noise;

[0024] Repeat the above updates until the following constraints are met:

[0025]

[0026] Where: ε is the set threshold;

[0027] When the update condition is met, the update iteration is stopped, the IMFs are obtained to draw a histogram, and the appropriate IMF is intuitively selected according to the Gaussian normal distribution to reconstruct the mode and obtain the processed voltage data signal.

[0028] Furthermore, S3 introduces Tent chaos mapping, Gaussian mutation strategy and hyperbolic tangent strategy to improve ASO and construct IASO algorithm. The specific implementation is as follows:

[0029] The population initialization based on Tent chaotic mapping is calculated as follows:

[0030]

[0031] Among them, x n is the value of the nth iteration; x0∈[0,1]; α is a control parameter;

[0032] Make the obtained mapping value within the value range of the ASO initial population. Assuming that the value range of ASO is [a, b], in order to make the mapping value within this range, calculate it by the following formula:

[0033] y n =a+(ba)x n

[0034] The core expressions of the ASO algorithm speed and position update are as follows:

[0035]

[0036] in, are the velocity and position of the ith atom at the tth iteration, respectively;

[0037] The Gaussian mutation strategy is used to improve the local optimization of the ASO algorithm. When a mutation occurs, the original gene value is replaced by a random number with an average of μ and a variance of σ. The expression is as follows:

[0038]

[0039] The hyperbolic tangent function is introduced into the speed update formula to control the search step size, that is, scaling The value of adjusts the input of the hyperbolic tangent function by changing its size. If it is positive, will be closer to 1, making This term contributes more to the speed update, the search step size may become larger, and the reduction It will make The closer it is to 0, the smaller the search step size changes.

[0040] The improved IASO algorithm updates the core expressions for algorithm speed and position as follows:

[0041]

[0042] Where: k p is the position variation intensity coefficient, and its value range is (0,1].

[0043] Furthermore, the weights and biases of BiLSTM are optimized and iterated through IASO, and the optimal weight and bias parameters are found for prediction using MSE. The specific implementation is as follows:

[0044] BiLSTM structure construction: connect the same input sequence to the two LSTMs in the forward and backward directions respectively, then connect the hidden layers of the two networks together and connect them to the output layer together;

[0045] Forget gate operation: receive the implicit state h of the previous moment t-1 and the current input x t , concatenate them, and use the sigmoid activation function to get a vector f between 0 and 1 t :

[0046] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0047] Among them, W f is the weight matrix of the forget gate; b f is the bias term; σ is the sigmoid function to determine which information in the cell state is discarded;

[0048] The input gate updates the cell state and determines which new information is added to the cell state. The updated cell state c t Its mathematical expression is as follows:

[0049]

[0050] The output gate controls the final output and first receives h t-1 and x t After a fully connected layer and a sigmoid function, a value between 0 and 1 is input to determine which parts of the cell state will be output. The mathematical expression is as follows:

[0051] o t =σ(W o ·[h t-1 ,x t ]+b0)

[0052] Where: W o is the weight matrix of the output gate; b o is the bias term.

[0053] Furthermore, the updated cell state c t The final hidden state h is obtained by multiplying the output of the output gate with the tanh function. t As the current output and also as part of the input at the next time, its mathematical expression is as follows:

[0054] h t =o t *tanh(c t )

[0055] Calculate BiLSTM output: Output the output after forward LSTM and the output of the reverse LSTM Together, we decide to pass it into the hidden layer and get the output y of BiLSTM t , which is expressed as follows:

[0056]

[0057] Where: b y They are the forward LSTM output weight, the reverse LSTM output weight and the output bias of BiLSTM respectively;

[0058] In order to obtain the optimal weight and bias, in the initialization stage of the IASO population, the mean square error is used as the fitness function value to obtain the minimum fitness function. The weight and bias at this time are the optimal solution, and the mathematical expression is as follows:

[0059]

[0060] Where: Reconstructed signals for IMFs.

[0061] Beneficial effects:

[0062] 1. The present invention significantly improves the noise suppression capability by combining wavelet hard threshold denoising with variational mode decomposition. After VMD decomposition, the normal distribution is used to screen and reconstruct the signal, remove abnormal modal components, retain key features reflecting battery degradation, and solve the signal distortion problem caused by modal aliasing in traditional methods.

