Method and system for predicting health degree of vehicle-mounted battery of electric vehicle based on neural network

Abstract

The invention discloses an electric vehicle vehicle-mounted battery health degree prediction method and system based on a neural network, and relates to the technical field of battery health degree prediction, and the method comprises the steps: collecting battery multi-source heterogeneous data, extracting key health factors through preprocessing, room temperature correction and staged feature engineering, and introducing a time window embedding strategy to construct a time sequence sample; a multi-module U-BiLSTM hybrid network is adopted as a core prediction model, and Bayesian optimization and an Adam optimizer are combined to complete hyper-parameter optimization and model training; and a prediction result is optimized through digital twinning correction, a physical compensation mechanism and residual service life label normalization processing. The method solves the problems that a traditional method is insufficient in deep feature extraction, low in time sequence information utilization rate and poor in prediction precision under complex working conditions, the aging state and the remaining service life of the battery can be accurately reflected, and reliable technical support is provided for safety management, operation and maintenance optimization and service life evaluation of the battery of the electric vehicle.

Description

Technical Field

[0001] This invention relates to the field of battery health prediction technology, and more specifically to a method and system for predicting the health of on-board batteries in electric vehicles based on neural networks. Background Technology

[0002] With the rapid development of the electric vehicle industry, the state of health (SOH) and remaining useful life (RUL) of on-board batteries are directly related to vehicle driving safety, range, and maintenance costs. Accurately predicting battery health has become a core requirement of battery management systems.

[0003] Existing technologies include battery health prediction methods that combine data acquisition with neural networks, but these methods mostly suffer from the following drawbacks: First, they rely solely on single-dimensional data or shallow feature extraction, making it difficult to capture the deep nonlinear features in the battery aging process; second, they have low utilization of historical information from long-term series data and do not fully consider the bidirectional dependencies of time-series data; third, they do not introduce effective time-series sample construction strategies, and the data preprocessing process is not perfect; fourth, the model training and optimization methods are simplistic, and the efficiency of hyperparameter optimization and gradient descent is insufficient; and fifth, the prediction results lack a direct correlation with RUL (Remaining Life of the Battery), making it difficult to meet the needs of accurate assessment of battery remaining life in actual operation and maintenance.

[0004] Therefore, how to propose a method and system for predicting the health of electric vehicle batteries based on neural networks, and overcome the shortcomings of existing technologies, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for predicting the health of electric vehicle on-board batteries based on neural networks. Through a multi-module U-BiLSTM network, a time window embedding strategy, and a multi-dimensional optimization mechanism, the accuracy, efficiency, and adaptability of battery health and remaining lifespan prediction are significantly improved. To achieve the above objectives, the present invention adopts the following technical solution: A neural network-based method for predicting the health of electric vehicle onboard batteries includes: Collect multi-source data related to the on-board battery of electric vehicles, and preprocess and correct the data for room temperature. The preprocessed and corrected multi-source data were standardized, and time-series samples were constructed using a time window embedding strategy. The dataset was then divided into stages based on outdoor temperature, key health factors were extracted and dimensionality was reduced to obtain a low-dimensional feature set. A Bayesian-optimized multi-module U-BiLSTM hybrid model was constructed, with the percentage of remaining battery life to the total battery lifespan as the label. The model was trained and optimized using a dataset divided into stages to obtain the final battery health prediction model. The low-dimensional feature set is input into the battery health prediction model to make preliminary predictions of SOH and RUL, and the prediction results are optimized by combining digital twin correction and physical compensation mechanisms.

[0006] Optionally, the multi-source data includes battery electrochemical parameters, environmental parameters, operating status parameters, indoor temperature data, and vibration acceleration signals. The battery electrochemical parameters include the voltage, current, internal resistance, and cycle count of individual battery cells and modules. The environmental parameters include ambient temperature and humidity, vibration intensity, and solar radiation intensity. The operating status parameters include charge / discharge power, driving speed, and driving mileage. The indoor temperature data is temperature data collected from multiple points inside the battery compartment. The vibration acceleration signals are collected by acceleration sensors installed in the horizontal and vertical directions of the battery pack.

[0007] Optionally, the preprocessing includes: missing value supplementation, outlier removal, and redundant data deletion; the missing value supplementation adopts the average of the two ends, linear interpolation in the middle, or polynomial fitting; the outlier removal adopts the Lada method; the room temperature correction includes taking a weighted average of the collected N battery compartment temperatures hourly to obtain hourly indoor temperature feature values, and then correcting the relevant parameters based on the hourly indoor temperature feature values.

