Battery health state prediction method based on multi-modal feature fusion and deep learning

Abstract

The invention discloses a battery health state prediction method based on multi-modal feature fusion and deep learning, and the method comprises the steps: obtaining the charging data, including current, voltage and temperature data, of each charging period during the multi-stage constant current charging of a battery; determining SOH (state of health) data of the battery; performing smooth filtering on the current and voltage data through a Savitzky-Golay filter, reconstructing a current curve, and reconstructing a voltage curve by adopting a linear interpolation method and a KF filter; constructing a comprehensive characteristic pool according to the current curve, the voltage curve and the temperature data, and directly adding the constant current value at the maximum charging amount into the comprehensive characteristic pool; screening features from the comprehensive feature pool and putting the features into a prediction feature pool; completely integrated empirical mode decomposition is carried out on the SOH data, and IMF components obtained through decomposition are put into a prediction feature pool; the method comprises the following steps: establishing a CNN (Content Number)-BILSTM-ATTENTION And performing single-step prediction and multi-step prediction on the SOH of the battery by using the trained model. According to the invention, high-precision prediction of the SOH of the electric vehicle battery can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of battery state of health prediction, and particularly to a battery state of health prediction method based on multi-modal feature fusion and deep learning. Background Art

[0002] Electric vehicles use lithium-ion batteries as energy storage devices, which can not only reduce the dependence on fossil fuels but also meet the carbon emission targets of the transportation industry. However, during the charge and discharge cycles, lithium-ion batteries gradually degrade in performance, manifested as a gradual decrease in available capacity, available energy, and available power. Therefore, how to accurately monitor and predict the state of health (SOH) of the battery using the data in the Battery Management System (BMS) has become an urgent problem to be solved.

[0003] Currently, many studies are dedicated to the analysis of battery degradation mechanisms, and extract the characteristic information of battery SOH degradation by processing the BMS monitoring data to construct a prediction model. Existing methods are mainly divided into model-based methods and data-driven methods. Model-based methods include electrochemical models and equivalent circuit models (ECMs). Electrochemical models describe the internal degradation mechanism by simulating the changes of key parameters during battery aging and use complex partial differential equations for simulation. However, due to the non-linear, strongly coupled, and time-varying characteristics of the battery SOH degradation process, this model has a high demand for health feature extraction and dimensionality reduction, which affects the accuracy of its application. In contrast, the ECM model is relatively simple and usually combines an adaptive filtering algorithm (such as the extended Kalman filter (EKF), smooth variable structure filter (SVSF)) to identify the degradation state by simplifying the parameters. However, due to the simplification of the model, the ECM often falls short when simulating complex battery degradation mechanisms and is difficult to accurately capture the changes in dynamic and static features, thus affecting the estimation accuracy.

[0004] As a popular research direction in recent years, the data-driven method shows significant advantages compared to the model-based method. The data-driven method can avoid complex electrochemical mechanisms and directly extract features from BMS data (such as charging current, voltage, temperature, etc.) and map them to the SOH state of the battery. The data-driven method does not rely on electrochemical knowledge and the state parameter identification process. With flexibility and strong non-linear adaptability, it can be directly applied to different systems, showing high adaptability and accuracy. Existing research has shown that deep learning models can achieve accurate estimation of SOH based on BMS feature data. However, the key to the data-driven method lies in selecting appropriate feature extraction methods and advanced prediction algorithms. Traditional degradation features are mostly extracted from charge and discharge data. However, in practical applications, the discharge process of electric vehicles is uncontrollable due to different application environments and usage states, the data is unstable, and the reliability of the extracted features is poor. The charging process is easier to control and has higher data stability. Many analysis methods such as incremental capacity analysis (ICA), differential voltage analysis (DVA), differential thermovoltammetry (DTV), etc. have been widely used to analyze the battery degradation mechanism. These methods usually obtain the corresponding curves through the differentiation of capacity, voltage, and temperature data, and select characteristic values such as the slope, peak, and its position of the curve. The application of such analysis methods is limited to the constant voltage and constant current charging environment and may be affected by noise. A new method based on noise-free reconstruction has been proposed to reduce the impact of noise on features.

[0005] In addition, with the growing demand for fast charging technology in the electric vehicle industry, multi-stage constant current charging technology has become an important performance indicator of batteries. However, non-constant current and high-speed charging pose new challenges to feature engineering. Researchers have tried to extract features from different cycles of the charging stage and build prediction models based on these features. Common regression models in the data-driven method include multiple linear regression, support vector regression (SVR), and relevance vector machine. In contrast, deep learning models not only have the ability of automatic feature extraction but also can learn long-term dependencies in time series data. Common methods include long short-term memory network (LSTM), recurrent neural network (RNN), and deep convolutional neural network (DCNN). In recent years, some research has combined bidirectional neural networks with attention mechanisms to construct a hybrid framework to make up for the limitations of single models and further improve the prediction accuracy of battery degradation. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for predicting the state of health of a battery based on multi-modal feature fusion and deep learning to achieve high-precision prediction of the SOH of an electric vehicle battery.

[0007] In the first aspect of the present invention below, a method for predicting the state of health of a battery based on multi-modal feature fusion and deep learning is provided. The method includes:

[0008] Obtain the charging data of each charging cycle during the multi-stage constant current charging of the battery, including current, voltage, and temperature data;

[0009] According to the charging data of each charging cycle of the battery, determine the state of health (SOH) data of each charging cycle of the battery, and perform filtering processing on the SOH data of each charging cycle of the battery to obtain the SOH data of the battery health state;

[0010] Perform smoothing filtering on the current and voltage data through a Savitzky-Golay filter, reconstruct the current curve based on the smoothed current data, and then use linear interpolation and a KF filter to process the smoothed voltage data to reconstruct the voltage curve;

[0011] According to the current curve, voltage curve, and temperature data, construct a comprehensive feature pool from three aspects: time series characteristics, temperature characteristics, incremental time and differential characteristics, and directly add the constant current value at the maximum charging amount to the comprehensive feature pool;

[0012] Use the Spearman correlation analysis method to quantify the correlation between each feature in the comprehensive feature pool and the SOH data of the battery health state, so as to screen features from the comprehensive feature pool and place them into the prediction feature pool;

[0013] Perform complete ensemble empirical mode decomposition (CEEMDAN) on the SOH data of the battery health state, and place the decomposed IMF components into the prediction feature pool;

[0014] Build a CNN-BILSTM-ATTENTION model, use the features in the prediction feature pool as input and the SOH data of the battery health state as output, and train the model;

[0015] Use the trained model to perform single-step prediction and multi-step prediction on the SOH of the battery health state.

[0016] In some embodiments, according to the charging data of each charging cycle of the battery, determining the state of health (SOH) data of each charging cycle of the battery includes:

[0017] Calculate the state of health (SOH) data of each charging cycle of the battery based on a variant of the SOC formula, and the formula is as follows:

[0018]

[0019] In the formula, ΔQ is the maximum capacitance difference of each charging cycle of the battery, ΔSOC is the maximum charging state difference of each charging cycle of the battery, and Q c represents the rated capacity of the battery.

