Machine Learning-Based Battery State Monitoring and Analysis Method and System

Abstract

The present invention provides a method and system for monitoring and analyzing battery state based on machine learning. First, real-time operation data during the charge and discharge cycles of the target battery is obtained, covering voltage, current, temperature sequences, and timestamp sequences. Then, multi-dimensional feature extraction is performed to generate an original feature set, and a standardized feature matrix is obtained through dynamic time series processing. Next, it is input into a pre-trained battery state monitoring model to generate degradation, health, and remaining life prediction parameters, thereby generating an anomaly detection result and a maintenance decision recommendation. Finally, based on the above results, the model parameter weights are adaptively adjusted to update the model prediction accuracy. This method can comprehensively and accurately monitor and analyze the battery state, improve the prediction accuracy, and provide a reliable basis for battery maintenance.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular, to a method and system for monitoring and analyzing battery state based on machine learning. Background Art

[0002] In today's era, as a key energy storage device, batteries are widely used in many fields such as electric vehicles, portable electronic devices, and energy storage power stations. With the continuous improvement of the requirements for battery performance and reliability in various devices, accurate monitoring and analysis of battery state have become crucial.

[0003] Existing technologies attempt to evaluate battery state through some conventional data analysis methods, but these methods often lack in-depth mining of the implicit features in battery operation data. For example, they usually only perform surface statistical analysis on the data, ignoring the deep correlations and temporal features between data. For instance, when analyzing battery charge and discharge data, the dynamic change laws of multiple variables such as current and temperature over time and their interactions are not fully considered, resulting in low prediction accuracy for battery health state and remaining life, and it is difficult to meet the demand for accurately grasping battery state in practical applications.

[0004] In terms of battery maintenance decision-making, most existing methods are based on fixed maintenance cycles or general maintenance standards, without fully considering the actual operating state and characteristics of each battery. This may not only lead to over-maintenance, increasing unnecessary costs and resource waste, but also may cause premature battery damage or performance degradation due to failure to handle abnormal situations of individual batteries in a timely manner, affecting the normal operation of the device. Summary of the Invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present invention, embodiments of the present invention provide a method for monitoring and analyzing battery state based on machine learning, and the method includes:

[0006] Obtain real-time operation data of a target battery during charge and discharge cycles, where the real-time operation data includes a voltage sequence, a current sequence, a temperature sequence, and a timestamp sequence;

[0007] Perform multi-dimensional feature extraction on the real-time operation data to generate an original feature set, and perform dynamic temporal processing on the original feature set to obtain a standardized feature matrix;

[0008] Input the standardized feature matrix into a pre-trained battery state monitoring model to generate degradation state parameters, health state parameters, and remaining life prediction parameters of the target battery;

[0009] Generate an anomaly detection result and a maintenance decision recommendation for the target battery according to the degradation state parameter, the health state parameter, and the remaining life prediction parameter;

[0010] Based on the anomaly detection result and the maintenance decision recommendation, adaptively adjust the parameter weights of the battery state monitoring model to update the prediction accuracy of the battery state monitoring model.

[0011] On the other hand, an embodiment of the present invention further provides a battery state monitoring and analysis system based on machine learning, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions, or codes, and the processor is used to execute the programs, instructions, or codes in the machine-readable storage medium to implement the above method.

