CN115510612A - Method for predicting remaining service life of lithium ion battery - Google Patents

Abstract

The invention relates to the field of lithium ion battery capacity detection, and discloses a method for predicting the remaining service life of a lithium ion battery, which comprises the following steps: respectively extracting V, I and T data in the charging process of each battery, and processing the data into matrix shapes required by two subnets; inputting data into an AFSC subnet, and adaptively giving a weight to each element in each charging characteristic spectrogram; the third step: inputting data into a ConvLSTM subnet to obtain hidden information of which the characteristic spectrogram fuses 20 kinds of cycle states; the contributions of the two subnets are fused through the two multilayer perceptrons, a high-accuracy early life prediction value is provided, and a predictor is guided to carry out RUL prediction; the life predictor remains muted until the percentage remaining useful life reaches a threshold value, and the predictor is activated to predict the remaining useful life. The method and the device improve the capability of the model for identifying the aging characteristics of the battery and provide high-precision prediction of the remaining service life of the battery.

Description

A method for predicting the remaining service life of lithium-ion batteries

technical field

The invention relates to the field of lithium-ion battery capacity detection, in particular to a method for predicting the remaining service life of a lithium-ion battery.

Background technique

In recent years, lithium-ion batteries have achieved significant advantages in terms of high energy density, reduced memory effect, low self-discharge rate, and long life cycle. It has played an important role in many fields such as electric vehicles, portable electronic products, aerospace, and intelligent power systems. However, with the passage of time, lithium-ion batteries will inevitably experience aging phenomena and performance degradation, which is manifested as a decrease in capacity and an increase in internal resistance. Battery aging can lead to battery leakage, insulation damage, and partial short circuits. If not detected in time, it can lead to more serious conditions. Prediction and health management of battery life is a method and technology to evaluate system reliability, detect early failures and predict failure progress under actual life cycle conditions, allowing users to make maintenance decisions in advance and prevent losses caused by unexpected failures . Therefore, accurate remaining service life prediction has very important practical significance for monitoring the battery health status, making timely failure replacement, and ensuring the safety of use.

However, battery capacity fading is a typical long-term series data, and it is a severe challenge to predict the later capacity fading trend using limited early cycle data. In the current prediction models, most models extract HIs from the charging and discharging process, such as internal resistance, discharge power, AC impedance, etc. These HIs are usually difficult to obtain online, and HIs such as discharge power are affected by random loads of equipment and vary. , which brings more data noise to the model, resulting in poor robustness. Not only that, in the early cycle of the battery, the changes of various HIs are very small, which makes the modeling work of early life prediction more challenging. Furthermore, the size of the battery dataset on which the model is trained greatly affects the generalization performance of the model.

To sum up, the current lithium-ion battery remaining service life prediction model has insufficient ability to solve the long-term dependence and gradient explosion problems caused by long-term sequence prediction, and is affected by random workload interference and cannot identify early cycle data. Key challenges such as small differences and predictions, small data sets used, etc., lead to key problems such as low prediction accuracy, weak robustness, and poor generalization performance of the model for the remaining service life. For this reason, we propose a lithium-ion battery Remaining useful life prediction method.

Contents of the invention

(1) Solved technical problems

Aiming at the deficiencies of the prior art, the present invention provides a method for predicting the remaining service life of a lithium-ion battery to solve the above problems.

(2) Technical solution

In order to achieve the above-mentioned purpose, the present invention provides the following technical solutions:

A method for predicting the remaining service life of a lithium-ion battery, comprising the following steps:

Step 1: Extract the V, I and T data of each battery during the charging process as the HI to measure the battery aging trend, and process the data into the matrix shape required by the two subnets;

Step 2: Input the data into the AFSC subnet, go through depth convolution, global attention, point convolution and local attention respectively, adaptively adjust the contribution of V, I and T to the model and assign corresponding weights, local attention The mechanism traverses the sampling data of different cycles under the same feature, and adaptively assigns a weight to each element in each charging feature spectrum;

Step 3: Input the data into the ConvLSTM subnet, firstly process the data through deep convolution, and adjust the input weight of the feature data adaptively in the global attention in 20 cyclic states. Before the local attention is embedded, the ConvLSTM passes " The "gate" structure determines the retention, forgetting and output of features in the input time series, and the obtained feature spectrum combines the hidden information of 20 cyclic states;

Step 4: The contributions of the two sub-networks are fused through two multi-layer perceptrons to provide a high-accuracy early life prediction value _NEOL and guide the predictor to perform RUL prediction;

Step 5: The lifetime predictor remains silent until the percentage remaining useful life N _RUL,% reaches a threshold of 10%, 7.5%, 5% or 2.5%, and the predictor is activated to predict the remaining useful life.