[0020] In some of these embodiments, the state of health (SOH) data for each charging cycle of the battery is filtered to obtain the SOH data of the battery, including:

[0021] The SOH of the battery for each charging cycle is filtered by a first-order low-pass filter (LPF), and the formula is as follows:

[0022] t n = ax n + (1 - a)y n-1

[0023]

[0024] In the formula, y n is the current filtering value, y n-1 is the previous filtering value, x n is the current sampling value, T s is the period, f c is the cut-off frequency, and the value range of a is [0, 1].

[0025] In some of these embodiments, the current and voltage data are smoothed and filtered by a Savitzky-Golay filter, including:

[0026] Set the window length and construct a Vandermonde matrix accordingly:

[0027]

[0028] In the formula, V is the Vandermonde matrix, k is the window length, and k = 2m + 1;

[0029] The smoothed and filtered data is obtained by weighted summation with the weights of the pseudo-inverse matrix:

[0030] A = (V T V) -1 V T

[0031]

[0032] In the formula, T is the transpose, A is the pseudo-inverse matrix of V, a j is the corresponding weight of A, and y i represents the smoothed and filtered data.

[0033] In some of these embodiments, the linear interpolation method and the KF filter are used to process the smoothed and filtered voltage data to reconstruct the voltage curve, including:

[0034] For each charging cycle, starting from the first data point of the smoothed voltage data, each sampling point is corrected. The first non-monotonically increasing point is marked as an abnormal point, and the previous point is indexed. Then, the first non-abnormal point after the abnormality is searched backward for interpolation correction, that is:

[0035]

[0036] where V i is the voltage value of the abnormal point, V i-1 is the voltage value of the point previous to V i , V n represents the voltage value of the first non-abnormal point after the abnormal point;

[0037] After linear interpolation, a recursive method is used to perform a smoothing operation through the KF filter to reconstruct the voltage curve:

[0038]

[0039] P t|t-1 = AP t-1|t-1 A T + Q

[0040] K t = P t|t-1 H T (HP t|t-1 H T + R) -1

[0041]

[0042] P t|t = (I - K t H)P t|t-1

[0043] where is the predicted state at time t, is the updated state at time t, P t|t-1 is the predicted state covariance matrix, P t|t is the updated state covariance matrix, K t represents the Kalman gain, Q and R respectively represent the process noise and the observation noise covariance matrix, A is the state transition matrix, describing the transformation of the system state over time, H is the observation matrix, mapping the state space to the observation space, z t is the actual value, and I is the identity matrix, used to adjust the covariance matrix.

[0044] In some of these embodiments, the timing features include constant current charging time (CCCT), equal voltage rise time (TEVR), isothermal rise time (VERT), slope of charging voltage (SCV), area under constant current charging voltage (ACCCV), area under constant current charging current (ACCCC), and area under constant current charging temperature (ACCCT).

[0045] The temperature features include highest constant current charging temperature (HCCCT), average constant current charging temperature (MCCCT), and lowest constant current charging temperature (LCCCT).

[0046] The incremental time and differential features include incremental capacity peak (ICP), incremental capacity peak position (ICPL), incremental capacity area (ICA), incremental capacity slope (ICS), differential voltage region (DVA), differential voltage slope (DVS), differential voltage valley (DVV), differential voltage valley position (DVVL), temperature difference region (DTA), temperature difference slope (DTS), temperature difference peak (DTP), and temperature difference peak position (DTPL).

[0047] In some of these embodiments, screening features from the comprehensive feature pool and placing them into the prediction feature pool includes:

[0048] Screening features from the comprehensive feature pool whose absolute value of correlation is greater than the first correlation threshold, and denoting them as the candidate feature set;

[0049] Using the Spearman correlation analysis method to calculate the correlation scores between each pair of features in the candidate feature set, and denoting it as the correlation matrix R;

[0050] Setting the second correlation threshold, and selecting the feature with the highest correlation with the state of health (SOH) data of the battery as the reference feature x_base from the candidate feature set; sequentially checking the correlation between other candidate features and the reference feature x_base: if the correlation score is less than the second correlation threshold, then retain the feature and remove it from the candidate feature set; if the correlation score is greater than or equal to the second correlation threshold, then remove it from the candidate feature set;

[0051] Selecting the next feature with the highest correlation with the state of health (SOH) data of the battery from the remaining candidate feature set as the new reference feature x_base, and repeating the previous step until the candidate feature set is empty.

[0052] In some of these embodiments, the CNN - BILSTM - ATTENTION model includes a feature extraction and fusion module, a temporal dependency extraction and prediction module, and a bidirectional attention mechanism module; where:

[0053] The feature extraction and fusion module uses a convolutional kernel of size 1 to capture local time series information, and through convolutional operations on the input features of multiple time steps, realizes the preliminary extraction and fusion of features.

[0054] The temporal dependence extraction and prediction module captures the bidirectional dependence of temporal data through two LSTM layers, namely the forward and backward LSTM layers;

[0055] The bidirectional attention mechanism module adaptively highlights the feature information highly relevant to the battery health state by weighting the output features, which is used to strengthen feature selection and improve the key feature recognition ability of the model in multi-step prediction tasks. The final output of the model generates the prediction result through a fully connected layer;

[0056] And the grid search algorithm is adopted to optimize the hyperparameter settings of the model.

[0057] In some of these embodiments, the single-step prediction and multi-step prediction include:

[0058] X t ={soh t-n+1 ,soh t-n+2 ,…,soh t-1 ,soh t}

[0059]

[0060]

[0061] In the formula, SOH t represents the health state of the battery at time t; X t represents a data segment, which represents the SOH data of the battery health state within time t and the previous n time steps; represents predicting the predicted value at the next time step t + 1 using the data segment X t , is the prediction result of the model for multiple time steps from future t + 1 to t + m based on the data segment X t .

[0062] According to the second aspect of the present invention, an electric vehicle is provided. This electric vehicle uses the battery health state prediction method based on multi-modal feature fusion and deep learning described in any item of the first aspect to predict the battery health state.

[0063] Generally speaking, compared with the prior art through the above technical solutions conceived by the present invention, the following beneficial effects can be achieved:

[0064] (1) This application proposes a voltage and current interpolation filtering method based on Savitzky-Golay and KF. By means of difference filtering technology, noise reduction is carried out on current, voltage and target prediction data to reduce the influence of noise interference on feature extraction. Subsequently, the incremental capacity analysis (ICA), differential voltage analysis (DVA), and differential thermovoltammetry analysis (DTV) curves are reconstructed using the concept of regional capacity and interpolation denoising, and key feature points such as slope and peak are extracted from them to enhance the expressiveness of the feature pool.

[0065] (2) This application proposes a multi-scale feature pool that integrates time period, frequency component and degradation mode. Multidimensional information is obtained based on ICA, DVA and DTV methods, and CEEMDAN decomposition is introduced to effectively suppress the influence of white noise on the prediction results. This scheme accurately characterizes the degradation characteristics of the battery under variable working conditions and improves the model's ability to judge the state of health of the battery.