[0012] Based on the above aspects, the embodiment of the present application first collects real-time operating data such as voltage, current, temperature and timestamp sequence in the charge and discharge cycle of the target battery in the data acquisition stage, which can fully and dynamically reflect the actual operating status of the battery under different working conditions. Secondly, the multi-dimensional feature extraction and dynamic timing processing link, by extracting multi-dimensional features of real-time operation data, generates an original feature set, and can deeply mine the key information hidden in the data. The dynamic timing processing further integrates these original features into a standardized feature matrix, which not only fully considers the time series characteristics of the battery operation data, but also adapts to the dynamic changes of battery performance over time. Compared with the traditional fixed mode feature processing method, it can better capture the subtle changes in the battery state, greatly improving the effectiveness and representativeness of the features. Furthermore, the standardized feature matrix is ​​input into the pre-trained battery state monitoring model to accurately generate degradation state parameters, health state parameters and remaining life prediction parameters. The battery state monitoring model is trained based on a large amount of data, has a high degree of generalization and accuracy, can effectively parse complex battery state information, and provides a quantitative and reliable basis for the full life cycle management of the battery. In terms of generating abnormal detection results and maintenance decision recommendations, based on comprehensive and accurate parameter analysis, it can quickly and accurately identify abnormal conditions in battery operation, and generate practical maintenance decision recommendations based on different battery states and potential risks, realizing personalized and intelligent maintenance management based on the actual state of the battery, greatly improving maintenance efficiency, reducing maintenance costs, and extending the service life of the battery. Finally, based on the abnormal detection results and maintenance decision recommendations, the parameter weights of the battery state monitoring model are adaptively adjusted, which can continuously optimize the performance of the battery state monitoring model according to the actual monitoring results and update the prediction accuracy in real time. Compared with the traditional fixed parameter model, this method can dynamically adapt to the complex changes in the battery state, continuously maintain high-precision monitoring and prediction capabilities, effectively avoid monitoring errors caused by model aging or gradual changes in battery performance, and provide a strong guarantee for long-term stable monitoring of the battery state. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic diagram of the execution flow of the battery status monitoring and analysis method based on machine learning provided in an embodiment of the present invention.

[0014] Figure 2 It is a schematic diagram of the hardware architecture of a battery status monitoring and analysis system based on machine learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1It is a schematic flowchart of a battery state monitoring and analysis method based on machine learning provided by an embodiment of the present invention. The battery state monitoring and analysis method based on machine learning will be introduced in detail below.

[0016] Step S110: Obtain the real-time operation data of the target battery during charge and discharge cycles. The real-time operation data includes a voltage sequence, a current sequence, a temperature sequence, and a timestamp sequence.

[0017] In this embodiment, taking the power battery of an electric vehicle as an example, during the operation of the electric vehicle, the battery continuously undergoes charge and discharge cycles.

[0018] To obtain the real-time operation data of the target battery, a distributed sensing node network can be constructed. Among them, the voltage acquisition node collects the battery voltage at a relatively high first sampling frequency (such as every 5 milliseconds). During vehicle acceleration, deceleration, or charging, the voltage will change differently, and these collected voltage values constitute the voltage sequence. The current acquisition node collects the current sequence at a second sampling frequency of every 3 milliseconds. When the vehicle is in different driving states (such as climbing, cruising, or rapid acceleration), the magnitude of the charging and discharging current of the battery will change significantly. The temperature acquisition node collects the temperature sequence at a third sampling frequency of every 10 milliseconds. Since the battery generates heat during high-power charging and discharging, the temperature acquisition points can capture the temperature changes at different positions of the battery.

[0019] Furthermore, the time synchronization node aligns the sampling moments of the above three acquisition nodes at the microsecond level. For example, when the vehicle rapidly accelerates at a certain moment, when the voltage acquisition node collects the voltage, the current acquisition node collects the corresponding large current value, and the temperature acquisition node also collects the temperature change of the battery at the same microsecond-level moment, ensuring that the timestamps of the three are consistent. The collected data is sent to the edge computing gateway through an encrypted transmission protocol for temporary caching and integrity verification. After the verification passes, the real-time operation data including the voltage sequence, current sequence, temperature sequence, and timestamp sequence can be read in batches.

[0020] Step S120: Extract multi-dimensional features from the real-time operation data to generate an original feature set, and perform dynamic time series processing on the original feature set to obtain a standardized feature matrix.

[0021] For voltage sequence feature extraction, during a single charging process, local maxima and minima in the voltage sequence are found through an extreme point detection algorithm. Suppose that during a certain period of charging, the local maximum is 4.2 volts and the local minimum is 4.0 volts. Calculating their absolute difference gives a voltage fluctuation extreme value of 0.2 volts. A first-order difference operation is performed to obtain a differential voltage sequence, and the arithmetic mean is calculated within a 5-second sliding time window to obtain the mean voltage change rate. At the same time, the duration of the voltage plateau is detected. If the maximum continuous time period during which the voltage change rate continuously falls below the plateau threshold of 0.01 volts / second is 20 seconds, then 20 seconds is the duration of the voltage plateau.

[0022] From the current sequence, if 1000 watt-hours of electrical energy is input during charging and 850 watt-hours of electrical energy is output during discharging, then the current charge-discharge efficiency is 85%. The current peaks are identified. If the first peak occurs at the 8th second and the second at the 20th second, the current peak interval period is 12 seconds. At the end of discharging, the current decay slope is calculated based on the change trend of the current from 30 amperes to 5 amperes.

