Lithium ion battery remaining service life prediction method based on DAE-ResBiLSTM-MHA

Abstract

The invention provides a lithium ion battery remaining service life prediction method based on DAE-ResBiLSTM-MHA. The lithium ion battery remaining service life prediction method comprises the following steps: step 1, constructing a lithium ion battery data set; step 2, carrying out normalization processing on the battery capacity value in the new data set; step 3, using a de-noising auto-encoder to remove noise introduced by noise interference and capacity regeneration in the acquisition process in the battery data set; 4, on the basis of the BiLSTM model, a ResBiLSTM-MHA model is built in combination with residual connection and a multi-head attention mechanism; 5, dividing a data set by adopting a leave-one-out method cross validation, namely randomly selecting one battery sample as a test set each time, and using the rest samples as a training set and a verification set; training the ResBiLSTM-MHA model by using the training set, and preliminarily evaluating the predictive ability of the model through the verification set; step 6, the RUL prediction performance of the DAE-ResBiLSTM-MHA model is evaluated; by applying the technical scheme, the robustness and accuracy of the model facing noise data and capacity regeneration phenomena can be improved.

Description

Technical Field

[0001] The invention relates to the technical field of lithium ion batteries, and in particular to a method for predicting the remaining service life of a lithium ion battery based on DAE-ResBiLSTM-MHA. Background Art

[0002] With the increase in the number of cars per capita and the number of fuel-powered cars, the demand for oil and natural gas has also risen sharply, which has led to an increase in automobile exhaust emissions and an upward trend, which has brought about serious energy crises and environmental problems. Therefore, the development of new energy electric vehicles can solve the above problems and has become a key issue of social concern. Lithium-ion batteries are the main power source of new energy electric vehicles. They have the advantages of long life, high energy density, wide operating temperature range, low self-discharge rate and low environmental pollution. They are one of the most popular new energy sources. At present, they have been widely used in electronic products, transportation, electric vehicles, energy storage systems, aerospace and military fields. However, with the increase in the number of charge and discharge cycles, the complex electrochemical reactions inside lithium-ion batteries cause irreversible performance degradation, which in turn leads to performance degradation and shortened service life, performance degradation of electrical equipment or system failure, and even combustion or explosion accidents. Therefore, it is crucial to study methods to accurately predict the remaining service life of lithium-ion batteries in the field of new energy vehicles. At present, researchers have developed some methods for predicting the remaining service life of lithium-ion batteries based on deep learning and have made significant progress. However, when these methods process battery data sets, they are still less accurate in prediction in the face of noise interference and capacity regeneration that may be encountered during the acquisition process. Summary of the invention

[0003] In view of this, the purpose of the present invention is to provide a method for predicting the remaining useful life of lithium-ion batteries based on DAE-ResBiLSTM-MHA, so as to improve the robustness and accuracy of the model in the face of noisy data and capacity regeneration phenomenon.

[0004] To achieve the above object, the present invention adopts the following technical solution: a method for predicting the remaining service life of a lithium-ion battery based on DAE-ResBiLSTM-MHA, comprising the following steps:

[0005] Step 1: Build a lithium-ion battery dataset;

[0006] Step 2: Normalize the battery capacity values ​​in the newly created data set;

[0007] Step 3: Apply a denoising autoencoder to remove the noise introduced by the noise interference and capacity regeneration phenomenon in the battery dataset during the acquisition process;

[0008] Step 4: Based on the BiLSTM model, the ResBiLSTM-MHA model is built by combining residual connection and multi-head attention mechanism;

[0009] Step 5: Use leave-one-out cross-validation to divide the data set, that is, randomly select one battery sample as the test set each time, and use the remaining samples as the training set and validation set; use the training set to train the ResBiLSTM-MHA model, and use the validation set to preliminarily evaluate the model's predictive ability;

[0010] Step 6: Evaluate the RUL prediction performance of the DAE-ResBiLSTM-MHA model.