[0066] (3) This application combines analysis methods such as ICA, DVA and DTV to extract the characteristic values of battery aging from different cycles of the charging process, and performs modal decomposition on the target data to expand the feature pool, avoiding the influence of capacity regeneration phenomenon on the prediction accuracy. By embedding temperature features in the feature pool, the model can learn the degradation laws under different temperature conditions and significantly improve the accuracy of SOH prediction.

[0067] (4) This application adopts an attention mechanism to embed feature information into the feature pool, enabling the model to adaptively focus on the features that have a greater impact on SOH under different temperature conditions, and realizing dynamic response and adjustment to temperature information, thereby enhancing the adaptability of the prediction model in a multi-temperature environment.

[0068] (5) This application designs a combined prediction model of spatio-temporal attention mechanism based on convolutional neural network (CNN) and bidirectional long short-term memory network (BiLSTM). First, 1D CNN is used to extract the local features of the time series and capture the subtle changes in the battery degradation process. Then, BiLSTM is used to strengthen the modeling ability of long-term dependencies, and the features of key time steps are weighted through the attention mechanism, thereby improving the stability of multi-step prediction. This scheme can effectively slow down the problem of error accumulation and achieve accurate multi-step prediction of the state of health of the battery. Brief Description of the Drawings

[0069] Figure 1 It is the overall framework diagram of a battery state of health prediction method based on multi-modal feature fusion and deep learning provided by an embodiment of this application;

[0070] Figure 2 It is the flow schematic diagram of a battery state of health prediction method based on multi-modal feature fusion and deep learning provided by an embodiment of this application;

[0071] Figure 3 A charging data curve graph provided by an embodiment of the present application;

[0072] Figure 4 An SOH low-frequency filtering graph provided by an embodiment of the present application;

[0073] Figure 5 A voltage-current curve correction graph provided by an embodiment of the present application;

[0074] Figure 6 A Spearman correlation analysis graph provided by an embodiment of the present application;

[0075] Figure 7 A CEEMDAN modal decomposition graph provided by an embodiment of the present application;

[0076] Figure 8 A schematic diagram of the model structure provided by an embodiment of the present application;

[0077] Figure 9 An overall result comparison graph provided by an embodiment of the present application;

[0078] Figure 10 A comparison graph of prediction period results provided by an embodiment of the present application;

[0079] Figure 11 An evaluation index comparison graph provided by an embodiment of the present application;

[0080] Figure 12 An absolute error comparison graph provided by an embodiment of the present application;

[0081] Figure 13 A prediction effect graph for different temperature zones provided by an embodiment of the present application;

[0082] Figure 14 A graph of changes in evaluation indicators provided by an embodiment of the present application;

[0083] Figure 15 An absolute error effect graph provided by an embodiment of the present application;

[0084] Figure 16 A performance test graph with prediction steps of 5, 10, and 15 provided by an embodiment of the present application; wherein, Figure 16 (a) therein represents a comparison graph of 5-step prediction results, Figure 16 (b) therein represents a comparison graph of 10-step prediction results, Figure 16 (c) therein represents a comparison graph of 15-step prediction results;

[0085] Figure 17 A multi-step prediction effect comparison graph provided by an embodiment of the present application;

[0086] Figure 18 This is a multi-step prediction absolute error comparison chart provided by the embodiments of the present application. Detailed implementation manners

[0087] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments provided in this application without creative efforts fall within the scope of protection of the present invention.

[0088] Obviously, the accompanying drawings in the following description are only some examples or embodiments of this application. For those of ordinary skill in the art, without creative efforts, this application can also be applied to other similar scenarios based on these drawings. In addition, it can also be understood that although the efforts made in this development process may be complex and lengthy, for those of ordinary skill in the art related to the content disclosed in this application, some design, manufacturing or production changes made on the basis of the technical content disclosed in this application are only conventional technical means and should not be understood as the content disclosed in this application being insufficient.

[0089] Referring to "embodiments" in this application means that the specific features, structures or characteristics described in connection with the embodiments can be included in at least one embodiment of this application. The phrase appears in various positions in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those of ordinary skill in the art explicitly and implicitly understand that the embodiments described in this application can be combined with other embodiments without conflict.

[0090] Unless otherwise defined, the technical terms or scientific terms involved in this application shall have the ordinary meanings understood by those with ordinary skills in the technical field to which this application belongs. The words such as "a", "an", "one", "the" and the like involved in this application do not indicate a limitation in quantity and may represent a singular or plural number. The terms "comprise", "include", "have" and any variations thereof involved in this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may further include steps or units not listed, or may further include other steps or units inherent to these processes, methods, products or devices. The similar words such as "connect", "be connected", "couple" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "plurality" involved in this application means two or more. "And / or" describes the association relationship of associated objects and indicates that three relationships may exist. For example, "A and / or B" may represent: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally represents an "or" relationship between the associated objects before and after. The terms "first", "second", "third" and the like involved in this application are only used to distinguish similar objects and do not represent a specific order for the objects.

[0091] In the existing field of predicting the state of health (SOH) of electric vehicle batteries, data-driven methods have shown significant advantages, but still face several key challenges in practical applications. Specifically as follows:

[0092] (1) The real electric vehicle charging data is severely interfered by noise

[0093] In the data collected by the actual electric vehicle battery management system (BMS), high noise interference affects the accurate prediction of the battery state of health. At the same time, existing methods cannot extract sufficient degradation features from the charging data and it is difficult to fully characterize the aging mode of the battery.

[0094] (2) Insufficient extraction of multi-scale features

[0095] When existing models construct the feature pool, it is difficult to fully consider the multi-scale features and frequency components of battery degradation, and often lack effective noise suppression means. Therefore, the richness and accuracy of the feature pool are insufficient, and it is difficult to accurately reflect the degradation characteristics of the battery under different working conditions.

[0096] (3) The difficulty of feature extraction and the capacity regeneration effect during the charging process

[0097] In the prior art, during the feature extraction process of charging data, it is vulnerable to the influence of the capacity regeneration effect, resulting in poor stability of the extracted features, and white noise cannot be effectively eliminated, which has an adverse impact on the prediction results.

[0098] (4) Lack of adaptive response to battery degradation characteristics at different temperatures

[0099] Different temperature conditions have an important impact on the battery degradation rate and mode. However, existing prediction models are difficult to adaptively focus on the key features under different temperature conditions, resulting in insufficient prediction stability.

[0100] (5) Error accumulation and insufficient capture of key features in multi-step prediction

[0101] Existing deep learning models have the problem of error accumulation in multi-step prediction, which affects the prediction stability. In addition, when the model extracts key features in the time series, it is difficult to effectively capture the subtle change information during the degradation process.