Preferably, the first step is specifically implemented according to the following steps:

S1: Set up an experimental data set consisting of 124 commercial lithium iron phosphate/graphite batteries with a rated capacity of 1.1Ah and a rated voltage of 3.3V;

S2: In a 48-channel Abin charge-discharge cabinet and a 30°C incubator, cycle to failure under 72 charging strategies and a fixed discharge rate;

S3: Randomly divided into training set and test set according to the ratio of 8:2;

S4: The first 5 cycles of each battery are regarded as the measure of the healthiest time of the battery, which is recorded as the initial state, and the data of the 15 cycles before the starting point of the prediction are recorded as the real-time state.

Preferably, the steps of the second step are as follows:

S1: Depth convolution is performed on the original 3-dimensional matrix. One convolution kernel is responsible for one channel. Take one of the channels as an explanation, denoted as X ^{(i, V)} , and zero-fill X ^{(i, V)} . Take out the convolved sub-matrix from X ^{(i, V)} , denote it as X ^{(i, V)} (n), and its convolution process can be expressed as:

和

分别是来自第k₁个2-D卷积核的权重和偏置，⊙为哈达玛积，

为X^(i,V)(n)与第k₁个卷积核的运算结果，为了对每条曲线进行单独卷积，卷积核的宽度和步进被设为1，每次卷积的卷积核的数量为1，而X^(i,V)(n)与卷积核的形状相同，输出的第k₁个2-D特征映射可以表示为：in,

and

are the weight and bias from the k _1st 2-D convolution kernel, ⊙ is the Hadamard product,

is the operation result of X ^(i,V) (n) and the kth _1st convolution kernel. In order to perform separate convolution on each curve, the width and step of the convolution kernel are set to 1, and each convolution The number of convolution kernels is 1, and X ^(i,V) (n) has the same shape as the convolution kernel, and the output k _- th 2-D feature map can be expressed as:

The same operation is performed on the I and T channels, and the output results will be batch normalized, max pooled and LeakyRectified Linear Unit activation. The four layers are usually operated as a convolution unit. The Leaky ReLU activation function (α=0.05) is as follows :

S2: The output volume is defined as V, and the global average pooling layer is used to obtain the attention scores s _n of the three channels to adaptively adjust their contribution to the model, and the feature map of the nth channel in V is recorded as For V _n , s _n is calculated as follows:

利用全局权重因子自适应赋予了模型对不同特征映射的权重，使得模型能专注在重要变量的特征提取上，计算如下所示：Normalize the resulting attention scores to get the final global weighting factor

The global weight factor is used to adaptively endow the model with weights for different feature maps, so that the model can focus on the feature extraction of important variables. The calculation is as follows:

By connecting g _n (n=1,2,...,K) will get the weighted output volume

进行补零中并取第m个卷积核大小的子矩阵记为

则点卷积的卷积结果可表示为：S3: yes

Perform zero padding and take the sub-matrix of the size of the mth convolution kernel as

Then the convolution result of point convolution can be expressed as:

是

与第k₂个卷积核的运算结果，卷积前后形状相同，

和

分别是来自第k₂个3-D卷积核的权重和偏置，输出的第k₂个2-D特征映射可以表示为：

yes

The operation result of the kth _2nd convolution kernel has the same shape before and after convolution,

and

are the weights and biases from the k2th ₃ -D convolution kernel, respectively, and the output k2th ₂ -D feature map can be expressed as:

的注意力权值，通过两个全连接层生成注意力权重矩阵A，对应的元素A_i,j可表示为：S4: Denote the output volume of the previous convolutional layer as F, define a matrix A with the same shape as F, and the elements A _{i, j} in A are the elements in the corresponding feature map

The attention weights of , the attention weight matrix A is generated through two fully connected layers, and the corresponding elements A _{i, j} can be expressed as:

Among them, δ and ω are weights, b and c are offsets, n _Fc is the number of neurons, the subscript is the index of the element in the matrix, g(·) and f(·) represent the hyperbolic tangent function and the Sigmoid function respectively , respectively as follows:

The Sigmoid function controls the elements of the weight matrix within 0-1, through the Hadamard product control of the two matrices.

