A method for estimating the health status of lithium batteries

Abstract

The present invention discloses a method for estimating the health state of a lithium battery. The method transforms the coordinate space of a complete charging voltage data sequence and a fragment charging voltage data sequence into a voltage-differential voltage coordinate space, uses the complete charging voltage data sequence after the coordinate space transformation as training data, and uses the fragment charging voltage data sequence after the coordinate space transformation as verification data; trains an adaptive sampling deep neural network based on the training data and the health state label to obtain a health state prediction model; inputs the verification data into the health state prediction model to obtain a health state prediction value sequence, extracts multiple sub-prediction value sequences from the health state prediction value sequence in sequence, calculates the variance of each sub-prediction value sequence, and averages the health state prediction values ​​of the sub-prediction value sequences whose variance is lower than the variance threshold to obtain the final lithium battery health state estimation value. The battery health state estimation method can accurately predict the health state of a lithium battery based on the fragment voltage data sequence.

Description

Technical Field

[0001] The present invention relates to the field of lithium-ion battery state estimation, and in particular to a lithium battery health state estimation method. Background Art

[0002] Lithium-ion batteries are widely used in energy storage batteries for power systems and power batteries for electric vehicles and engineering vehicles due to their high energy density, long cycle life, and low self-discharge rate. During use, lithium-ion batteries are often in a complex system of electrochemical, thermal, and mechanical multi-physical fields. Sensors cannot directly monitor the state of health (SOH) that represents the overall degradation of the battery. Accurate estimation of the state of health (SOH) can not only determine the aging state of the battery, thereby assisting in estimating the state of charge (SOC) and determining the efficiency of the battery during use, but also help determine the battery replacement node, assist in battery recycling, and determine the risk of thermal runaway of the battery. Existing health state estimation methods usually use the constant current charging data of the battery, including basic physical quantities such as current, voltage, and temperature, and use this as the data basis to extract data features in different forms, and construct a mapping relationship between data features and health status through machine learning and deep learning.

[0003] Patent document with publication number CN117741484A discloses a method for estimating the health status of a lithium battery based on the characteristics of a charging voltage curve. The method comprises: obtaining a charging voltage data sequence and a health status data sequence of each charging and discharging cycle of the lithium battery; determining a constant current charging time period, and calculating a difference ΔV between the charging voltage curve and the average charging voltage in the constant current charging time period. i ; Analyze ΔV using Pearson correlation coefficient i The correlation with the current maximum charging capacity of the battery; collecting training samples to train the convolutional neural network-long short-term memory network to obtain a lithium battery SOH estimation model; using the lithium battery SOH estimation model to estimate the SOH of the lithium battery.

[0004] However, the feature extraction method used in the invention requires that the estimated data be charging data in the complete charging range, which is very different from the charging situation in actual use and has obvious shortcomings.

[0005] Patent document with publication number CN117783884A discloses a method for predicting the health status of a lithium battery based on fragmented charging data, the method comprising: obtaining the health status, charging voltage, charging current, and charging energy data of the lithium-ion battery during constant current and constant voltage charging in each charge and discharge cycle; calculating the incremental energy-state of charge (IES) curve of the lithium-ion battery; extracting the maximum value of the IES curve of the lithium-ion battery and the standard deviation of the IES curve of charging data under four different initial states of charge as health indicators; establishing a health status estimation model based on bidirectional LSTM-reduction based on the health indicators and health status training data of the lithium-ion battery; and extracting IES features from the fragmented charging data of the lithium-ion battery and sending them to the LSTM-reduction model to obtain the health status result of the lithium-ion battery.

[0006] However, the invention uses many types of battery data and needs to be filtered, which will distort the data. In addition, the fragmented charging data is used to fit the actual charging situation, but the polarization of the battery fragment charging is not considered, which is obviously insufficient. Summary of the invention

[0007] The present invention provides a lithium battery health state estimation method, which can accurately predict the health state of a lithium battery based on a segment voltage data sequence.

