Lithium battery remaining service life prediction system and method

Abstract

The invention belongs to the technical field of battery life prediction, and discloses a system and a method for predicting the residual service life of a lithium battery, and the specific technical scheme is as follows: firstly, selecting a main index with relatively high capacity correlation from battery state indexes such as discharge time, discharge energy and average discharge voltage by using a Pearson's correlation coefficient; carrying out fusion by adopting a convolutional neural network; secondly, an attention mechanism is introduced into the graph convolutional network, and time sequence dependence feature extraction is carried out on the fused data; and finally, combining the output of the convolutional neural network and the output of the graph convolutional network, performing time sequence dependence enhancement by using a bidirectional long-short-term memory network, thereby obtaining a predicted value of the battery capacity, and calculating the remaining service life according to a capacity threshold value. According to the method, feature information extraction in the time dimension is further enhanced, the mean absolute error, the root-mean-square error, the mean absolute percentage error and the absolute error are small, and the prediction precision is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of battery life prediction, and in particular relates to a method for predicting the remaining service life of a lithium battery. Background Art

[0002] With the rapid development of industrial intelligence, lithium-ion batteries have become the basic energy guarantee for electronic devices and electric vehicles due to their advantages such as high energy density, long service life and low self-discharge characteristics. However, after frequent charge and discharge cycles, the battery will gradually degrade and its performance will continue to decline, affecting the normal use of the equipment and even causing safety hazards and significant economic losses to the energy storage system. Therefore, accurately assessing the remaining service life of the battery is of great significance for the safe use of the equipment.

[0003] At present, the remaining service life prediction methods of batteries are mainly divided into model-based methods and data-driven methods. The model-based method simulates the degradation process of lithium-ion batteries by constructing equivalent circuit models or electrochemical models to predict the remaining service life of batteries.

[0004] Xia et al. used an equivalent circuit model to characterize the aging of lithium-ion batteries, and analyzed the influence of parameters on the change of battery health status through accelerated aging experiments. Vennam et al. established a nonlinear electrothermal aging model, which closely combined the health status with the equivalent circuit model and provided a battery status evaluation method. Laribi et al. proposed a lithium-ion battery electrochemical impedance spectrum estimation method based on a fractional-order equivalent circuit model to simulate the aging process of lithium-ion batteries. Chen et al. proposed a method that integrates the empirical exponential model and the polynomial regression model to track the degradation trend of the battery during the cycle life cycle. These methods require degradation to conform to certain physical and chemical laws, and lack the flexibility to dynamically adjust according to real-time degradation conditions.

[0005] Based on the data-driven method, by establishing the mapping relationship between training samples and training labels, the network parameters are updated in real time to achieve the purpose of extracting features. Support vector machines, convolutional neural networks (CNN) and recurrent neural networks are widely used in time series prediction. Saurabh et al. used bilinear equations to describe the degradation process of battery capacity decay and used a two-dimensional convolutional neural network to obtain capacity decay prediction. Han et al. proposed a domain-adapted long short-term memory (LSTM) network model to adaptively learn the nonlinear mapping relationship from raw monitoring data to battery capacity, thereby improving the prediction accuracy. Ma et al. used a bidirectional long short-term memory (BiLSTM) network to solve the problem of insufficient information capture ability of a unidirectional long short-term memory network. In response to the poor performance of Transformer in long-term prediction, Tang Ziwei et al. designed a global timestamp and constructed a remaining service life model based on the Informer network. Kim et al. expressed the dependency between features and time through an adjacency matrix and obtained a prediction of the battery charge state by a graph convolutional network. Zhang Hongsheng et al. used a wide learning system to reduce the network depth and improve the efficiency of model training under limited computing resources and controllable training time.

[0006] Batteries are affected by usage and environmental conditions, and usually experience capacity regeneration, which leads to fluctuations in the degradation trajectory. Ignoring the degradation fluctuations caused by capacity regeneration will reduce the accuracy of the prediction results. Yuan et al. used the variational mode decomposition method to smooth the battery capacity data, reducing the impact of the capacity regeneration phenomenon. However, variational mode decomposition has the problems of high computational complexity and parameter sensitivity. Therefore, Xie et al. introduced the adaptive noise complete set empirical mode decomposition method to improve computational efficiency and reduce parameter sensitivity. However, the modal decomposition method usually uses a single feature for prediction, which makes it difficult to capture the complex degradation mechanism of the battery, resulting in poor life prediction results.

[0007] The deep learning method based on multi-feature fusion can fully extract the feature information in the battery degradation process and has obvious advantages. Xu et al. used the CNN-LSTM network to calculate the feature importance and select the main features for training. Zhang Ying et al. used the random forest algorithm to screen out high-contribution features and introduced long-term peeking to enhance the input gate of LSTM, which effectively improved the model's ability to handle long-term dependencies. Gu et al. used the principal component analysis method to reduce the dimension of features and reduced the amount of calculation of the model. Wang et al. introduced the feature fusion strategy into the graph structure to fuse multi-state indicators and enhance the feature fusion ability of the network. These methods failed to take into account the dependency between feature fusion and the time dimension, and did not consider the computational complexity under limited resources, which affected the accuracy and efficiency of the prediction.