[0011] In a preferred embodiment, the step 1 specifically includes: downloading the NASA data set and the CACLE data set; wherein the NASA data set is divided into 9 groups, with 3 to 4 batteries in each group, and the rated capacity is 2Ah; the charging stage adopts a constant current mode CC of 1.5A until the battery voltage reaches 4.2V, and then switches to a constant voltage mode CV until the charging current drops to 20mA; the discharging stage is carried out under a constant current mode CC of 2A until the battery voltage drops to a preset value; the experiment stops when the lithium-ion battery reaches a service life failure threshold, that is, stops when the rated capacity drops by 30%; select the four lithium-ion batteries CS_2 in the CACLE data set as another battery data set used to verify the generalization of the model; its charging mode adopts a constant current mode CC until the battery voltage reaches 4.2V, and then switches to a constant voltage mode CV until the charging current drops to 20mA; the discharging stage is carried out under a constant current mode until the battery voltage drops to 2.7V; when the battery reaches the service life failure threshold, that is, stops when the rated capacity drops by 30%.

[0012] In a preferred embodiment, the step 2 specifically includes: using the minimum-maximum normalization method to normalize the data, and the calculation formula is as follows:

[0013]

[0014] Where C is the input vector, min(C) is the minimum value of the input vector, max(C) is the maximum value of the input vector, and u is the normalized input vector.

[0015] In a preferred embodiment, the step 3 specifically includes:

[0016] Step 31: The denoising autoencoder DAE consists of an input layer, a hidden layer, and an output layer. Before the data is input into the denoising autoencoder, Gaussian noise of the same dimension is added to the input vector to generate a noisy input vector, as shown below:

[0017]

[0018] where u and d is the dimension of the input vector; δ∈N(0,σ 2 );σ is the standard deviation of Gaussian noise;

[0019] Step 32: The encoder compresses the input vector into a low-dimensional intermediate vector e, whose expression is shown in formula (3);

[0020] Then, the decoder reconstructs the input vector from the intermediate vector, which is expressed as shown in formula (4);

[0021]

[0022] where e∈R m is the output value of the hidden layer; m is the number of neurons in the hidden layer; is the output vector of the decoder; W1∈R m×d and W2∈R d×m is the weight matrix; b∈R m and c∈R d is the bias vector; f and g are nonlinear activation functions.

[0023] In a preferred embodiment, the step 4 specifically includes:

[0024] Step 41: BiLSTM is composed of two LSTM neural networks with opposite directions. The LSTM neural network selectively retains and updates the information in the hidden state by introducing a gating mechanism, including a forget gate, an input gate, and an output gate. The calculation formulas of the input gate, the forget gate, and the output gate are as follows:

[0025] f t =φ(W f ·[h t-1 ,x t ]+b f ) (5)

[0026] i t =φ(W i ·[h t-1 ,x t ]+b i ) (6)

[0027]

[0028] o t =φ(W o ·[h t-1 ,x t ]+b o ) (8)

[0029] h t =o t *tanh(C t ) (9)

[0030] Where φ is the sigmoid activation function; tanh is the hyperbolic tangent activation function; W f ,W i , W C and W o are the weight matrices of the forget gate, input gate, internal state, and output gate respectively; b f ,b i ,b C and b o They represent the bias vectors of the forget gate, input gate, internal state and output gate respectively; h t ,x t and C t , respectively represent the hidden, input and memory states of the LSTM unit at time t; * represents the element-by-element multiplication of the matrix;

[0031] Step 42: The ResBiLSTM network contains two layers of ResBiLSTM, each layer of ResBiLSTM consists of a BiLSTM and a residual connection; there is a layer of normalization LN after each layer of ResBiLSTM; the calculation formula of each layer of ResBiLST-M is as follows:

[0032]

[0033] where x (i) represents the feature vector of the i-th input sample; μ represents the feature vector x (i) The mean of (i) The standard deviation of Represents the feature vector after layer normalization (LN); L is the processing of the input state in the LSTM network; is the forward output vector of the forward LSTM input at time t after residual connection and LN; is the reverse output vector of the input of the reverse LSTM at time t after residual connection and LN; y t is the concatenation of the forward and reverse output vectors at time t;

[0034] Step 43: Add a multi-head attention mechanism MHA after the ResBiLSTM network. In the multi-head attention mechanism, first, the output data of the ResBiLSTM network is calculated through a linear layer to obtain three intermediate variable matrices, and then the attention matrix is ​​calculated. Finally, multiple attention mechanism modules are accumulated in parallel, so that the model can obtain multi-dimensional correlation information between data. The calculation formula of the multi-head attention mechanism is as follows:

[0035] Q=W Q X,K=W K X,V=W V ·X (14)

[0036]

[0037] MultiHead(Q,K,V)=Concat(head1,...,head h )W O (16)

