Battery health state prediction method based on improved autotransformer model

Abstract

The invention discloses a battery health state prediction method based on an improved autotransformer model, and belongs to the technical field of lithium ion battery life prediction. The method comprises the following steps: firstly, carrying out standardization processing on original data, filtering battery state of health (SOH) data by utilizing a wavelet denoising technology, and overcoming the defect of poor anti-noise capability of a traditional autotransformer; and then, extracting SOH related characteristics from time, voltage, current, temperature, charge state and historical SOH data by using the de-noised data through a one-dimensional convolution method of univariate and multivariate prediction in the autotransformer, and taking the extracted SOH related characteristics as input for improving the autotransformer. Secondly, establishing an autocorrelation model and a feedforward model, and respectively learning seasonal and trend behaviors of the battery data; and finally, combining the outputs of the two correction models to predict the SOH of the battery. According to the method, the data trend characteristics and the seasonal characteristics of the battery data are analyzed, so that the accuracy of predicting the health state of the battery is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium-ion battery life prediction, and in particular to a battery health state prediction method based on an improved autotransformer model. Background Art

[0002] An important indicator of battery life is the battery health status. Accurate prediction of battery health status provides important information for the maintenance and replacement of lithium-ion batteries, and is an important prerequisite for ensuring the safety of electric vehicles. Traditional prediction methods can be roughly divided into model-based methods, direct measurement methods, and machine learning-based methods. Model-based methods have clear physical meanings. Compared with model-based methods, direct measurement methods usually have lower computational complexity; machine learning-based methods do not need to analyze the specific physical meaning of the degradation mechanism, but only need to extract health features from the charge and discharge data that characterize battery degradation to effectively predict the battery health status.

[0003] However, the parameter calculation of model-based methods is extremely complex and time-consuming, making them difficult to apply in actual vehicles. At the same time, these methods are usually only applicable to specific working scenarios and are challenging in industrial and commercial applications. The accuracy of direct measurement methods requires a harsh test environment and hardware equipment to ensure, which is difficult to achieve in battery application systems. The prediction ratio of the original transformer model to the time series based on machine learning is kept at about 1:1. However, if the battery health status prediction is only a 1:1 prediction, in the case of a large data series, the prediction is only a small data set and lacks useful prediction information. In order to solve this problem, the autotransformer model was proposed, but the autotransformer has poor noise resistance to data and emphasizes the seasonal part of the learning data, ignoring the trend part of the data. In addition, battery data is full of noise and seasonal and trend data, so the autotransformer cannot be used directly for prediction. At present, there are methods such as correlation vector machine, Gaussian process regression, and feedforward neural network to predict the battery health status. Summary of the invention

[0004] The present invention provides a lithium battery health status prediction method to solve the problem mentioned in the background technology that the battery data has the characteristics of noise, seasonality and trend data and cannot be directly predicted using autotransformer, thereby improving the applicability of the autotransformer model.

[0005] To achieve the above technical objectives, the details are as follows:

[0006] A method for predicting a battery health state based on an improved autotransformer model comprises the following steps:

[0007] S1. Convert the collected raw battery data into battery health status data. The expression is as follows:

[0008]

[0009] In the formula, C n is the battery capacity during the charging period; t 1 and t 2 are the start and end time of charging respectively; Δt is the fixed sampling interval; I represents the negative current of the battery during charging; SOC t1 SOC is the initial state of charge for charging. t2 The charging state is the end of charging; C i is the current battery capacity; C 0 is the rated capacity of the battery.

[0010] S2. Standardize the battery health status data and perform wavelet denoising on the standardized data. The steps are as follows:

[0011] The present invention uses the standardscaler tool to standardize the preprocessed battery health status data, and then uses the standardized data as input for wavelet denoising, the steps are as follows:

[0012] S2.1. Standardize the battery health status data. The expression is as follows:

[0013]

[0014] Where z is the current result of the standardized battery health status data; N is the sequence length of the battery health status data; x is the original data of the battery; x i2 is the i2th data point in the battery health status data sequence;

[0015] S2.2, preprocessing the standardized battery health status data;

[0016] The preprocessing method is: clean the data and exclude abnormal data, that is, remove all data with battery health status (SOH) values ​​greater than 1 and data values ​​with data fluctuations greater than 0.5%;

[0017] S2.3, normalizing the preprocessed battery health status data;

[0018] The normalization operation is to reduce the data to the interval (0, 1), and the expression is as follows:

[0019]

[0020] Where Z' is the original data value; Z min is the minimum value in the data; Z max is the maximum value in the data;