[0008] Optionally, the hourly indoor temperature characteristic values ​​include: ; in, For the battery compartment Temperature characteristic value at any given time; for Time of the first Temperature of each sampling point inside the battery compartment; For the first The weights of each data collection point are assigned using the analytic hierarchy process (AHP).

[0009] Optionally, the correction of relevant parameters based on hourly indoor temperature characteristic values ​​includes: ; in, for The parameter values ​​to be corrected are collected at all times; These are the parameter values ​​after temperature correction; for outdoor dry-bulb temperature at all times; The optimal operating temperature threshold for the battery; To ensure a safety margin, adjustments should be made based on actual operating conditions.

[0010] Optionally, the step-by-step division of the dataset based on outdoor temperature includes: The dataset for each battery usage cycle is divided into two stages: early and late stage and middle stage, based on outdoor temperature; the key health factors include the time of occurrence of the lowest discharge voltage, average discharge voltage, temperature sensitivity index, corrected charge and discharge power characteristic value, and vibration acceleration signal characteristics; the dimensionality reduction process adopts the PCA algorithm.

[0011] Optionally, the multi-module U-BiLSTM hybrid model includes an encoder, decoder, and regressor connected in sequence; The encoder includes a two-dimensional convolutional neural network, residual blocks, and a two-dimensional deep convolutional separable neural network; The decoder includes deconvolutional layers, pooling layers, and batch normalization layers; The regressor comprises a BiLSTM network and a fully connected layer.

[0012] Optionally, the training and optimization of the model includes: The K-fold cross-validation method was used to divide the phased dataset into training and test sets. The data in the early and late stages were divided into part a according to time series, and the data in the middle stage were divided into part b according to time series. The learning rate, kernel size, number of hidden layer nodes, and time window step size hyperparameters were optimized using the Bayesian optimization algorithm. The model training process is optimized by using mean squared error as the loss function and combining it with the Adam optimizer. The normalization formula for the battery RUL label is: ; in, FPT is the time when the battery begins to operate, and FPT is the time when the battery transitions from a healthy state to a degraded state. T represents the moment when the battery completely fails, and T represents the current operating time. The label value is the normalized value.

[0013] Optionally, the update step size of the Adam optimizer includes: ; ; ; ; ; in, For the first moment estimate of the current time step, For the second moment estimate of the current time step, For gradient, , These are the correction values ​​for the estimates of the first and second moments, respectively. The initial learning rate, , , It is a constant. , These are the dynamic constraint values ​​for the current and next time steps, respectively.

[0014] Optionally, a neural network-based on-board battery health prediction system for electric vehicles includes: Data acquisition module: used to collect multi-source data related to the on-board battery of electric vehicles, and to preprocess and correct the data for room temperature. Data processing module: It is used to standardize the preprocessed and corrected multi-source data, construct time series samples using a time window embedding strategy, divide the dataset into stages according to outdoor temperature, extract key health factors and reduce dimensionality to obtain a low-dimensional feature set. Model training module: Used to build a Bayesian-optimized multi-module U-BiLSTM hybrid model, using the percentage of battery remaining lifespan to the total lifespan as the label, and using the phased dataset to train and optimize the model to obtain the final battery health prediction model. Prediction correction module: It is used to input the low-dimensional feature set into the battery health prediction model to make preliminary predictions of SOH and RUL, and optimize the prediction results by combining digital twin correction and physical compensation mechanisms.

[0015] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for predicting the health of electric vehicle on-board batteries based on neural networks, which has the following beneficial effects: This invention proposes a neural network-based method for predicting the health of electric vehicle on-board batteries, comprising: collecting multi-source data related to electric vehicle on-board batteries and preprocessing and correcting the data for room temperature; standardizing the preprocessed and corrected multi-source data, constructing time-series samples using a time window embedding strategy, dividing the dataset into stages based on outdoor temperature, extracting key health factors and reducing dimensionality to obtain a low-dimensional feature set; constructing a Bayesian-optimized multi-module U-BiLSTM hybrid model, using the percentage of remaining battery life to the total lifespan as a label, training and optimizing the model using the staged dataset to obtain the final battery health prediction model; inputting the low-dimensional feature set into the battery health prediction model for preliminary prediction of SOH and RUL, and optimizing the prediction results by combining digital twin correction and physical compensation mechanisms.