[0102] Therefore, this application provides a long-term health state monitoring method for electric vehicle battery data based on a data-driven method, and solves the above technical problems through the following technical solutions:

[0103] (1) This application proposes a voltage and current interpolation filtering method based on Savitzky-Golay and KF. Through the difference filtering technology, noise reduction is performed on the current, voltage, and target prediction data to reduce the influence of noise interference on feature extraction. Subsequently, the incremental capacity analysis (ICA), differential voltage analysis (DVA), and differential thermovoltammetry analysis (DTV) curves are reconstructed using the regional capacity concept and interpolation denoising, and key feature points such as slopes and peaks are extracted from them to enhance the expressiveness of the feature pool.

[0104] (2) This application proposes a multi-scale feature pool that fuses time period, frequency components, and degradation modes. Based on the ICA, DVA, and DTV methods, multi-dimensional information is obtained, and CEEMDAN decomposition is introduced to effectively suppress the influence of white noise on the prediction results. This solution accurately characterizes the degradation characteristics of the battery under variable working conditions and improves the model's ability to judge the battery health state.

[0105] (3) This application combines analysis methods such as ICA, DVA, and DTV to extract the characteristic values of battery aging from different cycles of the charging process, and performs modal decomposition on the target data to expand the feature pool, avoiding the influence of the capacity regeneration phenomenon on the prediction accuracy. By embedding temperature features in the feature pool, the model can learn the degradation laws under different temperature conditions and significantly improve the accuracy of SOH prediction.

[0106] (4) This application adopts an attention mechanism to embed feature information into the feature pool, enabling the model to adaptively focus on the features that have a greater impact on SOH under different temperature conditions, and achieving dynamic response and adjustment to temperature information, thereby enhancing the adaptability of the prediction model in a multi-temperature environment.

[0107] (5) This application designs a combined prediction model based on a convolutional neural network (CNN) and a bidirectional long short-term memory network (BiLSTM) spatio-temporal attention mechanism. First, 1D CNN is used to extract the local features of the time series to capture the subtle changes in the battery degradation process. Then, BiLSTM is utilized to strengthen the modeling ability of long-term dependencies, and the features of key time steps are weighted through the attention mechanism, thereby improving the stability of multi-step prediction. This solution can effectively slow down the problem of error accumulation and achieve accurate multi-step prediction of the battery health state.

[0108] As Figure 1 and Figure 2 shown, this application proposes a combined prediction model based on complete ensemble empirical mode decomposition (CEEMDAN) and a spatio-temporal attention mechanism neural network to achieve high-precision prediction of the SOH of electric vehicle batteries. The specific steps are as follows:

[0109] 1. Electric vehicle data collection

[0110] Obtain the charging data of 20 commercial electric vehicles with the same battery system under monitoring. The running time span of the vehicles is all two years (about 29 months). The vehicle numbers are #1, #2,..., #20. The battery model is a ternary lithium battery. The charging equipment receives the battery charging data through controller area network (CAN) communication during the charging process. The data coding frequency is 8s, the rated capacity of the battery is 145Ah, and the voltage and current of the battery pack are recorded as the total voltage and current. The charging current is defined as negative. Table 1 lists the main items of the battery charging data related to battery health assessment and shows its resolution.

[0111] Table 1 Battery resolution table

[0112]

[0113] Partial charging data of the vehicle is as Figure 3 shown. The vehicle adopts a multi-stage constant current charging strategy. During the charging process, the current remains constant at different stages. The advantage of this technology is that it can reduce the charging time and improve the cycle life of the battery. It can also be preliminarily found from the figure that there is a highly coupled relationship between SOC and battery capacity, and their increase and decrease relationships are extremely similar.

[0114] 2. Battery SOH data calculation and preliminary processing

[0115] Since the vehicle operation data is greatly affected by driving habits and environmental conditions, resulting in uncertainties in battery health prediction, this application will calculate the battery health state using charging-related data. First, it is observed that the charging segment intervals of some charging cycles of the vehicle are very small, with the starting value of SOC greater than 50% and the ending value less than 90%. Considering the errors caused by missing information, it is necessary to extract the charging data that meets the requirements from all 1500 cycles of the vehicle. By analyzing the charging protocol and past research, the charging cycle data with the SOC ranging from 30% to 80% and the terminal voltage segment ranging from 3.7V to 4.1V is used as the identification standard to eliminate the redundant information in the charging data and mark the time cycle.

[0116] For the extracted standard data, it is necessary to obtain the battery SOH by calculating the battery capacity in the cycle state, so as to predict the decay trajectory of the battery health. It is difficult for the real electric vehicle cycle to complete an absolute charge-discharge process. Therefore, in order to obtain the standard capacity under the limited state, this application uses a variant of the SOC formula shown in Equation (1) to calculate the battery SOH.

[0117]

[0118] Where ΔQ refers to the maximum capacitance difference in each charging cycle, ΔSOC is the maximum charging state difference in each charging cycle, and Q c represents the rated capacity. By outputting the SOH of the sample vehicle through the maximum sampling interval within the cycle, it can effectively avoid the error accumulation caused by the noise fluctuation of the system-reported data. Then, the final cycle SOH data is filtered through a first-order low-pass filter (Low Pass Filter, LPF) shown in Equation (2) to only reduce the high-frequency noise interference and retain the true characteristics of the battery cycle SOH changing with time. The core parameter of the LPF is the cut-off frequency f c , and this algorithm retains the data signal within the cut-off frequency while attenuating the signal outside the cut-off frequency. The sample SOH changes with the cycle as Figure 4 shown.

[0119]

[0120] Among them, the first formula and the second formula are equivalent, and y n is the filtering value this time, y n-1 is the previous filtering value, x n is the sampling value this time, T s is the sampling period of the sample, and f c is the cut-off frequency, and the value range of a is [0,1].

[0121] 3. Charging Data Filtering and Correction

[0122] Multi-level constant current charging data, which contains complex current change situations, not only generates noise interference but also has a great impact on data analysis extraction and prediction. The working principle of the Savitzky-Golay filter is based on local least squares fitting within a time window, smoothing the data within a given window, preserving the data change trend and reducing interference. By processing multi-level constant current charging data with the Savitzky-Golay filter, while retaining the important time-domain features of stable constant current data, unnecessary noise interference is removed. The specific operation is shown in Equation (3). First, since the amount of time data in a single cycle is still relatively large, the window length is set to 101 here, and a Vandermonde matrix is constructed accordingly. Finally, the smoothed data is obtained by weighted summation of the weights of the pseudo-inverse matrix.

[0123]

[0124] A=(V T V) -1 V T

[0125]

[0126] where m represents the upper boundary of the measurement points within the window, k = 2m + 1, V is the Vandermonde matrix established from the source data, A is the pseudo-inverse matrix of V, a j is the corresponding weight of A, and y i represents the smoothed data.

[0127] From Figure 5 it can be seen that even when the battery is charged with a constant current, there will still be fluctuating noise for a long time. After passing through the SG filter, the influence of the noise current can be greatly reduced, and the error value can be corrected through differential comparison while retaining the features.

