CN114943372A - Method and device for predicting life of proton exchange membrane based on Bayesian recurrent neural network - Google Patents

Abstract

The invention discloses a method and a device for predicting the service life of a proton exchange membrane based on a Bayesian recurrent neural network, which can quantify the uncertainty of a fuel cell in the operation process, and specifically comprises the following steps: preprocessing original data by using a principal component analysis method; establishing different cyclic neural networks, predicting different characteristic data, and selecting a proper neural network according to a prediction result; replacing fixed parameters in the deep learning neural network model with random variables, and quantifying uncertainty of each characteristic data through probability density distribution to establish a Bayesian neural network; fourthly, training network parameters of the Bayesian neural network by using the training set data; inputting the former operation data of the fuel cell into the circulation neural network to obtain the operation result of the characteristic data. The invention can not only perform interval estimation on the voltage of the fuel cell, but also perform point estimation on the voltage of the fuel cell, thereby effectively improving the fault tolerance rate and the accuracy of the life prediction of the fuel cell.

Description

Method and device for predicting life of proton exchange membrane based on Bayesian recurrent neural network

Technical Field

The invention relates to the technical field of power generation of proton exchange membrane fuel cells, in particular to a method and a device for predicting the service life of a proton exchange membrane based on a Bayesian circulating neural network.

Background

The fuel cell is a novel electrochemical energy storage cell, effectively utilizes hydrogen energy, and efficiently and environmentally converts chemical energy into electric energy through electrochemical reaction. Among various types of fuel cells, the proton exchange membrane fuel cell has the characteristics of low noise, zero pollution, no corrosion, long service life and the like, also has the advantages of large output current, low working temperature, high energy efficiency, quick start, compact structure and the like, and is widely applied to portable power supplies, motor vehicle power supplies and medium and small power generation systems. Meanwhile, a power station established on the basis of the proton exchange membrane fuel cell can be connected with a power grid dispatching system during the peak period of power utilization, so that the load of the power grid is balanced.

However, the short lifetime, fast performance decay, high maintenance cost and expensive hydrogen fuel have significantly hindered the deployment and commercialization of PEMFCs. In order to predict the remaining service life (RUL) of the pem fuel cell before a failure occurs and to arrange to maintain the fuel cell system in time to prolong the service life, RUL prediction of the PENFC becomes an urgent problem to be solved, so research on the prediction method of the remaining service life of the pem fuel cell has great research significance for improving the durability and the actual remaining life of the PEMFC.

At present, various methods for predicting the residual service life of the proton exchange membrane fuel cell exist. The patent CN202110385900.2 relates to a fuel cell remaining service life prediction method based on deep learning, which comprises the steps of fitting fuel cell operation data after noise reduction processing, establishing a nonlinear mapping relation between current input data and target data, establishing a loss function, optimizing neural network model parameters through a feedback derivation method, obtaining an optimal neural network model, and realizing fuel cell RUL prediction through the most neural network model. Patent CN201910260003.1 discloses a fuel cell life prediction method based on deep convolutional neural network, which includes selecting characteristic variables as model input data for original data in experimental test data set; carrying out data normalization on the original data; extracting sample data, and dividing the sample data into multiple batches; setting parameters of the fuel cell life prediction model, and then performing convolution layer calculation; after the calculation of one convolution layer is completed, before entering the next convolution layer, performing maximum pooling on the obtained characteristic mapping matrix; after the calculation of the multilayer convolution layer, carrying out global average pooling, and carrying out full-connection operation on the multilayer characteristic mapping matrix of the single characteristic; and (4) outputting a prediction result through multi-round training according to batch division.

Although many scholars have already done a lot of work on predicting the remaining service life of proton exchange membrane fuel cells, such as using various neural networks such as DNN, RNN, CNN, etc., most prediction methods belong to point estimation prediction, and basically do not consider uncertainty factors of fuel cell operation data. Therefore, in order to quantify uncertainty factors in the operation data, the invention discloses a fuel cell service life prediction method and a fuel cell service life prediction device based on a Bayesian recurrent neural network.

Disclosure of Invention

The invention discloses a life prediction method and a life prediction device of a proton exchange membrane based on a Bayesian recurrent neural network, which are used for making up the deficiency of the life prediction of a fuel cell in the aspect of interval estimation.

In order to achieve the purpose, the invention designs a life prediction method of a proton exchange membrane based on a Bayesian recurrent neural network, which is characterized by comprising the following steps:

1) performing data preprocessing on a fuel cell operation sample data set, extracting main characteristic data in the data set by adopting a principal component analysis method to realize data dimension reduction, and dividing the processed data set into a training set and a test set;

2) establishing three different recurrent neural networks to respectively predict the characteristic numbers selected in the step 1), and selecting the optimal recurrent neural network according to the prediction result;

3) replacing fixed parameters in the classical deep learning neural network with random variables, quantifying uncertainty factors contained in various types of characteristic data through probability density distribution, and establishing a Bayesian neural network;

4) taking the data in the training set as the input of a Bayesian neural network, training network parameters until a global optimal solution which enables a loss function to reach the minimum is found;

5) preprocessing the operation data of the fuel cell by using a principal component analysis method, predicting the selected characteristic data by using the optimal cyclic neural network selected in the step 2), and taking the predicted result as the input of the trained Bayesian neural network to obtain the distribution and interval of the predicted voltage of the fuel cell at each moment.