The size of the information flow into the next layer of network, the output output volume L ^l is as follows:

L ^l = A⊙F ^l .

Preferably, the local attention in the third step is to further extract useful information from the feature spectrum processed by the convolutional LSTM, so that the model focuses on the common features that all data frames have.

Preferably, the specific steps of the third step are as follows:

The input matrix first undergoes two deep convolution unit operations and then embeds the global attention mechanism to increase the dimension of the obtained input volume into a 5-dimensional tensor to meet the input of ConvLSTM, the gate function in the ConvLSTM cell and the key to data stream transmission The equation is as follows:

f _t = σ(W _Xf *X _t +W _hf *h _t-1 +b _f );

_{i t} = σ(W _Xi *X _t +W _hg *h _t-1 +bi );

g _t = tanh(W _Xg *X _t +W _hg *h _t-1 +b _g );

o _t = σ(W _Xo *X _t +W _ho *h _t-1 +b _o );

C _t = f _t ⊙C _t-1 +i _t ⊙g _t ;

h _t = o _t ⊙ tanh(C _t );

Among them, f represents the output of the forget gate, i represents the output of the input gate, o represents the output of the output gate, g represents the candidate memory, C represents the cell state of ConvLSTM, h represents the output of the hidden layer, X represents the input, W _{X ~} and W _{h ~} It is a 2-D convolution kernel, the subscript t represents the corresponding moment, the symbol "*" represents the convolution operation, and "⊙" represents the Hadamard product. The output of ConvLSTM becomes the output of subnetwork 2 after adding local attention.

Preferably, the fourth step specifically includes:

通过两个多层感知器融合两个子网的贡献值，对电池的剩余使用寿命进行预测，计算如下：Concatenate the outputs of the two subnets and define as

The remaining service life of the battery is predicted by fusing the contribution values of the two subnetworks through two multi-layer perceptrons, which are calculated as follows:

(3) Beneficial effects

Compared with the prior art, the method for predicting the remaining service life of the lithium-ion battery provided by the present invention has the following beneficial effects:

1. The remaining service life prediction method of lithium-ion batteries is based on battery charging process data modeling, which eliminates the interference of random workloads on the model, making the model more robust, practical and generalizable.

2. The method for predicting the remaining service life of lithium-ion batteries uses directly measurable V, I and T data as HI. The data is easy to obtain and is strongly correlated with battery life, which is conducive to real-time monitoring of battery health and life prediction.

3. The method for predicting the remaining service life of lithium-ion batteries proposes for the first time a predictor that can be used for both early prediction of battery life and RUL.

4. In this lithium-ion battery remaining service life prediction method, the AFSC subnet is good at adaptively adjusting the input weight of early cycle data from early cycle data and performing fusion between different features; the ConvLSTM subnet is good at capturing time and space from late cycle data feature, which reduces the redundancy of memory cells, effectively captures the long-term dependencies of long sequence data, and reduces the phenomenon of gradient explosion. The fusion of the two contributes to accurate prediction of both early and late RUL simultaneously.

5. In this lithium-ion battery remaining service life prediction method, the addition of the attention mechanism improves the model's ability to extract important information and discard useless information, further improving the prediction accuracy of the model.

Description of drawings

FIG. 1 is a schematic flow diagram of a method for predicting remaining service life according to an embodiment of the present invention;

Fig. 2 is the capacity fading curve diagram of 124 battery data sets used in the embodiment of the present invention;

Fig. 3 is the regularity difference of the embodiment of the present invention V, I and T under different cycles;

Fig. 4 is the data input structure of two sub-networks of the embodiment of the present invention;

FIG. 5 is a schematic structural diagram of a battery remaining service life prediction model according to an embodiment of the present invention;

FIG. 6 is a schematic diagram of the structure of ConvLSTM according to the embodiment of the present invention;

Figure 7 shows the early prediction performance of the model for battery life under different prediction starting points in the embodiment of the present invention;

Fig. 8 shows the RUL prediction performance of the model under different warning points according to the embodiment of the present invention.

detailed description

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Example

Please refer to Figures 1-8, the method for predicting the remaining service life of a lithium-ion battery provided by an embodiment of the present invention includes the following steps:

Step 1, extract the V, I and T data of each battery during the charging process as the HI to measure the battery aging trend, and process the data into the matrix shape required by the two sub-networks.

