CN117452261A - A method for predicting cycle life of lithium batteries - Google Patents

Abstract

The invention discloses a lithium battery cycle life prediction method, which takes into account the different characteristics of the lithium battery during the charge and discharge process. First, a recursive graph is used as a time series imaging method to convert charging data of different cycles into multi-dimensional data, from discharge to Extract relevant features from the data and analyze their correlation with battery cycle life; secondly, use a depth-separable three-dimensional convolution model to reduce parameter calculations and speed up model training, and introduce a 3D channel attention (3DCA) module to reduce the model complexity and increase the interactivity between channels; finally, an ablation experiment was conducted to explore the impact of different time series imaging methods on the accuracy of model prediction results. Compared with the existing technology, the present invention can accurately predict cycle life with fewer window periods. Simultaneous ablation experimental results prove that the time series imaging method can increase the data dimension and provide more information, which effectively improves the accuracy of the model.

Description

A method for predicting the cycle life of lithium batteries

Technical field

The present invention relates to the technical field of lithium batteries, and in particular to a method for predicting the cycle life of lithium batteries.

Background technique

Lithium batteries are widely used in many fields due to their long service life and high energy density. However, battery life is generally affected by battery operating and environmental conditions, such as charging rate, operating voltage, current, and temperature. During battery use, its performance degrades as the remaining battery life (RUL) changes. Moreover, even lithium batteries produced in the same batch may have different final cycle lives. Using a battery with a substandard cycle life creates a safety hazard because the battery will degrade prematurely. Therefore, in the early testing of the battery, it is necessary to conduct a comprehensive and repeated evaluation of the battery's cycle life to ensure that the battery meets the standard requirements. But like the testing of many complex systems, battery life testing usually takes months or even years, which greatly increases the time cost of the experiment.

To solve this problem, methods for predicting battery cycle life using external data have been proposed. These methods aim to predict battery cycle life using data collected in previous cycles, thereby significantly reducing the time required for battery life experiments and providing more opportunities for battery production and use. Therefore, accurately predicting battery cycle life is a key goal for battery failure prediction and shortening experimental time.

Contents of the invention

Purpose of the invention: In view of the problems existing in the existing technology, the present invention provides a lithium battery cycle life prediction method, which can improve the accuracy of the cycle life prediction model.

Technical solution: The present invention proposes a lithium battery cycle life prediction method, which includes the following steps:

Step 1: Collect lithium battery cycle life data and divide the data; for the voltage, current, and temperature VIT data during the charging process, use deep learning methods to automatically extract features, and use the recursive graph RP to transform the one-dimensional VIT data during the charging process respectively. are three two-dimensional matrices, and these two-dimensional matrices are spliced in the depth direction to form a three-dimensional matrix;

Step 2: Analyze the incremental capacity (IC) curve and discharge capacity Q-V curve of the discharge process, introduce the earth's movement distance EMD to calculate the curve changes of the battery between different cycles, and finally extract the key discharge characteristics from the curve and observe Correlation of key discharge characteristics with battery cycle life;

Step 3: For the modeling of the charging part data, build a 3D depth-separable convolution model, splice the VIT three-dimensional matrix under different cycles into a four-dimensional matrix, and reduce the calculation amount of parameters through depth separation convolution and point-by-point convolution. And the 3D channel attention module is introduced to strengthen the data interaction between VIT channels to ensure that the interaction between some channels can be achieved without increasing the model training parameters;

Step 4: For the modeling of the discharge part data, build a CBR layer and an MP layer to perform convolution and pooling operations on the data; the CBR layer structure consists of a convolution layer, a batch normalization layer and a ReLU activation layer; the The MP layer structure consists of a maximum pooling layer and several convolutional layers with different strides. The maximum pooling layer and the convolutional layer with a stride of 2 reduce the number of channels by half and maintain the number of input and output channels after splicing. constant;

Step 5: Feature fusion is performed on the feature vectors output by the charging process and discharging process models, and the final output is the lithium battery cycle life prediction result. Finally, an ablation experiment is performed to verify the effectiveness of the time series imaging method.

Further, in step 1, the two collected lithium battery data sets are divided and the RP diagram is used to extract the VIT data characteristics of the charging process, including the following steps:

Step 1.1: After removing problematic batteries during the data collection process, divide the two data sets into training sets, verification sets and test sets according to a certain proportion;

Step 1.2: Divide the two data sets into multiple intervals according to the number of cycles, and divide the training data, verification data and test data of each interval in proportion;

Step 1.3: The voltage, current and temperature during battery charging are regarded as functions of capacity ratio. These data are composed of one-dimensional data of different cycles; RP chart is a time series imaging method that combines the charging between different charging cycles. Data is converted into multi-dimensional images to obtain more complete lithium battery degradation information;