Preferably, the step 1) of extracting data features includes:

11) according to the fuel cell operation sample data, recording an original sample-index data matrix as follows:

wherein x ₁ ,x ₂ ,…,x _p Respectively, index data of operation of the fuel cell, i.e. x _p ＝[x _1p ,x _2p ,…x _np ] ^T N is the number of samples, and p is the number of indexes of the fuel cell during operation;

12) standardizing the operation sample data of the fuel cell, wherein Var () represents the mean square error of the sample data:

13) calculating a sample data correlation coefficient matrix:

wherein r is _ij The closer | is to 1, the variable x _i ,x _j The more similar, the less similar if closer to 0;

14) calculating the eigenvalues and corresponding eigenvectors of the correlation coefficient matrix R:

characteristic value: lambda [ alpha ] ₁ ,λ ₂ ...λ _p

Feature vector: a is ₁ ,a ₂ ...a _p ；

15) Calculating the contribution rate of the sample data:

16) extracting data characteristics of the original data set by using a principal component analysis method, and selecting the characteristics with the accumulative contribution rate of 90% as the input characteristics of a subsequent Bayes recurrent neural network: the contribution rates are arranged according to a descending order, the first k principal components are taken to satisfy

Wherein T is ₀ ＝90％

And the data set after the data characteristic extraction is according to the following steps of 8: 2 into a training set and a test set.

Preferably, the specific step of selecting the optimal recurrent neural network in the step 2) includes:

21) three mainstream recurrent neural networks were selected: RNN, LSTM, GRU recurrent neural networks;

22) respectively predicting characteristic data of the fuel cell during operation by using three kinds of circulating neural networks, and comparing the prediction results of the three kinds of networks;

23) and comparing and analyzing the prediction results of the three networks by taking the mean square error, the root mean square error and the average absolute error as indexes, and selecting the optimal recurrent neural network according to the comparison result.

Preferably, the step 3) of establishing the bayesian neural network specifically comprises the following steps:

31) assuming that a network parameter in the neural network is ω, p (ω) is a prior distribution of the parameter, and a given training sample is denoted as D ═ X, Y, where X is input data and Y is label data, a value of the network parameter ω is related to the training sample, and ω obeys a certain distribution and is denoted as p (ω | D), a predicted value is expressed as:

wherein p (ω | D) is a posterior probability, p (D | ω) is a likelihood function, and p (D) is an edge function; because omega is a random variable, the predicted value is also a random variable and is used for representing the uncertainty of the predicted value; y is ^* Is a predicted value; Δ ω is the minimum infinitesimal of the network parameter ω;

32) since the posterior probability p (ω | D) in the bayesian neural network cannot be directly solved, the posterior probability needs to be approximated by variational inference and a required loss function is constructed;

using a variation inference method, the true a posteriori probability p (ω | D) is approximated by a distribution q (ω | θ) controlled by a set of parameters θ, approximated by a gaussian distribution, i.e. with θ equal to (μ, σ), and each network parameter ω is then approximated by (μ, σ) _i Compliance parameter is (mu) _i ,σ _i ) (ii) a gaussian distribution of; the difference of the KL divergence measures q (ω | θ) and p (ω | D) is used:

θ ^* ＝argmin KL[q(ω|θ||p(ω|D)]，

in the formula, theta ^* Characterizing the difference between q (ω | θ) and p (ω | D), in relation to the parameter θ;

according to the formula of KL divergence, the following is further deduced:

since the edge function p (D) is independent of θ, the loss function is written as:

in the formula, E _q(ω|θ) Represents a mathematical expectation obeying a q (ω | θ) distribution; l (D, θ) is a loss function whose value varies with the parameter θ;

33) using Monte Carlo to simulate the loss function approximation to obtain:

in the formula, theta _i Is a parameter omega _i The gaussian distribution obeyed; x _j Is j input data; y is _j Is X _j Corresponding tag data;

34) scaling the model; suppose ^ divide the entire dataset into M batches, and average each small batch of data, resulting in:

in the formula, L _i ^EQ (D _i θ) represents the loss function for each small batch of data, i.e.

35) Updating network parameters by a gradient descent method, namely:

in the formula, theta' is the updated parameter theta; v _θ Is the gradient with respect to the parameter θ; alpha is the learning rate; l is the loss function L (D, theta).