[0023] Regarding the temperature sequence, assume that the temperature at the center of the battery surface rises rapidly while that at the edge rises slowly, and the temperature gradient distribution is obtained through data from multiple acquisition points. If the temperature rises from 30 degrees Celsius to 35 degrees Celsius within 10 minutes, the temperature rise rate is 0.5 degrees Celsius per minute. If the temperature rise rate threshold is 0.4 degrees Celsius per minute, it can be determined whether it is exceeded. The temperature standard deviation at each position is calculated to obtain the temperature uniformity index.

[0024] From the timestamp sequence, the charging and discharging cycle length is 2.5 hours for the interval from one full charge to the next charge, and the resting stage duration is 15 minutes for the interval from the end of discharging to the start of the next charge. The cycle interval consistency coefficient is obtained through the statistics of the time intervals of multiple charging and discharging cycles.

[0025] These features are normalized and concatenated to generate an original feature set. Then, dynamic time series processing is performed, and missing values are filled and noise is filtered for each feature dimension of the original feature set. For example, the current charge-discharge efficiency may have missing values due to acquisition interference, and interpolation is used to fill and filter them. According to the statistical distribution characteristics of the features, those that are relatively stable like the charging and discharging cycle length are classified into the static feature subset, and those that change dynamically like the current charge-discharge efficiency are classified into the dynamic feature subset.

[0026] The mean smoothing process is performed on the static feature subset with a sliding window of 3 cycles. For the dynamic feature subset, taking the temperature gradient distribution as an example, the first-order seasonal difference operation is performed to eliminate the trend component and retain the fluctuation component. Then, the fast Fourier transform is performed to obtain the dominant frequency component to construct a reference phase signal. After calculating the phase offset, time-domain translation compensation is performed to make the phase difference less than the preset value. Finally, amplitude normalization and timestamp resampling are performed to obtain the dynamically aligned feature subset. The static smoothed feature subset and the dynamically aligned feature subset are cross-dimensionally fused according to the timestamp sequence to obtain a standardized feature matrix.

[0027] Step S130, input the standardized feature matrix into a pre-trained battery state monitoring model to generate the degradation state parameter, health state parameter, and remaining life prediction parameter of the target battery.

[0028] In this embodiment, the standardized feature matrix is input into a pre-trained battery state monitoring model. The convolutional time series encoding layer captures local patterns to generate a primary spatio-temporal feature map. Taking an electric vehicle battery as an example, this layer can identify the local change patterns of battery characteristics under different driving modes.

[0029] The attention weight assignment layer strengthens the key time points of the primary spatio-temporal feature map. For example, the time point during deep discharge of the battery is crucial for judging the battery state, and this layer will increase the weight of the features corresponding to this time point to generate a weighted spatio-temporal feature map.

[0030] The bidirectional recurrent prediction layer divides the weighted spatio-temporal feature map according to the time step. Assuming the time step is 5 seconds, the forward input sequence is the process from charging to discharging, and the reverse is from discharging to charging. The degradation increment of each time step is calculated by the forward recurrent neural network unit and accumulated to obtain the degradation cumulative amount. The life residual component is calculated by the reverse recurrent neural network unit and weighted aggregated to obtain the life residual coefficient. At the same time, cross-layer connection and non-linear transformation are performed to output the correction factor.

[0031] In the multi-task output layer, the degradation cumulative amount and the life residual coefficient are linearly weighted and spliced, and cross-channel feature scaling is performed to obtain standardized features. The standardized degradation features are input into the first fully connected branch, and the initial degradation state parameter is generated through the degradation state activation function. The standardized residual features are input into the second fully connected branch to generate the initial life prediction parameter. The cross-attention interaction between the two generates a health state interaction weight matrix, and the standardized degradation features are weighted and aggregated to obtain a health state intermediate vector, which is input into the third fully connected branch and the initial health state parameter is generated through the health state activation function. Operations such as time series alignment compensation, residual coupling, non-linear interpolation fusion, and correction module adjustment are performed, and finally output consistency constraints are performed to obtain the final degradation state parameter, health state parameter, and remaining life prediction parameter.

[0032] Step S140: Generate an anomaly detection result and a maintenance decision recommendation for the target battery according to the degradation state parameter, the health state parameter, and the remaining life prediction parameter.