[0128] In addition, in the charging mode, the terminal voltage of the battery should satisfy: when t i >t i-1 then,

[0129] However, due to the errors of the measurement device, the actually measured terminal voltage of the battery may be:

[0130]

[0131] To solve the problem of abnormal voltage fluctuations, the linear interpolation method and the KF filter are used to reconstruct the voltage curve. For each cycle, starting from the first data point, each sampling point is corrected. The first non-monotonically increasing point is marked as an outlier, and its previous point is added with an index. Then, the first non-outlier point after the outlier is searched backward for interpolation correction. Finally, through KF filtering, it is ensured that it is smoothly monotonically increasing within the cycle and important features are retained, laying a foundation for subsequent curve reconstruction. Assume that there is an outlier V i :

[0132]

[0133] After linear interpolation, a recursive method is used to smooth the operation through the KF filter as shown in Equation (4):

[0134]

[0135] where V n represents the first non-outlier point after the outlier point. is the predicted state at time, is the updated state at time, P t|t-1 is the predicted state covariance matrix, P t|t is the updated state covariance matrix, K t represents the Kalman gain, and Q and R represent the process noise and the observation noise covariance matrix respectively. The reconstructed curve is as Figure 5 shown. From the results in the figure, it can be seen that through the reconstruction method, the noisy voltage value can be completely corrected, and the process of reducing the jumping voltage at the end of the battery charging state can be repaired. Finally, based on the corrected voltage and current data, the DQ / DV curve, DV / Dt curve, and DT / Dt curve are extracted.

[0136] 4. Feature Pool Establishment

[0137] The feature engineering extraction logic of this application refers to the comprehensive analysis of all features of lithium batteries that affect the battery health status (SOH). Referring to the health diagnosis framework of the research, this application constructs an electric vehicle-related feature pool from three aspects: the basic time series features, temperature features, and incremental time and differential features of the source data. At the same time, due to the strong adaptability and generalization ability of the complete ensemble empirical mode decomposition (CEEMADAN), this application uses the CEEMADAN method to enrich the feature set and improve the prediction accuracy. Finally, the appropriate prediction feature pool is selected through feature screening. The extracted features have been proven to have a strong correlation with battery health as shown in Table 2.

[0138] Table 2 Comprehensive Feature Pool Table

[0139]

[0140]

[0141] 5. Establish a prediction feature pool

[0142] For the extracted features, it is also necessary to reduce the dimensionality of the features to avoid overfitting, so that the model method has high accuracy and robustness. In this application, the Spearman correlation analysis method is used to screen features. Due to its universality, non-parametric nature and robustness, and its applicability to non-linear relationship data, it is widely used in the correlation analysis part of many studies on battery datasets. Based on the above advantages, this application uses Spearman correlation analysis to quantify the association relationship between SOH and each feature, and the calculation equation is Equation (5). After calculating the Spearman correlation coefficient, it is necessary to check whether there is autocorrelation and dependence among the features to avoid the impact of information redundancy on the model accuracy. The correlation calculation results are as Figure 6 shown.

[0143]

[0144] In the formula: x i and y i represent the reference sequence and the comparison sequence respectively, and n represents the length of the sample sequence. Among them, the soh of the battery and the extracted feature set are successively defined as the constructed reference sequences. The value of the correlation coefficient ranges from -1 to +1, and the larger the absolute value, the stronger the dependence. The correlation coefficient is shown in the figure. Generally speaking, in statistics, when the absolute value of the correlation coefficient is lower than 0.25, the variables are considered to be weakly correlated. When the absolute value of the correlation coefficient is higher than 0.75, there is a strong correlation between the two variables. Therefore, in this embodiment, the thresholds are set to 0.25 and 0.75 respectively to ensure the rationality of feature screening. The screening steps are shown in Table 3, and finally 6 are selected from the comprehensive feature pool and placed into the prediction feature pool.

[0145] Table 3 Feature screening process table

[0146]

[0147] At the same time, due to the strong adaptability and generalization ability of the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), the present invention uses the CEEMDAN method to enrich the feature set and improve the prediction accuracy. The decomposition results are shown in Figure 7 . As Figure 7 shown, this application selects 6 IMF components.

[0148] 6. Establish a CNN-BILSTM-ATTENTION model

[0149] This application proposes an electric vehicle battery state of health (SOH) prediction model based on a feature extraction and fusion module, a temporal dependency extraction and prediction module, and a bidirectional attention mechanism. The overall framework is as shown in Figure 8 . This model consists of three core modules: First is the feature extraction and fusion module, which aims to extract key features from high-noise, multi-dimensional battery data. By using a convolutional kernel of size 1, the model can accurately capture local time series information with the smallest sliding window. Meanwhile, 64 filters are configured, and the ReLU activation function is used to enhance the non-linear expression ability of features, effectively improving the representation effect of complex features. 1D CNN has high noise resistance in processing time series data, and can quickly filter out invalid information and extract representative time series patterns. This module receives the input feature pool from 13 filtered prediction features and performs convolution operations on sequence data with a length of 15 time steps, thereby realizing the preliminary extraction and fusion of features and providing a high-quality feature basis for subsequent temporal relationship modeling.

[0150] Second is the temporal dependency extraction and prediction module. This module introduces BiLSTM to model the complex dependencies of battery SOH over time. By simultaneously capturing the bidirectional dependencies of time series data through forward and backward LSTM layers, it not only utilizes past historical information but also combines future time series data for global temporal feature learning. The module uses the Bidirectional(LSTM) function, which applies bidirectional LSTM to the feature sequence output by the convolutional layer. The number of LSTM units is set to 50 to fully process the high-dimensional feature sequence output by the convolutional layer. Compared with unidirectional LSTM, BiLSTM can learn complex temporal dependencies more comprehensively. Especially in the case where the change of battery health state is non-linear and non-stationary, it can effectively improve the ability to capture long-term dependency features, thereby improving the accuracy of SOH prediction.

[0151] Finally, there is a bidirectional attention mechanism module, which is used to strengthen feature selection and improve the model's key feature recognition ability in multi-step prediction tasks. The attention mechanism adaptively highlights the feature information highly correlated with the battery health state by weighting the output features of the convolution and temporal extraction modules. This module takes into account the importance distribution of features in both time and space, significantly improving the stability and accuracy of the model in multi-step prediction tasks. The attention mechanism outputs the feature information as a one-dimensional feature vector through the Flatten function. It is also controlled by the parameter SINGLE_ATTENTION_VECTOR. When it is True, a single vector is used to simplify the calculation. The final output of the model generates the prediction result through a fully connected layer (Dense). During the training process, the Dropout layer is used to avoid overfitting, and Batch Normalization is used to accelerate the convergence of the model. To optimize the model performance, the learning rate is initialized to 0.001, and the number of units in the hidden layer is 99. During training, the batch size of the model is set to 12, and the number of training epochs is set to 90. In addition, 10% of the data (validation_split = 0.1) is used as the validation set during the training process to monitor the performance of the model on the validation set and avoid overfitting.