[0016] This invention (1) improves data quality by integrating multi-dimensional data of electrochemistry, environment, operating status, temperature and vibration, combined with room temperature correction and standardization; and constructs time-series samples by embedding time windows to fully preserve the temporal correlation of data. (2) The multi-module U-BiLSTM network extracts deep nonlinear features through the residual blocks of the encoder and the deep convolutional structure, optimizes the feature expression of the decoder, and captures bidirectional temporal dependencies, solving the problems of insufficient feature extraction and low utilization of temporal information in traditional models. (3) Combining Bayesian hyperparameter optimization and Adam gradient descent optimization, it balances optimization efficiency and convergence stability; staged data processing and K-fold cross-validation improve the model's adaptability to different temperature conditions. (4) Through digital twin correction, physical compensation mechanism and RUL label normalization and inverse transformation, it achieves accurate prediction of SOH and actual remaining service life, meeting the actual needs of operation and maintenance. (5) Through the collaborative mode of real-time inference at the edge and iterative update in the cloud, it ensures that the model can continuously adapt to the battery aging process and changes in operating conditions, and maintain high prediction performance in the long term. This invention solves the problems of low prediction accuracy, poor adaptability, and insufficient practicality of traditional methods under complex working conditions. It can provide reliable technical support for electric vehicle battery safety management, operation and maintenance optimization, and life assessment, and has broad application prospects. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of a neural network-based method for predicting the health of an electric vehicle's onboard battery, provided by the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention discloses a method for predicting the health of an electric vehicle's onboard battery based on a neural network, such as... Figure 1 As shown, it includes: Collect multi-source data related to the on-board battery of electric vehicles, and preprocess and correct the data for room temperature. The preprocessed and corrected multi-source data were standardized, and time-series samples were constructed using a time window embedding strategy. The dataset was then divided into stages based on outdoor temperature, key health factors were extracted and dimensionality was reduced to obtain a low-dimensional feature set. A Bayesian-optimized multi-module U-BiLSTM hybrid model was constructed, with the percentage of remaining battery life to the total battery lifespan as the label. The model was trained and optimized using a dataset divided into stages to obtain the final battery health prediction model. The low-dimensional feature set is input into the battery health prediction model to make preliminary predictions of SOH and RUL, and the prediction results are optimized by combining digital twin correction and physical compensation mechanisms.

[0021] Furthermore, the multi-source data includes battery electrochemical parameters, environmental parameters, operating status parameters, indoor temperature data, and vibration acceleration signals. The battery electrochemical parameters include the voltage, current, internal resistance, and cycle count of individual battery cells and modules. The environmental parameters include ambient temperature and humidity, vibration intensity, and solar radiation intensity. The operating status parameters include charge / discharge power, driving speed, and driving mileage. The indoor temperature data is temperature data collected from multiple points within the battery compartment. The vibration acceleration signals are collected by acceleration sensors installed in the horizontal and vertical directions of the battery pack.

[0022] Furthermore, the preprocessing includes: missing value supplementation, outlier removal, and redundant data deletion; the missing value supplementation adopts the average of the two ends, linear interpolation in the middle, or polynomial fitting; the outlier removal adopts the Lada method; the room temperature correction includes taking a weighted average of the collected N battery compartment temperatures hourly to obtain hourly indoor temperature feature values, and then correcting the relevant parameters based on the hourly indoor temperature feature values.

[0023] Furthermore, the hourly indoor temperature characteristic values ​​include: ; in, For the battery compartment Temperature characteristic value at any given time; for Time of the first Temperature of each sampling point inside the battery compartment; For the first The weights of each data collection point are assigned using the Analytic Hierarchy Process (AHP).

[0024] Furthermore, the correction of relevant parameters based on hourly indoor temperature characteristic values ​​includes: ; in, for The parameter values ​​to be corrected are collected at all times; These are the parameter values ​​after temperature correction; for outdoor dry-bulb temperature at all times; The optimal operating temperature threshold for the battery; To ensure a safety margin, adjustments should be made based on actual operating conditions.

[0025] Furthermore, the formula for the standardization process is as follows: ,in, The original data after preprocessing and correction. The mean of the original data. The standard deviation of the original data. The data is standardized; the time-series samples constructed using the time window embedding strategy are represented as follows: ; in, For the standardized data at time t, The data at time t-S+1 is standardized, where S is the total sliding step size of the time window.