Preferably, the specific steps of step 4) include:

41) inputting the data processed by the principal component analysis algorithm in the step 1) into the Bayesian neural network;

42) from a Gaussian distribution (. mu.) _i ,σ _i ) Middle sampling to obtain network parameter omega _i ；

43) Separately calculate log q (ω) _i |θ _i )、log p(ω _i )、log p(Y _j |ω,X _j ) Thereby obtaining a loss function

44) Repeatedly updating the parameter θ '═ θ' - α _θ L until the loss function takes a minimum value.

Preferably, the specific steps of step 5) include:

51) preprocessing the operation data of the fuel cell by using a principal component analysis method, and selecting the characteristic that the accumulated contribution rate reaches 90% to realize data dimension reduction;

52) predicting the selected characteristic data by using the optimal recurrent neural network selected in the step 2);

53) because the uncertain factors of the fuel cell operation data are quantified by using probability density distribution, the result of the predicted voltage obeys the distribution rule, and a confidence interval is manually fetched;

54) and taking the prediction result of the circulating neural network as the input of the trained Bayes neural network to obtain the distribution and interval of the predicted voltage of the fuel cell at each moment.

The invention also provides a device for predicting the service life of the proton exchange membrane fuel cell based on the Bayesian recurrent neural network, which is characterized by comprising a data set preprocessing unit, a recurrent neural network selecting unit, a Bayesian neural network establishing unit, a model training unit and a fuel cell service life predicting unit:

the data set preprocessing unit: preprocessing a given data set, extracting main data features in the data set by adopting a principal component analysis method, realizing data dimension reduction, and simultaneously processing the processed data according to the following steps of 8: 2, dividing the ratio into a training set and a testing set;

the selection unit of the recurrent neural network comprises: respectively predicting characteristic data of the fuel cell during operation by using RNN, LSTM and GRU, comparing and analyzing prediction results of the three networks by using mean square error, root mean square error and average absolute error as indexes, and selecting an optimal recurrent neural network according to the comparison result;

the Bayesian neural network establishing unit: replacing fixed parameters in the deep learning neural network with random variables, and quantifying uncertainty through probability density distribution to establish a Bayesian neural network;

the model training unit: taking the data in the training set as the input of a Bayesian neural network, training network parameters until a global optimal solution which enables a loss function to reach the minimum is found;

the fuel cell life prediction unit: and taking the fuel cell characteristic data predicted by the optimal circulating neural network as the input of the Bayesian neural network to obtain the predicted voltage interval of the fuel cell at each moment.

Further, in the bayesian neural network establishing unit, the bayesian neural network is established according to the following steps:

31) assuming that a network parameter in the neural network is ω, p (ω) is a prior distribution of the parameter, and a given training sample is denoted as D ═ X, Y, where X is input data and Y is label data, a value of the network parameter ω is related to the training sample, and ω obeys a certain distribution and is denoted as p (ω | D), a predicted value is expressed as:

where p (ω | D) is a posterior probability, p (D | ω) is a likelihood function, and p (D) is an edge function; because omega is a random variable, the predicted value is also a random variable and is used for representing the uncertainty of the predicted value; y is ^* Is a predicted value; Δ ω is the minimum infinitesimal of the network parameter ω;

32) since the posterior probability p (ω | D) in the bayesian neural network cannot be directly solved, the posterior probability needs to be approximated by variational inference and a required loss function is constructed;

using a method of variational inference, the true a posteriori probability p (ω | D) is approximated by a distribution q (ω | θ) controlled by a set of parameters θ, approximated using a gaussian distribution, i.e. let θ be (μ, σ), then each network parameter ω is approximated by (μ, σ) and the parameters p (ω | D) are approximated by a set of parameters θ _i Compliance parameter is (mu) _i ,σ _i ) (ii) a gaussian distribution of; the difference of KL divergence measures q (ω | θ) and p (ω | D) is used:

θ ^* ＝argmin KL[q(ω|θ||p(ω|D)]，

in the formula, theta ^* Characterizing the difference between q (ω | θ) and p (ω | D), in relation to the parameter θ;

according to the formula of KL divergence, the following is further deduced:

since the edge function p (D) is independent of θ, the loss function is written as:

in the formula, E _q(ω|θ) Represents a mathematical expectation obeying a q (ω | θ) distribution; l (D, θ) is a loss function whose value varies with the parameter θ;

33) using Monte Carlo to simulate the loss function approximation to obtain:

in the formula, theta _i Is a parameter omega _i The gaussian distribution obeyed; x _j Is j input data; y is _j Is X _j Corresponding tag data;

34) scaling the model; suppose ^ divide the entire dataset into M batches, and average each small batch of data, resulting in:

in the formula, L _i ^EQ (D _i θ) represents a loss function for each small batch of data, i.e.

35) Updating network parameters by a gradient descent method, namely:

in the formula, theta' is the updated parameter theta; v _θ Is the gradient with respect to the parameter θ; alpha is the learning rate; l is the loss function L (D, theta).