Step 1 is specifically implemented according to the following steps:

In step 1.1, the experimental data set consists of 124 commercial lithium iron phosphate/graphite batteries with a rated capacity of 1.1Ah and a rated voltage of 3.3V. These batteries were cycled to failure under 72 charging strategies and fixed discharge rates in a 48-channel Abin charge-discharge cabinet and a 30°C incubator. The data set is randomly divided into training set and test set according to the ratio of 8:2, which means that there will be 99 battery data with 52 charging strategies and 25 battery data with 20 charging strategies, which are used for RUL estimation and RUL respectively. Validation of model generalization performance.

Figure 2 presents the capacity decay curves for 124 battery datasets.

In step 1.2, the V, I, and T data of the battery show different degrees of deviation as the cycle progresses. Moreover, batteries with different lifespans have shown obvious differences in the cycle of the initial state of the battery, and these regular differences are strongly correlated with the lifespan of the battery. Therefore, the input raw data should reflect the differences in different loop states. The first 5 cycles of each battery are regarded as the measure of the healthiest time of the battery, which is recorded as the initial state. The data of 15 cycles before the predicted starting point is recorded as the real-time state, and is used to compare with the initial state of the battery. This data input structure can be used for both early prediction of battery life and real-time prediction of remaining service life in the later period, depending on the selection of the starting point of prediction, and the flexibility of the model is very high. For example, when using the V, I and T data of the first 20 cycles for early life prediction, the prediction starting point = 20, the data structure is [1,5]+[6,20] (initial state + real-time state), and the numbers in the interval represent The sequence number of the loop to use. As for RUL prediction, for a battery with a cycle life of 1000, if the prediction starting point=800, the data structure is [1,5]+[786,800].

Figure 3 shows the regularity difference of V, I and T under different cycles.

Figure 4 shows the data input structure of the two subnets.

Step 2, input the data into the AFSC subnet, and go through the steps of deep convolution, global attention, point convolution and local attention, respectively, to adaptively adjust the contribution of V, I and T to the model and assign corresponding weights; local attention The force mechanism traverses the sampling data of different cycles under the same feature, and adaptively assigns a weight to each element in each charging characteristic spectrum, so that the AFSC subnetwork has a high degree of control over the variation rules at each time sampling point. sensitivity.

Step 2 is specifically implemented according to the following steps:

Step 2.1, perform deep convolution on the original 3D matrix, and one convolution kernel is responsible for one channel. Take one of the channels (voltage) for illustration, denoted as X ^(i,V) . In order to avoid loss of edge data, zero padding is performed on X ^{(i, V)} . Take out the convolved sub-matrix from X ^{(i, V)} , denote it as X ^{(i, V)} (n), and its convolution process can be expressed as:

和

分别是来自第k₁个2-D卷积核的权重和偏置，⊙为哈达玛积。

为X^(i,V)(n)与第k₁个卷积核的运算结果。为了对每条曲线进行单独卷积，卷积核的宽度和步进被设为1，每次卷积的卷积核的数量为1。而X^(i,V)(n)与卷积核的形状相同。输出的第k₁个2-D特征映射可以表示为：in,

and

are the weight and bias from the k _1th 2-D convolution kernel, respectively, and ⊙ is the Hadamard product.

is the operation result of X ^(i,V) (n) and the kth _1st convolution kernel. To convolve each curve individually, the kernel width and stride are set to 1, and the number of kernels per convolution is 1. And X ^{(i, V)} (n) has the same shape as the convolution kernel. The output k1th ₂ -D feature map can be expressed as:

The same operation is performed on the I and T channels, and the output results will be subjected to batch normalization, maximum pooling, and LeakyRectified Linear Unit (Leaky ReLU) activation to alleviate the gradient explosion during training. The above four layers are usually operated as a convolutional unit. The Leaky ReLU activation function (α=0.05) is as follows:

In step 2.2, the output volume of the previous step is defined as V, and the global average pooling layer is used to obtain the attention scores s _n of the three channels to adaptively adjust their contributions to the model. Take the feature map of the nth channel in V and record it as V _n , and s _n is calculated as follows:

利用全局权重因子自适应赋予了模型对不同特征映射的权重，使得模型能专注在重要变量的特征提取上，计算如下所示：Normalize the resulting attention scores to get the final global weighting factor

The global weight factor is used to adaptively endow the model with weights for different feature maps, so that the model can focus on the feature extraction of important variables. The calculation is as follows:

用作下一层的输入。By connecting g _n (n=1,2,...,K) will get the weighted output volume

used as input for the next layer.