[0026] Furthermore, the step-by-step division of the dataset based on outdoor temperature includes: The dataset for each battery usage cycle is divided into two stages: early and late stage and middle stage, based on outdoor temperature; the key health factors include the time of occurrence of the lowest discharge voltage, average discharge voltage, temperature sensitivity index, corrected charge and discharge power characteristic value, and vibration acceleration signal characteristics; the dimensionality reduction process adopts the PCA algorithm.

[0027] Furthermore, the multi-module U-BiLSTM hybrid model includes an encoder, decoder, and regressor connected in sequence; The encoder includes a two-dimensional convolutional neural network (shallow feature extraction), residual blocks, and a two-dimensional deep convolutional separable neural network (deep feature extraction). The decoder includes deconvolutional layers, pooling layers, and batch normalization layers; The regressor comprises a BiLSTM network and a fully connected layer.

[0028] Furthermore, the training and optimization of the model includes: The K-fold cross-validation method was used to divide the phased dataset into training and test sets. The data in the early and late stages were divided into 6 parts according to time series, and the data in the middle stage were divided into 5 parts according to time series. The learning rate, kernel size, number of hidden layer nodes, and time window step size hyperparameters were optimized using the Bayesian optimization algorithm. The model training process is optimized using mean squared error (MSE) as the loss function and combined with the Adam optimizer. The normalization formula for the battery RUL label is: ; in, FPT is the time when the battery begins to operate, and FPT is the time when the battery transitions from a healthy state to a degraded state. T represents the moment when the battery completely fails, and T represents the current operating time. The label value is the normalized value.

[0029] Furthermore, the update step size of the Adam optimizer includes: ; ; ; ; ; in, For the first moment estimate of the current time step, For the second moment estimate of the current time step, For gradient, , These are the correction values ​​for the estimates of the first and second moments, respectively. The initial learning rate, , , It is a constant. , These are the dynamic constraint values ​​for the current and next time steps, respectively.

[0030] Furthermore, the digital twin correction simulates the electrochemical-thermal coupling process by constructing a battery digital twin model, and performs parameter calibration on the preliminary prediction results by combining actual operating data; the physical compensation mechanism is used to correct the single-pack or dual-pack evaluation matrix to avoid upward fluctuations in SOH.

[0031] Furthermore, it also includes: outputting prediction results through real-time inference at the edge, and aggregating global data in the cloud to iteratively update the model, thereby achieving dynamic adaptive prediction.

[0032] Furthermore, the edge device performs real-time inference every 5 minutes, and the cloud gathers global data weekly to update the training model, achieving incremental updates of the edge device model parameters through OTA (Over-The-Air) technology.

[0033] In a specific implementation, a neural network-based method for predicting the health of an electric vehicle's onboard battery includes five core steps: data acquisition and preprocessing correction, time-series sample construction and feature engineering, model training and optimization, prediction result correction, and dynamic updating, as detailed below: Step 1: Multi-source data acquisition and preprocessing correction (1) Collect multi-source heterogeneous data of electric vehicle on-board batteries. Collect electrochemical parameters such as voltage, current, internal resistance, and cycle number of battery cells and modules in real time through BMS. Collect environmental parameters such as ambient temperature and humidity, vibration intensity, and solar radiation intensity, as well as operating status parameters such as charging and discharging power, driving speed, and driving mileage through IoT controller. Collect vibration acceleration signals through acceleration sensors installed in the horizontal and vertical directions of the battery pack to capture the aging effects caused by battery structure vibration. At the same time, collect indoor temperature and outdoor dry bulb temperature data at multiple points in the battery compartment. All data are transmitted in real time through CAN bus.

[0034] (2) Preprocess the collected data: use the average of the two ends, linear interpolation in the middle or polynomial fitting to supplement the missing data; use the Lada method to remove abnormal data to avoid the interference of extreme values ​​on model training; delete redundant data to simplify the computational complexity.

[0035] (3) Room temperature correction based on indoor temperature data: The temperature of N data collection points in the battery compartment is weighted and averaged hourly to obtain hourly indoor temperature characteristic values. The weights are assigned using the analytic hierarchy process (AHP), and the calculation formula is as follows: ; in, Let J be the temperature characteristic value inside the battery compartment at time j; Let be the temperature of the i-th sampling point at time j; Let be the weight of the i-th collection point.

