Power battery pack state evaluation method and system

Abstract

The invention relates to the technical field of power battery packs, and discloses a power battery pack state evaluation method and system, and the method comprises the steps: collecting an operation characteristic data set of a power battery pack under various working conditions, and constructing a hybrid evaluation model; performing parameter configuration on the double-extended Kalman filter to obtain an improved filter; acquiring real-time voltage, current and temperature data, and inputting the real-time voltage, current and temperature data into the improved filter for state calculation to obtain a charge state value and an internal parameter value; inputting the charge state value and the internal parameter value to perform multi-branch adversarial neural network analysis to obtain a fault mode type and a fault severity; according to the method, the multi-factor health index is calculated, particle swarm optimization fitting is carried out on the degradation curve of the power battery pack, the residual service life prediction value is obtained, more accurate residual service life prediction is further achieved, the precision in a wide temperature range is remarkably improved, and the difference between monomers in the power battery pack can be effectively processed.

Description

Technical Field

[0001] The present application relates to the technical field of power battery packs, and in particular to a power battery pack status assessment method and system. Background Art

[0002] With the rapid development of the electric vehicle industry, the performance of power battery packs, as the core components of electric vehicles, directly affects the vehicle's range, safety and service life. Accurately evaluating the state of charge (SOC), state of health (SOH) and remaining useful life (RUL) of power battery packs has become a key technology in battery management systems. Traditional battery status assessment methods mainly rely on Kalman filter algorithms and equivalent circuit models, but these methods are heavily dependent on accurate battery model parameters, and battery parameters will continue to change with the aging process and working environment (such as temperature changes), resulting in a decrease in assessment accuracy.

[0003] In recent years, data-driven methods such as neural networks and machine learning techniques have been widely used in the field of battery state assessment. These methods can adaptively adjust model parameters to match actual battery behavior. However, pure data-driven methods have disadvantages such as poor interpretability, limited generalization ability, and reliance on large amounts of training data. They are difficult to provide reliable assessment results, especially when faced with multiple failure modes and complex operating conditions. In addition, existing methods often only focus on a single state of the battery (such as SOC or SOH) and lack the ability to comprehensively evaluate multiple states of the battery pack and identify failure modes. Summary of the invention

[0004] The present application provides a method and system for evaluating the state of a power battery pack, thereby achieving a more accurate prediction of the remaining service life, significantly improving the accuracy within a wide temperature range, and being able to effectively handle the differences between cells in the power battery pack.

[0005] A first aspect of the present application provides a power battery pack status assessment method, the power battery pack status assessment method comprising: Collect the operating characteristic data set of the power battery pack under various working conditions and build a hybrid evaluation model; Based on the hybrid evaluation model, parameters of the double extended Kalman filter are configured to obtain an improved filter; Collecting real-time voltage, current and temperature data of the power battery pack and inputting them into the improved filter for state calculation to obtain state of charge value and internal parameter value; The state of charge value and the internal parameter value are input into a multi-branch adversarial neural network analysis to obtain the fault mode type and fault severity; A multi-factor health index is calculated according to the fault mode type and the fault severity, and a particle swarm optimization fitting is performed on the degradation curve of the power battery pack to obtain a remaining service life prediction value.

[0006] A second aspect of the present application provides a power battery pack status assessment system, the power battery pack status assessment system comprising: The acquisition module is used to collect the operating characteristic data set of the power battery pack under various working conditions and build a hybrid evaluation model; A parameter configuration module, used for configuring the parameters of the double extended Kalman filter based on the hybrid evaluation model to obtain an improved filter; A state calculation module, used to collect real-time voltage, current and temperature data of the power battery pack, and input the data into the improved filter for state calculation to obtain a state of charge value and an internal parameter value; An analysis module, used for inputting the state of charge value and internal parameter value into a multi-branch adversarial neural network analysis to obtain a fault mode type and a fault severity; The fitting module is used to calculate the multi-factor health index according to the fault mode type and the fault severity, and perform particle swarm optimization fitting on the degradation curve of the power battery pack to obtain a remaining service life prediction value.

[0007] Compared with the prior art, the present application has the following beneficial effects: by combining the third-order RC equivalent circuit model with the neural network to form a hybrid evaluation model, the interpretability of the physical model is retained, and the nonlinear characteristics are captured with the help of the neural network, overcoming the limitations of a single model, making the battery state evaluation more accurate, and being able to adapt to battery aging and operating condition changes. Based on the improved double extended Kalman filter, the dynamic adaptive adjustment of process noise and measurement noise parameters is realized, which effectively copes with the change of voltage prediction error, enhances the robustness of the algorithm under complex working conditions, and improves the accuracy of SOC estimation and internal parameter identification. The multi-branch adversarial neural network structure is used to achieve accurate identification of the four main fault modes of capacity attenuation, internal resistance increase, polarization aggravation and imbalance deterioration, and can accurately assess the severity of the fault, providing the battery management system with timely fault warning capabilities. The weight coefficient of the health index is dynamically adjusted according to the identified fault mode type, a targeted degradation model is constructed, and parameter identification is performed through the particle swarm optimization algorithm, achieving a more accurate prediction of the remaining service life, providing a scientific basis for battery maintenance and replacement decisions. By introducing the temperature compensation function and battery pack consistency coefficient matrix, the accuracy of the evaluation method is significantly improved in a wide temperature range, and it can effectively handle the differences between cells in the battery pack and adapt to the actual electric vehicle operating environment. Unlike traditional methods that only focus on single cells, this method fully considers the characteristics of the power battery pack level, including the consistency between cells and temperature distribution, and provides a more comprehensive battery pack status evaluation capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0009] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.

[0010] Figure 1 is a flow chart of a method for evaluating a power battery pack state provided by an embodiment of the present invention; Figure 2 It is a schematic block diagram of the structure of a power battery pack status assessment system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0011] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0012] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0013] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0014] It should be further understood that the term "and / or" used in the specification and appended claims of this application refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations. Figure 1 , an embodiment of the power battery pack state evaluation method in the embodiment of the present application includes: Step 100: Collecting operating characteristic data sets of the power battery pack under various operating conditions and constructing a hybrid evaluation model; It is understandable that the execution subject of the present application may be a power battery pack status assessment system, or a terminal or a server, which is not specifically limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0015] Specifically, within the preset temperature range, i.e., from -20°C to 60°C, at intervals of 10°C, data is collected on the operating status of the power battery pack at different charge and discharge rates. During the data collection process, high-precision voltage, current and temperature sensors are used to ensure the accuracy and reliability of the measured data. By collecting key parameters such as the terminal voltage, charge and discharge current, and ambient temperature of the battery under different charge and discharge rate conditions, the original electrical characteristic data covering the entire life cycle of the power battery pack is obtained. The original electrical characteristic data is subjected to noise removal, and the Butterworth low-pass filter is used to suppress high-frequency noise. At the same time, the mean filter algorithm is applied to smooth the temperature data to reduce the impact of measurement errors. The data is detected for outliers, and the triple standard deviation method is used to screen abnormal data points, and the linear interpolation of the previous and next data points is used to correct them to obtain the operating characteristic data set. Based on the operating characteristic data set, a third-order RC equivalent circuit model containing a series internal resistance, three parallel RC networks, and an open-circuit voltage source is constructed. The model can accurately describe the dynamic voltage response characteristics of the battery, in which the series internal resistance represents the ohmic internal resistance loss of the battery, and the three parallel RC networks are used to characterize the rapid polarization characteristics, medium-frequency polarization characteristics and low-frequency diffusion characteristics of the battery, respectively, so that the model can cover the dynamic characteristics of different time scales from milliseconds to seconds. The parameters of the third-order RC equivalent circuit model are identified by the least squares fitting method or the particle swarm optimization algorithm to ensure that the model can better fit the voltage and current response relationship of the battery, thereby obtaining the basic electrical model of the battery. On the basis of the basic electrical model of the battery, a consistency coefficient matrix that characterizes the differences of each single cell in the battery pack is introduced. This matrix is ​​used to describe the differences in the internal resistance, polarization resistance and polarization capacitance of each single cell, and the parameters are identified by the optimization method, so that it can accurately characterize the unbalanced characteristics inside the battery pack. Since the parameters of the power battery will drift with temperature changes, a temperature compensation function is established based on the parameter changes at different temperatures to correct the influence of temperature on the battery parameters. The temperature compensation function adopts an exponential form, in which the reference temperature is set to 25°C, and the relevant parameters are obtained by regression analysis of experimental data to obtain their optimal values. After the introduction of the consistency coefficient matrix and the correction of the temperature compensation function, the compensated physical electrical model is obtained. A three-layer feedforward neural network is constructed to provide additional nonlinear compensation capabilities. The input layer of the neural network receives the battery's charge and discharge current, temperature, and historical state of charge data. The hidden layer consists of three layers of fully connected networks, containing 64, 32, and 16 neurons respectively. The ReLU activation function is used to enhance the network's learning ability for nonlinear relationships. The output layer is used to predict the polarization voltage of the power battery, so that the neural network can compensate for complex nonlinear effects that cannot be accurately portrayed by the physical model. During the training process of the neural network, the stochastic gradient descent algorithm is used for optimization. The learning rate is set to 0.001 and the batch size is 128 to ensure that the model can converge efficiently and avoid falling into the local optimal solution.The training data is extracted from the operating characteristic data set, and the cross-validation method is used to optimize the neural network model to improve its generalization ability. In order to achieve high-precision evaluation of the power battery pack status, the compensated physical electrical model and the nonlinear characteristic compensation model are integrated to form a hybrid evaluation model. In the hybrid evaluation model, the compensated physical electrical model provides a basic electrical characteristic description of the power battery to ensure the interpretability of the model under different working environments, while the nonlinear characteristic compensation model is used to capture complex nonlinear characteristics that are difficult to characterize with the physical model, thereby improving the overall prediction accuracy of the model. The model integration adopts a parallel architecture, that is, the physical electrical model and the neural network compensation model are calculated independently, and finally fused at the output end to obtain a more accurate prediction result of the power battery status.