进行补零中并取第m个卷积核大小的子矩阵记为

则点卷积的卷积结果可表示为：Step 2.3, yes

Perform zero padding and take the sub-matrix of the size of the mth convolution kernel as

Then the convolution result of point convolution can be expressed as:

是

与第k₂个卷积核的运算结果，卷积前后形状相同。

和

分别是来自第k₂个3-D卷积核的权重和偏置。输出的第k₂个2-D特征映射可以表示为：Similarly,

yes

The operation result of the kth _2nd convolution kernel has the same shape before and after convolution.

and

are the weights and biases from the k _- th 3-D convolution kernel, respectively. The output k2th ₂ -D feature map can be expressed as:

的注意力权值。本文中通过两个全连接层生成注意力权重矩阵A，对应的元素A_i,j可表示为：Step 2.4, denote the output volume of the previous convolutional layer as F. In order to explore the contribution of the data at each time sampling point in the variable curve that can be directly measured to the model, a matrix A with the same shape as F is defined, and the elements A _i,j in A are the elements in the corresponding feature map

attention weight. In this paper, the attention weight matrix A is generated through two fully connected layers, and the corresponding elements A _{i, j} can be expressed as:

Among them, δ and ω are weights, b and c are biases, and n _Fc is the number of neurons. The lower corner is the index of the element in the matrix. g(·) and f(·) represent the hyperbolic tangent function and the Sigmoid function respectively, which are expressed as follows:

The Sigmoid function controls the elements of the weight matrix within 0-1, and controls the size of the information flow entering the next layer of the network through the Hadamard product of the two matrices. The output volume L ^l of the output is as follows:

L ^l ＝A⊙F ^l (11)

FIG. 5 shows a schematic structure of a new lithium-ion battery remaining service life prediction model based on AFSC-ConvLSTM disclosed in the present invention.

Step 3: Input the data into the ConvLSTM subnetwork, firstly process the data through deep convolution, and adjust the input weight of the feature data adaptively in the global attention in 20 cyclic states; before the local attention is embedded, the ConvLSTM passes the "gate "The structure determines the retention, forgetting and output of features in the input time series, and the obtained feature spectrum combines the hidden information of 20 cyclic states. Local attention is to further extract useful information from the feature spectrum processed by the convolutional LSTM, so that the model "focuses" on the common features that all data frames have.

Step 3 is specifically implemented according to the following steps:

Step 3.1, similar to subnetwork 1, the input matrix is embedded in the global attention mechanism after two deep convolution unit operations, which will not be repeated here. Increase the dimension of the obtained input volume into a 5-dimensional tensor to meet the input of ConvLSTM. The key equations of the gate function and data flow transmission in the ConvLSTM cell are as follows:

f _t ＝σ(W _Xf *X _t +W _hf *h _t-1 +b _f ) (12)

i _t ＝σ(W _Xi* X _t +W _hg *h _t-1 +b _i ) (13)

g _t ＝tanh(W _Xg *X _t +W _hg *h _t-1 +b _g ) (14)

o _t ＝σ(W _Xo *X _t +W _ho *h _t-1 +b _o ) (15)

C _t ＝f _t ⊙C _t-1 +i _t ⊙g _t (16)

h _t ＝o _t ⊙tanh(C _t ) (17)

Among them, f represents the output of the forget gate; i represents the output of the input gate; o represents the output of the output gate; g represents the candidate memory; C represents the cell state of ConvLSTM; h represents the output of the hidden layer; X represents the input; W _{X ~} and W _{h ~} is a 2-D convolution kernel; the subscript t indicates the corresponding moment. The symbol "*" represents the convolution operation, and "⊙" represents the Hadamard product. The output of ConvLSTM becomes the output of subnetwork 2 after adding local attention, which will not be described here.

Figure 6 shows the structure of ConvLSTM.