[0036] (4) The obtained indoor temperature characteristic values ​​are used to correct key operating parameters such as charging and discharging power. The correction formula is as follows: ; in, The original parameter values ​​collected at time j; These are the corrected parameter values; Let j be the outdoor dry-bulb temperature. The appropriate operating temperature threshold for the battery (e.g., 25°C); For safety, adjust to between 0.9 and 1.1 based on actual operating conditions.

[0037] Step 2: Construction of time-series samples and phased feature engineering

[0038] (1) Standardize all preprocessed and corrected data to eliminate dimensional differences and data distribution bias. The standardization formula is: ; in, The original data are preprocessed and corrected, where μ is the mean of the original data and σ is the standard deviation of the original data. This is the standardized data.

[0039] (2) A time-series sample is constructed using a time window embedding strategy, transforming the standardized continuous time-series data into input samples that the model can process. The time-series sample is represented as follows: ; in, For the standardized data at time t, The data at time t-S+1 is standardized, and S is the total sliding step size of the time window, which is set to 10-30 according to the data acquisition frequency and battery aging cycle.

[0040] (3) Based on the outdoor temperature, the time series sample dataset of each battery usage cycle (e.g., one year) is divided into two stages: the beginning and end stage and the middle stage. The beginning and end stage is the period when the outdoor temperature is low (below 5℃) or high (above 35℃), and the middle stage is the period when the outdoor temperature is relatively stable (5℃-35℃).

[0041] (4) Extract key health factors from the phased dataset, including the time of the lowest discharge voltage, average discharge voltage, temperature sensitivity index, corrected charge and discharge power characteristic values, and time-domain statistical characteristics (mean, variance, peak value) of vibration acceleration signal. Use PCA algorithm to reduce the dimensionality of the extracted key health factors, remove redundant features, and obtain a low-dimensional feature set to improve model training efficiency and generalization ability.

[0042] Step 3: Training and Optimization of Multi-Module U-BiLSTM Hybrid Model

[0043] (1) Construct a Bayesian-optimized multi-module U-BiLSTM hybrid model, which combines the deep feature extraction capability of the multi-module U network with the bidirectional temporal dependency capture advantage of the BiLSTM network. The specific structure is as follows: 1) Encoder: This consists of a sequentially connected 2D convolutional neural network, residual blocks, and a 2D deep convolutional separable neural network. The 2D convolutional neural network is used to extract shallow spatial features of the data; the residual blocks achieve multi-scale feature fusion and reduce the risk of gradient vanishing; and the 2D deep convolutional separable neural network is used to efficiently extract deep nonlinear features.

[0044] 2) Decoder: This consists of sequentially connected deconvolutional layers, pooling layers, and batch normalization layers. The deconvolutional layers restore the feature dimensions, the pooling layers enhance the feature aggregation capability, and the batch normalization layers accelerate model convergence and reduce the risk of overfitting.

[0045] 3) Regressor: This consists of a sequentially connected BiLSTM network and a fully connected layer. The BiLSTM network captures the forward and backward temporal dependencies of the data, improving the utilization of historical information in long-term series; the fully connected layer maps features to SOH and RUL prediction values.

[0046] (2) The K-fold cross-validation method is used to divide the dataset into training and test sets after the phased process: the data in the early and late stages are divided into 6 parts according to time (3 parts in the early stage and 3 parts in the late stage), and the data in the middle stage are divided into 5 parts according to time. Each part is used as the test set in turn, and the rest are used as the training set to ensure the generalization ability of the model under different temperature conditions.

[0047] (3) The percentage of remaining battery life (RUL) to the total battery life is used as the label for normalization. The label conversion formula is: ; in, The FPT is the time when the battery starts operating, and the FPT is the time when the battery transitions from a healthy state to a degraded state (determined by the battery capacity decaying to 80% of its rated capacity). The point at which the battery completely fails (determined by the battery capacity decaying to 60% of its rated capacity), where T is the current operating time. The label value is the normalized value.

[0048] (4) Input the training set into the hybrid model for training: The encoder extracts shallow and deep features from the low-dimensional feature set. After the decoder optimizes the feature representation, the features are input into a BiLSTM network to capture bidirectional temporal dependencies. Finally, the predicted value is output through a fully connected layer. A Bayesian optimization algorithm is used to optimize hyperparameters such as learning rate, kernel size, number of hidden layer nodes, and temporal window stride. The mean squared error (MSE) is used as the loss function, and the loss function formula is: ; Where L is the loss function, This is the measured value (SOH or RUL normalized value) of the i-th sample. Let m be the predicted value of the i-th sample and m be the number of samples. By combining the Adam optimizer to optimize the model training process, the step size is dynamically adjusted and updated by comprehensively considering the first moment estimation and second moment estimation of the gradient, thereby improving the convergence speed and stability of the model.