[0008] A specific embodiment of the present invention provides a method for estimating the health status of a lithium battery, comprising:

[0009] The obtained time-voltage coordinate space of the complete charging voltage data sequence and the fragment charging voltage data sequence is transformed into the voltage-differential voltage coordinate space, and the complete charging voltage data sequence after the coordinate space transformation is used as training data, and the fragment charging voltage data sequence after the coordinate space transformation is used as verification data;

[0010] Based on the training data and the obtained health status labels, an adaptive sampling deep neural network is trained to obtain a health status prediction model;

[0011] The verification data is input into the health status prediction model to obtain a health status prediction value sequence, and multiple sub-prediction value sequences are extracted from the health status prediction value sequence in turn, and the variance of each sub-prediction value sequence is calculated. The health status prediction values ​​of the sub-prediction value sequences whose variance is lower than the variance threshold are averaged to obtain the final lithium battery health status estimation value.

[0012] Preferably, the method for obtaining the voltage-differential voltage coordinate space comprises:

[0013] The voltage data points of the interval are converted into differential voltage by setting a fixed data interval, and the time-voltage coordinate space of the obtained complete charging voltage data sequence and the fragment charging voltage data sequence is converted into the voltage-differential voltage coordinate space by using the differential voltage and the voltage data points;

[0014] The complete charging voltage data sequence and the fragmentary charging voltage data sequence whose coordinate space is the time-voltage coordinate space are both composed of multiple pairs of time-voltage data points;

[0015] The complete charging voltage data sequence and the fragmentary charging voltage data sequence transformed into the voltage-differential voltage coordinate space are both composed of multiple pairs of voltage data points-differential voltages.

[0016] Preferably, the adaptive sampling deep neural network includes a first sliding window, and a plurality of subsequences of the same length are extracted from the training data through the first sliding window.

[0017] Preferably, a plurality of subsequences with the same length are extracted from the verification data through the first sliding window.

[0018] Preferably, the adaptive sampling deep neural network also includes a multi-layer neural network block, and the subsequence extracted from the training data is input into the multi-layer neural network block to obtain a predicted value, and the adaptive sampling deep neural network is fed back based on the health status label and the predicted value to update the parameters to obtain a health status prediction model.

[0019] Preferably, each neural network block includes a linear layer, a batch normalization layer, an activation function and Dropout, and the subsequences extracted from the training data are sequentially input into the linear layer, Dropout, batch normalization layer and activation function to obtain the output features of the current layer.

[0020] Preferably, a plurality of sub-prediction value sequences are sequentially extracted from the health status prediction value sequence, including:

[0021] The step size and width of the second sliding window are set, and multiple sub-prediction value sequences are extracted from the health status prediction value sequence by sliding the second sliding window. The data in the second sliding window after each sliding constitutes a sub-prediction value sequence, and each sub-prediction value sequence includes multiple health status prediction values.

[0022] Preferably, after the coordinate space transformation is completed, the change of the complete charging voltage data sequence after the coordinate space transformation is judged by an independent factor to be correlated with the change of the health state, so as to verify that the complete charging voltage data sequence after the coordinate space transformation can be used to train the neural network to obtain the health state prediction model; in the complete charging voltage data sequence after the coordinate space transformation, the first i Independent factors for each voltage data point for:

[0023]

[0024] in, For the i The differential voltage of the voltage-differential voltage data point, Convex hull of each data point after coordinate space transformation The closed boundary of Convex hull The closed boundary and the horizontal coordinate are i Voltage data points The intersection points correspond to the set of differential voltages, For The connecting line segment of two adjacent points on the closed boundary is the first e Points and e +1 point Connect line segments, k is the number of boundary points, n is the number of voltage-differential voltage data points, For the i voltage-differential voltage data points, and S is the set of all voltage-differential voltage data points.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] The present invention converts the time-voltage state space of the obtained charging voltage data sequence into a voltage-differential voltage coordinate space, so as to achieve the effect of evaluating the health status of the lithium battery without determining the starting time point, thereby solving the problem that the starting time point cannot be determined in practical applications, thus affecting the evaluation effect.