[0008] In summary, model-based methods cannot dynamically adjust the model structure according to environmental conditions, single-feature data-driven methods are difficult to fully reflect various battery degradation states, and multi-feature fusion methods do not fully consider the dependency between feature fusion and time. Summary of the invention

[0009] In order to solve the technical problems existing in the prior art, the present invention provides a neural network method of multi-feature fusion and timing dependency enhancement to predict the capacity state and obtain the remaining service life of the battery.

[0010] To achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a remaining service life prediction system for lithium batteries, in which capacity is a direct indicator of the battery health status, and its threshold determines the length of the remaining service life. The MFTDE method combines multi-feature fusion and timing dependency enhancement to predict the capacity state of the battery degradation process, and then obtain the remaining service life.

[0011] Construct a graph convolutional network, which is divided into four modules:

[0012] Feature screening module, which uses Pearson correlation coefficient to select features of battery status indicators;

[0013] CNN multi-feature fusion module, which performs feature fusion on the selected main indicators;

[0014] GCNAM temporal dependency extraction module introduces the attention mechanism into the graph convolutional network to extract the temporal dependency features of the degradation state;

[0015] The BiLSTM timing dependency enhancement module takes the output data of the CNN multi-feature fusion module and the GCNAM timing dependency extraction module as input, enhances the timing dependency relationship, predicts the capacity state during the battery degradation process, and obtains the remaining service life of the battery based on the capacity threshold.

[0016] A method for predicting the remaining service life of a lithium battery, the specific steps are as follows:

[0017] Step 1: Feature selection and fusion:

[0018] Feature selection: In the process of battery degradation, in addition to the capacity status indicator, discharge time, discharge energy, average discharge voltage, average discharge internal resistance and average discharge voltage rate can also reflect the degradation state of the battery. Too many feature indicators will increase the complexity of the network and reduce the efficiency of calculation. Therefore, the main features with high correlation with capacity are selected for battery remaining service life prediction in order to make full use of the battery status information.

[0019] The Pearson correlation coefficient is usually used to measure the linear correlation between two random variables, and its value is between -1 and 1. When the correlation coefficient is closer to -1, the negative correlation is stronger. On the contrary, when it is closer to 1, the positive correlation is stronger. When it is close to 0, it indicates that the two variables are weakly correlated or unrelated. Assume that X represents the battery capacity, {X 1 ,X 2 ,…X n} represents n observations of X, Y i represents the i-th state indicator, {Y i1 ,Y i2 ,…Y in} means Y i If there are n observation data, the Pearson correlation coefficient between capacity and other state indicators is calculated as follows:

[0020]

[0021] in, Represents X and Y i The Pearson correlation coefficient, Indicates the average value of capacity data, Represents the average value of the observed data of the i-th state indicator.

[0022] Calculated from Pearson's correlation coefficient and The state indicators with correlation coefficient greater than 0.9 are selected as the main indicators for training.

[0023] Multi-feature fusion: Assume that the selected λ state indicators are Y 1 ,Y 2 ,…,Y λ , Y 1 ,Y 2 ,…,Y λAs the input feature of CNN, it is fused to solve the problem of incomplete single feature information. When the convolution kernel size is S and the number of input channels is P, the output of the pth channel of the lth layer at time point t is shown in formula (2):

[0024]

[0025] Among them, V l (t+s,p) represents the value of the input data in the pth channel of the lth layer at time point t+s, × represents matrix multiplication, W lp (k) represents the weight of the sth convolution kernel in the lth layer on channel p, b l represents the bias term of the lth layer, and Re(·) represents the ReLU function.

[0026] The feature fusion data obtained by convolution is pooled to reduce the network's sensitivity to state noise and changes.

[0027] Step 2: Extracting temporal dependencies: The graph convolutional network is a neural network specifically designed to process graph structure data, which can efficiently extract the correlation in the graph structure. The attention mechanism increases the contribution of the main features by redistributing the weights of different features. Introducing the attention mechanism into the graph convolutional network not only extracts the temporal correlation of features, but also plays a role in strengthening important features.

[0028] For time series data, the attention mechanism calculates the correlation between the current element and other elements, normalizes the correlation into attention weights, and establishes temporal relationship dependencies. The output Z(t) of CNN is used as the input of the attention mechanism. and Denote query and key respectively, c,d∈{1,2,…,t}, then the attention matrix is ​​calculated as follows:

[0029]

[0030] in, express The transpose of A, × represents matrix multiplication, cd represents the normalized attention matrix, and F(·) represents the softmax function.

[0031] The graph convolutional network performs convolution operations on the features of the nodes and uses the connection relationship and neighborhood information between the nodes to learn and generate the representation of each node. The calculation of the graph convolutional network from the mth layer to the m+1th layer is as follows:

[0032]

[0033] Among them, B (m+1) represents the feature matrix of the m+1 layer, represents the degree matrix, represents the adjacency matrix, M+I represents adding the unit matrix I to the adjacency matrix M of the current feature so that it is connected to itself, B (m) represents the feature matrix of the mth layer, W (m) Represents the weight matrix of the mth layer.

[0034] For time series prediction tasks, the adjacency matrix constructed only by the time order between adjacent elements is a tridiagonal matrix, which only reflects the feature information provided by one or two adjacent elements. The adjacency matrix constructed by the attention mechanism takes into account the relevance of all elements to the current element, and is represented by A. cd Replace the original graph convolutional network The graph convolutional network based on the attention mechanism is obtained, and its calculation formula is as follows:

[0035] B (m+1) =Re(A cd B (m) W (m) ) (6)

[0036] By A cd Replace the original graph convolutional network It can reduce the dependence of nodes on adjacent information, enhance the weight of global important information, and effectively deal with the differences in node characteristics in time series.