[0152] 7. Finding the Optimal Hyperparameters

[0153] When selecting the model hyperparameters, the grid search algorithm is used for adaptive adjustment and selection to improve efficiency. The specific steps are as follows:

[0154] Step 1: First, process the dataset of electric vehicles, reconstruct the voltage, current, and temperature curves of the battery, and extract the necessary capacity degradation data. By calculating the capacity corresponding to each charge-discharge cycle, it is converted into the state of health (SOH) of the battery, and denoising techniques (Savitzky-Golay and KF filtering) are used to remove the noise interference in the data to ensure the data quality.

[0155] Step 2: Screen the key features from the source data features and the reconstructed curves, and then split the processed SOH data into a training set and a test set with a ratio of 7:3. At the same time, the complete ensemble empirical mode decomposition (CEEMDAN) method is used to eliminate the influence of the capacity regeneration phenomenon and enrich the feature set to ensure that the model can capture more complex patterns and trends.

[0156] Step 3: Train the model on the training set and optimize the hyperparameter settings of the model using the Grid Search algorithm. Through grid search, the finally found optimal hyperparameters are: the learning rate (lr) is 0.05142, the batch size (batch_sizes) is 38, the number of LSTM units (lstm_units) is 70, the number of units in the Dense layer (dense1) is 47, the number of epochs is 937, and the time step (look_back) is 6.

[0157] Step 4: Verify whether the network converges within the predetermined 937 iterations. If the model converges successfully, record and output the hyperparameter values and perform verification on the test set; if the model does not converge, return to Step 3, readjust the hyperparameters and perform training again.

[0158] Step 5: Evaluate the model performance through the test set, generate prediction results and relevant evaluation metrics (such as RMSE and MAPE) to verify the actual effect of the model in predicting the state of health of electric vehicle batteries, and conduct a comprehensive analysis of the prediction results.

[0159] This application uses the state of health of electric vehicle batteries as the prediction target. For the prediction step of the model, single-step prediction and multi-step prediction are considered successively. Assume:

[0160] X t ={soh t-n+1 ,soh t-n+2 ,…,soh t-1 ,soh t}

[0161]

[0162]

[0163] In Equation (6), SOH t represents the state of health of the electric vehicle battery at time t. X t represents a data segment, indicating the SOH data at time t and within the previous n time periods. represents using the data segment X t to predict the predicted value at the next time step t + 1, is the prediction result of the model for multiple time steps from t + 1 to t + m in the future based on the data segment X t .

[0164] First, consider that at a specific time T, use the historical SOH data and feature data from the initial time to this moment, input them into the trained prediction model, and the model outputs the SOH data for the future cycle time. Then the predicted generated data is compared with the recorded data Xt Fuse to generate a new historical sequence X t+1 , and recursively predict the future SOH data of the battery accordingly. This application uses 6 historical data for recursion to achieve accurate prediction of the SOH data of the test set. At the same time, this study extrapolates single-step prediction to 15-step multi-step prediction, which is more suitable for scenarios involving long-term prediction and trend analysis. By effectively combining single-step and multi-step prediction techniques, the accuracy and reliability of the prediction can be improved, providing support for battery prognosis decision-making.

[0165] 8. Model evaluation

[0166] To verify the correlation between eigenvalue and the prediction accuracy of the model, the root mean square error (RMSE) and mean absolute percentage error (MAPE) indexes in Equation (7) are selected to quantify the prediction error. The calculation formulas are as follows:

[0167]

[0168] In the equation, n represents the number of test samples, and y i is the actual health state of the sample, represents the predicted value of the sample. Among these indexes, the smaller the RMSE and MAPE are, the higher the prediction accuracy is, and the better the performance of the model is.

[0169] In summary, this application proposes a long-term health monitoring method for electric vehicle batteries based on a combined deep learning model. The superiority of the system is described below from the results of single-step prediction, prediction in different temperature ranges, and multi-step prediction.

[0170] 1. Comparison of single-step prediction results

[0171] To verify the effectiveness of the monitoring system, the SOH values of the cyclic changes of real electric vehicle batteries are predicted using the same selected features, and the prediction framework of the present invention is compared with other mainstream methods to verify its performance in predicting battery SOH. All methods use the first 554 cycle data as the training set, accounting for 70% of the total data set, and the remaining part as the test set. In addition, this study deletes the cycle data in some mutation stages to exclude unnecessary noise interference. The overall prediction results are shown in Figure 9 as follows, Figure 10It is possible to carefully compare the prediction results of various models. It can be preliminarily found from the figure that the prediction effects of the RNN and GRU models are relatively poor. In the subsequent result comparison, they will not be included in the comparison elements. By comparing the evaluation indicators of the models, it is found that compared with the RNN, GRU, LSTM, and Seq2seq models, the selected prediction model CBAG (i.e., the CNN-BILSTM-ATTENTION model optimized by the grid search algorithm) shows better prediction fitting. Its RMSE and MAPE are 0.00278 and 0.00279 respectively, which are the minimum values in the comparison. In Table 4, the SOH prediction results of four electric vehicles show that according to the detailed evaluation of the model performance indicators, the CBAG model is superior to the RNN and GRU models. The evaluation results show that the performance of the CBAG model has increased by more than 90%. Compared with the LSTM and Seq2Seq models, the average RMSE of the CBAG model has increased by 62% and 69% respectively, while the average MAPE results show that the accuracy has increased by 65% and 72% respectively.

[0172] Table 4. Comparison table of RMSE and MAPE results of different models for four electric vehicles

[0173]

[0174] At the same time, in order to study the influence of each block in the model on the model, the present invention respectively compares the performance among the CNN-Bilstm, Bilstm-Attention, and Bilstm models through ablation experiments. Figure 11 It details the comparison of the MAPE and RMSE results between each model. Figure 12The absolute error is represented by the difference between the predicted value and the actual value. Table 5 shows the comparison of the evaluation index results of the ablation experiment using the data of four different electric vehicles. The comparison results between the ablation experiment and the benchmark model show that the algorithm of this application has a better single-step prediction effect on the health of real electric vehicles than the current research models. Single-step prediction requires the model to make accurate predictions for the next time step. The BiLSTM module can capture more time-dependent information by processing the forward and backward data of the time series, thus performing more precisely in single-step prediction. Coupled with the attention mechanism's focus on important features, the model can more effectively filter out irrelevant information and further improve the performance of single-step prediction. In contrast, traditional RNN, LSTM, and GRU have limitations in capturing time-dependent relationships and cannot adjust weights as flexibly as the attention mechanism, so it is difficult to accurately predict the battery SOH. Table 5 shows that each module plays a crucial role in improving prediction performance. The performance of the CBAG model is always better than that of the simplified configuration and the benchmark model, with the RMSE increased by an average of 25% and the MAPE increased by an average of 29%. These findings emphasize the robustness of the CBAG model in accurately predicting the health of real electric vehicle batteries in single-step prediction.