Further, in the model training unit, a bayesian neural network is trained as follows:

41) inputting the data processed by the principal component analysis algorithm into the Bayesian neural network;

42) from a Gaussian distribution (. mu.) _i ,σ _i ) Middle sampling to obtain network parameter omega _i ；

43) Separately calculate log q (ω) _i |θ _i )、log p(ω _i )、log p(Y _j |ω,X _j ) Thereby obtaining a loss function

44) Repeatedly updating the parameter θ '═ θ' - α + _θ L until the loss function takes a minimum value.

The present invention further provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the computer program implements the method for predicting lifetime of a proton exchange membrane fuel cell based on a bayesian recurrent neural network described above.

Compared with the prior art, the invention has the beneficial effects that: firstly, the method uses the optimal cyclic neural network to predict the operation data of the fuel cell without using a large amount of measurement data; and secondly, quantifying uncertainty factors of the prediction data by using probability density distribution, introducing the uncertainty factors into the prediction of the residual service life of the fuel cell, and establishing a proton exchange membrane fuel cell service life prediction model based on the Bayesian recurrent neural network, so that the point estimation and interval estimation can be performed on the predicted voltage at each moment, and the fault tolerance rate and accuracy of the prediction of the service life of the fuel cell are improved.

In addition, the data set is subjected to data preprocessing by adopting a principal component analysis method, and main characteristic data is selected as the input of the Bayesian recurrent neural network, so that the complexity of the data set is reduced, and the model solving speed is higher.

Drawings

FIG. 1 is a flow chart of a life prediction method of a proton exchange membrane fuel cell based on a Bayesian recurrent neural network of the present invention;

FIG. 2 is a flow chart of the optimal recurrent neural network selection of the present invention;

fig. 3 is a flowchart of the bayesian neural network training process in the present invention.

Fig. 4 is a distribution diagram of the predicted voltage at some time points in the embodiment of the present invention, which is a distribution diagram of the predicted voltage at the time points of 400h, 430h, 460h, 490h, 520h, 550h, 580h, 610h, 640h, 670h, and 700h from left to right in sequence.

Detailed Description

The invention is described in further detail below with reference to the figures and specific embodiments.

The following detailed description of the embodiments of the present invention will be made with reference to the drawings and examples, which are provided for illustrating the present invention and are not intended to limit the scope of the present invention.

The invention discloses a life prediction method of a proton exchange membrane based on a Bayesian recurrent neural network, which comprises the following steps: preprocessing original data by using a Principal Component Analysis (PCA), selecting characteristic data with the accumulative contribution rate exceeding 90%, and carrying out data analysis according to the ratio of 8: 2, dividing the training set into a training set and a testing set; establishing different cyclic neural networks, respectively predicting different characteristic data, and selecting a proper neural network according to a prediction result; replacing fixed parameters in the deep learning neural network model with random variables, and quantifying uncertainty of each characteristic data through probability density distribution to establish a Bayesian neural network; fourthly, training network parameters of the Bayesian neural network by using the training set data; after the network training is finished, inputting the previous operation data of the fuel cell into a circulating neural network to obtain the operation result of the characteristic data, and inputting the characteristic data into a Bayesian neural network to finally obtain the interval of the predicted voltage of the fuel cell.

Examples

The embodiment provides a lifetime prediction method of a proton exchange membrane based on a Bayesian cycle neural network, which comprises five steps of sample data set preprocessing, namely, selection of characteristic data, selection of an optimal cycle neural network, construction of the Bayesian neural network, model training and lifetime prediction of a fuel cell, and specifically comprises the following steps:

1) and carrying out data preprocessing on a given sample data set, extracting main characteristic data in the data set by adopting a principal component analysis method, realizing data dimension reduction, and dividing the processed data set into a training set and a testing set.

2) Establishing three different recurrent neural networks including RNN, LSTM and GRU, predicting the characteristic numbers selected in the step 1) respectively, and selecting the optimal recurrent neural network according to the prediction result.

3) Fixed parameters in the classical deep learning neural network are replaced by random variables, uncertainty factors contained in various types of feature data are quantified through probability density distribution, and the Bayesian neural network is established.

4) And taking the data in the training set as the input of the Bayesian neural network, and training network parameters until a global optimal solution which enables the loss function to be minimum is found.

5) And predicting the characteristic data of the fuel cell by using the optimal cyclic neural network, taking the predicted result as the input of the trained Bayesian neural network, and giving the interval of the predicted voltage at each moment.

Step 1): preprocessing sample data by using a principal component analysis method, and extracting characteristic data, wherein the specific method comprises the following steps:

step 11) recording an original sample-index data matrix according to the fuel cell operation data as follows:

wherein x ₁ ,x ₂ ,…,x _p Respectively, index data of operation of the fuel cell, i.e. x _p ＝[x _1p ,x _2p ,…x _np ] ^T N is the number of samples, and p is the number of indexes when the fuel cell operates;

step 12) standardizing the operation sample data of the fuel cell, wherein Var () represents the mean square error of the sample data:

step 13) calculating a sample correlation coefficient matrix:

wherein r is _ij The closer | is to 1, the variable x _i ,x _j The more similar, the less similar if closer to 0.