[0016] Step 200: Based on the hybrid evaluation model, the double extended Kalman filter is parameterized to obtain an improved filter; Specifically, the state of charge, series internal resistance, polarization resistance, polarization capacitance and open circuit voltage parameters are combined to construct a state variable vector. By associating and mapping the state variable vector with the hybrid evaluation model, a nonlinear state equation and an observation equation are constructed, so that the double extended Kalman filter can accurately track the internal state of the battery under complex working conditions. In this process, the nonlinear state equation is used to describe the evolution of the power battery state over time, while the observation equation is used to characterize the relationship between the battery terminal voltage and the internal state variable, thereby ensuring that the filter uses the measurement data to correct and optimize the state. The initial value of the state variable vector is reasonably set, and the error covariance matrix is ​​configured to obtain the initial parameters of the double extended Kalman filter. The initial value of the state of charge is set according to the historical records of the battery management system. If there is no historical data, it is set to 50% by default to ensure that the filter can operate normally under unknown initial conditions. The initial values ​​of the series internal resistance, polarization resistance and polarization capacitance are set according to the nominal parameters of the battery, while the initial value of the open circuit voltage is calculated based on the state of charge and the known battery open circuit voltage curve. The error covariance matrix is ​​used to describe the uncertainty of the state variables, and its initial value is appropriately set to balance the estimation accuracy and convergence speed. If the initial value of the error covariance matrix is ​​too large, the estimation result of the filter will fluctuate greatly, and if the initial value is too small, the filter will not be able to correctly adjust the state variables under actual working conditions. Therefore, it is necessary to comprehensively consider the characteristics of the power battery and historical data for reasonable settings. After obtaining the initial parameters, the nonlinear state equation and observation equation are linearized to construct the state transfer Jacobian matrix and observation Jacobian matrix. Since the state evolution of the power battery is a nonlinear process, directly solving the state equation will lead to large calculation errors. Therefore, the Jacobian matrix is ​​used to locally linearize it so that the filter can achieve more stable numerical calculations during the state update process. The state transfer Jacobian matrix is ​​used to describe the rate of change of the state variable under small disturbances, while the observation Jacobian matrix is ​​used to characterize the sensitivity of the observation variable to the state variable. The calculation results of these two matrices directly affect the prediction and update accuracy of the filter. In order to improve computational efficiency and stability, numerical differentiation or model-based analytical derivation methods are used to calculate the Jacobian matrix, and optimization and adjustment are performed in combination with the physical characteristics of the power battery. State prediction and update calculations are performed based on the state transfer Jacobian matrix and the observation Jacobian matrix to obtain the optimal estimate of the state at the current moment. In the prediction stage, the filter uses the state estimate and state equation of the previous moment for time update, calculates the state prediction value at the current moment, and uses the state transfer Jacobian matrix to update the error covariance matrix. In the update stage, the filter calculates the state residual in combination with the observation value and observation equation at the current moment, and uses the observation Jacobian matrix to calculate the Kalman gain to adjust the state prediction value to make it as close to the true state as possible.After multiple rounds of iterative calculations, the optimal state estimation value at the current moment is obtained, including parameters such as state of charge, series internal resistance, polarization resistance, polarization capacitance and open circuit voltage. Based on the estimation result, the voltage prediction error is calculated, that is, the deviation between the predicted voltage calculated by the filter and the actual measured voltage. This deviation can reflect the estimation accuracy of the filter. In order to improve the adaptability of the filter, the voltage prediction error in the sliding window is monitored, and the process noise and measurement noise parameters are dynamically adjusted according to the change trend of the error, thereby optimizing the performance of the filter. In the filtering process, process noise is used to describe the random changes in the internal state of the system, while measurement noise is used to characterize sensor measurement errors and external interference. If the voltage prediction error continues to increase in the sliding window, it means that the model error of the filter is large. At this time, the elements of the process noise covariance matrix are increased to improve the adaptability of the filter so that it can adjust the state estimation value faster. On the contrary, if the voltage prediction error continues to decrease, it means that the estimation accuracy of the filter is high, and the process noise covariance is appropriately reduced to improve the stability of the estimation. At the same time, the measurement noise variance is dynamically adjusted according to the change of the measurement error, thereby enhancing the robustness of the filter under different working conditions. After the above optimization and adjustment, an improved dual extended Kalman filter is finally obtained.

[0017] During the filtering calculation process, the voltage prediction error in the sliding window is monitored in real time, and the process noise and measurement noise parameters are dynamically adjusted based on this to optimize the performance of the double extended Kalman filter. During the operation of the filter, the predicted value of the battery terminal voltage at the current moment is obtained after each state update. There is a certain error between the predicted value and the actual measured battery terminal voltage. In order to accurately evaluate the change trend of the error, a sliding window mechanism is introduced to continuously track the voltage prediction error in the recent period of time, and the root mean square error in the window is calculated to obtain the voltage error evaluation index, which effectively measures the degree of deviation between the predicted voltage and the actual measured voltage. A threshold of 10mV is set for the voltage error evaluation index and a comparative judgment of the error state is performed. When the voltage error evaluation index exceeds the threshold, it indicates that there is a deviation in the prediction ability of the current filter, and the filtering parameters are appropriately adjusted; when the error evaluation index is lower than the threshold, it indicates that the estimated error of the filter is small and is in a relatively stable working state. The noise adjustment amplitude is appropriately reduced to improve the stability of the estimation. Through the error state judgment process, the different working states of the filter can be effectively distinguished. After the error state judgment is completed, the relevant elements in the process noise covariance matrix are dynamically adjusted, especially the noise parameters related to the state of charge. Since the process noise covariance matrix determines the sensitivity of the filter to the change of state variables, when the voltage error evaluation index exceeds the set first target threshold, the corresponding element value in the process noise covariance matrix is ​​increased to improve the filter's responsiveness to the change of the state of charge, so that it can adjust the state estimation faster and reduce the cumulative error. On the contrary, when the voltage error evaluation index is lower than the first target threshold, it indicates that the prediction accuracy of the filter is high. At this time, the process noise covariance is appropriately reduced to reduce the estimation fluctuation and improve the stability and reliability of the estimation. Through the adaptive adjustment mechanism, the process noise covariance matrix can be optimized and adjusted according to the actual situation of the voltage prediction error, thereby improving the robustness and adaptability of the filter. At the same time, the measurement noise variance is adjusted inversely based on the error state judgment result to ensure that the filter reasonably allocates the weights of model prediction and measurement update under different error levels. When the voltage error evaluation index exceeds the second target threshold, it indicates that the measurement error is relatively small, while the model prediction error is large. At this time, the measurement noise variance value is reduced to enhance the weight of measurement update, so that the filter relies more on actual observation data for correction, thereby improving the accuracy of state estimation. On the contrary, when the voltage error evaluation index is lower than the second target threshold, it indicates that the measurement error is large, while the model prediction is relatively accurate. At this time, the measurement noise variance value is increased to enhance the stability of the model, reduce the sensitivity of the filter to the observation noise, and avoid drastic changes in the estimated value due to fluctuations in the measurement data. Through this inverse proportional adjustment mechanism, the filter adaptively balances the model prediction and measurement update according to different error levels, thereby improving the stability of the estimate while ensuring accuracy.The updated process noise covariance matrix and the updated measurement noise variance are applied to the state update equation of the double extended Kalman filter to ensure that the filter performs state estimation based on the latest noise parameters. In the new filtering iteration process, the filter uses the updated noise parameters for state prediction and update calculation, so that the state estimate can converge to the true value faster and reduce the accumulation of prediction errors. After multiple iterations, the improved double extended Kalman filter maintains high estimation accuracy under complex working conditions, while having better adaptability and stability.

