Fast battery life prediction method with key interval feature extraction

Abstract

The application relates to a key interval feature extraction fast battery life prediction method, and belongs to the technical field of new energy automobile power batteries. The purpose of the application is to obtain the dependency relationship between feature data by adopting a multi-head attention mechanism, to determine the weight of different time steps through a feedforward neural network, and finally to estimate the key interval feature extraction fast battery life prediction method of the remaining life of a lithium battery through a full connection layer and a prediction layer. The steps of the application are as follows: identification of the knee point and the knee rise of a lithium ion battery aging curve, aging feature extraction of a key feature interval based on a TE-DeTransformer architecture, and lithium ion battery life prediction based on key interval feature extraction. The application adopts a multi-head attention mechanism to obtain the dependency relationship between feature data, determines the weight of different time steps through a feedforward neural network, finally inputs the full connection layer into a battery life RUL prediction model, and the precision of the battery life prediction is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of new energy automobile power batteries. BACKGROUND

[0002] As a key component in the field of new energy electric vehicles and fixed energy storage, lithium ion batteries have the advantages of long cycle life, high energy density, no pollution, etc. With the passage of time, the amount of electricity provided by lithium ion batteries will decrease, which is called capacity degradation. When the battery capacity degradation reaches the failure threshold, that is, eighty percent of the initial capacity, the battery reaches its service life, thereby reducing the practicality and safety of the battery, and a new battery needs to be replaced. Therefore, the life prediction of the battery is of great significance to maintain the safe and reliable operation of the electric vehicle energy storage system, and accurate prediction of the battery life enables users to take optimal strategies to maintain the use state of the battery within sufficient time, thereby ensuring the safety of the electric vehicle battery and reducing the cost of the electric vehicle.

[0003] Generally speaking, since the performance of the battery is affected by many factors and changes over time, the prediction of the battery life often requires a large amount of experimental data and resources. Among these data, the attenuation of the battery capacity is a direct indicator of the battery life, and there are many key feature points in the capacity degradation curve, and the prediction of the battery life through the key feature data between these feature points has always been the focus of researchers. The method of predicting the battery life using key feature data provides new possibilities for the production and optimization of batteries, and is also of great significance to prolong the driving range and use safety of electric vehicles. Therefore, it is very important to extract aging features from the key interval data of the battery for the life prediction of the battery.

[0004] In order to ensure the safety and reliability of the battery, it is of great significance to establish an effective battery management system to monitor the state of the lithium ion battery, and its most basic function includes accurate prediction of the remaining useful life (RUL), which helps to predict maintenance and enables optimal maintenance strategies to be taken within sufficient time, thereby saving resources, reducing costs and ensuring safety. The prediction method of RUL can be divided into model-based method and data-driven method. The online accuracy of the model-based method is poor, and it cannot be updated adaptively, so the main limitation of this method is that its accuracy depends largely on the selection of the underlying battery model and the accuracy of the model parameters. Therefore, with the development of big data technology, data-driven methods have been widely used because they can map inputs to output data without having to choose a physical model. Existing data-driven methods mainly implement battery life prediction based on complete cycle aging data, but the work of extracting complete cycle aging data is very difficult in actual application. SUMMARY

[0005] The purpose of this invention is to use a multi-head attention mechanism to obtain the dependencies between feature data, determine the weights of different time steps through a feedforward neural network, and finally estimate the key interval features of the remaining life of lithium batteries through a fully connected layer and a prediction layer to form a fast battery life prediction method.

[0006] The steps of this invention are:

[0007] S1. Identification of the knee point and knee start of the lithium-ion battery aging curve

[0008] The Bacon-Watts model is used to determine the knee point in the battery capacity degradation curve, which is an unknown transition point in the battery aging curve. There are two straight lines to the left and right of the object. The relationship between the two lines is:

[0009] (1)

[0010] in, This represents the number of cycles remaining until the knee point of the battery capacity curve. It is a normally distributed random variable centered at zero, representing the residual; and These are the slopes of the two intersecting lines, one on the left and one on the right. This refers to the number of battery cycles. hour, is the leftmost intercept; tanh is the activation function; Parameters used to control abrupt changes in slope;

[0011] For parameters and Optimize and put Fix a small value to obtain the point of change around the mutation. This point in the cycle of change It is then defined as the knee point;

[0012] The double Bacon-Watts model was used to identify the knee origin. The model identifies two transformations in the data, namely...