Step 1.4: The calculation principle of the RP diagram is as follows: for a time series u _k (k=1,2,...,n) with a sampling time interval Δt and a length n, re-embed the dimension m and the delay time τ To construct this time series, the reconstructed expression is x _i =(u _i ,u _i+τ ,...,u _i+(m-1)τ ),i=1,2,...,n-(m -1)τ, then the Euclidean distance s _ij between the i-th and j-th data points x _i and x _j in the reconstruction space is expressed as formula (1):

s _ij =||x _i -x _j ||,i,j＝1,2,...,n-(m-1)τ (1)

Then calculate the recursion value R(i,j):

R(i,j)＝H(ε _i -s _ij ),i,j＝1,2,...,n (2)

In the formula, ε _i is the threshold, which is fixed or changed with i, and H(x) represents the Heaviside function;

Step 1.5: Convert the one-dimensional time series data during the battery charging process into two-dimensional voltage, current, and temperature RP-VIT data through the RP chart imaging method.

Further, the steps for feature extraction of the battery discharge process in step 2 include:

Step 2.1: Obtain the IC curve characteristics according to the Incremental Capacity (IC) analysis method. IC is calculated by the following formula (3):

In the formula, Q is the battery capacity in the current state, V is the voltage, I is the discharge current, t is the sampling time, IC is the step ratio voltage dQ/dV, as the voltage step window moves, the incremental capacity and The complete relationship between voltages;

Step 2.2: According to the variation curve of the discharge voltage between period i and period j, it is expressed as ΔQ _ij (V) = Q _i (V)-Q _j (V), where i and j both represent the number of cycles, and their sum The greater the difference in values, the more obvious the curve ΔQ _ij (V) will be;

Step 2.3: Introduce the earth moving distance EMD to measure the difference between two adjacent period ΔQ(V) curves.

Further, in step 2.3, the earth moving distance EMD measures the difference between two adjacent period ΔQ(V) curves, and the calculation steps are as follows:

Step 2.3.1: The EMD distance is defined as the estimate of the minimum cost of converting from one distribution P to another distribution Q;

Step 2.3.2: W(P,Q) represents the EMD between two probability distributions, which is defined as follows:

W(P,Q)＝inf _γ∈Π(P,Q) E _(x,y)～γ [||xy||] (4)

In the formula, inf means taking the lower bound of all possible joint probability distributions Π(P,Q), γ is the joint distribution, Π(P,Q) is the set of all joint distributions of P and Q, and P and Q are phases. The distribution of IC curves of two adjacent periods, ||xy|| is the distance between sample x and y, E _{(x, y) ~ γ} [||xy||] is the distance expectation;

Step 2.3.3: The calculation formula of distance expectation is as follows:

In the formula, i and j represent two different periods, and P _{(i, j)} represents the distribution of the IC curves of periods i and j.

Further, in step 3, a 3D depth separable convolution model is built to model the charging part data. The steps include:

Step 3.1: For one charging cycle, use the RP imaging method to convert VIT data into RP-VIT data, and splice multiple VIT three-dimensional matrices under different cycles into a four-dimensional matrix;

Step 3.2: Use three-dimensional convolution to extract features from the four-dimensional matrix. The calculation formula of the feature map is:

In the formula, W _out ×H _out ×C _out and W _in ×H _in ×C _in represent the width, height and number of channels at image output and input respectively, w × h × c represents the size of the convolution kernel, and s represents the step. Amplitude, p represents the filling value, k represents the number of scans, and the number of parameters in the convolution process is (w×h×c+1)×k;

Step 3.3: Use depth-separable three-dimensional convolution to reduce the difficulty of model training and reduce the amount of convolution calculations. Divide the feature maps of each charging cycle into a group, and perform convolution operations within the group. The convolution kernel in the group Generate feature maps;

Step 3.4: Add a point-by-point convolution layer after the depth convolution. The point-by-point convolution operation is similar to the conventional convolution operation. The convolution kernel size is 1×1×n, where n represents the depth direction of the three-dimensional matrix, that is The first n-cycle charging cycle; point-by-point convolution calculation will weight the previous step's map in the depth direction to generate a new feature map, and use multiple convolution kernels to generate multiple feature maps;

Step 3.5: Introduce the 3D Channel Attention (3DCA) module to learn these features. The 3DCA module highlights the specific areas related to the cycle life from the feature map based on the correlation between the input data and the battery cycle life, and combines The attention layer detects the saliency of different channel features in the input data;

Step 3.6: Add a 3D Global Average Pooling (3DGAP) layer after the 3DCA layer; 3DGAP is the average pooling in 3D.