[0063] 2. The present invention introduces Tent chaotic mapping to enhance population diversity, and combines Gaussian mutation and hyperbolic tangent strategy to balance global search and local development capabilities. The convergence speed of IASO in BiLSTM parameter optimization is further improved, the global search capability and optimization performance are enhanced, the number of iterations of IASO is reduced, and the risk of gradient disappearance is further avoided.

[0064] 3. The present invention adopts a bidirectional long short-term memory network (BiLSTM) to capture the forward and backward dependencies of the voltage sequence, and optimizes the network weights and bias parameters through IASO, which significantly improves the prediction accuracy under dynamic conditions without relying on complex mechanism models and maintains high prediction robustness under limited data conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a specific flow chart of the present invention;

[0066] Figure 2 IASO flow chart of the present invention;

[0067] Figure 3 This is a comparison chart of the training results of the present invention and various models; DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0069] like Figure 1-Figure 2 As shown, the present invention discloses a method for predicting the remaining service life of a proton exchange membrane fuel cell based on VMD-IASO-BiLSTM, and the method steps are as follows:

[0070] Step 1: Establish the original data set of PEMFC and perform WHT on the original data to remove noise signals;

[0071] Step 2: Use VMD to perform modal decomposition on the denoised data, use normal distribution to filter clean signals, and reconstruct new voltage signals;

[0072] Step 3: Introduce Tent chaotic mapping, Gaussian mutation strategy and hyperbolic tangent strategy to improve ASO and construct IASO;

[0073] Step 4: Use IASO to iteratively optimize the weights and biases of BiLSTM, and use MSE to find the optimal weight and bias parameters for prediction.

[0074] In step 1, the original data is subjected to WHT to remove the noise signal, which is specifically implemented as follows:

[0075] (11) The noisy signal y(t) is decomposed into wavelet coefficients at different scales using discrete wavelet transform. Let the number of decomposition layers be J. At the jth layer (1≤j≤J), the signal is decomposed into low-frequency coefficients (approximate coefficients) cA j,k and high frequency coefficient (detail coefficient) cD j,k The formula is as follows:

[0076]

[0077] Where: is the scaling function The conjugate function of ; is the wavelet function ψ j,k (t); k is the discrete time index. The formula is as follows:

[0078]

[0079] Where: Parent scale function The calculation formula of wavelet mother function ψ(t) is as follows:

[0080]

[0081] (12) Using the hard threshold processing method, select the appropriate λ for the detail coefficient cD of each layer j,k , coefficients after hard thresholding The calculation formula is:

[0082]

[0083] (13) The wavelet coefficients after hard threshold processing are used to reconstruct the signal through inverse wavelet transform. For the denoised signal, the reconstruction formula is:

[0084]

[0085] Step 2: Use VMD to perform modal decomposition on the denoised data, use normal distribution to filter clean signals, and reconstruct new voltage signals. The specific implementation is as follows:

[0086] (21) VMD redefines the intrinsic mode function of limited bandwidth with stricter constraints. The intrinsic mode component is defined as the component mode function of amplitude-frequency modulation. The mathematical expression is as follows:

[0087] u k (t) = A k (t)cos(φ k (t)) (7)

[0088] Where: A k (t) is the signal u k (t) envelope amplitude; φ k (t) is the instantaneous phase.

[0089] (22) Assume that the original signal f(t) is decomposed into k components. To ensure that the original signal is decomposed into multiple narrowband modal components with center frequencies and the sum of the estimated bandwidths of each mode is minimized, the constraint condition is that the sum of all modes is equal to the original signal. The constrained variational model formula of VMD is as follows:

[0090]

[0091] Where: u k is each mode function; ω k is the center frequency of each mode; δ(t) is the Dirac function.

[0092] (23) To solve the constrained optimal problem constructed above, the constrained variational problem is transformed into an unconstrained variational problem. By taking advantage of the quadratic penalty term and the Lagrange multiplier method, the augmented Lagrangian function is introduced, and the formula is as follows:

[0093]

[0094] Where: α is the penalty parameter; λ is the Lagrange multiplier.