Step 14) calculating the eigenvalues and corresponding eigenvectors of the correlation coefficient matrix R:

characteristic value: lambda [ alpha ] ₁ ,λ ₂ ...λ _p

Feature vector: a is ₁ ,a ₂ ...a _p

Step 15) calculating the contribution rate:

step 16) selecting important principal components:

the contribution rates are arranged according to a descending order, the first k principal components are taken to satisfy

Wherein T is ₀ ＝90％

Extracting data characteristics of an original data set by using a principal component analysis method, selecting the characteristics with the accumulative contribution rate of 90% as the input characteristics of a subsequent Bayes recurrent neural network, and processing the data according to the following ratio of 8: 2 into a training set and a test set.

In this embodiment, selecting the main components includes:

step 2) using three kinds of circulating neural networks to respectively predict the characteristics selected in the step one, and comparing the prediction effects, wherein the specific steps are as follows:

step 21) selecting three mainstream recurrent neural networks, namely RNN, LSTM and GRU recurrent neural networks;

step 22) using three kinds of circulating neural networks to respectively predict the characteristic data of the fuel cell during operation and comparing the prediction results of the three kinds of networks;

and step 23) comparing and analyzing the prediction results of the three networks by taking the mean square error, the root mean square error and the average absolute error as indexes, and selecting the optimal recurrent neural network according to the comparison result.

Through comparison of prediction effects and error analysis, the selected optimal recurrent neural network is GRU, and prediction errors of the GRU on certain characteristics of the fuel cell are as follows:

step 3) establishing a Bayesian neural network according to the following steps, as shown in FIG. 2:

step 31) assuming that a network parameter in the neural network is ω, p (ω) is prior distribution of the parameter, and a given training sample is denoted as D ═ X, Y, where X is input data and Y is label data, a value of the network parameter ω is related to the training sample, and ω obeys a certain distribution and is denoted as p (ω | D), then a predicted value is expressed as:

wherein p (ω | D) is a posterior probability, p (D | ω) is a likelihood function, and p (D) is an edge function; because omega is a random variable, the predicted value is also a random variable and is used for representing the uncertainty of the predicted value; y is ^* Is a predicted value; Δ ω is the minimum infinitesimal of the network parameter ω;

step 32), because the posterior probability p (omega | D) in the Bayesian neural network can not be directly solved, the posterior probability needs to be approximated by variational inference and a required loss function is constructed;

using a method of variational inference, the true a posteriori probability p (ω | D) is approximated by a distribution q (ω | θ) controlled by a set of parameters θ, approximated using a gaussian distribution, i.e. let θ be (μ, σ), then each network parameter ω is approximated by (μ, σ) and the parameters p (ω | D) are approximated by a set of parameters θ _i Compliance parameter is (mu) _i ,σ _i ) (ii) a gaussian distribution of; using the KL divergence measure difference of q (ω | θ) and p (ω | D), i.e. optimizing:

θ ^* ＝argmin KL[q(ω|θ||p(ω|D)]，

in the formula, theta ^* Characterizing the difference between q (ω | θ) and p (ω | D), in relation to the parameter θ;

according to the formula of KL divergence, the following is further deduced:

since the edge function p (D) is independent of θ, the loss function is written as:

in the formula, E _q(ω|θ) Represents a mathematical expectation obeying a q (ω | θ) distribution; l (D, θ) is a loss function whose value varies with the parameter θ;

step 33) approximating by a Monte Carlo simulation loss function to obtain:

in the formula, theta _i Is a parameter omega _i The gaussian distribution obeyed; x _j Is j input data; y is _j Is X _j Corresponding tag data;

step 34) in order to reduce the complexity of the model, the model needs to be scaled to a certain extent; suppose ^ divide the entire dataset into M batches, and average each small batch of data, resulting in:

in the formula, L _i ^EQ (D _i θ) represents a loss function for each small batch of data, i.e.

Step 35) updating the network parameters by a gradient descent method, namely:

in the formula, theta is the updated parameter theta; v _θ Is the gradient with respect to parameter θ; alpha is the learning rate; l is the loss function L (D, theta).

Step 4) training a Bayesian neural network according to the following steps, as shown in FIG. 3:

step 41) inputting the data processed by the principal component analysis algorithm into a network;

step 42) from a Gaussian distribution (μ) _i ,σ _i ) Middle sampling to obtain network parameter omega _i ；

Step 43) separately calculating log q (ω) _i |θ _i )、log p(ω _i )、log p(Y _j |ω,X _j ) Thereby obtaining a loss function

Step 44) repeatedly updating the parameter θ '═ θ' - α ∑ Δ ∑ _θ L until the loss function takes a minimum value.