[0175] Table 5. Comparison table of RMSE and MAPE results of ablation experiment for four electric vehicles

[0176]

[0177] 2. Description of the prediction effect of the model in different temperature ranges

[0178] The impact of high-temperature environment on vehicle batteries is an important research area. The high-temperature environment will accelerate the capacity decay of the battery, accelerate the battery aging process, and is accompanied by the risk of thermal runaway. Therefore, temperature is an influencing factor that cannot be ignored for battery health. It can be found from the data that the distribution of the highest and lowest temperatures of the battery is mainly between 15 - 40. To study whether the prediction model can still accurately predict the battery health in different temperature environments, first, according to the existing research, set the high-temperature working environment of the battery environment temperature > 35 degrees Celsius (about 87 degrees Fahrenheit), and then divide the high-temperature cycle and the normal cycle in the life cycle into Y1 - Y4. 70% of the data in each temperature range is designated as the training set by the prediction model, and the latter 30% of the data is used as the test set. As Figure 13 shown, the model fitting degree of the prediction model in the overall temperature range is 0.9686, with a good fitting degree. The overall RMSE and MAPE are 0.0063 and 0.0057 respectively. The prediction results in different temperature ranges are as Figure 14As shown, it can be found that the prediction accuracy gradually increases with the increase of the temperature range. When the temperature range is small, that is, within the first 149 high-temperature range cycles, the values of REMS and MAPE decrease from 0.00832 and 0.008812 respectively to 0.005543 and 0.004525 in the latter 231 high-temperature range cycles.

[0179] In addition, by comparing Figure 15 the absolute errors between the SOH prediction values and the true values in each temperature range, it can be found that the proposed model has good performance in predicting in different temperature environments. This prediction effect benefits from the CNN module extracting local features and combining with BiLSTM to deeply model time series data, enabling the model to better capture the complex non-linear relationships brought by temperature fluctuations. The attention mechanism can automatically identify and highlight the important features related to temperature changes during the training process, thereby enhancing the prediction stability under temperature conditions.

[0180] 3. Comparison of multi-step prediction results

[0181] Multi-step prediction based on single-step prediction is a complex but important research field. It has extensive applications in multiple fields. In actual battery capacity demand prediction, using a multi-step prediction model can help the energy management system allocate resources more effectively. Although the recursive prediction method is the simplest implementation method, its accuracy may be affected due to problems such as error accumulation. This study improves the prediction accuracy, efficiency and reliability of multi-step prediction through the CNN-BiLSTM-Attention prediction framework. To verify the stability and accuracy of the model in multi-step prediction, this study conducts performance tests with prediction steps of 5, 10 and 15, and the results are as Figure 16 . Among them, Figure 16 (a) in represents the comparison chart of 5-step prediction results, Figure 16 (b) in represents the comparison chart of 10-step prediction results, Figure 16 (c) in represents the comparison chart of 15-step prediction results.

[0182] As Figure 17 shown, as the number of time steps to be predicted increases, the performance of the SEQ2SEQ and LSTM models gradually decreases, while the performance of the CNN-BiLSTM-Attention model is relatively stable. It can be seen from the box Figure 18 that the absolute error of the prediction model of this application also stabilizes in a small fluctuation range.

[0183] First, in terms of RMSE, the CNN-BiLSTM-Attention model shows the most stable error growth in predictions from step 1 to step 15. Its RMSE value gradually increases from 0.002783 to 0.00505758, with a very smooth growth rate. This indicates that the model has good long-term prediction ability, and the error increases relatively slowly as the number of prediction steps increases. In contrast, when the prediction step length increases, the RMSE of the LSTM and Seq2Seq models fluctuates more violently, especially when predicting after step 10. The RMSE of the LSTM model reaches 0.010573344 at step 15, while the Seq2Seq model even has higher errors at some steps, such as the RMSE at step 10 being 0.016258908. This also shows that the CNN-BiLSTM-Attention model can better handle the problem of error accumulation in multi-step predictions. The MAPE results further prove the robustness and accuracy of the CNN-BiLSTM-Attention model. The MAPE of the CNN-BiLSTM-Attention model increases from 0.002791 to 0.00506873 at step 15, with a small increase, maintaining a high-precision prediction of the target value. In contrast, the MAPE of the LSTM and Seq2Seq models increases sharply as the step length increases. The MAPE of the LSTM reaches 0.010881246 at step 15, while the MAPE value of the Seq2Seq model reaches 0.013918079 at step 15. This shows that when dealing with the complex battery capacity demand prediction task, the CNN-BiLSTM-Attention model is more stable and reliable than other models. Multi-step prediction is a more complex task that requires the model to have both good long-term dependence modeling ability and effective feature screening ability. The CNN-BiLSTM-Attention model solves this problem through a triple mechanism: first, CNN extracts local features, which can capture the patterns in different time windows in multi-step predictions; second, BiLSTM improves the model's long-term dependence processing ability through bidirectional modeling; finally, the Attention module can selectively focus on key points in historical data in multi-step predictions, avoiding information redundancy and improving the stability of predictions. These advantages make the CNN-BiLSTM-Attention model perform more superiorly than models such as Seq2Seq, LSTM, RNN, and GRU in multi-step prediction tasks, especially when facing scenarios with large data fluctuations.

[0184] Generally speaking, the CNN-BiLSTM-Attention framework not only performs excellently in short-term prediction, but also maintains better stability in long-term step prediction, and can effectively reduce the problem of error accumulation in multi-step prediction. By comparing the RMSE and MAPE results of different models, the significant performance advantages of this model in multi-step prediction are further verified. Especially in the application of energy management systems that require high accuracy and robustness, it will have great advantages.

[0185] Finally, this application also provides an electric vehicle, which uses the battery health state prediction method based on multi-modal feature fusion and deep learning described in any one of the above method embodiments to predict the battery health state.

[0186] In summary, this application significantly improves the accuracy and stability of SOH prediction through innovative feature pool construction, temperature feature embedding, and multi-step prediction strategies, and has broad adaptability and promotion value under different working conditions.

[0187] It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combinations of these technical features do not conflict, they should be considered as the scope described in this specification. In addition, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, or two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of this application.

[0188] It is easy for those skilled in the art to understand that the above embodiments only represent several implementation manners of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the patent application. It should be pointed out that for those of ordinary skill in the art, without departing from the concept of this application, several modifications and improvements can still be made, and these all belong to the protection scope of this application. Therefore, the protection scope of this application patent shall be subject to the appended claims.

Claims

1. A battery health status prediction method based on multimodal feature fusion and deep learning, characterized in that: The method includes: Obtain charging data for each charging cycle during multi-stage constant current charging of the battery, including current, voltage and temperature data; Determine the health state SOH data of the battery for each charging cycle according to the charging data of each charging cycle of the battery, and filter the health state SOH data of the battery for each charging cycle to obtain the battery health state SOH data; The current and voltage data are smoothed by Savitzky-Golay filter, and the current curve is reconstructed based on the smoothed current data. The voltage data after smoothing is then processed by linear interpolation and KF filter to reconstruct the voltage curve. According to the current curve, voltage curve and temperature data, a comprehensive feature pool is constructed from three aspects: timing characteristics, temperature characteristics, incremental time and differential characteristics, and the constant current value at the maximum charge capacity is directly added to the comprehensive feature pool; The Spearman correlation analysis method is used to quantify the correlation between each feature in the comprehensive feature pool and the battery health status SOH data, so as to select features from the comprehensive feature pool and place them into the prediction feature pool; Perform fully integrated empirical mode decomposition (CEEMADAN) on the battery health status SOH data, and place the decomposed IMF components into the prediction feature pool; Establish a CNN-BILSTM-ATTENTION model, use the features in the prediction feature pool as input, and use the battery health status SOH data as output to train the model; The trained model is used to perform single-step and multi-step predictions on the battery state of health (SOH).