[0095] For all ω ≥ 0, the update formulas for each component and the corresponding center frequency are as follows:

[0096]

[0097] Where: are the Fourier transforms of f(t), u(t), and λ(t), ω is the frequency, n is the number of iterations, and γ represents the noise.

[0098] Repeat the above updates until the following constraints are met:

[0099]

[0100] Where: ε is the set threshold.

[0101] After the update condition is met, the update iteration is stopped and the IMFs are obtained.

[0102] (24) For IMFs, a histogram is drawn, and the appropriate IMF is intuitively selected according to the Gaussian normal distribution to reconstruct the mode and obtain the processed voltage data signal.

[0103] Step 3 introduces Tent chaotic mapping, Gaussian mutation strategy and hyperbolic tangent strategy to improve ASO and construct IASO. The specific implementation is as follows:

[0104] (31)Tent chaotic mapping has significant advantages in initializing the population of the algorithm. On the one hand, it can greatly enhance the diversity of the population. Through its unique randomness and ergodicity, it can make the initial solution evenly distributed in the search space, avoid falling into the trap of local optimal solutions, and provide a wider range of possibilities for searching for global optimal solutions. On the other hand, this optimization method can significantly improve the convergence speed of the algorithm, allowing efficient exploration from multiple directions at the beginning of the search, quickly locating the better area, reducing unnecessary iterative calculations, and improving the overall optimization efficiency. The calculation formula is as follows:

[0105]

[0106] Where: x n is the value of the nth iteration; x0∈[0,1]; α is a control parameter.

[0107] In order to make the obtained mapping value within the value range of the ASO initial population, assuming that the value range of ASO is [a, b], in order to make the mapping value within this range, it is calculated by the following formula:

[0108] y n =a+(ba)x n (13)

[0109] (32) ASO is a new intelligent algorithm based on the molecular dynamics model proposed in 2019. Its principle is to simulate the phenomenon that atoms in a molecular system composed of atoms are displaced due to the mutual interaction and system constraints. In a molecular system, there are two kinds of interaction forces between adjacent atoms: attraction and repulsion. At the same time, the global optimal atom will exert geometric constraints on other atoms. Among them, attraction can drive atoms to explore the entire search space extensively, while repulsion helps atoms to effectively develop potential areas. This algorithm has the advantages of strong optimization ability and fast convergence speed.

[0110] Assume that a molecular system is a d-dimensional space composed of s atoms. is the position of the i-th atom at the t-th iteration, which can be expressed as

[0111] Atomic motion follows classical mechanics. According to Newton's second law, the acceleration of an atom is related to its mass, and the acceleration is the result of the interaction between atoms and the geometric constraints imposed on it by the optimal atom. Therefore, the acceleration of the i-th atom at the t-th iteration is expressed as follows:

[0112]

[0113] Where: is the total force acting on the i-th atom in the d-dimensional space at the t-th iteration, which can be regarded as the random weighted sum of the forces of k atoms with better fitness function values ​​on the i-th atom. The expression is as follows:

[0114]

[0115] Where: rand j is a random number in [0,1]; K best is the set of k atoms with better fitness values; is the force between two atoms, expressed as follows:

[0116]

[0117] Where: is the range of the attractive force area and the repulsive force area that can be adjusted; n decreases with the increase of the number of iterations, so that the range of global search and local development are gradually reduced to the optimal value; β is the depth weighting; t max is the maximum number of iterations; h ij (t) is the distance between two atoms, expressed as follows:

[0118]

[0119] Where: h max 、h min are the upper and lower bounds of h; It's K best The distance range of atoms in the set from the i-th atom.

[0120] In the ASO algorithm, the algorithm focuses on strengthening the global search capability in the early stage. Therefore, in the early stage of iteration, each atom needs to establish an interaction relationship with a relatively large number of neighboring atoms with good fitness, so as to expand the search range and avoid falling into the local optimal solution too early. As the iteration process advances to the later stage, the focus of the algorithm shifts to enhancing the ability of local development, prompting the algorithm to converge to the optimal solution quickly. At this time, each atom only needs to interact with a small number of neighboring atoms with good fitness, so that it can search more finely in the current better area. The number of neighboring atoms with good fitness is represented by K, and its expression is:

[0121]

[0122] In ASO, it is assumed that each atom has a covalent bond with the optimal atom, so each atom is subject to the constraint force from the optimal atom, so is the covalent bond constraint force of the global optimal atom on the i-th atom in the d-dimensional space at the t-th iteration, and the expression is as follows:

[0123]

[0124] Where: η is the multiplier weight.