[0027] The present invention uses sampling segment charging voltage data as a validation set, and based on the polarization principle of the battery, selects a health status prediction value that is lower than the variance threshold, that is, a health status prediction value with higher stability, and takes the average to obtain a health status estimation value of the lithium battery, thereby realizing end-to-end health status estimation of charging data of different lengths and different starting points, and at the same time being able to solve the data offset caused by battery polarization during actual use, which is more in line with the actual situation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0029] Figure 1A flowchart of a lithium battery health status estimation method provided by a specific embodiment of the present invention;

[0030] Figure 2 A schematic diagram of a specific flow chart of a lithium ion health status estimation method provided by a specific embodiment of the present invention;

[0031] Figure 3 An experimental flow chart of battery data provided for a specific embodiment of the present invention;

[0032] Figure 4 A schematic diagram of an experimental charging mode provided for a specific embodiment of the present invention;

[0033] Figure 5 A diagram of training data and test data provided for a specific embodiment of the present invention;

[0034] Figure 6 A v-Δv graph of training data provided by a specific embodiment of the present invention;

[0035] Figure 7 An independent factor distribution diagram provided for a specific embodiment of the present invention;

[0036] Figure 8 A schematic diagram of a neural network unit structure provided for a specific embodiment of the present invention;

[0037] Fig. 9 A schematic diagram of the data length of test data provided by a specific embodiment of the present invention;

[0038] Fig.10 A schematic diagram of stability determination provided by a specific embodiment of the present invention;

[0039] Fig.11 The following are prediction results and comparison diagrams provided for specific embodiments of the present invention. DETAILED DESCRIPTION

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0041] When actually predicting the health status of a lithium battery, the input is mostly a fragmented charging voltage data sequence, and the actual starting time point of the fragmented charging data sequence cannot be used to locate the starting time point of its relatively complete charging process, and there is also polarization in battery charging. The above two points affect the accuracy of predicting the health status of the lithium battery. The specific embodiment of the present invention solves the problem that the starting time point cannot be accurately predicted by transforming the spatial state from the time-voltage coordinate space to the voltage-differential voltage coordinate space, predicts the health status of the lithium battery by voltage attenuation, and obtains the health status prediction value in the stable stage by the variance of the sub-prediction value sequence, thereby eliminating the unstable polarization process, making the predicted health status of the lithium battery more accurate; the neural network provided by the present invention uses an adaptive sampling method to extract multiple sub-sequences of the same length from the complete charging voltage data sequence and the fragmented charging voltage data sequence after the coordinate space is transformed to meet the needs of neural network model training and prediction. Compared with the prior art of zero-filling to achieve the same length, the data used by the present invention are all real data, and the model prediction obtained by training is more accurate. The method for estimating the health status of a lithium battery provided by the specific embodiment of the present invention, such as Figure 1 and Figure 2 As shown, including:

[0042] S1. Obtain a complete charging voltage data sequence and a fragment charging voltage data sequence of a lithium battery during a cyclic charging aging process, that is, a partial charging voltage data sequence at different charging starting points and different charging times, and a corresponding battery health status data sequence.

[0043] In a specific embodiment, if Figure 3 As shown, the embodiment of the present invention is carried out through constant current constant voltage aging cycle test (1C CC-CV) of lithium ion battery and segmented charging test after every 30 aging cycles, and obtains Figure 4 (a) shows the time-voltage coordinate space of the complete charging voltage data sequence, Figure 4 (b) shows a fragment of the charging voltage data sequence in the time-voltage coordinate space, as Figure 5 (a) and Figure 5 As shown in (b), a data sequence of the battery health status of a complete charging voltage data sequence and a fragmentary charging voltage data sequence is obtained.

[0044] The specific embodiment of the present invention provides j Voltage data sequence V j for:

[0045]

[0046] in, For the j The first voltage data sequence nVoltage data points are collected at a fixed frequency. In a specific embodiment, the data set has a sampling frequency of 1 Hz.

[0047] The specific embodiment of the present invention provides j The health status corresponding to the voltage data sequence SOH j for:

[0048]

[0049] in, Indicates that the battery is j A voltage data sequence, namely j The discharge capacity of the battery during a complete charge and discharge cycle. Indicates the initial discharge capacity of the battery.