[0021] S2.4. Perform wavelet denoising on the normalized battery health status data. The steps are as follows:

[0022] S2.4.1. Perform wavelet transform on the normalized battery health status data to obtain wavelet coefficients;

[0023] The principle of wavelet transform is to transform the basis function in Fourier transform into wavelet function for processing, that is, to transform an infinitely long sine function into an attenuated finite-length wavelet function; in wavelet transform, the magnitude of stretching corresponds to the frequency component of the signal, the stretching corresponds to the low-frequency component, and the compression corresponds to the high-frequency component; after multiplying the translation by the signal of each other scale, it is possible to obtain which frequency components are contained in which position of the signal; the role of this step is to perform wavelet transform on the battery health status data to obtain a new array with wavelet coefficients;

[0024] The expression of wavelet transform is as follows:

[0025]

[0026] In the formula, w j,k is the wavelet coefficient; f(t) is the array for wavelet transform, that is, z new ; is the wavelet basis function, where j is the frequency and tm is the time;

[0027] S2.4.2, performing multi-level wavelet decomposition operation on the wavelet coefficients obtained after wavelet transformation of the battery health status data;

[0028] The present invention performs 5-level wavelet decomposition;

[0029] Decompose the battery health status data into a low-frequency signal and a high-frequency signal by wavelet decomposition, perform five wavelet decomposition operations on the battery health status data in sequence, and after the low-frequency signal is obtained by the first decomposition, perform wavelet decomposition operations on the low-frequency signal obtained by the previous decomposition in sequence;

[0030] Signal data is the superposition of signals with different frequency components. The frequency components can be divided into low-frequency signals (LF) and high-frequency signals (HF). The low frequency of the signal represents the outline of the signal, and the high frequency represents the details of the signal change. Therefore, in practical applications, low frequency is more important than high frequency. High-pass filters and low-pass filters are used to decompose the data into high-frequency signals and low-frequency signals. Then the obtained low-frequency signal (LF) is decomposed again in multiple layers, and each layer of the decomposed LF is decomposed. LF and LF HF Perform wavelet transform and extract wavelet coefficients of each layer;

[0031] S2.4.3, performing a wavelet synthesis operation on the obtained wavelet coefficients after denoising to obtain denoised data;

[0032] The wavelet coefficient of noise is smaller than that of signal. By selecting a threshold, the wavelet coefficient greater than the threshold is considered useful data, and the wavelet coefficient less than the threshold is considered noise data.

[0033] The calculation formula for threshold selection is:

[0034]

[0035] Where cD1 is the first layer wavelet coefficient, N 1 is the signal data length;

[0036] The denoising method is: soft threshold method;

[0037] The wavelet synthesis operation is: the wavelet coefficients obtained each time are reconstructed into a denoised signal using an inverse wavelet transform (IDWT) to obtain denoised data.

[0038] S3, performing one-dimensional convolution on the denoised data to obtain health feature time series data related to the battery health status;

[0039] One-dimensional convolution only convolves the width but not the height. Through one-dimensional convolution operation, the column dimension of the data matrix can be changed to extract the characteristics of each data sample.

[0040] The one-dimensional convolution operation process is: multiply each row of data by a convolution kernel to get a number; if the column data is changed to several dimensions, multiply several convolution kernels, and then traverse each point row by row to obtain a new feature matrix; the calculation process is that the convolution kernel slides on the input data and multiplies and sums the values ​​at the corresponding positions;

[0041] The one-dimensional convolution expression is as follows:

[0042]

[0043] Where n is the position index of the output sequence; c out is the output channel index; b is the bias term; e is the position index of the convolution kernel; w is the convolution kernel weight; c in is the input channel index; d(·) is the value of the input data at position n+e-1 and channel;

[0044] S4, decomposing the health characteristic time series data related to the battery health status to obtain seasonal data and trend data, and inputting the two parts of data into the improved autotransformer model for prediction;

[0045] The improved autotransformer model includes: trend part encoder and seasonal part encoder;

[0046] The trend part encoder and seasonal part encoder include three modules: sequence decomposition module, self-attention mechanism, and feedforward neural network. Among them, sequence decomposition is used to decompose the health feature time series data related to the battery health status into seasonal data and trend data, and can also be used to transmit seasonal data and trend data.