[0125] In the above formula is the mass of the atom, expressed as follows:

[0126]

[0127] Where: Fit i (t) is the fitness value of atom i at the tth iteration; Fit best (t), Fit worst (t) are the minimum and maximum fitness values ​​of all atoms at the tth iteration, respectively.

[0128] Acceleration changes the speed and displacement of atoms, which is the core of the speed and position update of the ASO algorithm. The expression is as follows:

[0129]

[0130] Where: are the velocity and position of the ith atom at the tth iteration, respectively.

[0131] (33) Gaussian mutation improves the local optimization ability of the optimization algorithm on key issues. When a mutation occurs, the original gene value is replaced by a random number with a mean of μ and a variance of σ. Its expression is as follows:

[0132]

[0133] (34) Based on the original ASO, chaotic mapping is introduced into the initial population to increase the ergodicity and randomness of the population; a hyperbolic tangent function is introduced into the speed update formula to control the search step size, that is, to scale The value of adjusts the input of the hyperbolic tangent function by changing its size. If it is positive, will be closer to 1, making This term contributes more to the speed update, the search step size may become larger, and the reduction It will make As it gets closer to 0, the change in search step size becomes smaller, increasing local search capabilities. The introduction of Gaussian mutation strategy into the position update formula increases the randomness and diversity of the search, helps escape from the local optimum, and improves the algorithm's global search capabilities and optimization performance.

[0134] For IASO, the specific optimization strategies are as follows:

[0135] Modify formula (21) to get the following formula:

[0136]

[0137] Where: k p is the position variation intensity coefficient, and its value range is (0,1].

[0138] Step 4: Use IASO to iteratively optimize the weights and biases of BiLSTM, and use MSE to find the optimal weight and bias parameters for prediction. The specific implementation is as follows:

[0139] (41) BiLSTM is an improved recurrent neural network for processing sequence data. The idea is to connect the same input sequence to two LSTMs, one in the forward direction and the other in the backward direction, and then connect the hidden layers of the two networks together and connect them to the output layer. Compared with single-phase LSTM, BiLSTM can take into account the context information of the sequence at the same time, and in many natural language processing tasks, BiLSTM usually performs better than unidirectional LSTM. Like LSTM, BiLSTM also has a gating mechanism.

[0140] (411) The function of the forget gate is to decide which information to discard from the cell state. It receives the hidden state h at the previous moment. t-1 and the current input x t , then concatenate them and use the sigmoid activation function to get a vector f between 0 and 1 t , each element of which is related to the cell state C t-1 Each element of corresponds to, if f tIf an element in is close to 0, it means that most of the information in the corresponding cell state will be forgotten. t If an element in is close to 1, it means that most of the information in the corresponding cell state will be retained. Its mathematical expression is as follows:

[0141] f t =σ(W f ·[h t-1 ,x t ]+b f ) (twenty four)

[0142] Where: W f is the weight matrix of the forget gate; b f is the bias term; σ is the sigmoid function.

[0143] (412) The function of the input gate is to update the cell state and decide which new information to add to the cell state. It consists of two parts. The first part is to decide which information to update through the sigmoid function. Its mathematical expression is as follows:

[0144] i t =σ(W i ·[h t-1 ,x t ]+b i ) (25)

[0145] Where: W i is the weight matrix of the input gate; b i is the bias term.

[0146] The second part is to create a new candidate value vector c% through the tanh function t , it accepts h t-1 and x t Then it passes through a fully connected layer and tanh function. Its mathematical expression is as follows:

[0147]

[0148] Where: W c is the candidate weight matrix; b c is the bias term.

[0149] Updated cell state c t Its mathematical expression is as follows:

[0150]

[0151] (413) The function of the output gate is to control the final output. First, it receives h t-1 and x tAfter a fully connected layer and a sigmoid function, a value between 0 and 1 is input to determine which parts of the cell state will be output. The mathematical expression is as follows:

[0152] o t =σ(W o ·[h t-1 ,x t ]+b0) (28)

[0153] Where: W o is the weight matrix of the output gate; b o is the bias term.