Step 5) inputting the prediction result of the optimal cyclic neural network into the trained Bayes neural network to obtain the predicted voltage distribution and interval at each moment, and the specific steps are as follows:

step 51) preprocessing the operation data of the fuel cell by using a principal component analysis method, and selecting the characteristic that the accumulated contribution rate reaches 90 percent to realize data dimension reduction;

step 52) using the optimal recurrent neural network to predict the selected characteristic data;

step 53) because the uncertain factors of the fuel cell operation data are quantified by using probability density distribution, and the result of the predicted voltage follows certain distribution, a certain confidence interval needs to be manually selected;

and step 54) taking the prediction result of the circulating neural network as the input of the trained Bayes neural network to obtain the distribution and interval of the predicted voltage of the fuel cell at each moment.

Based on the method, the invention provides a life prediction device of a proton exchange membrane fuel cell based on a Bayesian circulation neural network, which comprises a data set preprocessing unit, a selection unit of the circulation neural network, a Bayesian neural network establishing unit, a model training unit and a life prediction unit of the fuel cell; wherein the content of the first and second substances,

a data set preprocessing unit: preprocessing a given data set, extracting main data features in the data set by adopting a principal component analysis method, realizing data dimension reduction, and simultaneously processing the processed data according to the following steps of 8: 2, dividing the ratio of the training set to the test set, and executing the process of selecting the characteristic data in the step 1);

the selection unit of the recurrent neural network: respectively predicting characteristic data of the fuel cell during operation by using RNN, LSTM and GRU, comparing and analyzing prediction results of the three networks by using mean square error, root mean square error and mean absolute error as indexes, selecting an optimal cyclic neural network according to the comparison result, and executing the selection process of the optimal cyclic neural network in the step 2);

a Bayesian neural network establishing unit: replacing fixed parameters in the deep learning neural network with random variables, carrying out uncertainty quantification through probability density distribution, establishing a Bayesian neural network, and executing the construction process of the Bayesian neural network in the step 3);

a model training unit: taking the data in the training set as the input of a Bayesian neural network, training network parameters until a global optimal solution which enables a loss function to be minimum is found, and executing the model training process in the step 4);

fuel cell life prediction unit: and (5) taking the fuel cell characteristic data predicted by the optimal cyclic neural network as the input of the Bayesian neural network to obtain the predicted voltage interval of the fuel cell at each moment, and executing the process of step 5) predicting the service life of the fuel cell.

The invention further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the life prediction method of the proton exchange membrane based on the bayesian recurrent neural network.

Finally, it should be noted that the above-mentioned embodiments are only for illustrating the patent solution and not for limiting, and those skilled in the art should understand that the technical solution of the patent can be modified or substituted with equivalent without departing from the spirit and scope of the patent solution, which shall be covered by the claims of the patent.

Claims (10)

1. A life prediction method of a proton exchange membrane based on a Bayesian recurrent neural network is characterized by comprising the following steps:

1) performing data preprocessing on a fuel cell operation sample data set, extracting main characteristic data in the data set by adopting a principal component analysis method, realizing data dimension reduction, and dividing the processed data set into a training set and a testing set;

2) establishing three different recurrent neural networks to respectively predict the characteristic numbers selected in the step 1), and selecting the optimal recurrent neural network according to the prediction result;

3) replacing fixed parameters in the classical deep learning neural network with random variables, quantifying uncertainty factors contained in various types of characteristic data through probability density distribution, and establishing a Bayesian neural network;

4) taking the data in the training set as the input of a Bayesian neural network, training network parameters until a global optimal solution which enables a loss function to reach the minimum is found;

5) preprocessing the operation data of the fuel cell by using a principal component analysis method, predicting the selected characteristic data by using the optimal cyclic neural network selected in the step 2), and taking the predicted result as the input of the trained Bayesian neural network to obtain the distribution and interval of the predicted voltage of the fuel cell at each moment.

2. The proton exchange membrane life prediction method based on the Bayesian recurrent neural network as recited in claim 1, wherein: step 1) the specific steps of data feature extraction include:

11) according to the fuel cell operation sample data, recording an original sample-index data matrix as follows:

wherein x ₁ ,x ₂ ,…,x _p Respectively, index data of operation of the fuel cell, i.e. x _p ＝[x _1p ,x _2p ,…x _np ] ^T N is the number of samples, and p is the number of indexes when the fuel cell operates;

12) standardizing the operation sample data of the fuel cell, wherein Var () represents the mean square error of the sample data:

13) calculating a sample data correlation coefficient matrix:

wherein r is _ij The closer | is to 1, the variable x _i ,x _j The more similar, the less similar if closer to 0;

14) calculating the eigenvalues and corresponding eigenvectors of the correlation coefficient matrix R:

characteristic value: lambda [ alpha ] ₁ ,λ ₂ ...λ _p

Feature vector: a is a ₁ ,a ₂ ...a _p ；

15) Calculating the contribution rate of the sample data:

16) extracting data characteristics of the original data set by using a principal component analysis method, and selecting the characteristics with the accumulative contribution rate of 90% as the input characteristics of a subsequent Bayes recurrent neural network: the contribution rates are arranged according to a descending order, the first k principal components are taken to satisfy

Wherein T is ₀ ＝90％

And the data set after the data characteristic extraction is according to the following steps of 8: 2 into a training set and a test set.