[0037] Step 3: Temporal dependency enhancement: The attention mechanism cannot fully capture all dependencies in the sequence, especially when processing long time series, some important local dependencies will be ignored. The BiLSTM network processes the outputs of CNN and graph convolutional network at the same time, which can make up for the shortcomings of the attention mechanism in capturing sequence dependencies. E (t) and the output B of the graph convolutional network (G) The concatenation is performed to obtain x, which is input into the BiLSTM network for temporal dependency enhancement.

[0038] BiLSTM computes the forward LSTM hidden state at each time step and the backward LSTM hidden state Then concatenate these two hidden states to get the final output H a , the hidden state and final output calculation formula are as follows:

[0039]

[0040] in, and Represent the outputs of the forward and backward output gates, respectively. and Represent the current memory unit in the forward and backward directions respectively, H t represents the output of the bidirectional long short-term memory network, express and splicing.

[0041] The BiLSTM network consists of a forward propagation LSTM and a backward propagation LSTM. The LSTM processes information through the forget gate, input gate, and output gate. The unidirectional LSTM only calculates the current hidden state h at each time step. t and the current memory cell C t , the specific calculation is shown in formulas (10) and (11):

[0042] h t =o t ⊙tanh(C t ) (10)

[0043]

[0044] Among them, f t represents the forget gate, ⊙ represents element-wise multiplication, C t-1 represents the state of the memory cell at the previous time step, represents the input gate, represents the candidate memory unit, o t Represents the output gate.

[0045] Forget the hidden state h of the previous time step t-1 and the current time step x t As input, the sigmoid activation function scales the value to between 0 and 1, which represents the probability that the output of the previous layer unit is forgotten. Among them, the output value 0 means that all information is forgotten, and the output value 1 means that all information is saved. The forget gate f t The calculation of is given by formula (12):

[0046] f t =σ(W f ×[h t-1 ,x t ]+b f ) (12)

[0047] Where σ(·) represents the sigmoid activation function, W f represents the weight matrix of the forget gate, × represents matrix multiplication, b f Represents the bias term of the forget gate.

[0048] The input gate selectively stores new information into candidate memory cells In , the calculation formulas for the input gate and candidate memory cells are as follows:

[0049]

[0050] in, represents the weight matrix of the input gate, represents the bias term of the input gate, W c represents the weight matrix of the candidate memory unit, b c represents the bias term of the candidate memory unit, and tanh(·) represents the tanh activation function.

[0051] The output gate determines which information needs to be output to the next time step. Its calculation formula is as follows:

[0052] o t =σ(W o ×[h t-1 ,x t ]+b o )(15)

[0053] Among them, W o represents the weight matrix of the output gate, b o Represents the bias term of the output gate.

[0054] BiLSTM can capture data features from both positive and negative directions at the same time, connect the current time point with past and future information, and make full use of the historical and future information of time series data.

[0055] Step 4: MFTDE network life prediction process: By integrating multi-state indicators to reflect the capacity degradation process, the remaining service life of the battery is predicted. The main steps of MFTDE network prediction are as follows:

[0056] (1) Input multiple status indicators of battery degradation;

[0057] (2) Use the Pearson correlation coefficient to screen different status indicators;

[0058] (3) Use CNN to perform multi-feature fusion on the selected battery degradation status indicators;

[0059] (4) Input the fused features into the attention graph convolutional network to extract temporal dependencies;

[0060] (5) Concatenate the outputs of step (3) and step (4) and input them into the BiLSTM network for temporal dependency enhancement;

[0061] (6) Use full connection on the output of BiLSTM to obtain capacity prediction;

[0062] (7) Calculate the predicted remaining service life L based on the capacity threshold pr, and its calculation formula is as follows:

[0063] L pr =L th -L st (16)

[0064] Among them, L th Indicates the number of charge and discharge cycles when the capacity reaches the failure threshold, L st Indicates the number of charge and discharge cycles at the prediction starting point.

[0065] Remaining useful life prediction using MFTDE network.

[0066] Compared with the prior art, the present invention has the following specific beneficial effects: the present invention proposes a deep learning network prediction model with multi-feature fusion and time series dependency enhancement, which selects features with strong correlation with the target variable by calculating the Pearson correlation coefficient between different indicators, reduces the interference of irrelevant information, and reduces the complexity of the model; the present invention is based on a multi-feature fusion network of CNN, overcomes the information loss caused by the prediction of a single capacity feature, introduces the attention mechanism into the graph convolutional network to establish a time series dependency modeling relationship of the fused features in the time dimension, which not only extracts the correlation of the features in the time series, but also plays a role in strengthening important features; the present invention uses BiLSTM to mix the fusion features and time series dependency features of different indicators, solves the problem of insufficient capture ability of unidirectional LSTM, further strengthens the feature information extraction in the time dimension, enables the model to better utilize high-frequency feature information, and can capture more feature information and increase learning efficiency compared with other inventions; compared with other prediction methods, the method of the present invention has smaller mean absolute error, root mean square error, mean absolute percentage error and absolute error, which greatly improves the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 This is a network structure diagram for predicting the remaining useful life of the present invention.

[0068] Figure 2 Schematic diagram of the structure of the multi-feature fusion network.

[0069] Figure 3 Schematic diagram of the structure of the attention mechanism.