[0013] (2)

[0014] in, and These are the points of change surrounding the mutation, and the points of change cycle. Defined as knee-up;

[0015] Battery aging data is categorized based on the number of cycles between the knee point and the knee rise.

[0016] (3)

[0017] wherein, is the number of cycles between the knee point and the knee start point;

[0018] S2, aging feature extraction based on key feature interval of TE-DeTransformer architecture

[0019] (1) Normalize the input battery aging data:

[0020] The aging data includes the battery state of health between the knee point and the knee start point, the battery constant voltage charging time, the battery constant current charging time, and the battery internal resistance impedance data; let the battery feature input sequence be Normalized as

[0021] (4)

[0022] wherein, is the initial rated capacity of the battery, is the normalized value of

[0023] (2) Denoise the battery aging data:

[0024] Let represent a sequence of feature data sample slices, add Gaussian noise to the normalized aging data input sequence, and then input the aging data containing noise to the autoencoder, and use the decoder to reconstruct the input from the feature space; this process is summarized as

[0025] (5)

[0026] (6)

[0027] (7)

[0028] wherein, and are the weights of the encoder and the decoder, respectively; and are the biases of the encoder and the decoder, respectively; and are activation functions, wherein the encoder selects the identity function and the decoder selects the ReLU function; represents Gaussian noise, is the mean value of the input battery aging data, is the standard deviation of the input battery aging data; is the output of the encoder, i.e., the hidden features;​ is the output of the decoder, i.e., the reconstructed input; is the vector contaminated by noise;

[0029] (3) By minimizing the reconstruction error, unsupervised learning can learn an effective mapping from the low-dimensional code to the high-dimensional representation in order to remove noise and reconstruct more accurate data, i.e.,

[0030] (8)

[0031] where, is the objective function; is the loss function; is the regularization parameter; is the length of the input feature data of the battery; is the Fibonacci criterion;

[0032] The number of weights of the model is reduced by parameter sharing, i.e., let , so as to accelerate the training;

[0033] (4) The transformer decoder extracts the degradation features of the battery from the reconstructed battery aging data, which has two sub-layers: a multi-head attention mechanism layer and a feedforward neural network layer;

[0034] (5) The relative position of each time step of the battery aging feature data is added to the input, which is specifically defined as

[0035] (9)

[0036] (10)

[0037] where, is the relative position of the aging feature in the sequence; is the position step; is the dimension of the encoder output;

[0038] (6) The cycle period information corresponding to the battery aging data feature is also introduced into the input data as a feature by using the time encoding layer. The definition of time encoding is similar to the position encoding PE, i.e.,

[0039] (11)

[0040] (12)

[0041] where, is the current sequence time; is the time step; is the dimension of the encoder output;

[0042] (7) Attention mechanism includes query , key and value vectors, , , The formula is

[0043] (13)

[0044] where, , and are the weights of projection respectively; is the parallel attention function of the layer;

[0045] (8) Scaling dot product attention reduces the problem of too large gradient change caused by random initialization weight by dividing the dot product result by a scaling factor, so as to accelerate the convergence and optimization of the model, which is expressed as

[0046] (14)

[0047] where, , is the function for calculating attention weight;

[0048] Connect scaling dot product attention SDA together and multiply by weight matrix to generate the final attention of multi-head attention mechanism MHA layer, the process is expressed as

[0049] (15)

[0050] (16)

[0051] where, is the trainable weight; represents the self-attention result of the head; the final attention is connected by function;

[0052] (9) From multi-head attention mechanism get , that is

[0053] (17)

[0054] (18)

[0055] where, and are the weight matrices input to the hidden layer and between the hidden layer and the output layer, corresponding to the linear mapping input to the hidden layer and from the hidden layer to the output layer, respectively; and are the bias terms of the hidden layer and the output layer, respectively;

[0056] S3, lithium ion battery life prediction based on key interval feature extraction

[0057] The prediction result of the battery capacity Given by the fully connected layer input by the FFN, that is,

[0058] (19)

[0059] wherein, is the weight; is the bias; is the hidden aging feature input extracted in the last step; is the mapping function of the prediction layer.