[0018] Step 300: collect the real-time voltage, current and temperature data of the power battery pack, and input them into the improved filter for state calculation to obtain the state of charge value and internal parameter value; It should be noted that the voltage, current and temperature data of the power battery pack are collected in real time to construct the original monitoring data sequence. In data collection, the acquisition of voltage signals requires high-precision voltage sensors to ensure that the measurement accuracy of the battery terminal voltage reaches the millivolt level; the acquisition of current signals relies on high-precision Hall current sensors, which can capture the current changes of the battery pack during the charging and discharging process and ensure the measurement accuracy at the milliampere level; the temperature signal is monitored using a thermistor array, and multiple temperature sensors are arranged at key positions of the battery pack to obtain the temperature distribution in different areas and ensure that the acquisition accuracy is not less than 0.5°C. All collected data are recorded at a fixed sampling rate and form a time-synchronized original monitoring data sequence. The original monitoring data sequence is processed using a Butterworth low-pass filter and a three-times standard difference constant detection method to obtain more stable and reliable input data. The Butterworth low-pass filter is used to remove high-frequency noise, thereby eliminating the interference caused by current ripple and transient signal fluctuations, making the voltage and current data smoother; the three-times standard deviation constant detection method is used to identify and remove abnormal data points that deviate from the normal range to ensure the stability and consistency of the data, and the linear interpolation method is used to compensate for the abnormal data to avoid the impact of data loss on subsequent calculations. After the preprocessing process, the data sequence after denoising and outlier correction is obtained. The state of charge is preliminarily calculated based on the ampere-hour integration method to obtain the initial value of the state of charge. The basic principle of the ampere-hour integration method is to use the battery's charge and discharge current and time interval to calculate the change in the battery's state of charge, and to perform cumulative updates combined with the previous state of charge to estimate the current state of charge. This method has high accuracy in a short period of time, but due to the battery's self-discharge, Coulomb efficiency, and the accumulation of current sensor errors, its long-term accuracy is affected to a certain extent, so it is only used as a preliminary estimate of the state of charge. The time update step of the improved filter is performed on the preprocessed data sequence and the initial value of the state of charge. In this step, the filter uses the state estimate at the previous moment and the input data at the current moment to calculate the state prediction value through the state transfer equation, and updates the prediction error covariance matrix in combination with the state transfer Jacobian matrix to obtain the predicted value of the state of charge and internal parameters of the power battery. The state prediction process not only considers the voltage and current changes of the battery, but also combines the temperature compensation mechanism to correct the parameter drift effect under different temperature conditions to ensure the prediction accuracy. In this process, the improved filter calculates the state transfer Jacobian matrix so that the prediction error covariance matrix is ​​adaptively adjusted as the system state changes, thereby improving the stability and robustness of the prediction. After completing the state prediction, the measurement update step is performed to correct the predicted state of charge and internal parameter values.In the measurement update phase, the filter calculates the residual between the predicted voltage and the actual measured voltage, and uses the observation Jacobian matrix to calculate the Kalman gain to determine how to weight the fusion of the predicted value and the observed data to correct the state estimate. Weighted fusion adjusts the filter's trust in the predicted results and measured data according to the relative size of the prediction error and the measurement error, thereby optimizing the accuracy of the state estimate. When the measurement noise is small, the filter tends to rely more on the measured value for correction, while when the measurement noise is large, it relies more on the model prediction value to improve the stability of the estimate. After the measurement update step, the corrected state of charge value and the internal parameter value of the power battery are obtained, and the optimization calculation is used to ensure that the state estimation result is more accurate and reliable. Physical constraints and temperature compensation are performed on the corrected state of charge value and internal parameter value. The physical constraints include upper and lower limit constraints on the state of charge value to ensure that it is always within a reasonable range. At the same time, the rationality of internal parameters such as internal resistance, polarization resistance and polarization capacitance is checked to prevent non-physical numerical drift. The temperature compensation mechanism is used to correct parameter offsets caused by changes in ambient temperature. After this process, the optimized and corrected state of charge value and internal parameter value are finally obtained.

[0019] Step 400: Input the state of charge value and the internal parameter value into a multi-branch adversarial neural network analysis to obtain the fault mode type and fault severity; Specifically, the state of charge value and internal parameter value are subjected to time series statistical analysis to extract the characteristic vector of the battery pack, which contains the state information of the battery pack at different time points, and effectively reflects the dynamic characteristics and working conditions of the battery. By analyzing the time series data of internal parameter values ​​such as state of charge, internal resistance, polarization resistance, polarization capacitance, etc., the operation mode and characteristics of the battery pack at different stages are identified. The results of the time series statistical analysis are calculated by the sliding average, variance and deviation of different time windows to obtain the power battery pack characteristic vector reflecting the health status of the battery pack. The power battery pack characteristic vector is input into the multi-branch adversarial neural network for analysis. The shared input layer of the network consists of a 64-node fully connected layer and a batch normalization layer. Through the fully connected layer, the input feature vector is compressed and mapped to a new space, so that the network can better handle complex nonlinear relationships. In this process, the batch normalization layer helps to standardize the input data and reduce the unevenness of data distribution, thereby speeding up the network training process and improving the convergence speed and accuracy. In the shared input layer, the ReLU activation function performs a nonlinear transformation on the data, allowing the network to capture the nonlinear characteristics in the feature vector, thereby enhancing the learning ability of the model. After processing this layer, the network obtains a preliminary feature map and passes it to each branch for further analysis. The preliminary feature map is input to four different branches in parallel, each of which corresponds to a failure mode: capacity decay, internal resistance increase, polarization aggravation, and unbalanced deterioration. In each branch, the network processes according to a specific failure mode and extracts feature information related to the mode. The output of each branch generates a set of feature representations that reflect the characteristics and manifestations of different failure modes. For example, the capacity decay branch focuses on the capacity loss of the battery, the internal resistance increase branch focuses on the change of the internal resistance of the battery, the polarization aggravation branch focuses on the polarization effect of the battery, and the unbalanced deterioration branch detects the performance differences between battery cells. Each branch works independently and efficiently extracts feature information related to its own failure mode in a local range. The attention mechanism is applied to the four sets of branch feature representations for weighted fusion. The attention mechanism dynamically adjusts the importance of each branch feature by calculating the correlation weight between each set of features and each fault mode, thereby strengthening the feature representation related to the fault mode. Through weighted fusion, the network captures the key information of each fault mode more accurately and suppresses the interference of irrelevant features, thereby improving the accuracy of fault mode recognition. After weighted fusion, the enhanced feature representation obtained contains the comprehensive health status of the battery pack and more comprehensively reflects the performance of the battery under different fault modes. The enhanced feature representation is input into the fully connected merging layer, and finally the probability distribution of each fault mode is obtained through the softmax output layer. The fully connected merging layer integrates the information of different branches to improve the expressiveness of the feature representation.The Softmax layer calculates the probability value of each fault mode based on the output of the network, and finally obtains a probability distribution of the fault mode type, which reflects the possibility of different fault modes currently faced by the battery pack. Through the softmax output layer, the network assigns a weight to each fault mode, indicating its probability of occurrence at the current moment. Threshold judgment and continuous monitoring are performed on the probability distribution of each fault mode to ensure the reliability of fault diagnosis. By setting a certain threshold, the network judges the severity and possibility of the fault mode. When the probability of a certain fault mode exceeds the set threshold and lasts for more than a certain period of time, it is determined that the fault mode has occurred, and the system will locate and evaluate the fault mode. At the same time, the fault degree evaluation algorithm is executed to calculate the parameter deviation and evaluate the difference between the current state of the battery pack and the normal operating state. By calculating the parameter deviation of each fault mode, the severity of the fault is quantified, and the level of the fault is judged according to the size of the deviation.

[0020] Step 500: Calculate a multi-factor health index according to the fault mode type and fault severity, and perform particle swarm optimization fitting on the degradation curve of the power battery pack to obtain a remaining service life prediction value.