[0070] Figure 4 It is a schematic diagram of the connection between element nodes. Figure 4 (a) is a tridiagonal matrix constructed only by the time sequence between adjacent elements, and the connection diagram between its element nodes is shown in Figure 1. Figure 4 (b) The adjacency matrix constructed by the attention mechanism takes into account the relevance of all elements to the current element, and is a schematic diagram of the connections between its element nodes.

[0071] Figure 5 Schematic diagram of the structure of a bidirectional long short-term memory network.

[0072] Figure 6 This is a schematic diagram of the LSTM structure.

[0073] Figure 7 Flowchart for remaining useful life prediction of MFTDE.

[0074] Figure 8 This is the curve diagram of the change of CALCE battery capacity.

[0075] Fig. 9 This is the relative change trend chart between the CS2_35 battery capacity and other status indicators. Fig. 9 (a) is the relative change trend diagram between capacity and time during discharge. Fig. 9 (b) is the relative change trend diagram between capacity and energy during discharge. Fig. 9 (c) is the relative change trend diagram between capacity and average voltage during discharge. Fig. 9 (d) is the relative change trend diagram between capacity and average internal resistance during discharge. Fig. 9 (e) is the relative change trend diagram between capacity and average voltage rate during discharge.

[0076] Fig.10 This is the CS2_35 battery error distribution diagram. Fig.10 (a) is the error distribution diagram of MFTDE, Fig.10 (b) is the error distribution diagram of GCNAM-BiLSTM. Fig.10 (c) is the error distribution diagram of CNN-BiLSTM. Fig.10 (d) is the error distribution diagram of CNN-GCNAM.

[0077] Fig.11 This is a comparison chart of the prediction results of different models. Fig.11 (a) is a comparison chart of the prediction results of different models of CS2_35. Fig.11 (b) is a comparison chart of the prediction results of different models of CS2_36. Fig.11 (c) is a comparison chart of the prediction results of different models of CS2_37. Fig.11 (d) is a comparison chart of the prediction results of different CS2_38 models. DETAILED DESCRIPTION

[0078] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0079] like Figure 1 As shown, a lithium battery remaining service life prediction system is shown. Capacity is a direct indicator of the battery health status, and its threshold determines the length of the remaining service life. The MFTDE method combines multi-feature fusion and time-series dependency enhancement to predict the capacity state of the battery degradation process and then obtain the remaining service life.

[0080] Construct a graph convolutional network, which is divided into four modules:

[0081] Feature screening module, which uses Pearson correlation coefficient to select features of battery status indicators;

[0082] CNN multi-feature fusion module, which performs feature fusion on the selected main indicators;

[0083] GCNAM temporal dependency extraction module introduces the attention mechanism into the graph convolutional network to extract the temporal dependency features of the degradation state;

[0084] The BiLSTM timing dependency enhancement module takes the output data of the CNN multi-feature fusion module and the GCNAM timing dependency extraction module as input, enhances the timing dependency relationship, predicts the capacity state during the battery degradation process, and obtains the remaining service life of the battery based on the capacity threshold.

[0085] The specific remaining service life prediction method has the following steps:

[0086] Step 1: Feature selection and fusion:

[0087] Feature selection: In the process of battery degradation, in addition to the capacity status indicator, discharge time, discharge energy, average discharge voltage, average discharge internal resistance and average discharge voltage rate can also reflect the degradation state of the battery. Too many feature indicators will increase the complexity of the network and reduce the efficiency of calculation. Therefore, the main features with high correlation with capacity are selected for battery remaining service life prediction in order to make full use of the battery status information.

[0088] The Pearson correlation coefficient is usually used to measure the linear correlation between two random variables, and its value is between -1 and 1. When the correlation coefficient is closer to -1, the negative correlation is stronger. On the contrary, when it is closer to 1, the positive correlation is stronger. When it is close to 0, it indicates that the two variables are weakly correlated or unrelated. Assume that X represents the battery capacity, {X 1 ,X 2 ,…X n} represents n observations of X, Y i represents the i-th state indicator, {Y i1 ,Y i2 ,…Y in} means Y i If there are n observation data, the Pearson correlation coefficient between capacity and other state indicators is calculated as follows:

[0089]

[0090] in, Represents X and Y i The Pearson correlation coefficient, Indicates the average value of capacity data, Represents the average value of the observed data of the i-th state indicator.

[0091] Calculated from Pearson's correlation coefficient and The state indicators with correlation coefficient greater than 0.9 are selected as the main indicators for training.

[0092] Multi-feature fusion: Assume that the selected λ state indicators are Y 1 ,Y 2 ,…,Y λ , Y 1 ,Y 2 ,…,Y λ As the input feature of CNN, it is fused to solve the problem of incomplete single feature information. The multi-feature fusion network structure is as follows: Figure 2 When the convolution kernel size is S and the number of input channels is P, the output of the pth channel of the lth layer at time point t is shown in formula (2):

[0093]

[0094] Among them, V l (t+s,p) represents the value of the input data in the pth channel of the lth layer at time point t+s, × represents matrix multiplication, W lp (k) represents the weight of the sth convolution kernel in the lth layer on channel p, b l represents the bias term of the lth layer, and Re(·) represents the ReLU function.

[0095] The feature fusion data obtained by convolution is pooled to reduce the network's sensitivity to state noise and changes.