[0038] Where W Q ,W k and W V are the weight matrices corresponding to the query, key, and value respectively; X is the input data; d k The dimension of the key; head i is the output of the i-th head; W O is the weight matrix of multi-head attention output;

[0039] Step 44: Use a fully connected layer to map the multi-head attention weighted feature vector to the capacity value;

[0040] Step 45: Calculate the loss function. The error sources of the DAE-ResBiLSTM-MHA model include the error between the input value and the reconstructed value and the error between the true value and the predicted value. MSE is used to calculate the loss function of the DAE-ResBiLSTM-MHA model, and the calculation formula is as follows:

[0041] Loss = L MSE +L REC

[0042]

[0043] Where L MSE and L REC are the loss functions of mean square error and reconstruction error respectively; N is the dimension of input vector; y t and are the label and predicted output of the tth sample respectively; λ is the regularization parameter.

[0044] In a preferred embodiment, the step 5 specifically includes:

[0045] Step 51: Install graphics card driver, CUDA and CUDNN acceleration library;

[0046] Step 52: Configure the training environment;

[0047] Step 53: Write a Python program;

[0048] Step 54: Train the DAE-ResBiLSTM-MHA model;

[0049] Step 55: The computer configuration is NVIDIA GeForce RTX 1050Ti graphics card and Intel Core i5-7300HQCPU@2.50GHz processor.

[0050] In a preferred embodiment, the step 6 specifically includes:

[0051] Step 61: Calculate the relative error RE, mean absolute error MAE and root mean square error RMSE. The expressions are as follows:

[0052]

[0053] in is the predicted value, y t is a real value, n represents the length of the sequence;

[0054] Step 62: Evaluate the performance of the model;

[0055] Step 63: Evaluate the RUL prediction ability of the DAE-ResBiLSTM-MHA model.

[0056] Compared with the prior art, the present invention has the following beneficial effects: improving the robustness and accuracy of the model when facing noise data and capacity regeneration phenomenon. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A flowchart of a preferred embodiment of the present invention;

[0058] Figure 2 A schematic diagram of the structure of a denoising autoencoder according to a preferred embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of the DAE-ResBiLSTM-MHA model structure of a preferred embodiment of the present invention;

[0060] Figure 4 The RUL prediction results of ResBiLSTM, ResBiLSTM-MHA, DAE-ResBiLSTM and DAE-ResBiLSTM-MHA of the preferred embodiment of the present invention on the NASA data set; wherein (a) is battery B0005, (b) is battery B0006, (c) is battery B0007, and (d) is battery B0018;

[0061] Figure 5The RUL prediction results of ResBiLSTM, ResBiLSTM-MHA, DAE-ResBiLSTM and DAE-ResBiLSTM-MHA of the preferred embodiment of the present invention on the CALCE data set; wherein (a) is battery CS2_35, (b) is battery CS2_36, (c) is battery CS2_37, and (d) is battery CS2_38;

[0062] Figure 6 A schematic diagram of battery aging trend of a NASA data set according to a preferred embodiment of the present invention;

[0063] Figure 7 A schematic diagram of battery aging trend of a CACLE data set according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0064] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0065] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0066] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or their combinations.

[0067] A method for predicting the remaining useful life of lithium-ion batteries based on DAE-ResBiLSTM-MHA, referring to Figure 1-7 The specific steps of the method of the present invention are described as follows:

[0068] Step 1: Download NASA and CACLE datasets to construct the lithium-ion battery dataset used in the present invention.

[0069] Step 2: Normalize the battery capacity values ​​in the newly created data set to convert the data to a uniform scale to improve the performance and stability of the model.

[0070] Step 3: Apply a denoising autoencoder to remove the noise introduced by noise interference during the acquisition process and capacity regeneration in the battery dataset, thereby improving data quality and achieving dimensionality reduction.

[0071] Step 4: Based on the BiLSTM model, the ResBiLSTM-MHA model is built by combining residual connection and multi-head attention mechanism.

[0072] Step 5: Use leave-one-out cross-validation to divide the data set, that is, randomly select one battery sample as the test set each time, and use the remaining samples as the training set and validation set. Use the training set to train the ResBiLSTM-MHA model, and use the validation set to preliminarily evaluate the model's predictive ability.

[0073] Step 6: Evaluate the RUL prediction performance of the DAE-ResBiLSTM-MHA model.