3. The Bayesian recurrent neural network-based proton exchange membrane life prediction method as recited in claim 1, wherein: step 2) the concrete steps of selecting the optimal recurrent neural network comprise:

21) three mainstream recurrent neural networks were selected: RNN, LSTM, GRU recurrent neural networks;

22) respectively predicting characteristic data of the fuel cell during operation by using three kinds of circulating neural networks, and comparing the prediction results of the three kinds of networks;

23) and comparing and analyzing the prediction results of the three networks by taking the mean square error, the root mean square error and the average absolute error as indexes, and selecting the optimal recurrent neural network according to the comparison result.

4. The proton exchange membrane life prediction method based on the Bayesian recurrent neural network as recited in claim 1, wherein: step 3) the specific steps of establishing the Bayesian neural network comprise:

31) assuming that a network parameter in the neural network is ω, p (ω) is a prior distribution of the parameter, and a given training sample is denoted as D ═ X, Y, where X is input data and Y is label data, a value of the network parameter ω is related to the training sample, and ω obeys a certain distribution and is denoted as p (ω | D), a predicted value is expressed as:

wherein p (ω | D) is a posterior probability, p (D | ω) is a likelihood function, and p (D) is an edge function; because omega is a random variable, the predicted value is also a random variable and is used for representing the uncertainty of the predicted value; y is ^* Is a predicted value; Δ ω is the minimum infinitesimal of the network parameter ω;

32) since the posterior probability p (ω | D) in the bayesian neural network cannot be directly solved, the posterior probability needs to be approximated by variational inference and a required loss function is constructed;

using a variation inference method, the true a posteriori probability p (ω | D) is approximated by a distribution q (ω | θ) controlled by a set of parameters θ, approximated by a gaussian distribution, i.e. with θ equal to (μ, σ), and each network parameter ω is then approximated by (μ, σ) _i Compliance parameter is (mu) _i ,σ _i ) (ii) a gaussian distribution of; the difference of KL divergence measures q (ω | θ) and p (ω | D) is used:

θ ^* ＝argmin KL[q(ω|θ||p(ω|D)]，

in the formula, theta ^* Characterizing the difference between q (ω | θ) and p (ω | D), in relation to the parameter θ;

according to the formula of KL divergence, the following is further deduced:

since the edge function p (D) is independent of θ, the loss function is written as:

in the formula, E _q(ω|θ) Represents a mathematical expectation obeying a q (ω | θ) distribution; l (D, θ) is a loss function whose value varies with the parameter θ;

33) using Monte Carlo to simulate the loss function approximation to obtain:

in the formula, theta _i Is a parameter omega _i The gaussian distribution obeyed; x _j Is j input data; y is _j Is X _j Corresponding tag data;

34) scaling the model; suppose ^ divide the entire dataset into M batches, and average each small batch of data, resulting in:

in the formula, L _i ^EQ (D _i θ) represents the loss function for each small batch of data, i.e.

35) Updating network parameters by a gradient descent method, namely:

in the formula, theta' is the updated parameter theta; v _θ Is the gradient with respect to the parameter θ; alpha is the learning rate; l is the loss function L (D, theta).

5. The Bayesian recurrent neural network-based proton exchange membrane life prediction method as recited in claim 3, wherein: the specific steps of step 4) include:

41) inputting the data processed by the principal component analysis algorithm in the step 1) into the Bayesian neural network;

42) from a Gaussian distribution (. mu.) _i ,σ _i ) Middle sampling to obtain network parameter omega _i ；

43) Separately calculate log q (ω) _i |θ _i )、log p(ω _i )、log p(Y _j |ω,X _j ) Thereby obtaining a loss letterNumber of

44) Repeatedly updating the parameter θ '═ θ' - α + _θ L until the loss function takes a minimum value.

6. The method for predicting the life of the proton exchange membrane based on the Bayesian recurrent neural network as recited in claim 1, wherein: the specific steps of step 5) include:

51) preprocessing the operation data of the fuel cell by using a principal component analysis method, and selecting the characteristic that the accumulated contribution rate reaches 90% to realize data dimension reduction;

52) predicting the selected characteristic data by using the optimal recurrent neural network selected in the step 2);

53) because the uncertain factors of the fuel cell operation data are quantified by using probability density distribution, the result of the predicted voltage obeys the distribution rule, and a confidence interval is manually fetched;

54) and taking the prediction result of the circulating neural network as the input of the trained Bayes neural network to obtain the distribution and interval of the predicted voltage of the fuel cell at each moment.