[0050] S2, such as Figure 6 As shown, the specific embodiment of the present invention transforms the time-voltage (tv) coordinate space of the obtained complete charging voltage data sequence and the fragment charging voltage data sequence into the voltage-differential voltage (v-Δv) coordinate space, uses the complete charging voltage data sequence after the coordinate space transformation as training data, and uses the fragment charging voltage data sequence after the coordinate space transformation as verification data. The coordinate space transformation method of the coordinate space is:

[0051]

[0052] in, Indicates i voltage data points of voltage-differential voltage data points, For the i The differential voltage of the voltage-differential voltage data points, where is a fixed data interval, and the one used in this embodiment is 50. This embodiment converts the interval voltage data points into differential voltages by setting a fixed data interval. The time-voltage coordinate space of the obtained complete charging voltage data sequence and the fragment charging voltage data sequence is converted into the voltage-differential voltage coordinate space through the differential voltage and voltage data points. The complete charging voltage data sequence and the fragment charging voltage data sequence converted into the voltage-differential voltage coordinate space provided by this embodiment are both composed of multiple pairs of voltage data points-differential voltages, thereby establishing a basic strategy for estimating the health status using only voltage data, without using additional data such as current, temperature, and electrochemical impedance spectrum, and without data filtering, which is closer to the sensor configuration of a single battery under actual conditions.

[0053] S3. The voltage data points of the complete charging voltage data sequence after the coordinate space transformation are converted into independent factors. The independent factors are used to analyze the correlation between the voltage data points and the health status and the independence of the data of different health statuses. The independent factors obtained by the conversion are as follows: Figure 7 As shown in the figure. In the complete charging voltage data sequence after the coordinate space transformation, i Independent factors for each voltage data point for:

[0054]

[0055] in, For the i The differential voltage of the voltage-differential voltage data point, Convex hull of each data point after coordinate space transformation The closed boundary of Convex hull The boundary and horizontal coordinate are i Voltage data points The intersection points correspond to the set of differential voltages, For The connecting line segment of two adjacent points on the boundary is the first e Points and e +1 point Connect line segments, k is the number of boundary points, n is the number of voltage-differential voltage data points, For the i voltage-differential voltage data points, S is the set of all voltage-differential voltage data points, such as Figure 6 shown.

[0056] like Figure 7 As shown, it can be observed through the independence indicator that within the voltage range, the independence factors at different terminal voltages show a regular increasing change as the health state of the battery declines, indicating that the voltage-differential voltage coordinate space obtained by the coordinate space transformation provided by the present invention can be well used to train the model to obtain a health state prediction model.

[0057] S4. An adaptive sampling deep neural network (ASDNN) is established, and the complete charging voltage data sequence after coordinate space transformation is used to train the adaptive sampling deep neural network to obtain a health status prediction model.

[0058] S401: The adaptive sampling deep neural network (ASDNN) includes a first sliding window, and the data after the coordinate transformation of the complete charging voltage data is adaptively sampled, that is, the first sliding window is used to achieve training for input data of different lengths, wherein the complete charging voltage data sequence after the coordinate transformation X full The format is:

[0059]

[0060] in, t It is the length of the complete charging voltage data.

[0061] The adaptive sampling step provided in the specific embodiment of the present invention is: given an input sequence ,in, x t Depend on and Composition, and a fixed size W The first sliding window is used to obtain the sequence X Extract multiple subsequences from . The length of each subsequence is , the format is ,in f is the starting position of the window, The subsequence Evenly distribute the original sequence Corresponding health status label , the window used in this example is 50.

[0062] S402: In a specific embodiment of the present invention, the correspondence between multiple subsequences after adaptive sampling and health status is learned through neural network training, and the neural network is a multi-layer perceptron, and the multi-layer perceptron includes a multi-layer neural network block, and the training sample is input into the multi-layer neural network block to obtain a predicted value, and a loss function is constructed based on the health status label and the predicted value, and the neural network is trained using the loss function, and the parameters are updated to obtain a health status prediction model. The neural network established for end-to-end health status estimation in a specific embodiment of the present invention is based on a multi-layer perceptron, which can learn complex nonlinear relationships and adapt to various data distributions and patterns, and has the advantages of automatic feature extraction and high regression capabilities.

[0063] The multi-layer perceptron provided in the specific embodiment of the present invention is composed of 7 layers of neural network blocks, each of which is composed of a linear layer, a batch normalization layer (BN LAYRR), an activation function (LEAKY RELU) and Dropout. Figure 8 shown.