[0021] Specifically, according to the operating data of the power battery, the relative decay rate of the state of charge value, the relative increase rate of the internal resistance, the relative change rate of the polarization resistance, the relative change rate of the polarization capacitance, and the temperature difference index in the battery pack are collected, and these data are normalized to construct a basic health parameter set. In data collection, high-precision sensors are used to continuously monitor the voltage, current and temperature of the battery pack, and the change trend of each parameter is calculated in combination with historical data. Among them, the relative decay rate of the state of charge value is used to measure the loss of the available capacity of the battery, the relative increase rate of the internal resistance reflects the aging degree of the internal material of the battery, the relative change rate of the polarization resistance and polarization capacitance describes the evolution of the polarization characteristics of the battery over time, and the temperature difference index in the battery pack is used to evaluate the thermal equilibrium state of the battery pack under different operating environments. The purpose of normalization is to convert parameters of different dimensions to the same numerical range to ensure that the degree of influence of different parameters in subsequent calculations will not be affected by the numerical scale, thereby improving the accuracy of the evaluation and the stability of the calculation. The basic health parameter set is weighted to each parameter based on the fault mode type to obtain a weight coefficient suitable for the current battery health state. Since different fault modes have different degrees of influence on the battery, a dynamic weighting method is used to assign different weights to each parameter in the basic health parameter set according to the type of fault mode. For example, in the capacity decay fault mode, the relative decay rate of the state of charge should have a larger weight, while the weights of the internal resistance change and polarization resistance change are relatively small; in the internal resistance increase fault mode, the relative increase rate of the internal resistance should be assigned a higher weight to more accurately reflect the health of the battery; in the polarization aggravation fault mode, the change rate of the polarization resistance and polarization capacitance has a greater impact on the battery performance, so the weights of these parameters should be increased; and in the unbalanced deterioration fault mode, the temperature difference index and the state of charge deviation in the battery pack have a greater impact, and their influence in health assessment needs to be increased. Through the dynamic weight allocation method based on the fault mode, it is ensured that the health index calculation under different fault modes can accurately reflect the actual degradation state of the battery pack. The weight coefficient is used to perform weighted fusion calculation on the basic health parameter set to obtain a multi-factor health index that describes the current health state of the power battery pack. The calculation process of the health index adopts a linear weighted method to weight and sum different health parameters according to their respective weights to ensure that the health index finally calculated can comprehensively reflect the overall health status of the battery pack. The numerical range of the health index is normalized to between 0 and 1, where 1 indicates that the battery is in the optimal state and 0 indicates that the battery has completely failed. After the health index is calculated, it is used for subsequent life prediction to estimate the remaining service life of the battery pack under the current operating conditions. In order to make a remaining life prediction, historical data similar to the current failure mode type is extracted from the historical database of the power battery pack, and a time series data set of the health index changing with the number of cycles is established.This dataset is used to train and optimize the degradation model to accurately characterize the long-term evolution trend of the battery health state. During the data extraction process, it is ensured that the selected historical data has a high similarity with the current battery pack operating conditions, ambient temperature, workload and other conditions to improve the prediction accuracy of the degradation model. After constructing the time series dataset, the corresponding degradation function model structure is selected according to the current fault mode type to obtain the degradation model to be fitted. The degradation model includes exponential decay model, power function model and polynomial regression model. Among them, the capacity decay is fitted by a double exponential model, the internal resistance increase is modeled by a power function, and the change trend of polarization characteristics and temperature imbalance is described by a polynomial function. The particle swarm optimization algorithm is applied to the degradation model to be fitted to optimize the model parameters to minimize the root mean square error between the model prediction value and the actual observation value. The particle swarm optimization algorithm is an optimization method based on swarm intelligence, which efficiently searches for the optimal solution in a high-dimensional parameter space. During the optimization process, each particle represents a set of parameter values ​​of the degradation model. The algorithm adjusts the position and velocity of the particle through multiple iterative steps to gradually approach the optimal solution. By continuously adjusting the parameters of the degradation model and calculating the error between the predicted value and the actual observed value, the particle swarm optimization algorithm can find the optimal fitting degradation model parameters, so that the degradation model can accurately describe the changing trend of the health index with the number of cycles. The optimal fitting degradation model parameters are used to predict the changing trend of the health index with the number of cycles, and the remaining service life prediction value is calculated. The remaining service life of the power battery pack is obtained by substituting the current health index into the optimized degradation model and forward calculating the time point when the health index drops to the set threshold. The threshold of the health index is set to 0.7 or 0.5 to correspond to different maintenance strategies. When the predicted health index reaches this threshold, it means that the battery needs to be replaced or maintained. The remaining service life prediction value is output in the form of cycle number or calendar time.

[0022] In the embodiment of the present application, a hybrid evaluation model combining a third-order RC equivalent circuit model with a neural network is used to retain the interpretability of the physical model and capture the nonlinear characteristics with the help of a neural network, thereby overcoming the limitations of a single model, making the battery state evaluation more accurate, and being able to adapt to battery aging and operating condition changes. Based on the improved dual extended Kalman filter, dynamic adaptive adjustment of process noise and measurement noise parameters is achieved, which effectively responds to changes in voltage prediction errors, enhances the robustness of the algorithm under complex working conditions, and improves the accuracy of SOC estimation and internal parameter identification. The multi-branch adversarial neural network structure is used to accurately identify the four main fault modes of capacity attenuation, internal resistance increase, polarization aggravation, and imbalance deterioration, and can accurately assess the severity of the fault, providing the battery management system with timely fault warning capabilities. The weight coefficient of the health index is dynamically adjusted according to the type of fault mode identified, a targeted degradation model is constructed, and parameter identification is performed through a particle swarm optimization algorithm, achieving a more accurate prediction of the remaining service life, providing a scientific basis for battery maintenance and replacement decisions. By introducing the temperature compensation function and battery pack consistency coefficient matrix, the accuracy of the evaluation method is significantly improved in a wide temperature range, and it can effectively handle the differences between cells in the battery pack and adapt to the actual electric vehicle operating environment. Unlike traditional methods that only focus on single cells, this method fully considers the characteristics of the power battery pack level, including the consistency between cells and temperature distribution, and provides a more comprehensive battery pack status evaluation capability.

[0023] In a specific embodiment, the process of executing step 100 may specifically include the following steps: Collect voltage and current data of the power battery pack within a preset temperature range and at different charge and discharge rates to obtain original electrical characteristic data; Perform noise removal and outlier detection processing on the original electrical characteristic data to obtain an operating characteristic data set; Based on the operating characteristic data set, a third-order RC equivalent circuit model containing a series internal resistance Rs, three parallel RC networks and an open-circuit voltage source Voc is constructed to obtain the basic electrical model of the battery. A consistency coefficient matrix that characterizes the differences between cells in a battery pack is introduced into the basic electrical model of the battery, and a temperature compensation function is established based on parameter changes at different temperatures to obtain a compensated physical electrical model. A three-layer feedforward neural network is constructed, with current, temperature and historical SOC values ​​as input and polarization voltage as output. The ReLU activation function and stochastic gradient descent algorithm are used for training to obtain a nonlinear characteristic compensation model. The compensated physical electrical model is integrated with the nonlinear characteristic compensation model, so that the compensated physical electrical model provides basic electrical characteristic description and the nonlinear characteristic compensation model provides nonlinear compensation, thereby obtaining a hybrid evaluation model.

[0024] Specifically, data is collected on the operating status of the battery pack within a preset temperature range. During the test, the set temperature range covers -20°C to 60°C, and the ambient temperature is gradually adjusted at intervals of 10°C. At the same time, different charge and discharge rates are controlled to obtain battery voltage and current data under various operating conditions. The charge and discharge rates include 0.1C, 0.5C, 1C, 2C and other different rates to ensure the comprehensiveness and representativeness of the data. During the data collection process, high-precision voltage and current sensors are used, with a voltage measurement accuracy of 1mV and a current measurement accuracy of 10mA to ensure the accuracy and stability of the data. By conducting standardized tests in a laboratory environment and combining the actual operating data of the vehicle, the original electrical characteristic data covering a wide range of operating conditions is obtained. The original data is subjected to noise removal and outlier detection processing to obtain an operating characteristic data set. A Butterworth low-pass filter is used to smooth the current and voltage signals to filter out high-frequency noise and ensure that the data is more stable. The triple standard deviation method is used for outlier detection, that is, the mean and standard deviation of the data are calculated, and abnormal data points that are beyond the range of three times the standard deviation are removed to eliminate extreme values ​​caused by sensor errors or short-time current pulses. The removed data points are compensated using the interpolation method to maintain the continuity and integrity of the data series. After this series of processing, an operating characteristic data set that accurately reflects the working characteristics of the battery is finally obtained. Based on the cleaned operating characteristic data set, a third-order RC equivalent circuit model is constructed to accurately describe the electrical characteristics of the battery. In this model, the equivalent circuit of the battery consists of a series internal resistor, three parallel RC networks and an open-circuit voltage source. Among them, the series internal resistor It is mainly used to describe the ohmic loss of the battery, that is, the transient voltage drop generated when the current passes through the battery; the parallel RC network consists of polarization resistance and polarization capacitance of three different time scales, which are used to describe the rapid polarization effect, medium-frequency polarization effect and low-frequency diffusion effect of the battery respectively. The mathematical description of this model is expressed as: in, is the open circuit voltage, is the terminal voltage, is the battery current, Represent the voltage drop under different polarization time scales, and the calculation formula is as follows: in, and Respectively The polarization resistance and polarization capacitance determine the polarization behavior of the battery. In the process of solving the model parameters, the least squares fitting method is used to calculate the optimal parameter values ​​based on the collected experimental data, so that the equivalent circuit model can accurately fit the voltage-current characteristics of the battery. However, there are certain consistency differences between the individual cells in the battery pack, which makes the parameters of the individual cells change asynchronously. Therefore, a consistency coefficient matrix is ​​introduced into the basic electrical model to characterize the differences between the individual cells in the battery pack. . This matrix is ​​used to describe the degree of deviation of different battery cells in internal resistance, polarization resistance and polarization capacitance, and to identify parameters through optimization methods so that the model can take into account the imbalance of the battery pack. At the same time, the battery parameters will drift significantly with temperature changes, so a temperature compensation function is established based on experimental data at different temperatures. Considering the influence of temperature on internal resistance and polarization parameters, an exponential compensation function is used: in, is the current temperature, is the reference temperature (set to 25°C), and is the undetermined coefficient, and its optimal value is obtained by experimental regression. After the introduction of the consistency coefficient matrix and the correction of the temperature compensation function, the compensated physical electrical model is obtained. Due to the polarization effect, hysteresis and dynamic nonlinear behavior of the battery, it is still difficult to fully fit the actual data by relying solely on the physical model. Therefore, a three-layer feedforward neural network is constructed to provide additional nonlinear characteristic compensation capabilities. The input layer of the neural network receives the current, temperature and historical state of charge data of the battery. The hidden layer consists of three layers of fully connected networks with 64, 32 and 16 nodes respectively, and the ReLU activation function is used to enhance the network's learning ability for nonlinear relationships. The output layer is used to predict the polarization voltage of the power battery, so that the neural network can compensate for complex nonlinear effects that cannot be accurately portrayed by the physical model. The network is trained using a stochastic gradient descent algorithm, with a learning rate set to 0.001 and a batch size of 128 to ensure that the model can converge efficiently and avoid falling into a local optimal solution. The training data is extracted from the operating characteristic data set, and the cross-validation method is used to optimize the neural network model to improve its generalization ability. The compensated physical electrical model is integrated with the nonlinear characteristic compensation model to construct a hybrid evaluation model. In the hybrid evaluation model, the compensated physical electrical model provides a basic electrical characteristic description of the power battery to ensure that the model has high interpretability under different working conditions, while the nonlinear characteristic compensation model is used to capture complex nonlinear characteristics that are difficult to describe with the physical model, thereby improving the overall prediction accuracy of the model. During the model integration process, a parallel architecture is adopted so that the physical electrical model and the neural network compensation model are calculated independently and finally fused at the output end to obtain a more accurate prediction result of the power battery status.