[0049] (5) Input the test set into the trained model for performance verification. If the mean square error (MSE) and mean absolute percentage error (MAPE) of each test set meet the preset error rate requirements (MSE≤0.01, MAPE≤5%), then it is determined as the final battery health prediction model.

[0050] Step 4: Correction of Prediction Results

[0051] (1) Input the low-dimensional feature set obtained in step 2 into the final battery health prediction model to obtain the preliminary prediction results of SOH and RUL.

[0052] (2) Construct a digital twin model of the battery to simulate the electrochemical-thermal coupling process, structural stress change and aging mechanism of the battery. Combine the actual operating data of the battery (voltage, current, temperature and vibration signal) to perform parameter calibration on the preliminary prediction results and correct the deviation between the model and the actual scenario.

[0053] (3) A physical compensation mechanism is introduced to correct the evaluation matrix for single or dual battery packs, preventing upward fluctuations in the SOH prediction value; simultaneously, combined with the inverse transformation of the RUL normalized label, the predicted normalized RUL value is converted into the actual remaining usage time (such as days, number of cycles), improving the practicality of the prediction results. Specifically: (I) First, based on the physical characteristics of battery cells and modules, an evaluation matrix is ​​constructed, containing four dimensions: capacity, internal resistance, cycle life, and temperature. The matrix rows correspond to the battery pack / dual-pack number (e.g., when a vehicle has four battery packs, the row indices are 1-4), and the columns correspond to core health indicators, specifically including: Actual capacity: The current usable capacity of the battery obtained through charge-discharge cycle testing (unit: Ah). DC internal resistance: The average internal resistance of a single battery cell measured by the pulse discharge method (unit: mΩ). Cyclic decay coefficient: The capacity decay rate at the current number of cycles (calculated as: (rated capacity - actual capacity) / rated capacity × 100%). Temperature influence coefficient: Cumulative temperature loss calculated based on historical temperature data (unit: °C) h), reflecting the accelerating effect of high or low temperature environments on battery aging.

[0054] The initial values ​​of the evaluation matrix are filled with battery factory parameters and real-time data collected by the BMS. For example, the initial evaluation matrix of a single battery pack, as shown in Table 1, can be represented as follows: Table 1

[0055] (ii) Physical rules for compensation and correction

[0056] Based on the physical mechanism of battery aging, three types of compensation rules are set to correct the abnormal parameters in the evaluation matrix, thereby constraining the SOH prediction value: (1) Capacity-internal resistance co-correction Battery capacity and internal resistance are strongly negatively correlated (internal resistance inevitably increases as capacity decays). If contradictory data of "increasing capacity and decreasing internal resistance" appears in the evaluation matrix, a collaborative correction is triggered. For example, if the predicted actual capacity of a single pack increases by 2% compared to the previous cycle, but the internal resistance decreases by 5% simultaneously, it is judged as data anomaly. The "internal resistance priority" principle is adopted: using the measured internal resistance as a benchmark, a reasonable capacity value is calculated by back-calculating the capacity-internal resistance fitting formula (based on battery factory test data, such as: actual capacity = rated capacity - 0.02 × (current internal resistance - initial internal resistance)) and replacing the abnormal capacity data in the evaluation matrix.

[0057] (2) Correction of loop number constraint

[0058] Battery SOH decreases monotonically with the number of cycles. A lower limit threshold for SOH is set for the number of cycles (e.g., 80% for 1000 cycles, 65% for 2000 cycles). If the predicted SOH value of a single pack in the evaluation matrix is ​​more than 10% higher than the lower limit threshold corresponding to the current number of cycles (e.g., 92% for 1000 cycles), cycle constraint correction is triggered: the theoretical SOH value is calculated using the cycle decay model (e.g., SOH = 100% - 0.02 × number of cycles / 100), and the predicted SOH value in the evaluation matrix is ​​corrected to "theoretical value + 3% error tolerance" (3% is the average deviation between measured data and the theoretical model), ensuring that the corrected value conforms to the cycle aging law.