[0033] In this embodiment, the degradation state parameter is compared with a preset degradation threshold range. For example, the preset range is [0, 0.2]. If 0.25 is calculated, it exceeds the range and a first anomaly mark is generated. The deviation degree of the health state parameter from the historical health state baseline calculation trend is calculated. If the set dynamic deviation threshold is 0.15 and 0.2 is calculated, a second anomaly mark is generated. The remaining life prediction parameter is compared with the median of the battery life distribution of the same batch. If the median of the same batch is 3 years, the life warning quantile is 0.4, and the predicted remaining life is 1 year, a third anomaly mark is generated.

[0034] Match the maintenance strategy knowledge base according to the anomaly mark combination pattern. If there are both the first and third anomaly marks, a high maintenance priority may be given. The maintenance operation type is to replace the battery cell, and the maintenance decision recommendation for the maintenance time window within 72 hours is given.

[0035] Step S150: Based on the anomaly detection result and the maintenance decision recommendation, adaptively adjust the parameter weights of the battery state monitoring model to update the prediction accuracy of the battery state monitoring model.

[0036] For example, select a weight update rule from the model parameter adjustment strategy library according to the anomaly mark type. If the first anomaly mark appears, the weights of the convolutional temporal encoding layer related to the degradation state are adjusted. Determine the learning rate and regularization strength according to the maintenance priority. If the maintenance priority is medium, the learning rate is adjusted from 0.001 to 0.003, and the regularization strength is adjusted from 0.05 to 0.1.

[0037] Construct a virtual feedback data set according to the maintenance operation type and the maintenance time window. For example, if the maintenance operation is balanced charging and the time window is 48 hours, construct the virtual feedback data set during this period and mix and sample it with the historical training data according to a certain ratio (such as 1:4). Based on the mixed sampling data, the updated learning rate, and the regularization strength, perform backpropagation optimization on the parameter weights of the convolutional temporal encoding layer, the attention weight allocation layer, and the bidirectional recurrent prediction layer, and at the same time freeze the parameter weights of the multi-task output layer, and iterate and update until the prediction error converges.

[0038] Based on the above steps, the embodiment of the present application first collects real-time operating data such as voltage, current, temperature and timestamp sequence in the charge and discharge cycle of the target battery in the data acquisition stage, which can fully and dynamically reflect the actual operating status of the battery under different working conditions. Secondly, the multi-dimensional feature extraction and dynamic timing processing link, by extracting multi-dimensional features of real-time operation data, generates an original feature set, and can deeply mine the key information hidden in the data. The dynamic timing processing further integrates these original features into a standardized feature matrix, which not only fully considers the time series characteristics of the battery operation data, but also can adapt to the dynamic changes of battery performance over time. Compared with the traditional fixed mode feature processing method, it can better capture the subtle changes in the battery state, greatly improving the effectiveness and representativeness of the features. Furthermore, the standardized feature matrix is ​​input into the pre-trained battery state monitoring model to accurately generate degradation state parameters, health state parameters and remaining life prediction parameters. The battery state monitoring model is trained based on a large amount of data, has a high degree of generalization ability and accuracy, can effectively parse complex battery state information, and provides a quantitative and reliable basis for the full life cycle management of the battery. In terms of generating abnormal detection results and maintenance decision recommendations, based on comprehensive and accurate parameter analysis, it can quickly and accurately identify abnormal conditions in battery operation, and generate practical maintenance decision recommendations based on different battery states and potential risks, realizing personalized and intelligent maintenance management based on the actual state of the battery, greatly improving maintenance efficiency, reducing maintenance costs, and extending the service life of the battery. Finally, based on the abnormal detection results and maintenance decision recommendations, the parameter weights of the battery state monitoring model are adaptively adjusted, which can continuously optimize the performance of the battery state monitoring model according to the actual monitoring results and update the prediction accuracy in real time. Compared with the traditional fixed parameter model, this method can dynamically adapt to the complex changes in the battery state, continuously maintain high-precision monitoring and prediction capabilities, effectively avoid monitoring errors caused by model aging or gradual changes in battery performance, and provide a strong guarantee for long-term stable monitoring of the battery state.

[0039] In a possible implementation, step S120 includes:

[0040] Step S121, extracting the voltage fluctuation extreme value, the voltage change rate mean value and the voltage platform duration from the voltage sequence.