[0064] The linear layer provided in the specific embodiment of the present invention is used to learn the weight relationship between features and provide the network with rich representation capabilities. l input neurons (input features), which make up the input vector At the same time, this layer has neurons, each neuron corresponds to a weight vector, where the t The weight vector is and a bias , they represent the connection to the previous layer neurons and the current layer The weights of the neurons in the current layer The bias of the neuron. neurons, whose output It can be calculated by the following formula:

[0065]

[0066] in, is a nonlinear activation function, is a batch normalization layer.

[0067] The batch normalization layer ( BN Layer) normalizes each feature in a small batch to maintain the distribution of each layer input stable during training, which reduces the problem of internal covariate drift and helps the propagation of gradients and network training. , corresponding to BN Output of the layer , the calculation process can be formalized as follows:

[0068] , ;

[0069] ;

[0070] ;

[0071] in, is batchsize; is a small constant to avoid division by zero; and are learnable parameters that are updated during the training process.

[0072] The activation function provided in the specific embodiment of the present invention introduces nonlinear transformations, allowing the network to learn and represent more complex nonlinear relationships. This embodiment uses the extended Leaky ReLU of the rectified linear unit. Given an input , its output is as follows:

[0073]

[0074] in, is a small constant. That is, when When is a positive number, the derivative is 1; when When is negative, the derivative is (When x=0, the derivative can be defined as or 1).

[0075] The Dropout provided in the specific embodiment of the present invention forces the model to learn a more robust feature representation by randomly shutting down some neurons during the training process, thereby reducing excessive reliance on certain specific neurons and improving the generalization performance of the model. , Dropout will use probability Randomly shut down some of the neurons. For each neuron , the probability of being retained is , and the retained neurons are divided by Scaling is performed to maintain the expected input value unchanged. This process can be expressed by the following formula:

[0076]

[0077] in, is a Bernoulli distribution, which means The probability of taking the value is 1. The probability of is 0.

[0078] S5. Input the verification data into the health status prediction model to obtain a health status prediction value sequence, extract multiple sub-prediction value sequences from the health status prediction value sequence in turn, and calculate the variance of each sub-prediction value sequence, and average the health status prediction values ​​of the sub-prediction value sequences whose variance is lower than the variance threshold to obtain the final lithium battery health status estimate to achieve stability judgment.

[0079] S501, extracting multiple subsequences of the same length from the fragment charging voltage data sequence transformed from the coordinate space in advance by an adaptive sampling method, so that the input test data of different lengths can all predict the health status, wherein: Fig. 9shows the length of multiple segment charging voltage data sequences data, and the segment charging voltage data sequences transformed in the coordinate space X segment The format is:

[0080]

[0081] in Indicates the length of the segment charging voltage data.

[0082] The adaptive sampling steps provided in the specific embodiment of the present invention are: given an input sequence and adaptive sampling of deep neural networks The first sliding window is used to extract multiple subsequences from the sequence , the length of each subsequence is , the format is ,in p is the starting position of the window, The window used in this example is 50.

[0083] S502: Each subsequence provided by the specific embodiment of the present invention After the trained neural network, the corresponding health status prediction value will be output , the same input sequence The obtained health status prediction value sequence is .

[0084] The trained model used in this embodiment has Dropout turned off during testing, so that each neuron participates in the inference process of the model.

[0085] S503: The predicted value sequence Through the width of The step length is A second sliding window is used, and multiple sub-prediction value sequences are extracted from the health status prediction value sequence by sliding the second sliding window. The data in the second sliding window after each sliding constitutes a sub-prediction value sequence, and each sub-prediction value sequence includes multiple health status prediction values. The variance of the sub-prediction value sequence in the window is calculated to obtain a variance sequence ,in is the predicted value series The data length.

[0086] like Fig.10As shown, the stability judgment provided by the specific embodiment of the present invention is to average the health state prediction values ​​of the sub-prediction value sequence whose variance is lower than the variance threshold to obtain the final lithium battery health state estimation value, so that the polarized part can be eliminated and the relatively stable part can be retained, thereby accurately predicting the health state of the lithium battery. In one embodiment, the third sliding window width used is The sliding step is 25. is 8, the variance threshold The final estimated health status of the lithium battery is 1e-6. SOH p Close to the real lithium battery health status value SOH real .

[0087] Fig.11 (a) shows the estimated health status of each data, where the absolute error is the difference between the estimated result and the actual result. , represents the data format used in the present invention, The estimation result obtained by the method of the present invention for the incremental capacity curve data form commonly used in other inventions shows that the absolute error of the data form provided by the present invention is small.