7. A proton exchange membrane fuel cell life prediction device based on Bayesian recurrent neural network is characterized in that: the device comprises a data set preprocessing unit, a selection unit of a cyclic neural network, a Bayesian neural network establishing unit, a model training unit and a fuel cell service life predicting unit:

the data set preprocessing unit: preprocessing a given data set, extracting main data features in the data set by adopting a principal component analysis method, realizing data dimension reduction, and simultaneously processing the processed data according to the following steps of 8: 2, dividing the ratio into a training set and a testing set;

the selection unit of the recurrent neural network comprises: respectively predicting characteristic data of the fuel cell during operation by using RNN, LSTM and GRU, comparing and analyzing prediction results of the three networks by using mean square error, root mean square error and average absolute error as indexes, and selecting an optimal recurrent neural network according to the comparison result;

the Bayesian neural network establishing unit: replacing fixed parameters in the deep learning neural network with random variables, and quantifying uncertainty through probability density distribution to establish a Bayesian neural network;

the model training unit: taking the data in the training set as the input of a Bayesian neural network, training network parameters until a global optimal solution which enables a loss function to reach the minimum is found;

the fuel cell life prediction unit: and taking the fuel cell characteristic data predicted by the optimal circulating neural network as the input of the Bayesian neural network to obtain the predicted voltage interval of the fuel cell at each moment.

8. The device for predicting the life of the proton exchange membrane fuel cell based on the Bayesian recurrent neural network as recited in claim 7, wherein: in the Bayesian neural network establishing unit, establishing a Bayesian neural network according to the following steps:

31) assuming that a network parameter in the neural network is ω, p (ω) is a prior distribution of the parameter, and a given training sample is denoted as D ═ X, Y, where X is input data and Y is label data, a value of the network parameter ω is related to the training sample, and ω obeys a certain distribution and is denoted as p (ω | D), a predicted value is expressed as:

wherein p (ω | D) is a posterior probability, p (D | ω) is a likelihood function, and p (D) is an edge function; because omega is a random variable, the predicted value is also a random variable and is used for representing the uncertainty of the predicted value; y is ^* Is a predicted value; Δ ω is the minimum infinitesimal of the network parameter ω;

32) since the posterior probability p (omega | D) in the Bayesian neural network can not be directly solved, the posterior probability needs to be approximated by variation inference and a required loss function is constructed;

using a variation inference method, the true a posteriori probability p (ω | D) is approximated by a distribution q (ω | θ) controlled by a set of parameters θ, approximated by a gaussian distribution, i.e. with θ equal to (μ, σ), and each network parameter ω is then approximated by (μ, σ) _i Compliance parameter is (mu) _i ,σ _i ) (ii) a gaussian distribution of; the difference of KL divergence measures q (ω | θ) and p (ω | D) is used:

θ ^* ＝argmin KL[q(ω|θ||p(ω|D)]，

in the formula, theta ^* Characterizing the difference between q (ω | θ) and p (ω | D), in relation to the parameter θ;

according to the formula of KL divergence, the following is further deduced:

since the edge function p (D) is independent of θ, the loss function is written as:

in the formula, E _q(ω|θ) Represents a mathematical expectation obeying a q (ω | θ) distribution; l (D, θ) is a loss function whose value varies with the parameter θ;

33) using Monte Carlo to simulate the loss function approximation to obtain:

in the formula, theta _i Is a parameter omega _i The gaussian distribution obeyed; x _j Is j input data; y is _j Is X _j Corresponding tag data;

34) scaling the model; assuming that the entire dataset is divided into M batches, and averaging each small batch of data, we get:

in the formula, L _i ^EQ (D _i θ) represents the loss function for each small batch of data, i.e.

35) Updating network parameters by a gradient descent method, namely:

in the formula, theta' is the updated parameter theta; v _θ Is the gradient with respect to parameter θ; alpha is the learning rate; l is the loss function L (D, theta).

9. The device for predicting the life of a proton exchange membrane fuel cell based on the Bayesian recurrent neural network as recited in claim 8, wherein: in the model training unit, training a Bayesian neural network according to the following steps:

41) inputting the data processed by the principal component analysis algorithm into the Bayesian neural network;

42) from a Gaussian distribution (. mu.) _i ,σ _i ) Middle sampling to obtain network parameters omega _i ；

43) Separately calculate log q (ω) _i |θ _i )、log p(ω _i )、log p(Y _j |ω,X _j ) Thereby obtaining a loss function

44) Repeatedly updating the parameter θ '═ θ' - α + _θ L until the loss function takes a minimum value.

10. A computer-readable storage medium, storing a computer program, characterized in that the computer program, when being executed by a processor, carries out the method of any one of claims 1 to 6.

Images