[0041] In this embodiment, during the driving and charging of an electric vehicle, the voltage sequence of the battery carries important battery state information. When extracting the extreme values of voltage fluctuations from the voltage sequence, when the vehicle is driving, for example, during acceleration and deceleration, the voltage of the battery will fluctuate. During a specific driving process, by analyzing the collected voltage sequence, the detected maximum voltage value is 4.2 volts, and the minimum voltage value is 3.8 volts. The difference of 0.4 volts between the two is the extreme value of voltage fluctuation. When calculating the average value of the voltage change rate, first perform a first-order difference operation on the voltage sequence to obtain a differential voltage sequence. During a specific 5-second sliding time window, for example, when the vehicle is driving steadily, the values of the differential voltage sequence are 0.01 volts, 0.02 volts, -0.01 volts, etc. The arithmetic average value of 0.01 volts of these values is determined as the average value of the voltage change rate. When detecting the duration of the voltage platform, when the vehicle is in the parked charging state, at a certain stage of charging, the voltage change rate continuously drops below the platform threshold of 0.01 volts / second, and the maximum continuous time period is 20 seconds. This 20 seconds is the duration of the voltage platform.

[0042] Step S122: Extract the current charge-discharge efficiency, the current peak interval period, and the current decay slope from the current sequence.

[0043] When extracting features from the current sequence, for the determination of the current charge-discharge efficiency, during a complete charge-discharge cycle, assume that the input electrical energy during charging is 1000 watt-hours, and the output electrical energy during discharging is 850 watt-hours. By calculation, the current charge-discharge efficiency is 850÷1000×100% = 85%. For the current peak interval period, during the driving of the vehicle, the current will change with the power demand of the vehicle. Identify the peak points in the current sequence. For example, the first current peak appears at the 8th second after the vehicle starts, and the second current peak appears at the 20th second during the vehicle's rapid acceleration. Then the current peak interval period is 20 - 8 = 12 seconds. At the end of discharging, as the battery power decreases, the current will gradually decay. According to the current change trend from 30 amperes to 5 amperes, the current decay slope is obtained through relevant mathematical calculations.

[0044] Step S123: Extract the temperature gradient distribution, the temperature rise rate threshold, and the temperature uniformity index from the temperature sequence.

[0045] During the charging and discharging process of the battery, the temperature changes at different positions of the battery are different. Since the temperature at the center of the battery rises faster and the temperature at the edge rises slower during high-power discharge of the battery (such as when the car is driving at high speed or accelerating suddenly) or rapid charging, the temperature gradient distribution is constructed through the data of multiple temperature acquisition points. During the charging process, if the temperature rises from 30 degrees Celsius to 35 degrees Celsius within 10 minutes, the calculated temperature rise rate is (35 - 30) ÷ 10 = 0.5 degrees Celsius per minute. If the preset temperature rise rate threshold is 0.4 degrees Celsius per minute, it can be judged whether the threshold is exceeded. Calculate the standard deviation of the temperature at each position of the battery. If the standard deviation is small, for example, 0.2 degrees Celsius, it means that the temperature uniformity is good, and this 0.2 degrees Celsius represents the temperature uniformity index.

[0046] Step S124, extract the charge-discharge cycle length, rest stage duration, and cycle interval consistency coefficient from the timestamp sequence.

[0047] The calculation of the charge-discharge cycle length is based on the time interval from the start of a complete charge to the start of the next charge. For example, after a car completes a charge and drives normally until the start of the next charge, this time interval is 2.5 hours, which is the charge-discharge cycle length. The rest stage duration refers to the rest time of the battery between charging and discharging. For example, after the car finishes discharging and does not start charging immediately due to some reason and rests for 15 minutes before starting to charge, this 15 minutes is the rest stage duration. The cycle interval consistency coefficient reflects the consistency of the time intervals between each charge-discharge cycle. It is obtained by calculating the ratio of the standard deviation to the mean value through statistical analysis of the time intervals of multiple charge-discharge cycles. Suppose the coefficient is 0.1 after multiple statistical calculations.

[0048] Step S125, perform normalized splicing on the voltage fluctuation extreme value, average voltage change rate, voltage platform duration, current charge-discharge efficiency, current peak interval period, current decay slope, temperature gradient distribution, temperature rise rate threshold, temperature uniformity index, charge-discharge cycle length, rest stage duration, and cycle interval consistency coefficient to generate the original feature set.