Further, the specific calculation process of the channel attention 3DCA module of the multi-channel three-dimensional convolution in step 3.5 is as follows:

Step 3.5.1: First, considering the aggregated feature y (y∈R ^C ), during the charging process, it is described by the following formula:

ω=σ(W _k y) (7)

In the formula, W _k is the parameter matrix of k×C, ω is the weight, and σ is the activation function;

Step 3.5.2: Calculate the weight ω _i by only considering the interaction between y _i and its adjacent node k, then its weight is rewritten as:

in the formula is the set k of adjacent channels _yi , all channels share the same learning parameters, which is implemented by a 1×1×1 convolution kernel with k kernel size, that is:

ω=σ(Conv _k (y)) (10)

In the formula, Conv _k represents a three-dimensional convolution with a kernel size of 1;

Step 3.5.3: The degree of interaction between channels is adjusted by changing the size of k. The size of kernel k is adaptively determined by:

In the formula, c represents the number of channels, |a| _odd represents the nearest odd distance from a. Through nonlinear mapping ψ, high-dimensional channels have longer distance interactions, while low-dimensional channels have shorter distance interactions.

Further, in step 5, the feature vectors output by the charging process and discharging process models are feature fused, and the final output is the lithium battery cycle life prediction result. The specific implementation steps are:

Step 5.1: Convert the features of different dimensions into one-dimensional feature vectors through the fully connected layer, and splice them to form the final new feature matrix;

Step 5.2: Predict the battery cycle life, split the battery window period into the first n charging cycles n _f and the latest m charging cycles m _l to complete the battery cycle life prediction, and finally conduct an ablation experiment to verify the time series Effectiveness of Imaging Methods.

Beneficial effects:

1. The present invention proposes a deep separable three-dimensional convolutional network model fused channel attention (DS-3DCA-CNN) model that takes into account the charging and discharging process for lithium battery life prediction. First, the recursive graph is used as a time series imaging method to convert the charging data of different cycles into multi-dimensional data; relevant features are extracted from the discharge data and their correlation with the battery cycle life is analyzed. Secondly, depth-separable three-dimensional convolution is used to reduce parameter calculations and speed up model training, and a 3D channel attention (3DCA) module is introduced to reduce model complexity while increasing the interactivity between channels. By analyzing the incremental capacity curve and voltage change curve during the discharge process of lithium batteries, the battery health indicators (HIs) degradation data are derived. These HIs are integrated with charging process features through multi-feature fusion using a CNN model. This comprehensive approach allows for a more thorough characterization of the battery aging process.

2. The present invention proposes a time series imaging method to increase the dimensionality of charging data. Convert one-dimensional time series into a recursive graph to construct a feature matrix to obtain more effective features.

3. Aiming at the problem of too many CNN model parameters and high coupling between measured values during the charging process, the present invention proposes a depth-separable three-dimensional convolution model (DS-3DCA-CNN) that fuses channel attention modules. This method reduces Parameter calculation during the three-dimensional convolution process enhances data interactivity between channels.

4. The present invention proposes a rapid prediction framework that considers charge and discharge characteristics, which can predict battery cycle life under different charging protocols. The effectiveness of this method was verified through different evaluation indicators and time series imaging ablation studies.

5. The present invention processes multi-channel three-dimensional data based on the RP imaging method and the DS-3DCA-CNN model, and can effectively predict the cycle life of the battery during use, thereby distinguishing healthy batteries from unhealthy batteries in advance.

Description of the drawings

Figure 1 is a schematic flow chart of feature extraction based on the discharge process provided by the present invention.

Figure 2 is a schematic structural diagram of the DS-3DCA-CNN provided by the present invention.

Figure 3 is a schematic diagram of the lithium battery cycle life prediction process provided by the present invention.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings.

Combining Figure 1, Figure 2 and Figure 3, the present invention discloses a lithium battery cycle life prediction method, which divides the data set into a training set, a verification set and a test set according to a certain proportion, and divides the two data sets according to the battery cycle life. into multiple intervals, and divide the training data, validation data, and test data in each interval proportionally. The deep learning method is used to automatically extract the characteristics of VIT data during the charging process from the data. Specifically, the one-dimensional voltage, current and temperature data during the charging process are converted into a two-dimensional matrix through a recursive graph to increase the amount of data information, and Three two-dimensional matrices are spliced in the depth direction to form a three-dimensional matrix. Analyze the incremental capacity (IC) curve and discharge capacity (Q-V) curve of the discharge process, and introduce the EMD distance to calculate the curve changes of the battery between different cycles, thereby obtaining the EMD_IC curve and EMD_QV curve, and extracting key points from the curve discharge characteristics. Build a 3D depthwise separable convolution model to model the charging part of the data, reduce the calculation amount of parameters through depthwise separation convolution and point-by-point convolution, and introduce a 3D channel attention module to enhance the data interaction between VIT channels . Build a CBR layer and an MP layer to model the discharge part of the data. The CBR structure consists of a convolution layer, a batch normalization layer and a ReLU activation layer to ensure that the network has a certain regularization effect. The MP structure mainly consists of a maximum pool It consists of a layer and several convolutional layers with different strides. Finally, the feature vectors output by the charging process and discharging process models are feature fused, and the final output is the lithium battery cycle life prediction result. Finally, an ablation experiment is added to verify the effectiveness of the time series imaging method used. Specific steps are as follows:

Step 1: Collect lithium battery cycle life data and divide the data. For the voltage, current, and temperature (VIT) data of the charging process, deep learning methods are used to automatically extract features. Recurrence Plots (RP) are used to convert the one-dimensional VIT data during the charging process into three two-dimensional matrices to increase the amount of data information under different cycles, and these two-dimensional matrices are spliced in the depth direction to form a three-dimensional matrix. .