[0096] Temporal dependency extraction: Graph convolutional network is a neural network specially designed for processing graph structure data. It can efficiently extract the correlation in the graph structure. The attention mechanism improves the contribution of the main features by redistributing the weights of different features. Introducing the attention mechanism into the graph convolutional network not only extracts the temporal correlation of features, but also plays a role in strengthening important features.

[0097] For time series data, the attention mechanism calculates the correlation between the current element and other elements, normalizes the correlation into attention weights, and establishes temporal relationship dependencies. Figure 3 As shown. The output Z(t) of CNN is used as the input of the attention mechanism. and Denote query and key respectively, c,d∈{1,2,…,t}, then the attention matrix is ​​calculated as follows:

[0098]

[0099] A cd =F(A' cd ) (4)

[0100] in, express The transpose of A, × represents matrix multiplication, cd represents the normalized attention matrix, and F(·) represents the softmax function.

[0101] The graph convolutional network performs convolution operations on the features of the nodes and uses the connection relationship and neighborhood information between nodes to learn and generate the representation of each node. The calculation of the graph convolutional network from the mth layer to the m+1th layer is as follows:

[0102]

[0103] Among them, B (m+1) represents the feature matrix of the m+1 layer, represents the degree matrix, represents the adjacency matrix, M+I represents adding the unit matrix I to the adjacency matrix M of the current feature so that it is connected to itself, B (m) represents the feature matrix of the mth layer, W (m) Represents the weight matrix of the mth layer.

[0104] For time series prediction tasks, the adjacency matrix constructed only by the time order between adjacent elements is a tridiagonal matrix, and the connections between its element nodes are as follows: Figure 4 As shown in (a), the matrix only reflects the feature information provided by one or two adjacent elements.

[0105] The adjacency matrix constructed by the attention mechanism takes into account the relevance of all elements to the current element. The connections between its element nodes are as follows: Figure 4 As shown in (b), use A cd Replace the original graph convolutional network The graph convolutional network based on the attention mechanism is obtained, and its calculation formula is as follows:

[0106] B (m+1)=Re(A cd B (m) W (m) ) (6)

[0107] By A cd Replace the original graph convolutional network It can reduce the dependence of nodes on adjacent information, enhance the weight of global important information, and effectively deal with the differences in node characteristics in time series.

[0108] Temporal dependency enhancement: The attention mechanism cannot fully capture all dependencies in the sequence, especially when processing long time series, some important local dependencies will be ignored. The BiLSTM network processes the outputs of CNN and graph convolutional network at the same time, which can make up for the shortcomings of the attention mechanism in capturing sequence dependencies. E (t) and the output B of the graph convolutional network (G) After concatenation, x is obtained and input into the BiLSTM network for temporal dependency enhancement. The BiLSTM network structure is as follows Figure 5 shown.

[0109] The BiLSTM network computes the forward LSTM hidden state at each time step and the backward LSTM hidden state Then concatenate these two hidden states to get the final output H t , the hidden state and final output calculation formula are as follows:

[0110]

[0111] in, represents the output of the forward output gate, represents the output of the backward output gate, represents the forward current memory unit, Represents the current memory unit, H t represents the output of the bidirectional long short-term memory network, express and splicing.

[0112] The BiLSTM network consists of a forward propagation LSTM and a backward propagation LSTM. LSTM processes information through a forget gate, an input gate, and an output gate. Its structure is as follows: Figure 6 As shown. For a unidirectional LSTM, it only calculates the current hidden state h at each time step t and the current memory cell C t , the specific calculation is shown in formulas (10) and (11):

[0113] ht =o t ⊙tanh(C t ) (10)

[0114]

[0115] Among them, f t represents the forget gate, ⊙ represents element-wise multiplication, C t-1 represents the state of the memory cell at the previous time step, represents the input gate, represents the candidate memory unit, o t Represents the output gate.

[0116] Forget the hidden state h of the previous time step t-1 and the current time step x t As input, the sigmoid activation function scales the value to between 0 and 1, which represents the probability that the output of the previous layer unit is forgotten. Among them, the output value 0 means that all information is forgotten, and the output value 1 means that all information is saved. The forget gate f t The calculation of is given by formula (12):

[0117] f t =σ(W f ×[h t-1 ,x t ]+b f ) (12)

[0118] Where σ(·) represents the sigmoid activation function, W f represents the weight matrix of the forget gate, × represents matrix multiplication, b f Represents the bias term of the forget gate.

[0119] The input gate selectively stores new information into candidate memory cells In , the calculation formulas for the input gate and candidate memory cells are as follows:

[0120]

[0121] in, represents the weight matrix of the input gate, represents the bias term of the input gate, W c represents the weight matrix of the candidate memory unit, b c represents the bias term of the candidate memory unit, and tanh(·) represents the tanh activation function.

[0122] The output gate determines which information needs to be output to the next time step. Its calculation formula is as follows:

[0123] o t=σ(W o ×[h t-1 ,x t ]+b o ) (15)

[0124] Among them, W o represents the weight matrix of the output gate, b o Represents the bias term of the output gate.

[0125] BiLSTM can capture data features from both positive and negative directions at the same time, connect the current time point with past and future information, and make full use of the historical and future information of time series data.