[0154] Finally, the updated cell state c t The final hidden state h is obtained by multiplying the output of the output gate with the tanh function. t As the current output and also as part of the input at the next time, its mathematical expression is as follows:

[0155] h t =o t *tanh(c t ) (29)

[0156] Through the operation of these gates, LSTM can effectively remember key information and flexibly update and output information when processing sequence data. For BiLSTM, its output is passed through the output of the forward LSTM. and the output of the reverse LSTM Together, we decide to pass it into the hidden layer and get the output y of BiLSTM t , which is expressed as follows:

[0157]

[0158] Where: b y They are the forward LSTM output weight, the reverse LSTM output weight and the output bias of BiLSTM respectively.

[0159] (414) In order to obtain the optimal weight and bias, in the initialization stage of the IASO population, the mean square error is used as the fitness function value. When the minimum fitness function is obtained, the weight and bias at this time are the optimal solution, and its mathematical expression is as follows:

[0160]

[0161] Where: Reconstructed signals for IMFs.

[0162] In order to better reflect the prediction results of the model, the following cases will significantly demonstrate the feasibility of the experiment.

[0163] The data set used in this experiment comes from the public data set of the French FCLAB Research Federation. The voltage signal in the data set is selected as the feature data to obtain the impact of voltage on the remaining service life of PEMFC, and then make data predictions. Since the original data set is too large, this experiment modified the data set and sampled the first 200 features for the experiment.

[0164] By comparing with the existing prediction models, it is found that the test set training accuracy of VMD-IASO-BiLSTM is the highest compared with other methods. The training results are shown in Table 1. Figure 3 Table 1Comparison results of prediction of each model

[0165]

[0166] The above description of the embodiments enables professionals and technicians in the field to implement or use the present invention. Various modifications to the embodiments will be apparent to professionals and technicians. 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 range consistent with the principles and novel features disclosed in the present invention.

Claims

1. A method for predicting the remaining service life of a proton exchange membrane fuel cell based on VMD-IASO-BiLSTM, characterized in that: The method comprises the following steps: S1 establishes the original PEMFC data set and removes the noise signal of the original data through WHT; S2 uses VMD to perform modal decomposition on the denoised data, uses normal distribution to filter clean signals, and reconstructs new voltage signals; S3 introduces Tent chaos mapping, Gaussian mutation strategy and hyperbolic tangent strategy to improve ASO and construct IASO algorithm; S4 uses the IASO algorithm to iteratively optimize the weights and biases of the BiLSTM model, uses MSE to find the optimal weight and bias parameters, and trains and predicts through the improved BiLSTM model.

2. The method for predicting the remaining useful life of a proton exchange membrane fuel cell based on VMD-IASO-BiLSTM according to claim 1, characterized in that: S2 uses VMD to perform modal decomposition on the denoised data, uses normal distribution to filter clean signals, and reconstructs new voltage signals. The specific implementation is as follows: VMD redefines the intrinsic mode function of limited bandwidth with stricter constraints. The intrinsic mode component is defined as the component mode function of amplitude-frequency modulation. The mathematical expression is as follows: u k (t)=A k (t)cos(φ k (t)) Where: A k (t) is the signal u k (t) envelope amplitude; φ k (t) is the instantaneous phase; Assume that the original signal f(t) is decomposed into k components. To ensure that the original signal is decomposed into multiple narrowband modal components with center frequencies, and the sum of the estimated bandwidths of each mode is minimized, the constraint condition is that the sum of all modes is equal to the original signal. The constrained variational model formula of VMD is as follows: Where: u k is each mode function; ω k is the center frequency of each mode; δ(t) is the Dirac function.