[0060] The beneficial effects of the present application are:

[0061] 1. Considering that the battery aging period is long, and the cost of obtaining the full life cycle data is large, the Bacon-Watts method and double Bacon-Watts method are used respectively to identify two key feature points of the knee point and the knee rise in the battery full cycle capacity attenuation curve, and the key feature interval of the battery aging curve is determined based on the two key feature points, so that data processing is carried out for different aging areas;

[0062] 2. Considering that the attenuation trajectory of the battery capacity attenuation curve before the knee point is approximately linear, which leads to that the aging features extracted in this interval are not obvious, thereby affecting the accuracy and rapidity of the battery life prediction. The present application extracts the data of the battery state of health, the battery constant voltage charging time, the battery constant current charging time, the battery internal resistance impedance, etc. between the knee point and the knee rise point, and adopts a feature extraction method based on unsupervised learning to denoise the battery aging feature data, automatically extracts the associated hidden features among these aging feature data, and then reconstructs the low-dimensional hidden feature data reflecting the battery aging, so as to realize the prediction of the battery life and guarantee the safe use of the battery;

[0063] 3. Considering that traditional neural networks cannot capture time information from input sequences well, traditional neural networks regard time series data as a set of independent data, ignoring the correlation between extracted aging features and time series, resulting in inability to obtain the evolution trend and periodic changes of data, and making the accuracy of battery life prediction low. The present application adopts the TE-DeTransformer method, which is a fusion network architecture that simultaneously introduces the time information and relative position of data into the Transformer algorithm, respectively encodes the cycle period information of the actual use of the battery corresponding to the data extracted in the key feature interval and the relative position information of each time step in the data. The multi-head attention mechanism is used to obtain the dependency relationship between the feature data, and the weights of different time steps are determined through the feedforward neural network, and finally input into the battery life RUL prediction model through the full connection layer, which improves the accuracy of battery life prediction. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 is to improve the recognition of the knee point and the knee starting point of the battery capacity aging curve in the Bacon-Watts model;

[0065] Figure 2 is the overall block diagram of the fast battery life prediction based on key interval feature extraction;

[0066] Figure 3 is the life prediction result of the battery CS2_35 in the CALCE data set;

[0067] Figure 4 is a comparison diagram of the prediction of the method proposed by the present application and the evaluation index of other methods. DETAILED DESCRIPTION

[0068] It has become a new possibility to predict the battery life using more obvious aging features in a small amount of aging data. According to the decay rate of different regions of the battery aging curve, lithium-ion batteries can exhibit linear, sub-linear or super-linear aging trajectories, and the "knee point" and "knee starting point" can be used as the dividing points of the battery aging trajectory. Therefore, by using the "knee point" and "knee starting point", the future capacity loss can be reliably predicted using the battery aging data in a small number of cycles. In addition, considering that some batteries in the battery data set have abnormal aging trajectories, and the cycle period of the key interval corresponding to these batteries is relatively short, the battery aging feature data in the key interval is obviously insufficient, and the traditional random selection of training set and test set will lead to sample bias to abnormal samples, which will lead to inaccurate prediction results.

[0069] In addition, two limitations that need to be addressed in the future are the prediction of battery life using less data and the impact of various machine learning and deep learning algorithms on prediction accuracy and speed. Therefore, the fusion algorithm will prove to be an effective and reliable prediction method, and how to improve the prediction accuracy and generalization of the fusion algorithm to meet the standards of practical application is a problem that needs to be solved.

[0070] The present application provides a fast battery life prediction based on key interval feature extraction. First, due to the long aging period of the battery and the high cost of obtaining full life cycle data, the present application classifies the battery data in the battery data set that does not meet the minimum aging period required for data extraction, i.e. the aging period between two points is less than 100 cycles, and the battery data used for testing and training, i.e. the aging period between two points is more than 100 cycles, by identifying the "knee point" and "knee start" points in the aging curve. Then, in order to avoid the influence of the battery aging characteristics before the knee start point on the accuracy and speed of the prediction model, the present application only uses the data between the "knee point" and "knee start" two key feature points in the battery aging curve, uses the TE-DeTransformer algorithm which is the fusion of Temporal Transformer Network (TTN) and unsupervised learning Denoising Autoencoder (DEA) to analyze and extract the battery aging characteristics in the key interval, reconstructs the low-dimensional hidden feature data of battery aging, uses the multi-head attention mechanism to obtain the dependency relationship between the feature data, and determines the weight of different time steps through the feedforward neural network, and finally estimates the remaining life of the lithium battery through the full connection layer and the prediction layer. The TE-DeTransformer method can be calculated in parallel, so that its accuracy and computational complexity for battery life prediction are superior to other neural network methods, and the method has high robustness and generalization.