Step 1.1: After removing some problematic batteries during the data collection process, divide the first 140-battery data set into a training set, a verification set, and a test set in proportions of about 80%, 10%, and 10%; for the second Due to the small sample size, a total of 45 batteries were divided into 45 batteries according to the ratio of 60%, 20% and 20% to verify the effectiveness of the method and the proposed model.

Step 1.2: Since the battery cycle life distribution in the two data sets is uneven, approximately between 140 and 2240, the model may be biased during training. Therefore, in order to ensure that the model can learn the average battery life with different cycles Characteristics, divide the two data sets into multiple intervals according to the number of cycles, and divide the training data, verification data and test data of each interval in proportion.

Step 1.3: Consider the voltage, current, and temperature during battery charging as functions of capacity ratio. These data consist of one-dimensional data at different cycles. As a time series imaging method, RP chart can convert charging data between different charging cycles into multi-dimensional images to obtain more complete lithium battery degradation information.

Step 1.4: The calculation principle of the RP diagram is as follows: for a time series u _k (k=1,2,...,n) with a sampling time interval Δt and a length n, re-embed the dimension m and the delay time τ To construct this time series, the reconstructed expression is x _i =(u _i ,u _i+τ ,...,u _i+(m-1)τ ),i=1,2,...,n-(m -1)τ. Then the Euclidean distance s _ij between the i-th and j-th data points x _i and x _j in the reconstruction space can be expressed as formula (1):

s _ij =||x _i -x _j ||,i,j＝1,2,...,n-(m-1)τ (1)

Then calculate the recursion value R(i,j):

R(i,j)＝H(ε _i -s _ij ),i,j＝1,2,...,n (2)

In the formula, ε _i is the threshold, which can be fixed or changed with i, and H(x) represents the Heaviside function;

Step 1.5: In order to reduce calculation and complexity, the charging capacity is divided into 110 equal parts (range 0~1.1Ah), that is, every 0.01Ah is used as a data sampling point for voltage, current and temperature. This means that there are a total of 110 sampling points as input data for a single charging cycle.

Step 1.6: Convert the one-dimensional time series data of voltage cycle 10, current cycle 150, 300 and temperature cycle 450 into two-dimensional voltage, current, temperature (RP-VIT) data through RP imaging method during battery charging. This can highlight the main features of the data and provide a prerequisite for learning from the feature extraction mechanism in the field of computer vision.

Step 2: For feature extraction during battery discharge, analyze the incremental capacity (IC) curve and discharge capacity (Q-V) curve of the discharge process, and introduce the Earth Mover's Distance (EMD distance) to calculate the battery in different cycles Finally, five key discharge characteristics were extracted from the curve and the correlation between these characteristics and battery cycle life was observed.

Step 2.1: Obtain the IC curve characteristics according to the Incremental Capacity (IC) analysis method. This characteristic indicates the electrochemical reaction inside the battery and reflects the aging process of the battery. IC can be calculated by the following formula (3):

In the formula, Q is the battery capacity in the current state, V is the voltage, I is the discharge current, t is the sampling time, and IC is the step ratio voltage (dQ/dV). As the voltage step window moves, the complete relationship between incremental capacity and voltage can be obtained.

Step 2.2: It can be seen from the transformation of equation (3) that IC is negatively correlated with the derivative of voltage with respect to time, indicating that the gentler the voltage change, the greater the IC value. Figure 1 shows different feature extraction curves during the discharge process. As shown in Figure 1(a) and (b), compared with extracting features directly from the Q-V curve, the features of the IC curve are more significant in different discharge cycles of the battery. The most obvious feature is the peak drift of the IC curve. . It can be seen that as the number of battery cycles increases, the peak height decreases and moves laterally to the left. Therefore, the peak coordinate (PICC) of the IC curve in Figure 1(b) is considered to be the key feature of the IC curve. The horizontal axis mark of its PICC is A1, and the vertical axis mark is B1.

Step 2.3: Figure 1(c) shows the variation curve of discharge voltage between period i and period j, which can be expressed as ΔQ _ij (V) = Q _i (V)-Q _j (V), where i and j both represent The greater the difference between the number of cycles and the sum, the more obvious the curve ΔQ _ij (V) will be.