[0059] (3) Temperature loss compensation correction

[0060] High temperatures accelerate the decomposition of the SEI film and the shedding of active materials in batteries, while low temperatures lead to lithium dendrite precipitation; both exacerbate battery aging. If the cumulative value of the temperature influence coefficient in the evaluation matrix exceeds a preset threshold (e.g., 1500℃), the battery will be affected. h), but the SOH prediction value does not show a corresponding decrease (e.g., if the temperature influence coefficient increases by 20%, the SOH only decreases by 1%), triggering temperature compensation correction: introducing a temperature loss factor (calculation formula: temperature loss factor = temperature influence coefficient / 1000 × 0.05), correcting the SOH prediction value to "original prediction value - temperature loss factor", for example, if the temperature influence coefficient is 1800℃. At time h, the temperature loss factor = 1800 / 1000 × 0.05 = 0.09, and the SOH prediction value is revised from 90% to 89.1% to match the additional aging loss caused by temperature.

[0061] (III) Revised SOH Calculation Logic

[0062] After revising the evaluation matrix, the final SOH value is calculated using a weighted summation method. The weights are based on the degree of influence of each indicator on battery health (assigned using the analytic hierarchy process, AHP). The formula is as follows: ; in, =0.4 (Volume is a core health indicator with the highest weight). =0.3 (Internal resistance is an important auxiliary indicator). =0.2、 =0.1, and + + + =1. The SOH value calculated by this formula integrates multi-dimensional physical parameters and avoids upward fluctuations through compensation rules, ensuring consistency with the actual aging state of the battery.

[0063] Step 5: Dynamic Adaptive Update

[0064] (1) Deploy the final battery health prediction model at the edge, perform real-time inference every 5 minutes, and output the current battery SOH value, actual remaining RUL time and health risk level to provide real-time decision support for the vehicle battery management system.

[0065] (2) The cloud aggregates global operating data of all connected vehicles, including multi-source battery data under multiple operating conditions, and information on the deviation between predicted results and actual aging data. The model is iterated and trained once a week to update the model parameters to adapt to battery aging patterns, changes in operating conditions, and individual differences of different vehicle models. The incremental update of the model parameters at the edge is achieved through OTA (Over-The-Air) technology to ensure that the model maintains high prediction accuracy in the long term and achieves dynamic adaptive prediction.

[0066] In a specific implementation, an electric vehicle on-board battery health prediction system based on neural networks includes: Data acquisition module: used to collect multi-source data related to the on-board battery of electric vehicles, and to preprocess and correct the data for room temperature. Data processing module: It is used to standardize the preprocessed and corrected multi-source data, construct time series samples using a time window embedding strategy, divide the dataset into stages according to outdoor temperature, extract key health factors and reduce dimensionality to obtain a low-dimensional feature set. Model training module: Used to build a Bayesian-optimized multi-module U-BiLSTM hybrid model, using the percentage of battery remaining lifespan to the total lifespan as the label, and using the phased dataset to train and optimize the model to obtain the final battery health prediction model. Prediction correction module: It is used to input the low-dimensional feature set into the battery health prediction model to make preliminary predictions of SOH and RUL, and optimize the prediction results by combining digital twin correction and physical compensation mechanisms.

[0067] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0068] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A neural network-based method for predicting the state of health of an electric vehicle's on-board battery, characterized in that, The application relates to a battery health degree prediction method and device. The application comprises the following steps: Collecting multi-source data related to an electric vehicle battery on board, and pre-processing and room temperature correcting the data; Standardizing the pre-processed and corrected multi-source data, adopting a time window embedding strategy to construct a time sequence sample, then dividing the data set into stages according to outdoor temperature, extracting key health factors and performing dimension reduction processing to obtain a low-dimensional feature set; Constructing a multi-module U-BiLSTM hybrid model based on Bayesian optimization, taking the percentage of battery remaining service life in the whole life cycle as a label, training and optimizing the model by using the data set divided into stages, and obtaining a final battery health degree prediction model; 2.The neural network-based battery state of health prediction method for electric vehicles of claim 1, wherein, Inputting the low-dimensional feature set into the battery health degree prediction model to preliminarily predict SOH and RUL, and optimizing the prediction result by combining a digital twin correction and a physical compensation mechanism. 3.The neural network-based electric vehicle on-board battery state of health prediction method of claim 1, wherein, The multi-source data comprises battery electrochemical parameters, environmental parameters, running state parameters, indoor temperature data and vibration acceleration signals, the battery electrochemical parameters comprise voltage, current, internal resistance and cycle number of battery monomers and modules, the environmental parameters comprise environmental temperature and humidity, vibration intensity and solar radiation intensity, the running state parameters comprise charging and discharging power, driving speed and driving mileage, the indoor temperature data are temperature data collected at multiple points in the battery cabin, and the vibration acceleration signals are collected by acceleration sensors installed in the horizontal and vertical directions of the battery pack. 4.The neural network-based battery state of health prediction method for electric vehicles of claim 3, wherein, The pre-processing comprises missing value supplement, abnormal value elimination and redundant data deletion, the missing value supplement adopts two-end average value, middle linear interpolation or polynomial fitting mode, the abnormal value elimination adopts a Ralda method, and the room temperature correction comprises weighting and averaging the indoor temperature of N battery cabins collected at different times to obtain a time-by-time indoor temperature characteristic value, and then correcting related parameters based on the time-by-time indoor temperature characteristic value. ； wherein, is the temperature of the battery compartment at the time instant t; is the temperature characteristic value at the time instant t; is the temperature of the battery compartment at the time instant t; is the temperature of the i-th battery compartment at the time instant t; is the temperature of the i-th battery compartment at the time instant t; is the temperature of the i-th battery compartment at the time instant t; is the weight of the i-th acquisition point.