[0088] Fig.11 (b) shows the stability of the estimation result. The percentage in the figure refers to the percentage of test data within a certain error range of the estimation result. The mean absolute error (MAE) and root mean square error (RMSE) of the estimation result of the method provided by the embodiment of the present invention are 0.0144 and 0.0182 respectively.

Claims

1. A method for estimating the health status of a lithium battery, characterized in that: include: The obtained time-voltage coordinate space of the complete charging voltage data sequence and the fragment charging voltage data sequence is transformed into the voltage-differential voltage coordinate space, and the complete charging voltage data sequence after the coordinate space transformation is used as training data, and the fragment charging voltage data sequence after the coordinate space transformation is used as verification data; Based on the training data and the obtained health status labels, an adaptive sampling deep neural network is trained to obtain a health status prediction model; Input the verification data into the health status prediction model to obtain a health status prediction value sequence, extract multiple sub-prediction value sequences from the health status prediction value sequence in turn, calculate the variance of each sub-prediction value sequence, and average the health status prediction values ​​of the sub-prediction value sequences whose variances are lower than the variance threshold to obtain the final lithium battery health status estimation value; The method for obtaining a voltage-differential voltage coordinate space includes: The voltage data points of the interval are converted into differential voltage by setting a fixed data interval, and the time-voltage coordinate space of the obtained complete charging voltage data sequence and the fragment charging voltage data sequence is converted into the voltage-differential voltage coordinate space by using the differential voltage and the voltage data points; The complete charging voltage data sequence and the fragmentary charging voltage data sequence whose coordinate space is the time-voltage coordinate space are both composed of multiple pairs of time-voltage data points; The complete charging voltage data sequence and the fragmentary charging voltage data sequence transformed into the voltage-differential voltage coordinate space are both composed of a plurality of pairs of voltage-differential voltage data points; The adaptive sampling deep neural network includes a first sliding window, through which a plurality of subsequences of the same length are extracted from the training data; Extracting multiple subsequences of the same length from the verification data through a first sliding window; After the coordinate space transformation is completed, the change of the complete charging voltage data sequence after the coordinate space transformation is judged by an independent factor to be correlated with the change of the health state, so as to verify that the complete charging voltage data sequence after the coordinate space transformation can be used to train the neural network to obtain a health state prediction model; In the complete charging voltage data sequence after the coordinate space transformation, i Independent factors for each voltage data point for: Among them, i The differential voltage of the voltage data points, Convex hull of each coordinate point after coordinate space transformation The closed boundary of Convex hull The boundary and horizontal coordinate are i Voltage data points The set of differential voltages corresponding to the intersection coordinate points, is the connecting line segment of two adjacent coordinate points on the boundary, that is, e Coordinate points and e +1 coordinate point The connecting line segments of k is the number of boundary points, n is the number of voltage data points, For the i The coordinate point corresponding to each voltage data point; The segment charging voltage data sequence after the coordinate space transformation is used as the verification data, and the coordinate space transformation method of the coordinate space is: ,in, Indicates i Voltage data points, For the i The voltage data points correspond to the differential voltage, where is a fixed data interval.

2. The lithium battery health status estimation method according to claim 1, characterized in that: The adaptive sampling deep neural network also includes a multi-layer neural network block, which inputs a subsequence extracted from the training data into the multi-layer neural network block to obtain a predicted value, feeds back the adaptive sampling deep neural network based on the health status label and the predicted value, and updates the parameters to obtain a health status prediction model.

3. The lithium battery health status estimation method according to claim 2, characterized in that: Each neural network block includes a linear layer, a batch normalization layer, an activation function and Dropout. The subsequences extracted from the training data are sequentially input into the linear layer, Dropout, batch normalization layer and activation function to obtain the output features of the current layer.

4. The lithium battery health status estimation method according to claim 1, characterized in that: Multiple sub-prediction value sequences are sequentially extracted from the health status prediction value sequence, including: The step size and width of the second sliding window are set, and multiple sub-prediction value sequences are extracted from the health status prediction value sequence by sliding the second sliding window. The data in the second sliding window after each sliding constitutes a sub-prediction value sequence, and each sub-prediction value sequence includes multiple health status prediction values.

Images