[0071] The following will describe the present application in detail in conjunction with the drawings:

[0072] First, the identification of the "knee point" and "knee start" of the lithium ion battery aging curve

[0073] First, the identification of the "knee point" in the battery capacity degradation curve. For the "knee point", the concept followed by the present application is that the "knee point" is the intersection of two lines with different slopes, and the two slopes are the characteristics of the two stages of battery capacity degradation.

[0074] The present application uses the Bacon-Watts model to determine the "knee point" in the battery capacity degradation curve, which does not depend on the gradient and is robust to noise. The Bacon-Watts model refers to an unknown transition point in the battery aging curve Two straight lines are on the left and right, and the relationship between the two straight lines can be described as

[0075] (1)

[0076] wherein, is the number of remaining cycles from the knee point of the battery capacity curve; is a normal distribution random variable centered at zero, representing the residual error; and are the slopes of the left and right intersecting lines, respectively; is the cycle number of the battery, when the cycle number , is the intercept of the leftmost end; tanh is the activation function; is a parameter for controlling the slope mutation.

[0077] The parameters and are optimized, and is fixed to a small value to obtain the change point around the mutation , and the change cycle point is defined as the "knee point".

[0078] The "knee point" does not remind the user that the battery is transitioning to a period of rapid capacity decay, but only that the battery capacity is undergoing accelerated degradation. From the user's point of view, it is important to identify the start of the accelerated degradation of the battery capacity. Therefore, the present application newly defines a "knee rise" point, which is the starting point of the increase in the degradation rate of the battery capacity.

[0079] The present application uses the double Bacon-Watts model to identify the "knee rise" point, which is different from the Bacon-Watts model in that the model can identify two transitions in the data, not one, i.e.

[0080] (2)

[0081] wherein the parameters of equation (2) are similar to those of equation (1), and are the change points around the mutation, and the change cycle point is defined as the "knee rise".

[0082] By calculating the number of cycles between the "knee point" and the "knee rise", the battery aging data is classified, i.e.

[0083] (3)

[0084] wherein, is the number of cycles between the "knee point" and the "knee rise" point.

[0085] In order not to affect the prediction accuracy, only the period data satisfying the condition that the number of battery aging cycles between the "knee point" and the "knee rise" is greater than 100 cycles are subjected to feature extraction and battery life prediction. The identification of the "knee point" and the "knee rise" is shown in Figure 1

[0086] Second, the aging feature extraction based on the key feature interval of TE-DeTransformer architecture

[0087] The fusion network algorithm divides the key interval feature extraction process of the battery capacity curve into three parts.

[0088] Firstly, the input battery aging data is normalized. The aging data includes the battery state of health between the "knee point" and the "knee rise", the battery constant voltage charging time, the battery constant current charging time, the battery internal resistance impedance and other data. The uneven distribution of battery aging data may affect the extraction of key features by TE-DeTransformer, so the input charge and discharge cycle data needs to be normalized. Let the battery feature input sequence be , which is normalized to

[0089] (4)

[0090] wherein, is the initial rated capacity of the battery, is the normalized value of .

[0091] Then, the battery aging data is denoised. The input is the capacity sequence. Due to the capacity regeneration phenomenon of the battery in the actual charge and discharge process, the original battery aging data usually contains noise, which will affect the accuracy of the battery RUL prediction. Therefore, the battery aging data needs to be denoised. The present application uses an unsupervised learning method to denoise the battery aging data and extract low-dimensional features from the original battery aging data, thereby minimizing the reconstruction error.

[0092] Let represent characteristic data sample slice sequence, the normalized aging data input sequence is added with Gaussian noise, and then the aging data containing noise is input to the autoencoder, and the decoder is used to reconstruct the input from the feature space.

[0093] This process can be summarized as

[0094] (5)

[0095] (6)​

[0096] (7)

[0097] where, and are the weights of the encoder and decoder, respectively; and are the biases of the encoder and decoder, respectively; and are the activation functions, where the encoder chooses the identity function and the decoder chooses the ReLU function; represents the Gaussian noise, is the mean of the input battery degradation data, is the standard deviation of the input battery degradation data; is the output of the encoder, i.e., the hidden features; is the output of the decoder, i.e., the reconstructed input; is the vector contaminated by noise.