5. The neural network-based on-board battery state of health prediction method for electric vehicles of claim 3, wherein, The time-by-time indoor temperature characteristic value comprises: ； wherein, is the parameter value to be corrected at the moment; is the parameter value after temperature correction; is the outdoor dry-bulb temperature at the moment; is the threshold value of the appropriate working temperature of the battery; is the safety factor.

6. The neural network-based electric vehicle on-board battery state of health prediction method of claim 1, wherein, The correction of related parameters based on the time-by-time indoor temperature characteristic value comprises: The division of the data set into stages according to outdoor temperature comprises:

7. The neural network-based electric vehicle on-board battery state of health prediction method of claim 1, wherein, Dividing the data set of each battery use cycle into two stages of initial and final stages and a middle stage according to outdoor temperature, and the key health factors comprise discharge voltage lowest point appearance time, average discharge voltage, temperature sensitivity index, corrected charging and discharging power characteristic value and vibration acceleration signal characteristic. The multi-module U-BiLSTM hybrid model comprises a sequentially connected encoder, a decoder and a regressor; The encoder comprises a two-dimensional convolutional neural network, a residual block and a two-dimensional deep convolutional separable neural network; The decoder comprises an inverse convolution layer, a pooling layer and a batch normalization layer; 8.The neural network-based battery state of health prediction method for electric vehicles of claim 1, wherein, The regressor comprises a BiLSTM network and a full connection layer. The training and optimization of the model comprises: Dividing the data set into training set and test set by adopting a K-fold cross-validation method, dividing the initial and final stage data into a part a and the middle stage data into a part b according to time sequence; Optimizing learning rate, convolution kernel size, hidden layer node number and time window step size hyperparameters by a Bayesian optimization algorithm; 9.The neural network-based battery state of health prediction method for electric vehicles of claim 8, wherein, Taking mean square error as a loss function, and optimizing the model training process by combining an Adam optimizer. ； ； ； ； ； wherein, is a first moment estimate for the current time step, is a second moment estimate for the current time step, is a gradient, , are correction values for the first and second moment estimates, respectively, is an initial learning rate, , , is a constant, , are dynamic constraint values for the current and next time steps, respectively. 10.A neural network-based on-board battery state-of-health prediction system for an electric vehicle, the system comprising: The update step of the Adam optimizer comprises: The data acquisition module is used for collecting multi-source data related to the battery on board the electric vehicle and pre-processing and room temperature correcting the data; The data processing module is used for standardizing the pre-processed and corrected multi-source data, constructing time sequence samples by adopting a time window embedding strategy, dividing the data set in stages according to the outdoor temperature, extracting key health factors and performing dimension reduction processing, and obtaining a low-dimensional feature set; The model training module is used for constructing a multi-module U-BiLSTM hybrid model optimized by Bayes, taking the percentage of the remaining service life of the battery in the whole life cycle as a label, training and optimizing the model by using the data set divided in stages, and obtaining a final battery health degree prediction model; The prediction correction module is used for inputting the low-dimensional feature set into the battery health degree prediction model to preliminarily predict SOH and RUL, and optimizing the prediction result by combining a digital twin correction mechanism and a physical compensation mechanism.

Images