[0047] The steps of prediction are as follows:

[0048] S4.1. Input the health characteristic time series data related to the battery health status into the sequence decomposition module to obtain the trend part and seasonal part of the health characteristic time series data related to the battery health status. The expression is as follows:

[0049]

[0050] Where AvgPool is the moving average; padding(x') means padding to keep the sequence length unchanged; x s The seasonal part of the data; x trend is the trend part of the data; x' is the characteristic time series data;

[0051] S4.2. Input the result of S4.1 into the trend part encoder and the seasonal part encoder for encoding operation respectively. The steps are as follows:

[0052] S4.2.1. Input the trend part and seasonal part of the health feature time series data related to the battery health status into the trend part encoder and seasonal part encoder in the improved autotransformer model respectively, obtain the delayed rolling data of the two parts of data through the same self-attention mechanism operation, and obtain the output of the autocorrelation mechanism through the delayed rolling data;

[0053] The delayed rolling data expression is as follows:

[0054]

[0055] In the formula, R xx (τ) is the autocorrelation function, which represents the time series x z The sequence x lagged by its time delay τ z-τ The time lag similarity between them; τ represents the time delay, which is the time point when the sequence is shifted backward; L represents the length of the time series; x t represents the value of the time series at time point t; x z-τ Represents the value of the time series at time point zt;

[0056] The mechanism for calculating the autocorrelation through delayed rolling data is: select the topk delays with the largest autocorrelation values, perform softmax normalization on the topk delays, and then output the results. The expression is as follows:

[0057]

[0058] Where argTopk represents the selection of k time delays with the largest autocorrelation value; R Q,K (τ) represents the autocorrelation function between the query Q and the key K; k represents the number of selected delays, k = [c × LogL], where c is a hyperparameter and L is the length of the time series; Roll(V i ,τ i ) means rolling the value V on the time axis by τ i time point; Denotes the delay τ i The corresponding autocorrelation weights;

[0059] The present invention uses a multi-head autocorrelation mechanism for output, and the expression is as follows:

[0060]

[0061] Where W output Represents the output weight matrix, which is used to linearly transform the outputs of multiple heads; head i1 i-th 1 The output of each head is calculated through the autocorrelation mechanism; AutoCorrelation represents the autocorrelation mechanism;

[0062] S4.2.2. Add the outputs of the autocorrelation mechanism of the trend part and the seasonal part of the health characteristic time series data related to the battery health status to the corresponding trend part data and seasonal part data to obtain the long-term change trend in the time series and the periodic change in the time series respectively;

[0063] The expression for adding the trend part of the health characteristic time series data related to the battery health status and the corresponding trend part data is as follows:

[0064] Combined(x trend )=x trend +AutoCorrelationgx trend (Q, K, V);

[0065] The expression for adding the seasonal part of the health feature time series data related to the battery health status and the corresponding seasonal part data is as follows:

[0066] Combined(x s )=x s+AutoCorrelationgx s (Q, K, V);

[0067] S4.2.3, input the result of S4.2.2 into two identical feedforward neural networks through the sequence decomposition module, and input the obtained result into the decoder through the sequence decomposition module;

[0068] The feedforward neural network contains two linear transformations and one activation function operation, expressed as follows:

[0069] Y = ReLU(X 1 W 1 +b 1 )W 2 +b 2

[0070] Where W 1 and W 2 are the first weight matrix and the second weight matrix of the two linear transformations; b 1 and b 2 are the first bias vector and the second bias vector; ReLU() is the activation function; X 1 The input is the time series data of health features with attention mechanism;

[0071] S4.3. Input the result of S4.1 and S4.2 into the decoder for decoding operation. The steps are as follows:

[0072] S4.3.1. Input the result of S4.1 into the self-attention mechanism to obtain the output of the self-correlation mechanism;

[0073] S4.3.2, input the result of S4.2 into the self-attention mechanism, and at the same time input the result of S4.3.1 into the self-attention mechanism through the sequence decomposition module to enhance the long-term change trend and periodic changes in the time series;

[0074] Enhanced to: Use the result of S4.3.1 as Q in the AutoCorrelation module, and the result of S4.2 as K, V in the AutoCorrelation module to perform AutoCorrelation operation, that is, the result of S4.3.1 and W Q The result of matrix multiplication is Q, while S4.2 and W K and W V The result of the multiplication is used as K, V for input, W Q , W K and W V are three trainable parameter matrices;

[0075] The expression is as follows:

[0076]

[0077] In the formula, x t2 represents the enhanced output; x t1 represents the result of S4.3.1, including: the trend part and seasonal part of the health characteristic time series data related to the battery health status; Y represents the result of S4.2; x t Represents the AutoCorrelation module of S4.3.2, where Indicates that the result of S4.3.1 is taken as Q, K in the AutoCorrelation module Y ,V Y It means taking the result of S4.2 as K and V in the AutoCorrelation module;

[0078] S4.3.3. The results of S4.3.2 are respectively outputted through a feedforward neural network to decode the trend part and seasonal part of the health characteristic time series data related to the battery health status.