3. The method for predicting the remaining useful life of a proton exchange membrane fuel cell based on VMD-IASO-BiLSTM according to claim 2, characterized in that: In order to solve the optimal problem of construction constraints, the constrained variation problem is transformed into an unconstrained variation problem. By taking advantage of the quadratic penalty term and the Lagrange multiplier method, the augmented Lagrangian function is introduced. The formula is as follows: Where: α is the penalty parameter; λ is the Lagrange multiplier; For all ω ≥ 0, the update formulas for each component and the corresponding center frequency are as follows: Where: are the Fourier transforms of f(t), u(t), and λ(t), respectively, ω is the frequency, n is the number of iterations, and γ represents the noise; Repeat the above updates until the following constraints are met: Where: ε is the set threshold; When the update condition is met, the update iteration is stopped, the IMFs are obtained to draw a histogram, and the appropriate IMF is intuitively selected according to the Gaussian normal distribution to reconstruct the mode and obtain the processed voltage data signal.

4. The method for predicting the remaining useful life of a proton exchange membrane fuel cell based on VMD-IASO-BiLSTM according to claim 1, characterized in that: S3 introduces Tent chaos mapping, Gaussian mutation strategy and hyperbolic tangent strategy to improve ASO and construct IASO algorithm. The specific implementation is as follows: The population initialization based on Tent chaotic mapping is calculated as follows: Among them, x n is the value of the nth iteration; x0∈[0,1]; α is a control parameter; Make the obtained mapping value within the value range of the ASO initial population. Assuming that the value range of ASO is [a, b], in order to make the mapping value within this range, calculate it by the following formula: y n =a+(b-a)x n The core expressions of the ASO algorithm speed and position update are as follows: in, are the velocity and position of the ith atom at the tth iteration, respectively; The Gaussian mutation strategy is used to improve the local optimization of the ASO algorithm. When a mutation occurs, the original gene value is replaced by a random number with an average of μ and a variance of σ. The expression is as follows: Introducing the hyperbolic tangent function into the velocity update formula to control the search step size and scale The value of adjusts the input of the hyperbolic tangent function by changing its size. If it is positive, will be closer to 1, making This term contributes more to the speed update, the search step size may become larger, and the reduction It will make The closer it is to 0, the smaller the change in search step size becomes; The improved IASO algorithm updates the core expressions for algorithm speed and position as follows: Where: k p is the position variation intensity coefficient, and its value range is (0,1].

5. The method for predicting the remaining useful life of a proton exchange membrane fuel cell based on VMD-IASO-BiLSTM according to claim 4, characterized in that: The weights and biases of BiLSTM are optimized and iterated through IASO, and the optimal weight and bias parameters are found for prediction using MSE. The specific implementation is as follows: BiLSTM structure construction: connect the same input sequence to the two LSTMs in the forward and backward directions respectively, then connect the hidden layers of the two networks together and connect them to the output layer together; Forget gate operation: receive the implicit state h of the previous moment t-1 and the current input x t , concatenate them, and use the sigmoid activation function to get a vector f between 0 and 1 t : f t =σ(W f ·[h t-1 ,x t ]+b f ) Among them, W f is the weight matrix of the forget gate; b f is the bias term; σ is the sigmoid function to determine which information in the cell state is discarded; The input gate updates the cell state and determines which new information is added to the cell state. The updated cell state c t Its mathematical expression is as follows: The output gate controls the final output and first receives h t-1 and x t After a fully connected layer and a sigmoid function, a value between 0 and 1 is input to determine which parts of the cell state will be output. The mathematical expression is as follows: o t =σ(W o ·[h t-1 ,x t ]+b0) Where: W o is the weight matrix of the output gate; b o is the bias term.

6. The method for predicting the remaining useful life of a proton exchange membrane fuel cell based on VMD-IASO-BiLSTM according to claim 5, characterized in that: The updated cell state c t The final hidden state h is obtained by multiplying the output of the output gate with the tanh function. t As the current output and also as part of the input at the next time, its mathematical expression is as follows: h t =o t *tanh(c t ) Calculate BiLSTM output: Output the output after forward LSTM and the output of the reverse LSTM Together, we decide to pass it into the hidden layer and get the output y of BiLSTM t , which is expressed as follows: Where: b y They are the forward LSTM output weight, the reverse LSTM output weight and the output bias of BiLSTM respectively; In order to obtain the optimal weight and bias, in the initialization stage of the IASO population, the mean square error is used as the fitness function value to obtain the minimum fitness function. The weight and bias at this time are the optimal solution, and the mathematical expression is as follows: Where: Reconstructed signals for IMFs.

Images