[0074] The method for constructing a new data set in step 1 is as follows: download NASA and CACLE data sets. The NASA battery data is divided into 9 groups, with 3 to 4 batteries in each group, and the rated capacity is 2Ah. Considering that the daily battery use environment is at room temperature, and there can only be one variable in the control variable method, the discharge current needs to be kept constant. Therefore, the first group of four lithium-ion batteries (B0005, B0006, B0007 and B0018) of NASA are selected as the battery data set used in this invention. The charging, discharging and impedance tests are performed at room temperature. The charging stage adopts a constant current (CC) mode of 1.5A until the battery voltage reaches 4.2V, and then switches to a constant voltage (CV) mode until the charging current drops to 20mA. The discharge stage is carried out in a constant current mode of 2A until the battery voltage drops to a preset value. The experiment stops when the lithium-ion battery reaches the service life failure threshold, that is, when the rated capacity drops by 30% (from 2Ahr to 1.4Ahr). In order to keep the battery cycle test conditions basically consistent, four lithium-ion batteries of CS_2 in CACLE (CS2_35, CS2_36, CS2_37 and CS2_38) are selected as another battery data set used in this invention to verify the generalization of the model. Its charging mode adopts constant current (CC) mode until the battery voltage reaches 4.2V, and then switches to constant voltage (CV) mode until the charging current drops to 20mA. The discharge phase is carried out in constant current mode until the battery voltage drops to 2.7V. When the battery reaches the service life failure threshold, that is, the rated capacity drops by 30% (from 1.1Ahr to 0.77Ahr), it stops. The battery aging trend diagrams of the NASA and CACLE battery data sets used in the present invention are shown as follows: Figure 5 and Figure 6-7 The battery data normalization process in step 2 is as follows:

[0075] The present invention adopts the minimum-maximum normalization method to normalize the data, and the calculation formula is as follows:

[0076]

[0077] Where C is the input vector, min(C) is the minimum value of the input vector, max(C) is the maximum value of the input vector, and u is the normalized input vector.

[0078] The method of applying the denoising autoencoder in step 3 is as follows:

[0079] Step 31: The denoising autoencoder (DAE) consists of an input layer, a hidden layer, and an output layer. Its network structure is as follows Figure 2 Before inputting the data into the denoising autoencoder, Gaussian noise of the same dimension is added to the input vector to generate a noisy input vector as shown below:

[0080]

[0081] where u and d is the dimension of the input vector; δ∈(0,σ 2 ), σ is the standard deviation of Gaussian noise.

[0082] Step 32: The encoder compresses the input vector into a low-dimensional intermediate vector e, whose expression is shown in formula (3).

[0083] Then, the decoder reconstructs the input vector from the intermediate vector, which is expressed as shown in formula (4).

[0084]

[0085] where e∈R m is the output value of the hidden layer; m is the number of neurons in the hidden layer; is the output vector of the decoder; W1∈R m×d and W2∈R d×m is the weight matrix; b∈R m and c∈R d is the bias vector; f and g are nonlinear activation functions.

[0086] The details of building the ResBiLSTM-MHA model in step 4 are as follows:

[0087] Step 41: BiLSTM is a time series recurrent neural network, which consists of two LSTM neural networks in opposite directions. LSTM neural network selectively retains and updates the information in the hidden state by introducing a gating mechanism, including three gate structures: forget gate, input gate and output gate. The calculation formulas of the input gate, forget gate and output gate are as follows:

[0088] f t =φ(W f·[h t-1 ,x t ]+b f ) (5)

[0089] i t =φ(W i ·[h t-1 ,x t ]+b i ) (6)

[0090]

[0091] o t =φ(W o ·[h t-1 ,x t ]+b o ) (8)

[0092] h t =o t *tanh(C t ) (9)

[0093] Where φ is the sigmoid activation function; tanh is the hyperbolic tangent activation function; W f ,W i , W C and W o are the weight matrices of the forget gate, input gate, internal state, and output gate respectively; b f ,b i ,b C and b o They represent the bias vectors of the forget gate, input gate, internal state and output gate respectively; h t ,x t and C t , respectively represent the hidden, input and memory states of the LSTM unit at time t; * represents the element-by-element multiplication of the matrix.