[0079] S5. The decoded outputs of the trend part and the seasonal part are combined in reverse order to obtain the prediction result.

[0080] Beneficial effects of the present invention:

[0081] The present invention performs multi-level wavelet denoising on the original battery data and uses an improved autotransformer model, so that the model can obviously achieve good results in noise removal and division of data trend characteristics and seasonal characteristics, thereby improving the applicability of the autotransformer and the accuracy of lithium-ion battery health status prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 is a flow chart of the steps of the present invention;

[0083] Figure 2 It is the overall process architecture diagram of the method of the present invention;

[0084] Figure 3 It is a conversion result diagram of the health status of the present invention;

[0085] Figure 4 is a comparison diagram of denoising results in an embodiment of the present invention;

[0086] Figure 5 It is a model flow structure diagram of the present invention;

[0087] Figure 6 It is a comparison chart of the prediction results of the present invention, the traditional autotransformer model and the linear model. DETAILED DESCRIPTION

[0088] The present invention is further described in detail below in conjunction with specific embodiments.

[0089] like Figure 1 and Figure 2 As shown, a battery health status prediction method based on an improved autotransformer model includes the following steps:

[0090] S1, such as Figure 3 As shown, the collected raw battery data is converted into battery health status data, and the expression is as follows:

[0091]

[0092] In the formula, C n is the battery capacity during the charging period; t 1 and t 2 are the start and end time of charging respectively; Δt is the fixed sampling interval; I represents the negative current of the battery during charging; SOC t1 SOC is the initial state of charge for charging. t2 The charging state is the end of charging; C i is the current battery capacity; C 0 is the rated capacity of the battery.

[0093] S2. Standardize the battery health status data and perform wavelet denoising on the standardized data. The steps are as follows:

[0094] The present invention uses the standardscaler tool to standardize the preprocessed battery health status data, and then uses the standardized data as input for wavelet denoising, the steps are as follows:

[0095] S2.1. Standardize the battery health status data. The expression is as follows:

[0096]

[0097] Where z is the current result of the standardized battery health status data; N is the sequence length of the battery health status data; x is the original data of the battery; x i2 is the i2th data point in the battery health status data sequence;

[0098] S2.2, preprocessing the standardized battery health status data;

[0099] The preprocessing method is: clean the data and exclude abnormal data, that is, remove all data with battery health status (SOH) values ​​greater than 1 and data values ​​with data fluctuations greater than 0.5%;

[0100] S2.3, normalizing the preprocessed battery health status data;

[0101] The normalization operation is to reduce the data to the interval (0, 1), and the expression is as follows:

[0102]

[0103] Where Z' is the original data value; Z min is the minimum value in the data; Z max is the maximum value in the data;

[0104] S2.4. Perform wavelet denoising on the normalized battery health status data. The steps are as follows:

[0105] S2.4.1. Perform wavelet transform on the normalized battery health status data to obtain wavelet coefficients;

[0106] The principle of wavelet transform is to transform the basis function in Fourier transform into wavelet function for processing, that is, to transform an infinitely long sine function into an attenuated finite-length wavelet function; in wavelet transform, the magnitude of stretching corresponds to the frequency component of the signal, the stretching corresponds to the low-frequency component, and the compression corresponds to the high-frequency component; after multiplying the translation by the signal of each other scale, it is possible to obtain which frequency components are contained in which position of the signal; the role of this step is to perform wavelet transform on the battery health status data to obtain a new array with wavelet coefficients;

[0107] The expression of wavelet transform is as follows:

[0108]

[0109] In the formula, w j,k is the wavelet coefficient; f(t) is the array for wavelet transform, that is, z new ; is the wavelet basis function, where j is the frequency and tm is the time;

[0110] S2.4.2, performing multi-level wavelet decomposition operation on the wavelet coefficients obtained after wavelet transformation of the battery health status data;

[0111] The present invention performs 5-level wavelet decomposition;

[0112] Decompose the battery health status data into a low-frequency signal and a high-frequency signal by wavelet decomposition, perform five wavelet decomposition operations on the battery health status data in sequence, and after the low-frequency signal is obtained by the first decomposition, perform wavelet decomposition operations on the low-frequency signal obtained by the previous decomposition in sequence;

[0113] Signal data is the superposition of signals with different frequency components. The frequency components can be divided into low-frequency signals (LF) and high-frequency signals (HF). The low frequency of the signal represents the outline of the signal, and the high frequency represents the details of the signal change. Therefore, in practical applications, low frequency is more important than high frequency. High-pass filters and low-pass filters are used to decompose the data into high-frequency signals and low-frequency signals. Then the obtained low-frequency signal (LF) is decomposed again in multiple layers, and each layer of the decomposed LF is decomposed. LF and LF HF Perform wavelet transform and extract wavelet coefficients of each layer;

[0114] S2.4.3, performing a wavelet synthesis operation on the obtained wavelet coefficients after denoising to obtain denoised data;

[0115] The wavelet coefficient of noise is smaller than that of signal. By selecting a threshold, the wavelet coefficient greater than the threshold is considered useful data, and the wavelet coefficient less than the threshold is considered noise data.