According to the ΔQ(V) curve, when the values of i and j are the same, the peak values of the ΔQ(V) curve (PQVC) of batteries with different cycle life are different, and the peak value of the battery with a longer cycle life is smaller, indicating that it contains information about battery degradation. Therefore, the ordinate of the PQVC peak is selected as the key feature, recorded as B2;

Step 2.4: Introduce EMD distance to measure the difference between two adjacent period ΔQ(V) curves. Figure 1(d) shows the change trend of the EMD-ΔQ(V) curve between different i-j units after the introduction of EMD to represent changes in different cycles. In the first 10 cycles, the curve showed an accelerated upward trend and reached the first peak. then decreased slightly and continued to increase after approximately 50 cycles (although the chemical and degradation mechanisms are unknown). In addition, the EMD-ΔQ(V) curve has different trends for batteries with different cycle lives, and therefore is also used as a key feature to describe battery degradation;

Step 2.4.1: The EMD distance is defined as the estimate of the minimum cost of transforming from one distribution P to another distribution Q. Compared with Kullback-Leibler (KL) divergence and Jensen-Shannon (JS) divergence, EMD can reflect the distance between two distributions even if there is no overlap or the overlap is small.

Step 2.4.2: W(P,Q) represents the EMD between two probability distributions, which is defined as follows:

W(P,Q)＝inf _γ∈Π(P,Q) E _(x,y)～γ [||xy||] (4)

In the formula, inf means taking the lower bound of all possible joint probability distributions Π(P,Q), γ is the joint distribution, ∏(P,Q) is the set of all joint distributions of P and Q, and P and Q are phases. The distribution of IC curves of two adjacent periods, ‖xy‖ is the distance between sample x and y, and E _{(x, y) ~ γ} [||xy||] is the distance expectation.

Step 2.4.3: The calculation formula of distance expectation is as follows:

In the formula, i and j represent two different periods, and P _{(i, j)} represents the distribution of the IC curves of periods i and j.

Step 2.5: Generate EMD-IC curve from IC curve. Figure 1(e) shows the EMD change trend of IC curves between different batteries. It can be seen from the figure that as the number of battery cycles increases, the EMD-IC curve gradually decreases. After 400 cycles, the acceleration ratio of most batteries shows a downward trend, and the downward trend of the curve is different for each battery. It can be seen that the EMD-IC curve has a certain correlation with battery life, and it is also used as a key feature to describe battery degradation.

Step 3: For modeling of the charging part data, build a 3D depth separable convolution model. Splice the VIT three-dimensional matrices under different cycles into a four-dimensional matrix. The calculation amount of parameters is reduced through depth separation convolution and point-wise convolution, and a 3D channel self-attention module is introduced to strengthen the data interaction between VIT channels to ensure that some channels can be reached without increasing model training parameters. interactions between.

Step 3.1: For one charging cycle, after converting VIT data to RP-VIT data, there are three 110×110 matrices. This matrix form is very similar to a multi-channel image, corresponding to a 110-pixel square image with three RGB components (i.e., VIT image). When predicting battery cycle life, the previous n charging cycles are used as the prediction window, and a four-dimensional 110×110×3×n matrix is obtained.

Step 3.2: Use three-dimensional convolution to extract features from the four-dimensional matrix. The calculation formula of the feature map is:

In the formula, W _out ×H _out ×C _out and W _in ×H _in ×C _in represent the width, height and number of channels at image output and input respectively, w × h × c represents the size of the convolution kernel, and s represents the step. Amplitude, p represents the filling value, k represents the number of scans, and the number of parameters in the convolution process is (w×h×c+1)×k;

Step 3.3: Compared with 2DCNN, 3DCNN has one more dimension in calculation, resulting in more parameters in the convolution process, which will bring a greater burden to the calculation of the model. Therefore, depth-separable three-dimensional convolution is used. Reduce the difficulty of model training and greatly reduce the amount of convolution calculations. The specific method is to divide the feature map of each charging cycle into a group, perform convolution operations within the group respectively, and the convolution kernel in the group generates the feature map.

Step 3.4: Since depth convolution has no information between different cycles, a point convolution layer needs to be added after depth convolution. The point-by-point convolution operation is similar to the conventional convolution operation, and its convolution kernel size is 1×1×n, where n represents the depth direction of the three-dimensional matrix, that is, the charging cycle of the first n cycle. Therefore, this convolution calculation will weightedly combine the images of the previous step in the depth direction to generate a new feature map, and use multiple convolution kernels to generate multiple feature maps.

Step 3.5: In order to automatically evaluate the significance and correlation of various features during battery charging to predict battery life, the multi-channel three-dimensional convolutional channel attention (3D Channel Attention, 3DCA) module is introduced to learn these features. The 3DCA module highlights the specific areas related to the cycle life from the feature map based on the correlation between the input data and the battery cycle life, and combines the attention layer to detect the significance of different channel features in the input data.

Step 3.5.1: First, considering the aggregate feature y (y∈R ^C ), during the charging process, it can be expressed by the following formula:

ω=σ(W _k y) (7)

In the formula, W _k is the parameter matrix of k×C, ω is the weight, and σ is the activation function.