[0098] The goal of unsupervised learning is to minimize the reconstruction error, which is the Mean Squared Error (MSE) between the input data and its reconstruction. By minimizing the reconstruction error, unsupervised learning can learn an effective mapping from the low-dimensional encoding to the high-dimensional representation in order to remove the noise and reconstruct more accurate data. That is,

[0099] (8)

[0100] where, is the objective function; is the loss function; is the regularization parameter; is the length of the input feature data of the battery; is the Fibonacci criterion.

[0101] Since the unsupervised learning model has a symmetric structure, the number of weights of the model is reduced by parameter sharing, i.e., letting , thereby accelerating the training.

[0102] After extracting the reconstructed battery degradation data through unsupervised learning, the decoder of the TE-DeTransformer extracts the degradation features of the battery. The Transformer decoder extracts the degradation features of the battery from the reconstructed battery degradation data, which contains two layers: the Multi-Head Self-Attention (MHA) layer and the Feed-Forward Network (FFN) layer.

[0103] To make full use of the position information in the reconstructed input sequence, the present application proposes a position encoding (PE) to add the relative position of each time step of the battery aging feature data to the input, which can be defined as

[0104] (9)

[0105] (10)

[0106] wherein, is the relative position of the aging feature in the sequence; is the position step; is the dimension of the encoder output.

[0107] Unlike the traditional Transformer, the present application uses a time encoding layer to also introduce the cycle information corresponding to the battery aging data features into the input data as features. The definition of the temporal encoding (TE) is similar to the position encoding PE, that is,

[0108] (11)

[0109] (12)

[0110] wherein, is the current sequence time; is the time step; is the dimension of the encoder output.

[0111] The multi-head attention mechanism layer adopts a multi-head attention mechanism to capture the dependency between the battery aging features, which has the advantage of ignoring the distance of these aging features in the input sequence. The attention mechanism includes query , key value and value vectors. , , The calculation formula is

[0112] (13)

[0113] wherein, , and are the weights of the projection; is the parallel attention function of the layer.

[0114] Scaling Dot-Product Attention (SDA) accelerates model convergence and optimization by dividing the dot product result by a scaling factor to reduce excessive gradient changes that may result from randomly initialized weights. This can be expressed as:

[0115] (14)

[0116] in, , The function is used to calculate attention weights.

[0117] Will The scaled dot product attention layers (SDAs) are concatenated and multiplied by a weight matrix to generate the final attention layer of the multi-head attention mechanism (MHA). This process can be represented as follows:

[0118] (15)

[0119] (16)

[0120] in, These are trainable weights; Indicates the first The self-attention results of each head; final attention through Function linking.

[0121] Following the multi-head attention mechanism (MHA) layer, a feedforward neural network (FFN) layer performs multi-layer mapping on the input battery aging feature vector to learn the complex nonlinear relationships in the input sequence and generate new codes. These codes contain more abstract features, including a nonlinear transformation and a linear transformation, and can be derived from the previous multi-head attention mechanism. get ,Right now

[0122] (17)

[0123] (18)

[0124] in, and It is the weight matrix between the input to the hidden layer and the hidden layer to the output layer, which correspond to the linear mapping from the input to the hidden layer and the linear mapping from the hidden layer to the output layer, respectively. and These are the deviation terms for the hidden layer and the output layer, respectively.

[0125] By the above steps, the hidden features between the battery aging features can be extracted to predict the remaining life of the battery. The overall framework of the fast battery life prediction method based on key interval feature extraction proposed in the present application is shown in Figure 2

[0126] Third step, lithium ion battery life prediction based on key interval feature extraction

[0127] In order to realize the prediction of the battery life, the hidden features extracted in the above step are taken as the prediction input, and the prediction result of the battery capacity is given by the full connection layer input by the FFN, that is,

[0128] (19)

[0129] wherein, is the weight; is the bias; is the hidden aging feature input extracted in the previous step; is the mapping function of the prediction layer.