[0094] Step 42: In order to solve the BiLSTM network gradient disappearance problem, the residual connection is used in the BiLSTM network

[0095] The core idea of ​​the residual connection is to introduce a skip connection to pass the input vector directly to the subsequent layers of the network. The ResBiLSTM network of the present invention comprises two layers of ResBiLSTM, each layer of ResBiLSTM consists of a BiLSTM and a residual connection. In order to speed up the convergence speed of the model and improve the robustness and generalization ability of the model, there is a layer normalization (LN) behind each layer of ResBiLSTM. The calculation formula of each layer of ResBiLSTM is as follows:

[0096]

[0097] where x (i) represents the feature vector of the i-th input sample; μ represents the feature vector x (i) The mean of (i) The standard deviation of Represents the feature vector after layer normalization (LN); L is the processing of the input state in the LSTM network; is the forward output vector of the forward LSTM input at time t after residual connection and LN; is the reverse output vector of the input of the reverse LSTM at time t after residual connection and LN; y t is the concatenation of the forward and reverse output vectors at time t.

[0098] Step 43: The multi-head attention mechanism (MHA) can fully capture the dependencies between different positions in the sequence, and by processing key parts in parallel, enhance the model's ability to capture long-term dependencies, thereby improving prediction accuracy. Therefore, the present invention adds a multi-head attention mechanism (MHA) after the ResBiLSTM network. In the multi-head attention mechanism, first, the output data of the ResBiLSTM network is calculated through a linear layer to obtain three intermediate variable matrices, and then the attention matrix is ​​calculated, and finally, by accumulating multiple attention mechanism modules in parallel, the model can obtain multi-dimensional correlation information between data. The calculation formula of the multi-head attention mechanism is as follows:

[0099] Q=W Q X,K=W K X,V=W V ·X (14)

[0100]

[0101] MultiHead(Q,K,V)=Concat(head1,...,head h )W O (16)

[0102] Where W Q ,W k and W V are the weight matrices corresponding to the query, key, and value respectively; X is the input data; d k The dimension of the key; head i is the output of the i-th head; W O is the weight matrix of multi-head attention output.

[0103] Step 44: In order to predict the unknown capacity, a fully connected layer is used to map the feature vector after multi-head attention weighting to the capacity value. In summary, the DAE-ResBiLSTM-MHA model structure of the present invention is as follows: Figure 3 shown.

[0104] Step 45: Calculate the loss function. The error sources of the DAE-ResBiLSTM-MHA model include the error between the input value and the reconstructed value and the error between the true value and the predicted value. Mean Square Error (MSE) is the most commonly used error in the regression loss function. The present invention uses MSE to calculate the loss function of the DAE-ResBiLSTM-MHA model, and its calculation formula is as follows:

[0105] Loss = L MSE +L REC

[0106]

[0107] Where L MSE and L REC are the loss functions of mean square error and reconstruction error respectively; N is the dimension of input vector; y t and are the label and predicted output of the tth sample respectively; λ is the regularization parameter.

[0108] The method for training the model in step 5 of the present invention is as follows:

[0109] Step 51: Install graphics card driver, CUDA and CUDNN acceleration library.

[0110] Step 52: Configure the training environment.

[0111] Step 53: Write the Python Program.

[0112] Step 54: Train the DAE-ResBiLSTM-MHA model. The hyperparameters of the model are set in Table 1.

[0113] Step 55: The computer configuration is NVIDIA GeForce RTX 1050Ti graphics card and Intel Core i5-7300HQCPU@2.50GHz processor.

[0114] Table 1 Hyperparameter settings of DAE-ResBiLSTM-MHA model

[0115]

[0116] The method for evaluating the model in step 6 of the present invention is as follows:

[0117] Step 61: Calculate the relative error (RE), mean absolute error (MAE) and root mean square error (RMSE). The expressions of these three indicators are as follows:

[0118]

[0119] in is the predicted value, y t is a real value, and n represents the length of the sequence.

[0120] Step 62: Evaluate the performance of the model. Table 2 lists the RE, MAE, and RMSE performance index values ​​of the DAE-ResBiLSTM-MHA model of the present invention on NASA and CACLE. The results show that the DAE-ResBiLSTM-MHA model of the present invention exhibits extremely high RE, MAE, RMSE, and excellent generalization ability.

[0121] Table 2 Comparison of performance indicators of DAE-ResBiLSTM-MHA

[0122]

[0123] Step 63: In order to verify the effectiveness of the modules used in the algorithm in this paper, four groups of ablation experiments are designed based on the proposed model, including (1) ResBiLSTM model, (2) ResBiLSTM+MHA model (3) DAE+ResBiLSTM model and (4) DAE+ResBiLSTM+MHA model. The ablation experiment results are shown in Table 3. From the results in Table 3, it can be seen that all modules of the DAE+ResBiLSTM+MHA model contribute greatly to the overall performance. When DAE or MHA is removed, the model prediction performance decreases, indicating that data denoising and effective feature extraction are valuable for RUL prediction of lithium-ion batteries.