Step 3.5.2: For the above formula, the weight ω _i is calculated by only considering the interaction between y _i and its adjacent node k in each period i. Then its weight can be rewritten as:

in the formula is the set k of adjacent channels _yi , and all channels share the same learning parameters. Note that this method can be easily implemented with a 1×1×1 convolution kernel with k kernel size, i.e.

ω=σ(Conv _k (y)) (10)

In the formula, Conv _k represents a three-dimensional convolution with a kernel size of 1.

Step 3.5.3: The degree of interaction between channels can be adjusted by changing the size of k, and the size of kernel k can be determined adaptively

In the formula, c represents the number of channels, |a| _odd represents the nearest odd distance from a. By nonlinear mapping ψ, high-dimensional channels have longer distance interactions, while low-dimensional channels have shorter distance interactions.

Step 3.6: Add a 3D Global Average Pooling (3DGAP) layer after the 3DCA layer. Among them, 3DGAP is the average pooling in 3D. The pooling operation can retain the main features while reducing the parameters and calculation amount to prevent over-fitting. This method can greatly reduce the amount of data and computational complexity.

Step 4: For the modeling of the discharge part data, build the CBR layer and MP layer to perform convolution and pooling operations on the data. The CBR structure consists of a convolutional layer, a batch normalization layer and a ReLU activation layer (Rectified Linear Unit layer) to ensure that the network has a certain regularization effect. The MP structure mainly consists of a max-pooling layer and several convolutional layers with different strides. The max-pooling layer and the convolutional layer with a stride of 2 reduce the number of channels by half. , and keep the number of input and output channels unchanged after splicing.

Step 4.1: The CBR structure consists of a convolution layer, a batch normalization layer and a ReLU activation layer. Among them, the addition of the batch normalization layer can well change the disorder of the original data, speed up the convergence speed of the network, and have a certain regularization effect, while the ReLU layer can convert linear transformation into nonlinear transformation;

Step 4.2: The MP structure consists of a maximum pooling layer and several convolutional layers with different strides. The max pooling layer and the convolutional layer with a stride of 2 reduce the number of channels by half and keep the number of input and output channels unchanged after splicing.

Step 5: Feature fusion is performed on the feature vectors output by the charging process and discharging process models, and the final output is the lithium battery cycle life prediction result.

Step 5.1: After feature extraction from VIT data, different feature dimensions and sizes are generated, so a feature fusion module is constructed to ensure that the model can obtain comprehensive battery degradation information from different feature types. This module converts features of different dimensions into one-dimensional feature vectors through a fully connected layer and splices them to form a final new feature matrix.

Step 5.2: The prediction module is responsible for predicting the battery cycle life. In this module, the window period of the battery is split into the first n charging cycles n _f and the latest m charging cycles m _l to complete the cycle life prediction of the battery. Finally, an ablation experiment is performed to verify the effectiveness of the time series imaging method. .

Step 5.2.1: Under the same experimental settings, repeat the experiment 3 times on the two data sets, and conduct a total of 6 experiments, named Case_1, Case_2,..., Case_6, but the data set partitioning of each experiment is random. . In order to evaluate the accuracy of estimation, mean square error (MSE), mean square logarithmic error (MSLE) and mean absolute error (MAE) are used as evaluation indicators. The calculation method is as follows:

In the formula, n is the number of batteries, _yi is the real value, is the predicted value.

Step 5.2.2: The selection of different window periods will change the size of the model input data, which may ultimately affect the prediction accuracy of the model. Different first n charge and discharge cycles (n _f =10, n _f =30, n _f =50) were selected for comparative experiments. Since the input data has been normalized before being sent to model training, denormalization calculation needs to be performed first when calculating the prediction results. Table 1 and Table 2 give the experimental results of the two data sets under different window periods.

Table 1 Experimental results on data set 1

Table 2 Experimental results on data set 2

It is obvious that compared with the cases of n _f =10 and n _f =30, the error of the model is smallest when n _f =50, which means that when the window period is larger, the cycle life of the battery can be predicted more accurately.

Step 5.2.3: Use different models to predict the cycle life of different batteries to verify the accuracy and effectiveness of the proposed model in this field. Table 3 shows the comparison of the results of this model with several other mainstream machine learning models and deep learning models under different window periods.

Table 3 Comparison of different battery cycle life prediction methods

It can be seen that most models require more than 100 charge-discharge cycles to achieve equivalent results to DS-3DCA-CNN. This model accurately predicts cycle life with fewer window periods than other models.

Step 5.2.4: In order to verify the effectiveness of the time series imaging method, ablation experiments need to be performed. In addition to RP diagrams, there are other time series imaging methods, such as Gramian Angular Field (GAF) and Markov Transition Field (MTF), where GAF is divided into Gramian Angular Field and Field (Gram Angle SumField, GASF) and Gram Angle difference Field (GADF), these methods will be used as control groups for experimental analysis. For Data Set 1 and Data Set 2, the results of the two experiments are shown in Table 4.