2. The battery health status prediction method based on multimodal feature fusion and deep learning according to claim 1, characterized in that: According to the charging data of each charging cycle of the battery, the health status SOH data of each charging cycle of the battery is determined, including: The state of health (SOH) data of the battery for each charging cycle is calculated based on a variation of the SOC formula, as follows: Where ΔQ is the maximum capacitance difference of the battery in each charging cycle, ΔSOC is the maximum state of charge difference of the battery in each charging cycle, and Q c Represents the rated capacity of the battery.

3. The battery health status prediction method based on multimodal feature fusion and deep learning according to claim 1, characterized in that: The health status SOH data of each charging cycle of the battery is filtered to obtain the battery health status SOH data, including: The health status SOH of each charging cycle of the battery is filtered by a first-order low-pass filter LPF. The formula is as follows: y n =ax n +(1-a)y n-1 In the formula, y n is the filtering value, y n-1 is the previous filtered value, x n For this sampling value, T s is the period, f c is the cutoff frequency, and the value range of a is [0, 1].

4. The battery health status prediction method based on multimodal feature fusion and deep learning according to claim 1, characterized in that: The current and voltage data are smoothed by Savitzky-Golay filter, including: Set the window length and construct the Vandermonde matrix accordingly: Where V is the Vandermonde matrix, k is the window length, k = 2m + 1; The smoothed filtered data is obtained by weighted summation of the pseudo-inverse matrix: A=(V T V) -1 V T Where T is the transpose, A is the pseudo-inverse matrix of V, and a j is the corresponding weight of A, y i Represents the data after smoothing filtering.

5. The battery health status prediction method based on multimodal feature fusion and deep learning according to claim 1, characterized in that: The voltage data after smooth filtering is processed by linear interpolation and KF filter to reconstruct the voltage curve, including: For each charging cycle, each sampling point is corrected starting from the first data point of the smoothed filtered voltage data. The first non-monotonic increasing point is marked as an abnormal point, and the previous point is indexed. The first non-abnormal point after the abnormality is searched backward for interpolation correction, that is: Where V i is the voltage value of the abnormal point, V i-1 Yes V i The voltage value of the previous point, V n Represents the voltage value of the first non-abnormal point after the abnormal point; After linear interpolation, a recursive method is used to perform smoothing operation through the KF filter to reconstruct the voltage curve: P t|t-1 =AP t-1|t-1 From T +Q K t =P t|t-1 H T (HP t|t-1 H T +R) -1 P t|t =(I-K t H)P t|t-1 In the formula, is the predicted state at time t, is the updated state at time t, P t|t-1 is the predicted state covariance matrix, P t|t is the updated state covariance matrix, K t represents the Kalman gain, Q and R represent the process noise and observation noise covariance matrices respectively, A is the state transfer matrix, which describes the transition of the system state over time, H is the observation matrix, which maps the state space to the observation space, and z t are the actual values ​​and I is the identity matrix used to adjust the covariance matrix.

6. The battery health status prediction method based on multimodal feature fusion and deep learning according to claim 1, characterized in that: The timing characteristics include constant current charging time CCCT, isobaric rise time TEVR, isothermal rise time VERT, charging voltage slope SCV, area under constant current charging voltage ACCCV, area under constant current charging current ACCCC, and area under constant current charging temperature ACCCT; The temperature characteristics include the highest constant current charging temperature HCCCT, the average constant current charging temperature MCCCT, and the lowest constant current charging temperature LCCCT; The incremental time and differential characteristics include incremental capacity peak ICP, incremental capacity peak position ICPL, incremental capacity area ICA, incremental capacity slope ICS, differential voltage area DVA, differential voltage slope DVS, differential voltage valley DVV, differential voltage valley position DVVL, temperature difference area DTA, temperature difference slope DTS, temperature difference peak DTP, and temperature difference peak position DTPL.

7. The battery health status prediction method based on multimodal feature fusion and deep learning according to claim 1, characterized in that: Features are selected from the comprehensive feature pool and placed into the prediction feature pool, including: Select features whose absolute correlation values ​​are greater than a first correlation threshold from the comprehensive feature pool and record them as a candidate feature set; The Spearman correlation analysis method is used to calculate the correlation score between each pair of features in the candidate feature set, which is recorded as the correlation matrix R; A second correlation threshold is set, and the feature with the highest correlation with the battery health state SOH data is selected from the candidate feature set as the baseline feature x_base; the correlation between other candidate features and the baseline feature x_base is checked in turn: if the correlation score is less than the second correlation threshold, the feature is retained and removed from the candidate feature set; if the correlation score is greater than or equal to the second correlation threshold, it is removed from the candidate feature set; Select the next feature with the highest correlation with the battery health status SOH data from the remaining candidate feature set as the new baseline feature x_base, and repeat the previous step until the candidate feature set is empty.

8. The battery health status prediction method based on multimodal feature fusion and deep learning according to claim 1, characterized in that: The CNN-BILSTM-ATTENTION model includes a feature extraction and fusion module, a temporal dependency extraction and prediction module, and a bidirectional attention mechanism module; among them: The feature extraction and fusion module uses a convolution kernel of size 1 to capture local time series information and performs convolution operations on input features of multiple time steps to achieve preliminary feature extraction and fusion; The temporal dependency extraction and prediction module simultaneously captures the bidirectional dependencies of temporal data through two LSTM layers, forward and reverse; The bidirectional attention mechanism module weights the output features to adaptively highlight the feature information that is highly relevant to the battery health status. This is used to strengthen feature selection and improve the model's key feature recognition capabilities in multi-step prediction tasks. The model's final output generates a prediction result through a fully connected layer. A grid search algorithm is used to optimize the hyperparameter settings of the model.

9. The battery health status prediction method based on multimodal feature fusion and deep learning according to claim 1, characterized in that: Single-step and multi-step forecasts include: X t ={soh t-n+1 ,soh t-n+2 ,··,soh t-1 ,soh t } In the formula, SOH t represents the health status of the battery at time t; X t Represents a data segment, indicating the battery health status SOH data at time t and the previous n times; Indicates that the data segment X is used t Predict the predicted value for the next time step t+1, The model is based on the data segment X t Predict the prediction results for multiple time steps in the future from t+1 to t+m.

10. An electric vehicle, characterized in that: The electric vehicle adopts the battery health state prediction method based on multimodal feature fusion and deep learning as described in any one of claims 1 to 9 to predict the battery health state.

Images