[0124] Table 3 Ablation experiment results

[0125]

[0126] Step 64: Evaluate the RUL prediction capability of the DAE-ResBiLSTM-MHA model. In order to more intuitively describe the RUL prediction results of the model, the present invention visualizes the prediction curves of the ResBiLSTM, ResBiLSTM-MHA, DAE-ResBiLSTM and DAE-ResBiLSTM-MHA models based on the NASA dataset and the CACLE dataset. Figure 4The RUL prediction results of the four models on the NASA dataset are shown starting from the 17th cycle. Figure 5 The RUL prediction results of the four models on the CACLE dataset are shown starting from the 65th cycle. Obviously, the DAE-ResBiLSTM-MHA model of the present invention, which combines denoising autoencoder, residual connection and multi-head attention mechanism, can more accurately predict the remaining service life of lithium-ion batteries.

Claims

1. A method for predicting the remaining service life of lithium-ion batteries based on DAE-ResBiLSTM-MHA, characterized in that The following steps are involved: Step 1: Build a lithium-ion battery dataset; Step 2: Normalize the battery capacity values ​​in the newly created data set; Step 3: Apply a denoising autoencoder to remove the noise introduced by the noise interference and capacity regeneration phenomenon in the battery dataset during the acquisition process; Step 4: Based on the BiLSTM model, the ResBiLSTM-MHA model is built by combining residual connection and multi-head attention mechanism; Step 5: Use leave-one-out cross-validation to divide the data set, that is, randomly select one battery sample as the test set each time, and use the remaining samples as the training set and validation set; use the training set to train the ResBiLSTM-MHA model, and use the validation set to preliminarily evaluate the model's predictive ability; Step 6: Evaluate the RUL prediction performance of the DAE-ResBiLSTM-MHA model.

2. According to claim 1, a method for predicting the remaining service life of a lithium-ion battery based on DAE-ResBiLSTM-MHA is characterized in that: The step 1 specifically includes: downloading the NASA data set and the CACLE data set; wherein the NASA data set is divided into 9 groups, with 3 to 4 batteries in each group, and the rated capacity is 2Ah; the charging stage adopts a constant current mode CC of 1.5A until the battery voltage reaches 4.2V, and then switches to a constant voltage mode CV until the charging current drops to 20mA; the discharging stage is carried out under a constant current mode CC of 2A until the battery voltage drops to a preset value; the experiment stops when the lithium-ion battery reaches a service life failure threshold, that is, stops when the rated capacity drops by 30%; select the four lithium-ion batteries CS_2 in the CACLE data set as another battery data set used to verify the generalization of the model; its charging mode adopts a constant current mode CC until the battery voltage reaches 4.2V, and then switches to a constant voltage mode CV until the charging current drops to 20mA; the discharging stage is carried out under a constant current mode until the battery voltage drops to 2.7V; and stops when the battery reaches a service life failure threshold, that is, stops when the rated capacity drops by 30%.

3. The method for predicting the remaining useful life of a lithium-ion battery based on DAE-ResBiLSTM-MHA according to claim 1, characterized in that: The step 2 specifically includes: using the minimum-maximum normalization method to normalize the data, and the calculation formula is as follows: Where C is the input vector, min(C) is the minimum value of the input vector, max(C) is the maximum value of the input vector, and u is the normalized input vector.

4. The method for predicting the remaining useful life of a lithium-ion battery based on DAE-ResBiLSTM-MHA according to claim 1, characterized in that: The step 3 specifically includes: Step 31: The denoising autoencoder DAE consists of an input layer, a hidden layer, and an output layer. Before the data is input into the denoising autoencoder, Gaussian noise of the same dimension is added to the input vector to generate a noisy input vector, as shown below: where u and d is the dimension of the input vector; δ∈N(0,σ 2 );σ is the standard deviation of Gaussian noise; Step 32: The encoder compresses the input vector into a low-dimensional intermediate vector e, whose expression is shown in formula (3); Then, the decoder reconstructs the input vector from the intermediate vector, which is expressed as shown in formula (4); where e∈R m is the output value of the hidden layer; m is the number of neurons in the hidden layer; is the output vector of the decoder; W1∈R m×d and W2∈R d×m is the weight matrix; b∈R m and c∈R d is the bias vector; f and g are nonlinear activation functions.