[0126] Step 4: MFTDE network life prediction process: The present invention reflects the capacity degradation process by integrating multiple state indicators, and then predicts the remaining service life of the battery. The main steps of MFTDE network prediction are as follows:

[0127] (1) Input multiple status indicators of battery degradation;

[0128] (2) Use the Pearson correlation coefficient to screen different status indicators;

[0129] (3) Use CNN to perform multi-feature fusion on the selected battery degradation status indicators;

[0130] (4) Input the fused features into the attention graph convolutional network to extract temporal dependencies;

[0131] (5) Concatenate the outputs of step (3) and step (4) and input them into the BiLSTM network for temporal dependency enhancement;

[0132] (6) Use full connection on the output of BiLSTM to obtain capacity prediction;

[0133] (7) Calculate the predicted remaining service life L based on the capacity threshold pr , and its calculation formula is as follows:

[0134] L pr =L th -L st (16)

[0135] Among them, L th Indicates the number of charge and discharge cycles when the capacity reaches the failure threshold, L st Indicates the number of charge and discharge cycles at the prediction starting point.

[0136] The detailed flow chart of remaining useful life prediction using MFTDE network is as follows: Figure 7 shown.

[0137] Experiment and result analysis: The present invention uses the CALCE data set provided by the University of Maryland, which contains multiple types of batteries. In order to verify the effectiveness of the proposed method, CS2_35, CS2_36, CS2_37 and CS2_38 batteries of the CS2 type under the same environmental conditions are selected for verification and analysis. The experiments on CS2_35, CS2_36, CS2_37 and CS2_38 batteries include two steps: charging and discharging. During the charging process, first charge at a constant current rate of 1C under a constant current state until the voltage reaches 4.2V. Then switch to a constant voltage state and charge at a constant voltage of 4.2V until the charging current drops below 0.05A, and the charging is completed. During the discharge process, the voltage is reduced to below 2.7V, and the discharge is completed. Each completion of charging and discharging is a charge and discharge cycle. The specific experimental conditions are shown in Table 1. The capacity of CALCECS2_35, CS2_36, CS2_37 and CS2_38 batteries changes with the increase of the number of cycles, and the degradation curve is shown in Figure 8 shown.

[0138] Table 1 Experimental conditions of CALCE battery dataset

[0139]

[0140] Battery status indicator screening: Taking CS2_35 battery as an example, the relative change trend between capacity and other status indicators is as follows Fig. 9 As shown. The Pearson correlation coefficients between the battery capacity and other state indicators of CS2_35, CS2_36, CS2_37 and CS2_38 are shown in Table 2. As can be seen from Table 2, the correlation coefficients between discharge time, discharge energy and average discharge voltage and capacity are all stable and maintained above 0.9. The correlation coefficient between the average discharge internal resistance and capacity of the CS2_38 battery is -0.1763, which is a low correlation; the correlation coefficient between the average discharge voltage rate and capacity is less than 0.9. Through the above analysis, discharge time, discharge energy and average discharge voltage are finally selected as state indicators for capacity prediction.

[0141] Table 2 Pearson correlation coefficients of different battery discharge capacities and other characteristics

[0142]

[0143] Evaluation index: In order to evaluate the effectiveness of the MFTDE method, the present invention selects several commonly used evaluation indexes, including mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE) and absolute error (AE), and the calculation formula is as follows:

[0144]

[0145] AE=|L tr -L pr | (20)

[0146] Among them, r represents the starting point of the prediction, t represents the prediction length, v represents the current prediction point, and X tr The actual capacity data indicating the capacity, X pr Indicates the predicted capacity data of capacity, L tr Indicates the true remaining useful life.

[0147] Experimental parameter setting: The hyperparameters of the model largely determine the accuracy of the prediction results and the overall performance of the model. Therefore, it is crucial to set the model parameters reasonably. The main hyperparameters of the MFTDE network model are shown in Table 3.

[0148] Table 3 Main hyper parameters of MFTDE model

[0149]

[0150] Ablation experiment analysis: In order to verify the effectiveness of the MFTDE method of the present invention, the method was subjected to ablation experiment analysis. The CNN-GCNAM method represents the removal of the time-dependent feature module in the MFTDE method, the CNN-BiLSTM method represents the removal of the time-dependent feature extraction module in the MFTDE method, and the GCNAM-BiLSTM method represents the removal of the multi-feature fusion module in the MFTDE method.

[0151] The present invention compares the performance of different methods on CS2_35, CS2_36, CS2_37 and CS2_38 batteries through multiple evaluation indicators, and the comparison results of various indicators are shown in Table 4. It can be seen from Table 4 that under the same starting point conditions, the MAE, RMSE, MAPE and AE of the MFTDE method are all the lowest. Taking the CS2_35 battery as an example, compared with the CNN-GCNAM method, the MAE, RMSE, MAPE and AE of the MFTDE method are reduced by 70.04%, 62.81%, 59.60% and 1250.00%, respectively. Compared with the CNN-BiLSTM method, the MAE, RMSE, MAPE and AE are reduced by 25.26%, 20.03%, 23.69% and 350.00%, respectively. Compared with the GCNAM-BiLSTM method, the MAE, RMSE, MAPE and AE are reduced by 84.42%, 84.09%, 83.09% and 600.00%, respectively.