In step 4, the contributions of the two subnetworks are fused through two multi-layer perceptrons to provide a high-accuracy early life prediction value _NEOL and guide the predictor to perform RUL prediction.

Step 4 specifically includes:

通过两个多层感知器融合两个子网的贡献值，对电池的剩余使用寿命进行预测，计算如下：Concatenate the outputs of the two subnets and define as

The remaining service life of the battery is predicted by fusing the contribution values of the two subnetworks through two multi-layer perceptrons, which are calculated as follows:

The interpretation of the symbols is similar to Equation (8), except that f(·) in Equation (18) is expressed as the activation function shown in Equation (3).

Fig. 7 shows the early prediction performance of the model on battery life under different prediction starting points of the method disclosed in the present invention.

Step 5, the life predictor remains silent until the percentage remaining useful life N _RUL,% (relative to _NEOL ) reaches a threshold of 10%, 7.5%, 5% or 2.5%, and the predictor is activated to predict the remaining useful life .

Fig. 8 shows the RUL prediction performance of the model under different warning points of the method disclosed in the present invention.

The focus of the above embodiments of the present invention is to extract the V, I and T data in each battery charging process as HI; input HI in parallel into the AFSC and ConvLSTM sub-networks, and the AFSC adaptively adjusts the input weight of the early cycle data and performs different Fusion between features, ConvLSTM is good at capturing spatio-temporal features from late cycle data, capturing long-term dependencies in HI sequences; embedding global attention mechanism and local attention mechanism for the two subnets, and identifying V, For the key data frames in the I and T matrices, the local attention mechanism helps to select important features accurate to each data sampling point under the key data frames, which improves the ability of the model to identify battery aging characteristics; the output of the last two subnets is passed through two The output of the model built by a multi-layer perceptron provides a high-precision prediction of the remaining service life of the battery.

Although the embodiments of the present invention have been shown and described, those skilled in the art can understand that various changes, modifications and substitutions can be made to these embodiments without departing from the principle and spirit of the present invention. and modifications, the scope of the invention is defined by the appended claims and their equivalents.

Claims (6)

1. A lithium-ion battery remaining service life prediction method, is characterized in that, comprises the following steps: Step 1: Extract the V, I and T data of each battery during the charging process as the HI to measure the battery aging trend, and process the data into the matrix shape required by the two subnets; Step 2: Input the data into the AFSC subnet, go through depth convolution, global attention, point convolution and local attention respectively, adaptively adjust the contribution of V, I and T to the model and assign corresponding weights, local attention The mechanism traverses the sampling data of different cycles under the same feature, and adaptively assigns a weight to each element in each charging feature spectrum; Step 3: Input the data into the ConvLSTM subnet, firstly process the data through deep convolution, and adjust the input weight of the feature data adaptively in the global attention in 20 cyclic states. Before the local attention is embedded, the ConvLSTM passes " The "gate" structure determines the retention, forgetting and output of features in the input time series, and the obtained feature spectrum combines the hidden information of 20 cyclic states; Step 4: The contributions of the two sub-networks are fused through two multi-layer perceptrons to provide a high-accuracy early life prediction value _NEOL and guide the predictor to perform RUL prediction; Step 5: The life predictor remains silent until the percentage remaining useful life X _RUL,% reaches a threshold of 10%, 7.5%, 5% or 2.5%, and the predictor is activated to predict the remaining useful life. 2. The lithium-ion battery remaining service life prediction method according to claim 1, characterized in that: the first step is specifically implemented according to the following steps: S1: Set up an experimental data set consisting of 124 commercial lithium iron phosphate/graphite batteries with a rated capacity of 1.1Ah and a rated voltage of 3.3V; S2: In a 48-channel Abin charge-discharge cabinet and a 30°C incubator, cycle to failure under 72 charging strategies and a fixed discharge rate; S3: Randomly divided into training set and test set according to the ratio of 8:2; S4: The first 5 cycles of each battery are regarded as the measure of the healthiest time of the battery, which is recorded as the initial state, and the data of the 15 cycles before the starting point of the prediction are recorded as the real-time state. 3. The lithium-ion battery remaining service life prediction method according to claim 1, characterized in that: the steps of the second step are as follows: S1: Depth convolution is performed on the original 3-dimensional matrix. One convolution kernel is responsible for one channel. Take one of the channels as an explanation, denoted as X ^{(i, V)} , and zero-fill X ^{(i, V)} . Take out the convolved sub-matrix from X ^{(i, V)} , denote it as X ^{(i, V)} (n), and its convolution process can be expressed as:

in,

and

are the weight and bias from the k _1st 2-D convolution kernel, ⊙ is the Hadamard product,

is the operation result of X ^(i,V) (n) and the kth _1st convolution kernel. In order to perform separate convolution on each curve, the width and step of the convolution kernel are set to 1, and each convolution The number of convolution kernels is 1, and X ^(i,V) (n) has the same shape as the convolution kernel, and the output k _- th 2-D feature map can be expressed as:

The same operation is performed on the I and T channels, and the output results will be batch normalized, max pooled and LeakyRectified Linear Unit activation. The four layers are usually operated as a convolution unit. The Leaky ReLU activation function (α=0.05) is as follows :

S2: The output volume is defined as V, and the global average pooling layer is used to obtain the attention scores s _n of the three channels to adaptively adjust their contribution to the model, and the feature map of the nth channel in V is recorded as For V _n , s _n is calculated as follows:

Normalize the resulting attention scores to get the final global weighting factor

The global weight factor is used to adaptively endow the model with weights for different feature maps, so that the model can focus on the feature extraction of important variables. The calculation is as follows:

By connecting g _n (n=1,2,…,K) will get the weighted output volume

S3: yes

Perform zero padding and take the sub-matrix of the size of the mth convolution kernel as

Then the convolution result of point convolution can be expressed as:

yes

The operation result of the kth _2nd convolution kernel has the same shape before and after convolution,

and

are the weights and biases from the k2th ₃ -D convolution kernel, respectively, and the output k2th ₂ -D feature map can be expressed as:

S4: Denote the output volume of the previous convolutional layer as F, define a matrix A with the same shape as F, and the elements A _{i, j} in A are the elements in the corresponding feature map

The attention weights of , the attention weight matrix A is generated through two fully connected layers, and the corresponding elements A _{i, j} can be expressed as:

Among them, δ and ω are weights, b and c are offsets, n _Fc is the number of neurons, the subscript is the index of the element in the matrix, g(·) and f(·) represent the hyperbolic tangent function and the Sigmoid function respectively , respectively as follows:

The Sigmoid function controls the elements of the weight matrix within 0-1, and controls the size of the information flow entering the next layer of the network through the Hadamard product of the two matrices. The output volume L ^l of the output is as follows: L ^l = A⊙F ^l . 4. The lithium-ion battery remaining service life prediction method according to claim 1, characterized in that: in the third step, local attention is to further extract useful information from the feature spectrum after convolution LSTM processing, so that the model Focus on the common characteristics that all data frames share. 5. The lithium-ion battery remaining service life prediction method according to claim 1, characterized in that: the specific steps of the third step are as follows: The input matrix first undergoes two deep convolution unit operations and then embeds the global attention mechanism to increase the dimension of the obtained input volume into a 5-dimensional tensor to meet the input of ConvLSTM, the gate function in the ConvLSTM cell and the key to data stream transmission The equation is as follows: f _t = σ(W _Xf *X _t +W _hf *h _t-1 +b _f ); i _t = σ(W _Xi* X _t +W _hg *h _t-1 +b _i ); g _t = tanh(W _Xg *X _t +W _hg *h _t-1 +b _g ); o _t = σ(W _Xo *X _t +W _ho *h _t-1 +b _o ); C _t = f _t ⊙C _t-1 +i _t ⊙g _t ; h _t = o _t ⊙ tanh(C _t ); Among them, f represents the output of the forget gate, i represents the output of the input gate, o represents the output of the output gate, g represents the candidate memory, C represents the cell state of ConvLSTM, h represents the output of the hidden layer, X represents the input, W _{X ~} and W _{h ~} It is a 2-D convolution kernel, the subscript t represents the corresponding moment, the symbol "*" represents the convolution operation, and "⊙" represents the Hadamard product. The output of ConvLSTM becomes the output of subnetwork 2 after adding local attention. 6. The lithium-ion battery remaining service life prediction method according to claim 1, characterized in that: the fourth step specifically includes: Concatenate the outputs of the two subnets and define as

The remaining service life of the battery is predicted by fusing the contribution values of the two subnetworks through two multi-layer perceptrons, which are calculated as follows:

Images