[0130] In order to verify the effectiveness of the method proposed in the present application, the root mean square error (RMSE), the mean absolute error (MAE) and the relative error (RE) of the remaining charge and discharge times are used to verify the prediction accuracy of the method proposed in the present application, that is,

[0131] (20)

[0132] (21)

[0133] (22)

[0134] wherein, denotes the length of the sequence; denotes the length of the sample generated from the sequence for training; and denote the battery life prediction RUL value and the real battery life RUL value at the start time, respectively.

[0135] The battery life prediction result and the evaluation index are compared as shown in Figure 3 and Figure 4

[0136] Verification case

[0137] ​​​To further verify the effectiveness of the present application, a specific example is used to verify the method proposed by the present application. The present application verifies the method proposed by the present application based on a set of published battery data of CS2 series batteries of the Center for Advanced Life Cycle Engineering of the University of Maryland. The present application selects four batteries with a rated capacity of 1.1A / h (denoted as CS35, CS36, CS37 and CS38). The positive electrode material is lithium cobalt oxide (LiCoO2). The experiment is also carried out under the standard CC-CV protocol. The battery is charged at a constant current rate of 0.5C until the voltage reaches 4.2V. Then the battery is charged at a constant voltage, and when the current drops to 20mA, the charging stage stops. The battery is discharged at a constant current rate of 1C until the voltage drops to 2.7V.

[0138] Firstly, since the aging period of the battery is long, and the cost of obtaining the full life cycle data is large, the present application identifies the "knee point" and "knee point" in the battery capacity attenuation curve, and determines the "knee point" and "knee point" by using Bacon-Watts method and double Bacon-Watts method respectively, and the results are shown in Figure 1 and Figure 2 The battery data with a cycle period of less than 100 cycles between the two points and the battery data with a cycle period of more than 100 cycles between the two points are extracted from the data set and used to test the trained data.

[0139] Then, considering that the attenuation trajectory of the battery capacity attenuation curve before the knee point is approximately linear, the aging characteristics extracted in this interval are not obvious. The present application uses TE-DeTransformer algorithm to extract the key interval characteristics of the battery aging curve, so as to time-encode the actual use cycle period information of the data extracted in the key feature interval, and position-encode the position information of each time step, further improving the efficiency and accuracy of the life prediction model, and the structure of TE-DeTransformer algorithm is shown in Figure 2 .

[0140] The multi-head attention mechanism is used to obtain the dependency relationship between the feature data, and the weights of different time steps are determined by the feedforward neural network, which is input into the battery life RUL prediction model through the full connection layer. The prediction layer predicts the remaining life of the lithium battery by extracting the hidden features of the key interval in the battery capacity attenuation in the previous step, and the prediction result is shown in Figure 3 .

[0141] In order to further illustrate the effectiveness of the method proposed by the present application, the method proposed by the present application is compared with the Multilayer Perceptron (MLP) method, and the comparison chart of RMSE, MAE and RE is shown in Figure 4The method has the RMSE value of 0.1020, the MAE value of 0.0804 and the RE value of 0.0933, and the RMSE value of the prediction of the MLP method is 0.1253, the MAE value is 0.0943 and the RE value is 0.2003. Compared with the MLP algorithm, the prediction of the residual life of the algorithm has smaller various error indexes, so the accuracy of the battery life prediction of the application is obviously improved compared with the MLP method.