Table 4 Comparison of experimental results of different time series imaging methods

Experimental results show that using time series imaging methods can improve the accuracy of the model compared to original one-dimensional data because more information is generated in the process of increasing the data dimensionality.

The above embodiments are only for illustrating the technical concepts and features of the present invention. Their purpose is to enable those familiar with this technology to understand the content of the present invention and implement it accordingly, and cannot limit the scope of protection of the present invention. All equivalent transformations or modifications made based on the spirit and essence of the present invention shall be included in the protection scope of the present invention.

Claims (7)

1. The lithium battery cycle life prediction method is characterized by comprising the following steps of:

step 1: collecting cycle life data of the lithium battery and dividing the data; for voltage, current and temperature VIT data in the charging process, automatically extracting features by adopting a deep learning method, respectively converting one-dimensional VIT data in the charging process into three two-dimensional matrixes by using a recursion graph RP, and splicing the two-dimensional matrixes in the depth direction to form a three-dimensional matrix;

step 2: analyzing an incremental capacity (Incremental Capacity, IC) curve and a discharge capacity Q-V curve of a discharge process, introducing an earth movement distance to calculate curve change conditions of the battery between different cycles, extracting key discharge characteristics from the curve, and observing the correlation of the key discharge characteristics and the cycle life of the battery;

step 3: for modeling of the charging part data, a 3D depth separable convolution model is built, VIT three-dimensional matrixes under different cycles are spliced into a four-dimensional matrix, the calculated amount of parameters is reduced through depth separation convolution and point-by-point convolution, and a 3D channel attention module is introduced to strengthen data interaction between VIT channels so as to ensure that interaction between part channels is achieved under the condition that model training parameters are not increased;

step 4: modeling the discharge part data, constructing a CBR layer and an MP layer, and performing rolling and pooling operation on the data; the CBR layer structure consists of a convolution layer, a batch normalization layer and a ReLU activation layer; the MP layer structure consists of a maximum pooling layer and a plurality of convolution layers with different steps, wherein the maximum pooling layer and the convolution layers with the step length of 2 halve the number of channels and keep the number of input and output channels unchanged after splicing;

step 5: feature vectors output by the charging process model and the discharging process model are subjected to feature fusion, the feature vectors are finally output as a lithium battery cycle life prediction result, and finally an ablation experiment is performed to verify the effectiveness of the time sequence imaging method.

2. The method for predicting the cycle life of a lithium battery according to claim 1, wherein the step 1 of dividing the two collected lithium battery data sets and extracting the VIT data features of the charging process by using the RP graph includes the following steps:

step 1.1: after the problem battery is removed in the data acquisition process, dividing the two data sets into a training set, a verification set and a test set according to a certain proportion;

step 1.2: dividing two data sets into a plurality of intervals according to the cycle times, and dividing training data, verification data and test data of each interval proportionally;

step 1.3: regarding the voltage, current and temperature in the battery charging process as a function of the capacity ratio, wherein the data consists of one-dimensional data of different periods; the RP image is used as a time sequence imaging method, and charging data among different charging periods is converted into a multi-dimensional image, so that more complete lithium battery degradation information is obtained;

step 1.4: the calculation principle of the RP graph is as follows: for a time sequence u of length n with a sampling time interval Δt _k (k=1, 2,) n), reconstructing the time series after embedding the dimension m and the delay time τ, the reconstruction expression being x _i ＝(u _i ,u _i+τ ,...,u _i+(m-1)τ ) I=1, 2,..n- (m-1) τ, then the i and j data points x in space are reconstructed _i And x _j Euclidean distance s between _ij Represented by formula (1):

s _ij ＝||x _i -x _j ||,i,j＝1,2,...,n-(m-1)τ (1)

the recursion value R (i, j) is then calculated:

R(i,j)＝H(ε _i -s _ij ),i,j＝1,2,...,n (2)

wherein ε _i As a threshold, H (x) represents a Heaviside function, fixed or varying with i;

step 1.5: and converting the one-dimensional time sequence data in the battery charging process into two-dimensional voltage, current and temperature RP-VIT data through an RP image imaging method.

3. The method for predicting the cycle life of a lithium battery according to claim 1, wherein the feature extraction for the battery discharging process in the step 2 comprises the steps of:

step 2.1: IC curve characteristics were obtained according to the incremental capacity (Incremental Capacity, IC) analysis method, and IC was calculated by the following formula (3):

wherein Q is the battery capacity in the current state, V is the voltage, I is the discharge current, t is the sampling time, IC is the stepping ratio voltage dQ/dV, and the complete relation between the increment capacity and the voltage is obtained along with the movement of the voltage step window;

step 2.2: according to the change curve of the discharge voltage between the period i and the period j, the discharge voltage is expressed as delta Q _i-j (V)＝Q _i (V)-Q _j (V) wherein i and j each represent the number of cycles, and the larger the difference in the sum, the curve DeltaQ _i-j The more pronounced (V);

step 2.3: the earth movement distance EMD is introduced to measure the difference between adjacent two periodic Δq (V) curves.