[0152] In order to intuitively compare the prediction errors, take the CS2_35 battery as an example, Fig.10 The comparison of the error distribution of the ablation experiment is shown, where the left ordinate represents the predicted number of charge and discharge cycles, and the right ordinate represents the probability density of the error value in a certain interval. Fig.10 (a) is the error distribution diagram of the capacity predicted by the MFTDE method, with 99.69% of the error values ​​within 0.05Ah; Fig.10 (b) shows the capacity prediction error distribution of the GCNAM-BiLSTM method, where 74.27% of the error values ​​are within 0.05Ah; Fig.10 (c) shows the error distribution of the predicted capacity by the CNN-BiLSTM method. The results show that 98.93% of the prediction errors are within 0.05Ah. Fig.10 (d) is the error distribution diagram of the predicted capacity by the CNN-GCNAM method, and 90.51% of the predicted error values ​​are within 0.05Ah. Fig.10 (a) to Fig.10 (d) The width of the kernel density curve of the MFTDE method is smaller than that of the CNN-GCNAM method, the CNN-BiLSTM method, and the GCNAM-BiLSTM method, indicating that the MFTDE method has a smaller error distribution range. When the error is close to 0, the kernel density peak of the MFTDE method reaches 65.696Ah-1, which is 263.17%, 50.06%, and 439.11% higher than that of the CNN-GCNAM method, the CNN-BiLSTM method, and the GCNAM-BiLSTM method, respectively. It can be seen that the MFTDE method can better fuse feature information and learn the temporal dependency between features.

[0153] In order to show the prediction effects of different methods in the ablation experiment, Fig.11 The degradation curves predicted by ablation experiments of CS2_35, CS2_36, CS2_37 and CS2_38 batteries were compared. Fig.11 It can be seen that the MFTDE method, CNN-BiLSTM method and CNN-GCNAM method perform better than the GCNAM-BiLSTM method in fitting the original capacity curve, indicating that the multi-feature fusion module effectively utilizes the multi-state degradation information of the battery. Compared with the MFTDE method, the curve fitting effect of the CNN-GCNAM method and the CNN-BiLSTM method is slightly worse, which verifies that the timing dependency extraction and enhancement module can better capture the timing feature information.

[0154] In summary, the MFTDE method can effectively improve the prediction effect by fusing multiple features and extracting and enhancing temporal dependencies.

[0155] Comparative experimental analysis: To further verify the effectiveness of the MFTDE method, it is compared with other methods, and the comparison results are shown in Table 5. Compared with the ISSA-LSTM single feature prediction method, the MFTDE method has smaller MAE, RMSE and AE when the prediction starting point is early. Taking the CS2_35 battery as an example, the MAE, RMSE and AE indicators are reduced by 47.37%, 11.36% and 100.00%, respectively. Compared with the GBLS Booster method, the MFTDE method has an earlier prediction starting point, and the MAE, RMSE and AE indicators are lower than those of the GBLS Booster method. Taking the CS2_36 battery as an example, the MAE, RMSE and AE of the MFTDE method when the prediction starting point is 200 are reduced by 21.00%, 18.57% and 300.00% respectively compared with the GBLS Booster method. Compared with the Attention-based CNN-BiLSTM method, under the same prediction starting point, the MAE, RMSE and AE indicators of the MFTDE method on four batteries are not less than those of the Attention-based CNN-BiLSTM method. The MAE, RMSE and AE indicators are reduced by at least 17.36%, 10.20% and 100.00%, respectively. From the above analysis, it can be seen that the MFTDE method still has smaller MAE, RMSE and AE under the condition of an earlier prediction starting point. The MFTDE method can more comprehensively utilize the multi-feature state information of battery degradation, fully extract the time dependency of features, and improve the accuracy of prediction.

[0156] Table 5 Comparison of prediction results of different methods

[0157]

[0158]

[0159] The present invention proposes a deep learning network prediction model with multi-feature fusion and time-series dependency enhancement, which solves the problem of state information loss caused by single feature and insufficient extraction of time-series dependency information. The main indicators are screened out by calculating the Pearson correlation coefficient between different indicators, reducing the interference of irrelevant information. The multi-feature fusion network based on CNN overcomes the information loss caused by the prediction of a single capacity feature. The attention mechanism is introduced into the graph convolutional network to establish the time-series dependency modeling relationship of the fused features in the time dimension. Finally, BiLSTM is used to mix the fusion features and time-series dependency features of different indicators, further strengthening the feature information extraction in the time dimension. Experiments show that compared with other prediction methods, the method of the present invention has smaller mean absolute error, root mean square error, mean absolute percentage error and absolute error, and improves the prediction accuracy.

[0160] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims

1. A system for predicting the remaining useful life of a lithium battery, characterized in that: Construct a graph convolutional network, which is divided into four modules: Feature screening module, which uses Pearson correlation coefficient to select features of battery status indicators; CNN multi-feature fusion module, which performs feature fusion on the selected indicators; GCNAM temporal dependency extraction module introduces the attention mechanism into the graph convolutional network to extract the temporal dependency features of the degradation state; The BiLSTM timing dependency enhancement module takes the output data of the CNN multi-feature fusion module and the GCNAM timing dependency extraction module as input, enhances the timing dependency relationship, predicts the capacity state during the battery degradation process, and obtains the remaining service life of the battery based on the capacity threshold.