[0116] The calculation formula for threshold selection is:

[0117]

[0118] Where cD1 is the first layer wavelet coefficient, N 1 is the signal data length;

[0119] The denoising method is: soft threshold method;

[0120] The wavelet synthesis operation is: the wavelet coefficients obtained each time are reconstructed into the denoised signal using the inverse wavelet transform (IDWT) to obtain the denoised data. The comparison between the noisy and noise-free data is shown in the figure below. Figure 4 shown.

[0121] S3, performing one-dimensional convolution on the denoised data to obtain health feature time series data related to the battery health status;

[0122] One-dimensional convolution only convolves the width but not the height. Through one-dimensional convolution operation, the column dimension of the data matrix can be changed to extract the characteristics of each data sample.

[0123] The one-dimensional convolution operation process is: multiply each row of data by a convolution kernel to get a number; if the column data is changed to several dimensions, multiply several convolution kernels, and then traverse each point row by row to obtain a new feature matrix; the calculation process is that the convolution kernel slides on the input data and multiplies and sums the values ​​at the corresponding positions;

[0124] The one-dimensional convolution expression is as follows:

[0125]

[0126] Where n is the position index of the output sequence; c out is the output channel index; b is the bias term; e is the position index of the convolution kernel; w is the convolution kernel weight; c in is the input channel index; d(·) is the value of the input data at position n+e-1 and channel;

[0127] S4, such as Figure 5 As shown, the health characteristic time series data related to the battery health status is decomposed to obtain seasonal data and trend data, and the two parts of data are input into the improved autotransformer model for prediction;

[0128] The improved autotransformer model includes: trend part encoder and seasonal part encoder;

[0129] The trend part encoder and seasonal part encoder include three modules: sequence decomposition module, self-attention mechanism, and feedforward neural network. Among them, sequence decomposition is used to decompose the health feature time series data related to the battery health status into seasonal data and trend data, and can also be used to transmit seasonal data and trend data.

[0130] The steps of prediction are as follows:

[0131] S4.1. Input the health characteristic time series data related to the battery health status into the sequence decomposition module to obtain the trend part and seasonal part of the health characteristic time series data related to the battery health status. The expression is as follows:

[0132]

[0133] Where AvgPool is the moving average; padding(x') means padding to keep the sequence length unchanged; x s The seasonal part of the data; x trend is the trend part of the data; x' is the characteristic time series data;

[0134] S4.2. Input the result of S4.1 into the trend part encoder and the seasonal part encoder for encoding operation respectively. The steps are as follows:

[0135] S4.2.1. Input the trend part and seasonal part of the health feature time series data related to the battery health status into the trend part encoder and seasonal part encoder in the improved autotransformer model respectively, obtain the delayed rolling data of the two parts of data through the same self-attention mechanism operation, and obtain the output of the autocorrelation mechanism through the delayed rolling data;

[0136] The delayed rolling data expression is as follows:

[0137]

[0138] In the formula, R xx (τ) is the autocorrelation function, which represents the time series x z The sequence x lagged by its time delay τ z-τ The time lag similarity between them; τ represents the time delay, which is the time point when the sequence is shifted backward; L represents the length of the time series; x t represents the value of the time series at time point t; x z-τ Represents the value of the time series at time point zt;

[0139] The mechanism for calculating the autocorrelation through delayed rolling data is: select the topk delays with the largest autocorrelation values, perform softmax normalization on the topk delays, and then output the results. The expression is as follows:

[0140]

[0141] Where argTopk represents the selection of k time delays with the largest autocorrelation value; R Q,K (τ) represents the autocorrelation function between the query Q and the key K; k represents the number of selected delays, k = [c × LogL], where c is a hyperparameter and L is the length of the time series; Roll(V i ,τ i ) means rolling the value V on the time axis by τ i time point; Denotes the delay τ i The corresponding autocorrelation weights;

[0142] The present invention uses a multi-head autocorrelation mechanism for output, and the expression is as follows:

[0143]

[0144] Where W output Represents the output weight matrix, which is used to linearly transform the outputs of multiple heads; head i1 i-th 1 The output of each head is calculated through the autocorrelation mechanism; AutoCorrelation represents the autocorrelation mechanism;

[0145] S4.2.2. Add the outputs of the autocorrelation mechanism of the trend part and the seasonal part of the health characteristic time series data related to the battery health status to the corresponding trend part data and seasonal part data to obtain the long-term change trend in the time series and the periodic change in the time series respectively;

[0146] The expression for adding the trend part of the health characteristic time series data related to the battery health status and the corresponding trend part data is as follows:

[0147] Combined(x trend )=x trend +AutoCorrelationgx trend (Q, K, V);

[0148] The expression for adding the seasonal part of the health feature time series data related to the battery health status and the corresponding seasonal part data is as follows:

[0149] Combined(x s )=x s +AutoCorrelationgx s (Q, K, V);

[0150] S4.2.3, input the result of S4.2.2 into two identical feedforward neural networks through the sequence decomposition module, and input the obtained result into the decoder through the sequence decomposition module;

[0151] The feedforward neural network contains two linear transformations and one activation function operation, expressed as follows:

[0152] Y = ReLU(X 1 W 1 +b 1 )W 2 +b 2

[0153] Where W 1 and W 2 are the first weight matrix and the second weight matrix of the two linear transformations; b 1 and b 2 are the first bias vector and the second bias vector; ReLU() is the activation function; X 1 The input is the time series data of health features with attention mechanism;

[0154] S4.3. Input the result of S4.1 and S4.2 into the decoder for decoding operation. The steps are as follows:

[0155] S4.3.1. Input the result of S4.1 into the self-attention mechanism to obtain the output of the self-correlation mechanism;

[0156] S4.3.2, input the result of S4.2 into the self-attention mechanism, and at the same time input the result of S4.3.1 into the self-attention mechanism through the sequence decomposition module to enhance the long-term change trend and periodic changes in the time series;

[0157] Enhanced to: Use the result of S4.3.1 as Q in the AutoCorrelation module, and the result of S4.2 as K, V in the AutoCorrelation module to perform AutoCorrelation operation, that is, the result of S4.3.1 and W Q The result of matrix multiplication is Q, while S4.2 and W K and W V The result of the multiplication is used as K, V for input, W Q , W K and W V are three trainable parameter matrices;

[0158] The expression is as follows:

[0159] ENHANCE t2 =AutoCorrelationgY,x t (Q xt1 ,K Y ,V Y )+x t1 ;

[0160] In the formula, x t2 represents the enhanced output; x t1 represents the result of S4.3.1, including: the trend part and seasonal part of the health characteristic time series data related to the battery health status; Y represents the result of S4.2; x t represents the AutoCorrelation module of S4.3.2, where Q xt1 Indicates that the result of S4.3.1 is taken as Q, K in the AutoCorrelation module Y ,V Y It means taking the result of S4.2 as K and V in the autocorrelation mechanism (AutoCorrelation) module;

[0161] S4.3.3. The results of S4.3.2 are respectively outputted through a feedforward neural network to decode the trend part and the seasonal part of the health characteristic time series data related to the battery health status.

[0162] S5. The decoded outputs of the trend part and the seasonal part are combined in reverse order to obtain the prediction result.

[0163] To verify the solution of the present invention, (MSE), (MAE) and (EMAX) were compared with the autotransformer and linear model. The 160,000 test sets were grouped into 96 data sets. After that, each set of data was put into the trained model to predict the 96 data after the data set. Finally, the 96 predicted data sets of each group were combined to obtain the prediction results and absolute errors of the data test set. Figure 6 As shown. Among them, in all EV data, the prediction results of the present invention can fit the real data curve well. In the error graph, the error of the self-converter is significantly larger than that of the other two models, while the error difference of the other two models is not large, but the error of the present invention is more concentrated in the area below 0.2. In the remaining 3 electric vehicle data, the prediction effect of the present invention is also higher than that of other models.

[0164] In addition, three evaluation criteria are shown in Table 1.

[0165] Table 1: Short time series prediction results

[0166]

[0167] It can be seen that the average MSE, MAE and EMAX of the self-transformation prediction results of the 5 EV data are 0.49%, 2.21% and 2.14% respectively. The average MSE, MAE and EMAX of the linear prediction results of the 5 EV data are 0.30%, 2.04% and 2.15% respectively. For our model, the average MSE, MAE and EMAX are 0.27%, 1.93% and 2.08% respectively. It can be seen that in the present invention, the average values ​​of these three indicators are lower than the other two indicators, and the prediction effect is better.