[0025] In a specific embodiment, the process of executing step 200 may specifically include the following steps: The state of charge, series internal resistance, polarization resistance, polarization capacitance and open circuit voltage parameters are combined to obtain a state variable vector, and the state variable vector and the hybrid evaluation model are associated and mapped to obtain a nonlinear state equation and an observation equation; The initial value of the state variable vector and the error covariance matrix are set to obtain the initial parameters of the double extended Kalman filter; Based on the initial parameters, the nonlinear state equation and observation equation are linearized to obtain the state transfer Jacobian matrix and the observation Jacobian matrix; Perform state prediction and update calculations according to the state transfer Jacobian matrix and the observation Jacobian matrix to obtain the optimal estimate of the state at the current moment, and calculate the voltage prediction error according to the optimal estimate of the state at the current moment; The voltage prediction error within the sliding window is monitored, and the process noise and measurement noise parameters are dynamically adjusted according to the voltage prediction error to obtain an improved filter.

[0026] Specifically, the state of charge, series internal resistance, polarization resistance, polarization capacitance, and open circuit voltage are combined to form a state variable vector. This state variable vector is defined as: in, Indicates the battery's state of charge. is the series internal resistance of the battery, are the polarization resistance in the third-order RC equivalent circuit model, are polarization capacitance, is the open circuit voltage. A nonlinear state equation is constructed to describe the changing relationship of these variables during the operation of the power battery. Based on the equivalent circuit model of the battery, the evolution equation of the state variables is derived: in, represents the input variables of the system, including charge and discharge current and temperature, is process noise, which describes the errors in the battery system caused by aging, temperature changes, and random disturbances. Specifically, the change in state of charge is represented by the Coulomb integral formula, while the evolution of the open circuit voltage depends on the empirical fitting function of the SOC curve. The relationship between the internal resistance and polarization parameters with temperature is described by the exponential decay model: in, is the current temperature, is the reference temperature, and is the temperature attenuation coefficient to be determined. The observation equation is used to describe the relationship between the measured terminal voltage and the state variable: in, To measure voltage, To measure noise. According to the RC equivalent circuit model, the measured voltage is expressed as: Through these nonlinear equations, a state estimation model is constructed so that the filter can accurately track the electrical characteristics of the battery. After the state equation and observation equation are established, the initial values ​​of the state variables are initialized and the error covariance matrix is ​​set. The initial state vector Set to the mean of historical operating data. If there is no historical data, use the nominal parameters for initialization. Error covariance matrix It is set to a diagonal matrix where the elements on the diagonal represent the uncertainty of the state variables, for example: in, Parameters such as represent the standard deviation of the corresponding state variables and are set according to the parameter tolerances provided by the battery manufacturer to ensure that the initial estimate of the filter has reasonable credibility. In order to realize the numerical calculation of the filter, the nonlinear state equation and observation equation are linearized to obtain the state transfer Jacobian matrix and observation Jacobian matrix. State transfer Jacobian matrix The state variables are calculated by the state equation The partial derivatives of are: Similarly, the observation Jacobian matrix It is composed of the partial derivatives of the observation equation with respect to the state variables: The results of these matrices are used in the prediction and update steps of the filter. In the state prediction stage, the state equation is used for time update: During the measurement update phase, the Kalman gain is calculated: And update the status: In the filtering process, process noise and measurement noise are dynamically adjusted to improve accuracy. By monitoring the voltage prediction error within the sliding window: like If the set threshold is exceeded, the process noise covariance will be increased. To enhance the system's sensitivity to state changes; if If it is below the threshold, then reduce To improve the estimation stability. At the same time, the measurement noise variance Adopt the inverse proportional adjustment strategy, that is, when When it is high, reduce To enhance the weight of the measured data; when When it is low, increase The improved filter can adaptively adjust the estimation accuracy under different working conditions, thereby improving the reliability and robustness of the power battery pack state estimation.

[0027] In a specific embodiment, the execution step monitors the voltage prediction error within the sliding window, and dynamically adjusts the process noise and measurement noise parameters according to the voltage prediction error to obtain an improved filter, which may specifically include the following steps: Monitor the voltage prediction error within the sliding window, and calculate the root mean square error within the window based on the voltage prediction error to obtain a voltage error evaluation index; A threshold of 10mV is set for the voltage error evaluation index and a comparison is made to obtain an error state judgment result; Based on the error state judgment result, the elements related to the state of charge in the process noise covariance matrix are adjusted, when the voltage error evaluation index exceeds the first target threshold, the corresponding element value is increased, and when the voltage error evaluation index is lower than the first target threshold, the corresponding element value is reduced to obtain an updated process noise covariance matrix; The measurement noise variance is adjusted inversely based on the error state judgment result. When the voltage error evaluation index exceeds the second target threshold, the noise variance value is reduced to enhance the measurement update weight. When the voltage error evaluation index is lower than the second target threshold, the noise variance value is increased to enhance the model stability, thereby obtaining an updated measurement noise variance. The updated process noise covariance matrix and the updated measurement noise variance are applied to the state update equation of the double extended Kalman filter to obtain an improved filter.

[0028] Specifically, during the operation of the filter, the predicted value of the battery terminal voltage at the current moment is calculated after each state update, and compared with the actual measured voltage to obtain the voltage prediction error. In order to evaluate the stability and change trend of the error, a sliding window mechanism is used to The error within a time step is statistically calculated, and the error level is measured by the root mean square error, which is calculated as follows: in, For the The measured voltage at each time step, is the predicted terminal voltage of the filter, is the size of the sliding window, which is set to 30 to 100 sampling points to balance short-term error fluctuations and long-term error trends. As a voltage error evaluation index, it is used to determine whether the estimated error of the filter is within a reasonable range. The voltage error evaluation index is compared with the set error threshold to determine whether the current filter estimation state needs to be adjusted. Set an error threshold mV, when When the threshold is exceeded, it means that the prediction error of the filter is large and the process noise needs to be increased to improve the adaptive ability. When it is lower than this threshold, it means that the filter estimation is relatively stable and the noise can be appropriately reduced to improve the calculation accuracy and reduce the volatility of the state estimation. Based on the error state judgment result, adjust the process noise covariance matrix , especially with the state of charge Related elements , in order to enhance the dynamic adaptability of the filter. Exceeding the first target threshold set When , in order to improve the sensitivity of the filter to SOC changes, the adjustment strategy is expressed as: in, is the adjustment coefficient, set to 0.1 to 0.5 to control the adjustment amplitude and prevent the noise parameter from changing too fast and causing the filter to oscillate. When it is lower than the first target threshold, it indicates that the filter estimate is relatively stable and should be appropriately reduced. , in order to reduce unnecessary fluctuations in state estimation, the adjustment is as follows: in, is the adjustment coefficient. The filter polarization resistance and polarization capacitance The process noise associated with the proposed method is similarly adjusted to ensure the adaptability of the overall state estimate. An inverse adjustment is made to balance the weights of measurement update and state prediction. Exceeding the second target threshold When , it means that the measurement error is small, but the state prediction error is large. Therefore, it is necessary to reduce the measurement noise variance to enhance the role of measurement data in state update, thereby improving the estimation accuracy. The adjustment strategy is as follows: in, To adjust the coefficient, the measurement noise variance is gradually reduced. When it is lower than the second target threshold, it means that the state prediction is more accurate and the measurement noise may be larger. In this case, it should be appropriately increased. , in order to reduce the impact of measurement data and improve the stability of the model, the adjustment method is as follows: in, To adjust the coefficients to ensure that the measurement noise is adjusted smoothly. Through this adjustment mechanism, the filter is ensured to rely on the measurement data for correction in the case of high error, and enhance the model prediction ability in the case of low error, thereby improving the stability and accuracy of the estimate. The updated process noise covariance matrix and the measurement noise variance Reapply the state update equation of the double extended Kalman filter to ensure that the filter can be adaptively adjusted under different working conditions. The update formula in the state prediction stage is as follows: In the measurement update phase, the Kalman gain is calculated: The application of the adaptive adjustment strategy enables the filter to dynamically adjust the noise parameters according to the changing trend of the voltage error, thereby improving the stability and accuracy of the power battery state estimation.