4. A battery cycle life prediction method according to claim 3, wherein the difference between two adjacent periodic Δq (V) curves is measured by the earth movement distance EMD in step 2.3, and the calculation steps are as follows:

step 2.3.1: the definition of EMD distance is an estimate of the minimum cost of transitioning from one profile P to another profile Q;

step 2.3.2: w (P, Q) represents the EMD between two probability distributions, defined as follows:

W(P,Q)＝inf _γ∈Π(P,Q) E _(x,y)～γ [||x-y||] (4)

wherein inf represents the infinitum of all possible joint probability distributions pi (P, Q), gamma is the joint distribution, pi (P, Q) is the set of all joint distributions of P and Q, P and Q are the distributions of adjacent two periodic IC curves, x-y is the distance between samples x and y, E _(x,y)～γ [||x-y||]Is a distance expectation;

step 2.3.3: the calculation formula for the distance expectation is as follows:

wherein i and j represent two different periods, P _(i,j) The distribution of the IC curves representing periods i and j.

5. The method for predicting the cycle life of a lithium battery according to claim 1, wherein the step 3 of modeling the charging part data by constructing a 3D depth separable convolution model includes the steps of:

step 3.1: for a charging period, converting VIT data into RP-VIT data by using an RP imaging method, and splicing a plurality of VIT three-dimensional matrixes under different cycles into a four-dimensional matrix;

step 3.2: the feature extraction is carried out on the obtained four-dimensional matrix by utilizing three-dimensional convolution, and the calculation formula of the feature mapping is as follows:

in which W is _out ×H _out ×C _out And W is _in ×H _in ×C _in The width, the height and the channel number of the image output and the input are respectively represented, w×h×c represents the size of a convolution kernel, s represents steps, p represents a filling value, k represents the scanning times, and the number of parameters in the convolution process is (w×h×c+1) ×k;

step 3.3: the difficulty of model training is reduced by adopting depth separable three-dimensional convolution, the convolution calculation amount is reduced, the characteristic images of each charging period are divided into a group, convolution operation is respectively carried out in the group, and the characteristic images are generated by convolution kernels in the group;

step 3.4: after the depth convolution, adding a point-by-point convolution layer, wherein the point-by-point convolution operation is similar to the conventional convolution operation, and the convolution kernel size is 1 multiplied by n, wherein n represents the depth direction of the three-dimensional matrix, namely the charging period of the first n cycles; the point-by-point convolution calculation carries out weighted combination on the previous step of graph in the depth direction to generate a new characteristic graph, and a plurality of characteristic graphs are generated by using a plurality of convolution kernels;

step 3.5: introducing a 3D channel attention (3D Channel Attention,3DCA) module to learn the characteristics, highlighting a specific area related to the cycle life from a characteristic diagram by the 3DCA module according to the correlation of the input data and the cycle life of the battery, and detecting the saliency of different channel characteristics in the input data by combining an attention layer;

step 3.6: adding a 3D global average pooling (3D Global Average Pooling,3DGAP) layer after the 3DCA layer; where 3DGAP is the average pooling in 3D.

6. The method for predicting the cycle life of a lithium battery according to claim 5, wherein the specific calculation process of the channel attention 3DCA module of the multi-channel three-dimensional convolution in step 3.5 is as follows:

step 3.5.1: first, consider the aggregate feature y, y ε R ^C During charging, it is described by the following formula:

ω＝σ(W _k y) (7)

w in the formula _k A parameter matrix of k multiplied by C, wherein omega is a weight, and sigma is an activation function;

step 3.5.2: consider only y for each period i _i Weight ω calculated by interaction with its neighboring node k _i The weight is rewritten as:

in the middle ofFor adjacent channels y _i All channels share the same learning parameters, implemented by a1 x 1 convolution kernel with a k kernel size, i.e.:

ω＝σ(Conv _k (y)) (10)

conv in _k Representing a three-dimensional convolution with a kernel size of 1;

step 3.5.3: the degree of interaction between channels is adjusted by varying the size of k, which is adaptively determined by:

wherein c represents the number of channels,indicating distance->The nearest odd distances, through the nonlinear mapping ψ, the high dimensional channels have longer distance interactions, while the low dimensional channels have shorter distance interactions.

7. The method for predicting the cycle life of a lithium battery according to claim 1, wherein in the step 5, feature vectors output by a charging process model and a discharging process model are subjected to feature fusion, and finally output as a predicted result of the cycle life of the lithium battery, the specific implementation steps are as follows:

step 5.1: converting the features of different dimensions into one-dimensional feature vectors through the full connection layer, and splicing the feature vectors to form a final new feature matrix;

step 5.2: predicting the cycle life of the battery, and dividing the window period of the battery into n charging periods n _f And the last m charging periods m _l Thus, the cycle life prediction of the battery is completed, and finally, an ablation experiment is performed to verify the effectiveness of the time series imaging method.