Claims

1. A fast battery lifetime prediction method for key interval feature extraction, characterized in that: The steps are as follows: S1. Identification of the knee point and knee rise in the aging curve of lithium-ion batteries The Bacon-Watts model is used to determine the knee point in the battery capacity decay curve. There are two straight lines to the left and right of an unknown transition point x1 in the battery aging curve. The relationship between the two straight lines is Y=α0+α1(x-x1)+α2(x-x1)tanh{(x-x1) / γ}+Z (1) Where Y is the remaining number of cycles from the knee of the battery capacity curve; Z is a zero-centered normally distributed random variable representing the residual; α1 and α2 are the slopes of the two intersecting lines, respectively; x is the number of battery cycles, and when the cycle period x = x1, α0 is the leftmost intercept; tanh is the activation function; and γ is the parameter that controls the abrupt change in slope. For parameter α i (i = 0, 1, 2) and x1 are optimized, and a small value of γ is fixed to obtain the change point x1 around the mutation. This change period point x1 is defined as the knee point. The double Bacon-Watts model was used to identify the knee origin. The model identifies two transformations in the data, namely... Y=α0+α1(x-x0)+α2(x-x0)tanh{(x-x0) / γ}+α3(x-x2)tanh{(x-x2) / γ}+Z (2) Here, x2 and x0 are the change points around the mutation, and the change period point x0 is defined as the knee. Battery aging data is categorized based on the number of cycles between the knee point and the knee rise. Δx=x1-x0 (3) Where Δx is the number of cycles between the knee point and the knee origin; S2. Aging Feature Extraction Based on Key Feature Regions of TE-DeTransformer Architecture (1) Normalize the input battery aging data: Aging data includes battery health status between the knee point and the knee start point, battery constant voltage charging time, battery constant current charging time, and battery internal resistance data; assuming the battery characteristic input sequence is... x = {x1, x2, ..., x} n }, normalized to Where C0 is the initial rated capacity of the battery, x N Let x be the normalized value; (2) Noise reduction of battery aging data: make This represents a sequence of m feature data sample slices. Gaussian noise is added to the normalized aging data input sequence, and then the noisy aging data is processed. The input is fed into an autoencoder, and a decoder reconstructs the input from the feature space; this process can be summarized as follows: Where W1 and W2 are the encoder and decoder weights, respectively; b1 and b2 are the encoder and decoder biases, respectively; φ(·) and , where the encoder chooses the identity function and the decoder chooses the ReLU function; N represents Gaussian noise, μ is the average value of the input battery aging data, σ is the standard deviation of the input battery aging data; z is the encoder output, i.e., the hidden feature; This is the output of the decoder, i.e., the reconstructed input; for The vector contaminated by noise; (3) By minimizing the reconstruction error, unsupervised learning can learn an efficient mapping from low-dimensional encoding to high-dimensional representation, in order to remove noise and reconstruct more accurate data. in, λ is the objective function; l is the loss function; λ is the regularization parameter; n is the length of the battery's input feature data. This is the Fibonacci criterion; By sharing parameters, the number of weights in the model can be reduced, i.e., W1 = W2, thereby accelerating training. (4) The Transformer decoder extracts the degradation features of the battery from the reconstructed battery aging data. It has two sub-layers: a multi-head attention mechanism layer and a feedforward neural network layer. (5) Add the relative position markers of each time step of the battery aging feature data to the input, specifically defined as follows: Where t is the relative position of the aging feature in the sequence; k is the position step size; d z The dimension of the encoder output; (6) The time-coding layer is used to incorporate the cycle information corresponding to the battery aging data features into the input data as a feature. The definition of time coding is similar to that of position coding (PE), i.e. Where i is the current sequence time; p is the time step; d z The dimension of the encoder output; (7) The attention mechanism includes a query Q, a key K, and a value V vector. The formulas for Q, K, and V are as follows: in, and These are the projection weights; H l-1 This is the parallel attention function for the (l-1)th layer; (8) Scaled Dot Product Attention reduces the potential for excessive gradient changes caused by random initialization of weights by dividing the dot product result by a scaling factor, thereby accelerating model convergence and optimization. This is expressed as... Where, d k =d z / m, the softmax function is used to calculate attention weights; The m scaled dot product attentions (SDAs) are concatenated and multiplied by a weight matrix to generate the final attention layer of the multi-head attention mechanism (MHA). This process is represented as follows: MultiHead(H l-1 )=Concat(head1,head2,…,head m )·W O (15) Among them, W O These are trainable weights; head j This represents the self-attention result of the j-th head; the final attention is concatenated using the Concat function. (9) From the multi-head attention mechanism H l-1 Obtain H l ,Right now H l =FFN(MultiHead(H l-1 )) (17) FFN(x)=ReL U(MHA*W3+b3)W4+b4 (18) Where W3 and W4 are the weight matrices between the input to the hidden layer and the hidden layer to the output layer, respectively, corresponding to the linear mapping from the input to the hidden layer and the linear mapping from the hidden layer to the output layer; b3 and b4 are the bias terms of the hidden layer and the output layer, respectively. S3. Lithium-ion battery lifetime prediction based on key interval feature extraction Battery capacity prediction results The fully connected layer input from the FFN is given, i.e. Among them, W predict For weight; b predict For deviation; H h is the hidden aging feature input extracted in the previous step; g(·) is the mapping function of the prediction layer.

Images