[0029] In a specific embodiment, the process of executing step 300 may specifically include the following steps: Collect the voltage, current and temperature data of the power battery pack in real time to obtain the original monitoring data sequence; The original monitoring data sequence is processed by using Butterworth low-pass filter and three times standard difference constant detection method to obtain the preprocessed data sequence; The state of charge is preliminarily calculated based on the ampere-hour integration method to obtain the initial value of the state of charge; The time updating step of the improved filter is performed on the preprocessed data sequence and the initial value of the state of charge, and the time updating step includes calculating the state prediction value according to the state at the previous moment and the current input, and updating the prediction error covariance matrix using the state transition Jacobian matrix to obtain the predicted value of the state of charge and the internal parameters; Performing a measurement update step of an improved filter on the predicted value, the measurement update step includes calculating the residual of the predicted voltage and the measured voltage, calculating the Kalman gain based on the observed Jacobian matrix, and updating the state estimate by weighted fusion of the predicted value and the observed residual to obtain a corrected state of charge value and internal parameter value; Physical constraints and temperature compensation are performed on the corrected state of charge value and internal parameter value, and finally the state of charge value and the internal parameter value are obtained.

[0030] Specifically, the key signals such as voltage, current and temperature of the battery pack are continuously monitored through real-time data acquisition. These collected data can provide the real-time working status information of the battery. After processing, the collected raw monitoring data form a complete data sequence. The raw monitoring data sequence is preprocessed. The Butterworth low-pass filter is used to remove noise from the voltage and current signals to filter out high-frequency noise. The low-pass filter can effectively smooth the data, remove unnecessary high-frequency fluctuations, and ensure the stability of the data. The triple standard deviation method is used for outlier detection. By calculating the mean and standard deviation of the data, abnormal data points that exceed the normal range are identified and eliminated. For example, if the current value at a certain moment is more than three times the standard deviation away from the average value, the value is considered to be an outlier and is corrected by the interpolation method of the previous and next data. The processed data sequence is used to calculate the state of charge of the battery. Since the state of charge of the battery changes over time, the current data estimates the state of charge of the battery through the ampere-hour integration method. The basic principle of the ampere-hour integration method is to use the battery's charge and discharge current and time interval to calculate the change in the battery's state of charge, and gradually accumulate to obtain an estimated value of the current state of charge. The change in SOC is calculated using the following formula: in, Indicates the current state of charge. is the current value at the previous moment, is the current efficiency (0.98 during charging and 1.0 during discharging), is the time interval, is the rated capacity of the battery. Through this method, a preliminary state of charge estimate is obtained as the initial value of the filter input. The time update step of the improved filter is performed on the preprocessed data sequence and the initial value of the state of charge. The time update step calculates the predicted value of the state by using the nonlinear state equation and updates the prediction error covariance matrix. In the Kalman filter framework, the state prediction formula is: in, is the state prediction value, is the state transfer function, which describes the transition from the state at the previous moment to the state at the current moment. is the input variable (such as current and temperature) at the previous moment. State transfer Jacobian matrix The update formula for the linearized state transfer equation and the prediction error covariance matrix is: in, is the predicted error covariance matrix, is the process noise covariance matrix, which reflects the error of the system model. This update step provides a preliminary state estimate for subsequent measurement updates by predicting the changes in the state and covariance matrix. After the state prediction is completed, the measurement update step is entered to calculate the residual between the current measurement value and the predicted value, and the observation Jacobian matrix is ​​used To adjust the filter estimate. The basic formula for measurement update is: in, is the Kalman gain, is the measured value, is the predicted voltage value, is the measurement noise covariance matrix. Through this weighted fusion mechanism, the filter reasonably weights the measured data and the predicted data to obtain the corrected state of charge and internal parameter estimates. After obtaining the corrected state of charge values ​​and internal parameter values, physical constraints and temperature compensation are performed to ensure that the state estimate has physical meaning and conforms to the actual working state of the battery. The physical constraints mainly include upper and lower limit constraints on the state of charge to ensure that it is always within the range of 0 to 100%. At the same time, the rationality of the battery's internal resistance, polarization resistance and other parameters is checked to prevent estimates that exceed physical boundaries. Temperature compensation takes into account the impact of temperature on battery performance, and uses temperature compensation functions to correct parameters such as internal resistance, polarization resistance and polarization capacitance. The form of the temperature compensation function is: in, is the temperature at the current moment, is the reference temperature, is the temperature coefficient, which describes the sensitivity of the internal resistance to temperature changes. Through temperature compensation, the filter can obtain more accurate battery state estimation under different temperature conditions.

[0031] In a specific embodiment, the process of executing step 400 may specifically include the following steps: Perform time series statistical analysis on the state of charge value and internal parameter values ​​to obtain the power battery pack feature vector; The power battery pack feature vector is input into the shared input layer of the multi-branch adversarial neural network. The shared input layer contains a 64-node fully connected layer and a batch normalization layer. The ReLU activation function is used for nonlinear transformation to obtain a preliminary feature map. The preliminary feature map is input in parallel into the capacity attenuation branch, the internal resistance increase branch, the polarization aggravation branch and the imbalance deterioration branch, and four groups of branch feature representations are obtained after being processed respectively; The attention mechanism is applied to the four groups of branch feature representations, and the weighted fusion is performed by calculating the correlation weights between the features and each fault mode to obtain the enhanced feature representation; The enhanced feature representation is input into the fully connected merging layer and the softmax output layer to obtain the probability distribution of each fault mode; The probability distribution of each fault mode is subjected to threshold judgment and continuous monitoring, and the fault severity assessment algorithm is executed to calculate the parameter deviation to obtain the fault mode type and fault severity.

[0032] Specifically, the state of charge (SOC) and internal parameters (such as series internal resistance, polarization resistance, polarization capacitance and open circuit voltage) are subjected to time series statistical analysis to extract the characteristics of the battery pack from the time series. These characteristics can reflect the health status of the battery and its performance under different working conditions. Time series statistical analysis involves calculating the statistical characteristics of each state variable in different time windows, such as mean, variance, skewness, kurtosis, etc., and generating a feature vector of the battery pack. For example, the mean and standard deviation of the state of charge reflect the charging level and stability of the battery, and the change in polarization resistance reveals the aging process of the battery. Through this step, the dynamic behavior of the battery is converted into a low-dimensional feature vector. The feature vector of the battery pack is input into the shared input layer of the multi-branch adversarial neural network. The goal of the neural network design is to extract the global features of the battery pack through the shared input layer and use it for multiple tasks. The shared input layer consists of a 64-node fully connected layer and a batch normalization layer, where the batch normalization layer is used to standardize the input data to reduce the learning instability caused by uneven data distribution during training. The ReLU activation function is used to perform nonlinear transformations on the input data, stimulating the learning ability of the neural network to better extract complex patterns in the data. After this layer of processing, the input feature vector is converted into a preliminary feature map containing the main information of the battery pack status. The preliminary feature map is input in parallel to four different branch networks, each of which specializes in a failure mode: capacity decay, internal resistance increase, polarization aggravation, and imbalance deterioration. Each branch network includes a series of convolutional layers and fully connected layers to extract feature representations related to a specific failure mode. For example, the capacity decay branch focuses on the trend of the state of charge and the degradation of the battery capacity, the internal resistance increase branch focuses on the impact of the increase in the battery internal resistance on the battery performance, the polarization aggravation branch focuses on the change of the polarization voltage of the battery during the charging and discharging process, and the imbalance deterioration branch detects the voltage difference and imbalance between battery cells. After each branch is processed, the output feature representation can reflect the manifestation of a specific failure mode and provide detailed information about the battery health status. These feature representations obtained from each branch are applied to the attention mechanism. The role of the attention mechanism is to dynamically adjust the influence of different features according to the correlation weight of each branch feature with the failure mode. By calculating the correlation between each feature vector and different fault modes, the attention mechanism can assign a weighted value to each feature to reflect its contribution to the current fault mode. Specifically, the weight coefficient is calculated using the following formula: in, Represents the feature vector and failure mode The similarity is calculated using cosine similarity or dot product. In this way, the weight of each branch is dynamically adjusted so that the network pays more attention to features that are highly relevant to the current fault mode. The feature representation after weighted fusion is passed to the fully connected merging layer and the softmax output layer. The function of the fully connected merging layer is to integrate the feature information of all branches to facilitate the next classification task. The softmax layer calculates the probability distribution of each fault mode and outputs the probability of occurrence of each fault mode. The formula of the softmax function is as follows: in, is the output of the merging layer, indicating the scores of the current battery pack corresponding to each failure mode. Is the battery pack corresponding to the failure mode The softmax output layer generates a probability value for each fault mode, indicating the possibility of the mode occurring. After obtaining the probability distribution of each fault mode, threshold judgment and continuous monitoring are performed. By setting an appropriate threshold, it is determined whether a fault alarm is triggered. In order to enhance the reliability of fault diagnosis, the fault severity is evaluated based on these probability values. The fault severity evaluation measures the severity of the fault by calculating the deviation of each battery parameter. The fault severity is calculated by the following formula: in, is the actual value of the current battery parameter, is the nominal value of the battery (the initial design value). The greater the deviation, the more serious the fault and the higher the severity of the fault mode.