5. The method for predicting the remaining useful life of a lithium-ion battery based on DAE-ResBiLSTM-MHA according to claim 1, characterized in that: The step 4 specifically includes: Step 41: BiLSTM is composed of two LSTM neural networks with opposite directions. The LSTM neural network selectively retains and updates the information in the hidden state by introducing a gating mechanism, including a forget gate, an input gate, and an output gate. The calculation formulas of the input gate, the forget gate, and the output gate are as follows: f t =φ(W f ·[h t-1 ,x t ]+b f ) (5) i t =φ(W i ·[h t-1 ,x t ]+b i ) (6) o t =φ(W o ·[h t-1 ,x t ]+b o ) (8) h t =o t *tanh(C t ) (9) Where φ is the sigmoid activation function; tanh is the hyperbolic tangent activation function; W f ,W i , W C and W o are the weight matrices of the forget gate, input gate, internal state, and output gate respectively; b f ,b i ,b C and b o They represent the bias vectors of the forget gate, input gate, internal state and output gate respectively; h t ,x t and C t , respectively represent the hidden, input and memory states of the LSTM unit at time t; * represents the element-by-element multiplication of the matrix; Step 42: The ResBiLSTM network contains two layers of ResBiLSTM, each layer of ResBiLSTM consists of a BiLSTM and a residual connection; there is a layer of normalization LN after each layer of ResBiLSTM; the calculation formula of each layer of ResBiLSTM is as follows: where x (i) represents the feature vector of the i-th input sample; μ represents the feature vector x (i) The mean of (i) The standard deviation of Represents the feature vector after layer normalization (LN); L is the processing of the input state in the LSTM network; is the forward output vector of the forward LSTM input at time t after residual connection and LN; is the reverse output vector of the input of the reverse LSTM at time t after residual connection and LN; y t is the concatenation of the forward and reverse output vectors at time t; Step 43: Add a multi-head attention mechanism MHA after the ResBiLSTM network. In the multi-head attention mechanism, first, the output data of the ResBiLSTM network is calculated through a linear layer to obtain three intermediate variable matrices, and then the attention matrix is ​​calculated. Finally, multiple attention mechanism modules are accumulated in parallel, so that the model can obtain multi-dimensional correlation information between data. The calculation formula of the multi-head attention mechanism is as follows: Q=W Q ·X,K=W K ·X,V=W V ·X (14) MultiHead(Q,K,V)=Concat(head1,...,head h )W O (16) Where W Q ,W k and W V are the weight matrices corresponding to the query, key, and value respectively; X is the input data; d k The dimension of the key; head i is the output of the i-th head; W O is the weight matrix of multi-head attention output; Step 44: Use a fully connected layer to map the multi-head attention weighted feature vector to the capacity value; Step 45: Calculate the loss function. The error sources of the DAE-ResBiLSTM-MHA model include the error between the input value and the reconstructed value and the error between the true value and the predicted value. MSE is used to calculate the loss function of the DAE-ResBiLSTM-MHA model, and the calculation formula is as follows: Where L MSE and L REC are the loss functions of mean square error and reconstruction error respectively; N is the dimension of input vector; y t and are the label and predicted output of the tth sample respectively; λ is the regularization parameter.

6. The method for predicting the remaining useful life of a lithium-ion battery based on DAE-ResBiLSTM-MHA according to claim 1, characterized in that: The step 5 specifically includes: Step 51: Install graphics card driver, CUDA and CUDNN acceleration library; Step 52: Configure the training environment; Step 53: Write a Python program; Step 54: Train the DAE-ResBiLSTM-MHA model; Step 55: The computer configuration is NVIDIA GeForce RTX 1050Ti graphics card and Intel Core i5-7300HQ CPU@2.50GHz processor.

7. The method for predicting the remaining useful life of a lithium-ion battery based on DAE-ResBiLSTM-MHA according to claim 1, characterized in that: The step 6 specifically includes: Step 61: Calculate the relative error RE, mean absolute error MAE and root mean square error RMSE. The expressions are as follows: in is the predicted value, y t is a real value, n represents the length of the sequence; Step 62: Evaluate the performance of the model; Step 63: Evaluate the RUL prediction ability of the DAE-ResBiLSTM-MHA model.

Images