[0168] From the EMAX point of view, the self-transformer is similar to our model, but has worse linearity than other models. Figure 6 It can be seen that at most time points, the error of the rest of the present invention is lower than that of the autocoupling.

[0169] In summary, the improved present invention has improved prediction effect compared with autocoupling and linear. Experiments have shown that the model has better fitting ability, reduces the number of required training rounds, and has a faster fitting speed. Through this part of the experiment, it is verified that the present invention can obtain better accuracy prediction results in SOH prediction.

Claims

1. A battery health status prediction method based on an improved autotransformer model, characterized in that: The following steps are involved: S1, converting the collected raw battery data into battery health status data; S2, standardizing the battery health status data, and performing wavelet denoising on the standardized data; S3, performing one-dimensional convolution on the denoised data to obtain health feature time series data related to the battery health status; S4, decomposing the health characteristic time series data related to the battery health status to obtain seasonal data and trend data, and inputting the two parts of data into the improved autotransformer model for prediction; S5. The decoded outputs of the trend part and the seasonal part are combined in reverse order to obtain the prediction result.

2. A battery health status prediction method based on an improved autotransformer model according to claim 1, characterized in that: The steps of standardizing the battery health status data and performing wavelet denoising on the standardized data are as follows: S2.

1. Standardize the battery health status data; S2.2, preprocessing the standardized battery health status data; S2.3, normalizing the preprocessed battery health status data; S2.

4. Perform wavelet denoising on the normalized battery health status data. The steps are as follows: S2.4.

1. Perform wavelet transform on the normalized battery health status data to obtain wavelet coefficients; The expression of wavelet transform is as follows: In the formula, w j,k is the wavelet coefficient; f(t) is the array for wavelet transform, that is, z new ; is the wavelet basis function, where j is the frequency and tm is the time; S2.4.2, performing multi-level wavelet decomposition operation on the wavelet coefficients obtained after wavelet transformation of the battery health status data; The present invention performs 5-level wavelet decomposition; Decompose the battery health status data into a low-frequency signal and a high-frequency signal by wavelet decomposition, perform five wavelet decomposition operations on the battery health status data in sequence, and after the low-frequency signal is obtained by the first decomposition, perform wavelet decomposition operations on the low-frequency signal obtained by the previous decomposition in sequence; S2.4.3, performing a wavelet synthesis operation on the obtained wavelet coefficients after denoising to obtain denoised data; The wavelet coefficient of noise is smaller than that of signal. By selecting a threshold, the wavelet coefficient greater than the threshold is considered useful data, and the wavelet coefficient less than the threshold is considered noise data. The calculation formula for threshold selection is: In the formula, cD1 is the first layer wavelet coefficient, N1 is the signal data length; The denoising method is: soft threshold method; The wavelet synthesis operation is: the wavelet coefficients obtained each time are reconstructed into a denoised signal using an inverse wavelet transform to obtain denoised data.

3. The method for predicting battery health status based on the improved autotransformer model according to claim 1, characterized in that: In the health feature time series data related to the battery health status obtained by performing one-dimensional convolution on the denoised data, the one-dimensional convolution is performed only on the width but not on the height. The one-dimensional convolution operation is used to change the column dimension of the data matrix and extract the features of each data sample; The one-dimensional convolution expression is as follows: Where n is the position index of the output sequence; c out is the output channel index; b is the bias term; e is the position index of the convolution kernel; w is the convolution kernel weight; c in is the input channel index; d(·) is the value of the input data at position n+e-1 and channel.

4. The method for predicting battery health status based on the improved autotransformer model according to claim 1, characterized in that: The health characteristic time series data related to the battery health status is decomposed to obtain seasonal data and trend data, and the two parts of data are input into the improved autotransformer model for prediction, and the improved autotransformer model includes: a trend part encoder and a seasonal part encoder; The trend part encoder and the seasonal part encoder include three modules: sequence decomposition module, self-attention mechanism, and feedforward neural network; among them, sequence decomposition is used to decompose the health feature time series data related to the battery health status into seasonal data and trend data, and can also be used to transmit seasonal data and trend data.

5. The method for predicting battery health status based on the improved autotransformer model according to claim 1, characterized in that: The steps of decomposing the health characteristic time series data related to the battery health status to obtain seasonal data and trend data, and inputting the two parts of data into the improved autotransformer model for prediction are as follows: S4.