[0033] In a specific embodiment, the process of executing step 500 may specifically include the following steps: The relative decay rate of the state of charge value, the relative increase rate of the internal resistance, the relative change rate of the polarization resistance, the relative change rate of the polarization capacitance and the temperature difference index in the battery pack are collected and normalized to obtain a basic health parameter set; Based on the failure mode type, the weight of each parameter in the basic health parameter set is assigned to obtain a weight coefficient; Perform weighted fusion calculation based on the basic health parameter set and weight coefficients to obtain a multi-factor health index that reflects the current health status of the power battery pack; Extract historical data similar to the current failure mode type from the power battery pack history database, establish a time series data set of health index changes with cycle number, and select the corresponding degradation function model structure according to the failure mode type to obtain the degradation model to be fitted; The particle swarm optimization algorithm is applied to the degradation model to be fitted. Through multiple iterative searches, the root mean square error between the model prediction value and the actual observation value is minimized to obtain the optimal fitting degradation model parameters. The optimal fitting degradation model parameters are used to predict the changing trend of the health index with the number of cycles, and the remaining service life prediction value is obtained.

[0034] Specifically, the relative decay rate of the state of charge value, the relative increase rate of the internal resistance, the relative change rate of the polarization resistance, the relative change rate of the polarization capacitance and the temperature difference index within the battery pack are collected. The relative decay rate of the state of charge value is used to indicate the change in battery capacity, the relative increase rate of the internal resistance reflects the degradation of the battery's electrical performance, the relative change rates of the polarization resistance and polarization capacitance indicate the change in the battery polarization effect, which is closely related to the battery's charge and discharge efficiency, while the temperature difference index within the battery pack characterizes the thermal balance within the battery. After collecting these data, they are normalized so that all parameters are in the same dimensional range to avoid calculation errors caused by unit differences. The normalization method uses linear normalization, and the formula is as follows: in, is the normalized data, is the original data, and are the minimum and maximum values ​​of the data, respectively. Through this step, the parameters of different dimensions are unified into the range of [0,1]. Based on the type of failure mode (such as capacity attenuation, increased internal resistance, increased polarization, etc.), weights are assigned to the parameters in the basic health parameter set to reflect the impact of different failure modes on the health status of the battery. Different failure modes have different effects on the battery pack, so the weight of each parameter will also vary depending on the failure mode. For example, in the capacity attenuation failure mode, the weight of the state of charge value relative to the attenuation rate is larger, while in the internal resistance increase failure mode, the weight of the internal resistance relative increase rate is larger. The allocation of weight coefficients is optimized through expert experience or machine learning methods to ensure that the contribution of each health parameter in a specific failure mode is correctly reflected. Assuming there are multiple failure modes, the weight coefficient under each failure mode is expressed by the following formula: in, It is The weight of the parameters, is the importance factor of the failure mode associated with this parameter, is the total number of failure modes is a normalization factor to ensure that the sum of all weights is 1. In this way, the weights of various parameters can be flexibly adjusted according to different failure modes and the actual performance of the battery health status. A weighted fusion calculation is performed based on the basic health parameter set and the weight coefficient to obtain a multi-factor health index that reflects the current health status of the power battery pack. This health index can comprehensively reflect the health status of the battery pack in multiple dimensions, including capacity, internal resistance, polarization effect, and temperature difference. The calculation formula for the multi-factor health index is: in, is the health index of the battery pack, It is The weight of the parameters, is the normalized Health parameter values, is the total number of parameters. Through weighted fusion, the influence of each battery performance parameter is combined to obtain an index that reflects the overall health status of the battery. This health index is used to monitor the health status of the battery in real time. Historical data similar to the current failure mode type is extracted from the historical database of the power battery pack, and a time series data set of the health index changing with the number of cycles is established. This data set can show the degradation trend of the battery under different usage conditions and provide a reference for subsequent degradation model fitting. In order to establish a more accurate degradation model, a suitable degradation function model is selected according to different failure mode types. Degradation function models include exponential decay models, power function models, and polynomial regression models. For example, the capacity decay is fitted using an exponential decay model: in, For the battery The capacity of the cycle, is the initial capacity, is the decay rate, is the number of cycles. The internal resistance increase and polarization resistance change are fitted using a power function model: in, For the battery The internal resistance during a cycle is is the initial internal resistance, and The fitting parameters determine the speed and law of the internal resistance growth. The particle swarm optimization algorithm is applied to the degradation model to be fitted to obtain the optimal degradation model parameters. The particle swarm optimization algorithm is an optimization method based on swarm intelligence. It gradually approaches the optimal solution by simulating the search of multiple particles in the parameter space. Each particle represents a set of model parameters. By calculating the root mean square error between the model prediction value and the actual observation value, the particle swarm optimization algorithm will continuously adjust the position of the particles to minimize the model error. The optimization process of the model is expressed as: in, is the actual observed value, is the model prediction value, is the total number of data points. Through particle swarm optimization, the optimal attenuation model parameters are finally obtained. Using the optimal fitting degradation model parameters, the trend of the health index changing with the number of cycles is predicted to obtain the predicted value of the remaining useful life (RUL). By calculating the change of the health index in the future and combining it with the set health threshold, the remaining useful life of the battery is predicted. The calculation formula for the remaining useful life is: in, is the health index at the current moment, For the set health threshold, is the rate at which the battery health index decays, obtained by fitting historical data.

[0035] The above describes the power battery pack state evaluation method in the embodiment of the present application. The following describes the power battery pack state evaluation system 10 in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the power battery pack status assessment system 10 includes: The acquisition module 11 is used to collect the operating characteristic data set of the power battery pack under various working conditions and build a hybrid evaluation model; A parameter configuration module 12 is used to configure the parameters of the double extended Kalman filter based on the hybrid evaluation model to obtain an improved filter; The state calculation module 13 is used to collect the real-time voltage, current and temperature data of the power battery pack and input them into the improved filter for state calculation to obtain the state of charge value and internal parameter values; An analysis module 14 is used to input the state of charge value and the internal parameter value into a multi-branch adversarial neural network analysis to obtain the fault mode type and fault severity; The fitting module 15 is used to calculate the multi-factor health index according to the fault mode type and fault severity, and perform particle swarm optimization fitting on the degradation curve of the power battery pack to obtain the remaining service life prediction value.

[0036] Through the synergy of the above components, the hybrid evaluation model that combines the third-order RC equivalent circuit model with the neural network not only retains the interpretability of the physical model, but also captures the nonlinear characteristics with the help of the neural network, overcomes the limitations of a single model, makes the battery state evaluation more accurate, and can adapt to battery aging and working condition changes. Based on the improved double extended Kalman filter, the dynamic adaptive adjustment of process noise and measurement noise parameters is realized, which effectively copes with the change of voltage prediction error, enhances the robustness of the algorithm under complex working conditions, and improves the accuracy of SOC estimation and internal parameter identification. The multi-branch adversarial neural network structure is used to accurately identify the four main fault modes of capacity attenuation, internal resistance increase, polarization aggravation and imbalance deterioration, and can accurately assess the severity of the fault, providing the battery management system with timely fault warning capabilities. The weight coefficient of the health index is dynamically adjusted according to the identified fault mode type, a targeted degradation model is constructed, and parameter identification is performed through the particle swarm optimization algorithm, achieving a more accurate prediction of the remaining service life, providing a scientific basis for battery maintenance and replacement decisions. By introducing the temperature compensation function and battery pack consistency coefficient matrix, the accuracy of the evaluation method is significantly improved in a wide temperature range, and it can effectively handle the differences between cells in the battery pack and adapt to the actual electric vehicle operating environment. Unlike traditional methods that only focus on single cells, this method fully considers the characteristics of the power battery pack level, including the consistency between cells and temperature distribution, and provides a more comprehensive battery pack status evaluation capability.