2. A method for predicting the remaining service life of a lithium battery, characterized in that: The specific steps are as follows: Step 1: Feature selection: Assume that X represents the battery capacity, {X1,X2,…X n } represents n observations of X, Y i represents the i-th state indicator, {Y i1 ,Y i2 ,…Y in } means Y i If there are n observation data, the Pearson correlation coefficient between capacity and other state indicators is calculated as follows: in, Represents X and Y i The Pearson correlation coefficient, Indicates the average value of capacity data, represents the average value of the observed data of the i-th state indicator; Calculated from the Pearson correlation coefficient and Select the state indicator with a correlation coefficient greater than 0.9 as the training indicator; Step 2: Multi-feature fusion: Assume that the selected λ state indicators are Y1, Y2, …, Y λ , Y1,Y2,…,Y λ As the input feature of CNN, when the convolution kernel size is S and the number of input channels is P, the output of the p-th channel of the l-th layer at time point t is shown in formula (2): Among them, V l (t+s,p) represents the value of the input data in the pth channel of the lth layer at time point t+s, × represents matrix multiplication, W lp (s) represents the weight of the sth convolution kernel in the lth layer on channel p, b l represents the bias term of the lth layer, Re(·) represents the ReLU function; Pool the feature fusion data obtained by convolution; Step 3: Temporal dependency extraction: For time series data, the attention mechanism calculates the correlation between the current element and other elements, normalizes the correlation into attention weights, establishes temporal relationship dependencies, and uses the CNN output Z(t) as the input of the attention mechanism. Indicates a query, Representing the key, c,d∈{1,2,…,t}, the attention matrix is ​​calculated as follows: A cd =F(A' cd ) (4) in, express The transpose of A, × represents matrix multiplication, cd represents the normalized attention matrix, and F(·) represents the softmax function; The graph convolutional network performs convolution operations on the features of the nodes and uses the connection relationship and neighborhood information between nodes to learn and generate the representation of each node. The calculation of the graph convolutional network from the mth layer to the m+1th layer is as follows: Among them, B (m+1) represents the feature matrix of the m+1 layer, represents the degree matrix, represents the adjacency matrix, M+I represents adding the unit matrix I to the adjacency matrix M of the current feature so that it is connected to itself, B (m) represents the feature matrix of the mth layer, W (m) represents the weight matrix of the mth layer; Use A cd Replace the original graph convolutional network The graph convolutional network based on the attention mechanism is obtained, and its calculation formula is as follows: B (m+1) =Re(A cd B (m) W (m) ) (6) By A cd Replace the original graph convolutional network Dealing with the differences in node features in time series; Step 4: Temporal dependency enhancement: transform the CNN output Z E (t) and the output B of the graph convolutional network (G) Concatenate to get x, and input it into the BiLSTM network for temporal dependency enhancement; The BiLSTM network computes the forward LSTM hidden state at each time step and the backward LSTM hidden state Then concatenate these two hidden states to get the final output H t , the hidden state and final output calculation formula are as follows: in, represents the output of the forward output gate, represents the output of the backward output gate, represents the forward current memory unit, Represents the current memory unit, H t represents the output of the bidirectional long short-term memory network, express and splicing; The BiLSTM network consists of a forward propagation LSTM and a backward propagation LSTM. The LSTM processes information through the forget gate, input gate, and output gate. The unidirectional LSTM only calculates the current hidden state h at each time step. t and the current memory cell C t , the specific calculation is shown in formulas (10) and (11): h t =o t ⊙tanh(C t )(10) Among them, f t represents the forget gate, ⊙ represents element-by-element multiplication, C t-1 represents the state of the memory cell at the previous time step, represents the input gate, represents the candidate memory unit, o t represents the output gate; Forget the hidden state h of the previous time step t-1 and the current time step x t As input, the value is scaled to between 0 and 1 through the sigmoid activation function, and the forget gate f t The calculation of is given by formula (12): f t =σ(W f ×[h t-1 ,x t ]+b f )(12) Where σ(·) represents the sigmoid activation function, W f represents the weight matrix of the forget gate, × represents matrix multiplication, b f Represents the bias term of the forget gate; Input Gate Selectively store new information in candidate memory cells In , the calculation formulas for the input gate and candidate memory cells are as follows: in, represents the weight matrix of the input gate, represents the bias term of the input gate, W c represents the weight matrix of the candidate memory unit, b c represents the bias term of the candidate memory unit, and tanh(·) represents the tanh activation function; Output gate o t The specific calculation formula is as follows: the t =σ(W o ×[h t-1 ,x t ]+b o (15) Among them, W o represents the weight matrix of the output gate, b o Represents the bias term of the output gate; BiLSTM captures data features from both positive and negative directions, connecting the current time point with past information and future information; Step 5: MFTDE network life prediction By integrating multi-state indicators to reflect the capacity degradation process, the remaining service life of the battery is predicted.

3. The method for predicting the remaining service life of a lithium battery according to claim 2, characterized in that: The specific steps are as follows: In step five, the specific steps are as follows: Step 51), input multiple status indicators of battery degradation; Step 52), using the Pearson correlation coefficient to screen different status indicators; Step 53), using CNN to perform multi-feature fusion on the selected battery degradation status indicators; Step 54), input the fused features into the attention graph convolutional network to extract temporal dependencies; Step 55), concatenating the outputs obtained in step 53) and step 54), and inputting them into the BiLSTM network for temporal dependency enhancement; Step 56), use full connection on the output of BiLSTM to obtain capacity prediction; Step 57), based on the capacity threshold, calculate the predicted remaining service life L pr , and its calculation formula is as follows: L pr =L th -L st (16) Among them, L th Indicates the number of charge and discharge cycles when the capacity reaches the failure threshold, L st Indicates the number of charge and discharge cycles at the prediction starting point; Remaining useful life prediction using MFTDE network.

Images