1. Input the health characteristic time series data related to the battery health status into the sequence decomposition module to obtain the trend part and seasonal part of the health characteristic time series data related to the battery health status. The expression is as follows: Where AvgPool is the moving average; padding(x') means padding to keep the sequence length unchanged; x s The seasonal part of the data; x trend is the trend part of the data; x' is the characteristic time series data; S4.

2. Input the result of S4.1 into the trend part encoder and the seasonal part encoder for encoding operation respectively. The steps are as follows: S4.2.

1. Input the trend part and seasonal part of the health feature time series data related to the battery health status into the trend part encoder and seasonal part encoder in the improved autotransformer model respectively, obtain the delayed rolling data of the two parts of data through the same self-attention mechanism operation, and obtain the output of the autocorrelation mechanism through the delayed rolling data; The delayed rolling data expression is as follows: In the formula, R xx (τ) is the autocorrelation function, which represents the time series x z The sequence x lagged by its time delay τ z-τ The time lag similarity between them; τ represents the time delay, which is the time point when the sequence is shifted backward; L represents the length of the time series; x t represents the value of the time series at time point t; x z-τ Represents the value of the time series at time point zt; The mechanism for calculating the autocorrelation through delayed rolling data is: select the topk delays with the largest autocorrelation values, perform softmax normalization on the topk delays, and then output the results. The expression is as follows: Where argTopk represents the selection of k time delays with the largest autocorrelation value; R Q,K (τ) represents the autocorrelation function between the query Q and the key K; k represents the number of selected delays, k = [c × LogL], where c is a hyperparameter and L is the length of the time series; Roll(V i ,τ i ) means rolling the value V on the time axis by τ i time point; Denotes the delay τ i The corresponding autocorrelation weights; The present invention uses a multi-head autocorrelation mechanism for output, and the expression is as follows: Where W output Represents the output weight matrix, which is used to linearly transform the outputs of multiple heads; head i1 The output of the i1th head is calculated by the autocorrelation mechanism; AutoCorrelation represents the autocorrelation mechanism; S4.2.

2. Add the outputs of the autocorrelation mechanism of the trend part and the seasonal part of the health characteristic time series data related to the battery health status to the corresponding trend part data and seasonal part data to obtain the long-term change trend in the time series and the periodic change in the time series respectively; The expression for adding the trend part of the health characteristic time series data related to the battery health status and the corresponding trend part data is as follows: Combined(x trend )=x trend +AutoCorrelationgx trend (Q,K,V); The expression for adding the seasonal part of the health feature time series data related to the battery health status and the corresponding seasonal part data is as follows: Combined(x s )=x s +AutoCorrelationgx s (Q,K,V); S4.2.3, input the result of S4.2.2 into two identical feedforward neural networks through the sequence decomposition module, and input the obtained result into the decoder through the sequence decomposition module; The feedforward neural network contains two linear transformations and one activation function operation, expressed as follows: Y=Re LU(X1W1+b1)W2+b2 Where W1 and W2 are the first weight matrix and the second weight matrix of the two linear transformations; b1 and b2 are the first bias vector and the second bias vector; ReLU() is the activation function; X1 is the input health feature time series data with attention mechanism; S4.

3. Input the result of S4.1 and S4.2 into the decoder for decoding operation. The steps are as follows: S4.3.

1. Input the result of S4.1 into the self-attention mechanism to obtain the output of the self-correlation mechanism; S4.3.2, input the result of S4.2 into the self-attention mechanism, and at the same time input the result of S4.3.1 into the self-attention mechanism through the sequence decomposition module to enhance the long-term change trend and periodic changes in the time series; Enhanced to: Use the result of S4.3.1 as Q in the AutoCorrelation module, and the result of S4.2 as K, V in the AutoCorrelation module to perform AutoCorrelation operation, that is, the result of S4.3.1 and W Q The result of matrix multiplication is Q, while S4.2 and W K and W V The result of the multiplication is used as K, V for input, W Q , W K and W V are three trainable parameter matrices; The expression is as follows: ENHANCEx t2 =AutoCorrelationgY,x t (Q xt1 ,K Y ,V Y )+x t1 ; In the formula, x t2 represents the enhanced output; x t1 represents the result of S4.3.1; Y represents the result of S4.2; x t represents the autocorrelation mechanism module of S4.3.2, where Q xt1 Indicates that the result of S4.3.1 is taken as Q, K in the autocorrelation mechanism module Y ,V Y It means taking the result of S4.2 as K,V in the autocorrelation mechanism module; S4.3.

3. The results of S4.3.2 are respectively outputted through a feedforward neural network to decode the trend part and the seasonal part of the health characteristic time series data related to the battery health status.

Images