[0037] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0038] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable an electronic device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0039] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for evaluating the state of a power battery pack, characterized in that: include: Collect the operating characteristic data set of the power battery pack under various working conditions and build a hybrid evaluation model; Based on the hybrid evaluation model, parameters of the double extended Kalman filter are configured to obtain an improved filter; Collecting real-time voltage, current and temperature data of the power battery pack and inputting them into the improved filter for state calculation to obtain state of charge value and internal parameter value; The state of charge value and the internal parameter value are input into a multi-branch adversarial neural network analysis to obtain the fault mode type and fault severity; A multi-factor health index is calculated according to the fault mode type and the fault severity, and a particle swarm optimization fitting is performed on the degradation curve of the power battery pack to obtain a remaining service life prediction value.

2. The power battery pack status assessment method according to claim 1, characterized in that: The method collects the operating characteristic data set of the power battery pack under various working conditions and constructs a hybrid evaluation model, including: Collect voltage and current data of the power battery pack within a preset temperature range and at different charge and discharge rates to obtain original electrical characteristic data; Performing noise removal and outlier detection processing on the original electrical characteristic data to obtain an operation characteristic data set; Based on the operating characteristic data set, a third-order RC equivalent circuit model including a series internal resistance Rs, three parallel RC networks and an open-circuit voltage source Voc is constructed to obtain a basic electrical model of the battery; A consistency coefficient matrix characterizing the differences between cells in a battery pack is introduced into the basic battery electrical model, and a temperature compensation function is established based on parameter changes at different temperatures to obtain a compensated physical electrical model; A three-layer feedforward neural network is constructed, with current, temperature and historical SOC values ​​as input and polarization voltage as output. The ReLU activation function and stochastic gradient descent algorithm are used for training to obtain a nonlinear characteristic compensation model. The compensated physical electrical model is integrated with the nonlinear characteristic compensation model, so that the compensated physical electrical model provides a basic electrical characteristic description and the nonlinear characteristic compensation model provides nonlinear compensation, thereby obtaining a hybrid evaluation model.

3. The power battery pack status assessment method according to claim 1, characterized in that: The method of configuring parameters of the double extended Kalman filter based on the hybrid evaluation model to obtain an improved filter includes: The state of charge, series internal resistance, polarization resistance, polarization capacitance and open circuit voltage parameters are combined to obtain a state variable vector, and the state variable vector and the hybrid evaluation model are associated and mapped to obtain a nonlinear state equation and an observation equation; Setting the initial value of the state variable vector and the error covariance matrix to obtain initial parameters of the double extended Kalman filter; Based on the initial parameters, the nonlinear state equation and the observation equation are linearized to obtain a state transfer Jacobian matrix and an observation Jacobian matrix; Performing state prediction and update calculation according to the state transfer Jacobian matrix and the observation Jacobian matrix to obtain the optimal estimate of the state at the current moment, and calculating the voltage prediction error according to the optimal estimate of the state at the current moment; The voltage prediction error within the sliding window is monitored, and process noise and measurement noise parameters are dynamically adjusted according to the voltage prediction error to obtain an improved filter.

4. The power battery pack status assessment method according to claim 3, characterized in that: The voltage prediction error within the monitoring sliding window is dynamically adjusted according to the voltage prediction error to obtain an improved filter, including: Monitoring the voltage prediction error within the sliding window, and calculating the root mean square error within the window according to the voltage prediction error to obtain a voltage error evaluation index; A threshold of 10 mV is set for the voltage error evaluation index and a comparison is performed to obtain an error state judgment result; Based on the error state judgment result, the elements related to the state of charge in the process noise covariance matrix are adjusted, when the voltage error evaluation index exceeds the first target threshold, the corresponding element value is increased, and when the voltage error evaluation index is lower than the first target threshold, the corresponding element value is reduced to obtain an updated process noise covariance matrix; Based on the error state judgment result, the measurement noise variance is adjusted inversely proportionally, when the voltage error evaluation index exceeds the second target threshold, the noise variance value is reduced to enhance the measurement update weight, and when the voltage error evaluation index is lower than the second target threshold, the noise variance value is increased to enhance the model stability, so as to obtain an updated measurement noise variance; The updated process noise covariance matrix and the updated measurement noise variance are applied to the state update equation of the double extended Kalman filter to obtain an improved filter.

5. The power battery pack status assessment method according to claim 1, characterized in that: The real-time voltage, current and temperature data of the power battery pack are collected and input into the improved filter for state calculation to obtain the state of charge value and internal parameter value, including: Collecting voltage, current and temperature data of the power battery pack in real time to obtain an original monitoring data sequence; The original monitoring data sequence is processed by using a Butterworth low-pass filter and a three-times standard difference constant detection method to obtain a preprocessed data sequence; The state of charge is preliminarily calculated based on the ampere-hour integration method to obtain the initial value of the state of charge; Performing a time update step of the improved filter on the preprocessed data sequence and the initial value of the state of charge, the time update step comprising calculating a state prediction value according to a state at a previous moment and a current input, and updating a prediction error covariance matrix using a state transition Jacobian matrix to obtain a prediction value of the state of charge and an internal parameter; Performing a measurement update step of the improved filter on the predicted value, the measurement update step comprising calculating a residual between the predicted voltage and the measured voltage, calculating a Kalman gain based on an observed Jacobian matrix, and updating a state estimate by weighted fusion of the predicted value and the observed residual to obtain a corrected state of charge value and an internal parameter value; Physical constraints and temperature compensation are performed on the corrected state of charge value and internal parameter value, and finally the state of charge value and the internal parameter value are obtained.

6. The power battery pack status assessment method according to claim 1, characterized in that: The input of the state of charge value and the internal parameter value into a multi-branch adversarial neural network analysis to obtain the fault mode type and fault severity includes: Performing time series statistical analysis on the state of charge value and the internal parameter value to obtain a power battery pack characteristic vector; Input the power battery pack feature vector into a shared input layer of a multi-branch adversarial neural network, wherein the shared input layer includes a 64-node fully connected layer and a batch normalization layer, and a ReLU activation function is used for nonlinear transformation to obtain a preliminary feature map; The preliminary feature map is input in parallel into the capacity attenuation branch, the internal resistance increase branch, the polarization aggravation branch and the imbalance deterioration branch, and four groups of branch feature representations are obtained after being processed respectively; Applying an attention mechanism to the four groups of branch feature representations, performing weighted fusion by calculating the correlation weights between the features and each fault mode, and obtaining an enhanced feature representation; Input the enhanced feature representation into the fully connected merging layer and the softmax output layer to obtain the probability distribution of each fault mode; The probability distribution of each fault mode is subjected to threshold judgment and continuous monitoring, and the fault severity assessment algorithm is executed to calculate the parameter deviation to obtain the fault mode type and fault severity.

7. The power battery pack status assessment method according to claim 1, characterized in that: The calculating of the multi-factor health index according to the fault mode type and the fault severity, and performing particle swarm optimization fitting on the degradation curve of the power battery pack to obtain the remaining service life prediction value includes: The relative decay rate of the state of charge value, the relative increase rate of the internal resistance, the relative change rate of the polarization resistance, the relative change rate of the polarization capacitance and the temperature difference index in the battery pack are collected and normalized to obtain a basic health parameter set; Based on the failure mode type, weights of the parameters in the basic health parameter set are assigned to obtain weight coefficients; Performing weighted fusion calculation according to the basic health parameter set and the weight coefficient to obtain a multi-factor health index reflecting the current health status of the power battery pack; Extracting historical data similar to the current fault mode type from the power battery pack historical database, establishing a time series data set of health index changes with cycle number, and selecting a corresponding degradation function model structure according to the fault mode type to obtain a degradation model to be fitted; Applying a particle swarm optimization algorithm to the degradation model to be fitted, minimizing the root mean square error between the model prediction value and the actual observation value through multiple iterative searches, and obtaining the optimal fitting degradation model parameters; The optimally fitted degradation model parameters are used to predict the changing trend of the health index with the number of cycles to obtain a predicted value of the remaining service life.

8. A power battery pack status assessment system, characterized in that: Used to execute the power battery pack state assessment method according to any one of claims 1 to 7, the power battery pack state assessment system comprises: The acquisition module is used to collect the operating characteristic data set of the power battery pack under various working conditions and build a hybrid evaluation model; A parameter configuration module, used for configuring the parameters of the double extended Kalman filter based on the hybrid evaluation model to obtain an improved filter; A state calculation module, used to collect real-time voltage, current and temperature data of the power battery pack, and input the data into the improved filter for state calculation to obtain a state of charge value and an internal parameter value; An analysis module, used for inputting the state of charge value and internal parameter value into a multi-branch adversarial neural network analysis to obtain a fault mode type and a fault severity; The fitting module is used to calculate the multi-factor health index according to the fault mode type and the fault severity, and perform particle swarm optimization fitting on the degradation curve of the power battery pack to obtain a remaining service life prediction value.

Images