[0049] Step S126, perform missing value filling and noise filtering processing on each feature dimension in the original feature set to generate a denoised feature set.

[0050] Each feature dimension in the original feature set may have incomplete data or be affected by noise interference. Missing value filling and noise filtering are performed on each feature dimension to generate a denoised feature set. For example, in the feature dimension of current charge-discharge efficiency, due to possible external electromagnetic interference or temporary failures of the acquisition device, the data acquisition at some moments is inaccurate, resulting in missing values. The missing values are filled by interpolation methods, such as linear interpolation, according to the values of surrounding valid data points. At the same time, for possible noise, such as small fluctuations in current data caused by electromagnetic interference in the vehicle's internal electrical system, filtering algorithms, such as low-pass filtering algorithms, are used to remove high-frequency noise, thereby obtaining the denoised current charge-discharge efficiency feature. The same methods of missing value filling and noise filtering are also applied to other feature dimensions, such as the average value of voltage change rate, temperature uniformity index, etc., and finally a denoised feature set is generated.

[0051] Step S127, according to the statistical distribution characteristics of each feature dimension in the denoised feature set, divide the denoised feature set into a static feature subset and a dynamic feature subset.

[0052] For example, features such as the length of the charge-discharge cycle and the duration of the static stage are relatively stable during the normal use of electric vehicles, and their values do not change drastically with each charge-discharge process. These features are divided into the static feature subset. While features such as current charge-discharge efficiency and temperature gradient distribution have obvious dynamic changes under different driving states (such as constant speed driving, acceleration, deceleration) and charge-discharge states of the vehicle. These features are divided into the dynamic feature subset.

[0053] Step S128, perform moving window mean smoothing on the static feature subset to generate a static smoothed feature subset.

[0054] Taking the length of the charge-discharge cycle as an example, set the moving window to 3 cycles and calculate the mean value within each window. Assume that the lengths of the charge-discharge cycles within the first window are 2.5 hours, 2.4 hours, and 2.6 hours respectively, and their mean value is (2.5 + 2.4 + 2.6) ÷ 3 = 2.5 hours. Calculate the mean value within each window in this way successively, so as to obtain the smoothed length value of the charge-discharge cycle. The same processing is also performed on other static features, and finally a static smoothed feature subset is generated.

[0055] Step S129, perform time series difference transformation and phase alignment processing on the dynamic feature subset to generate a dynamic aligned feature subset, and fuse the static smoothed feature subset and the dynamic aligned feature subset across dimensions according to the time stamp sequence to generate the standardized feature matrix.

[0056] Taking the dynamic feature of temperature gradient distribution as an example, first perform a first-order seasonal difference operation to eliminate the trend component and retain the periodic fluctuation component. During the driving of the vehicle, the temperature gradient distribution of the battery may be affected by various factors such as environmental temperature and driving conditions, and there is a certain trend change. This trend is removed through the difference operation. Then perform a fast Fourier transform to extract the dominant frequency component. Assume that the frequency corresponding to the dominant frequency component is 0.1 Hz. Based on this dominant frequency component, construct a reference phase signal. Calculate the cyclic phase difference between the phase component of the temperature gradient distribution in this feature dimension and the reference phase signal to obtain the phase offset. Due to various complex situations during the driving of the vehicle, the phase of the temperature gradient distribution collected at different times may shift. The phase offset is calculated to quantify this shift. According to this phase offset, perform time-domain translation compensation on the temperature gradient distribution so that the phase difference of its dominant frequency component is less than the preset phase tolerance. For example, the preset phase tolerance is 0.1 radian. Finally, perform amplitude normalization and timestamp resampling on the temperature gradient distribution after phase compensation to generate a dynamically aligned feature subset. Other dynamic features are processed according to the same steps.

[0057] Finally, cross-dimensionally fuse the static smoothed feature subset and the dynamically aligned feature subset according to the timestamp sequence to generate a standardized feature matrix. During the fusion process, in the order of timestamps, combine the feature values in the static smoothed feature subset with the corresponding feature values in the dynamically aligned feature subset to construct a new matrix, which is the standardized feature matrix. It synthesizes the feature information of the battery in different aspects and has better consistency and analyzability after processing, preparing for subsequent input into the battery state monitoring model.

[0058] In a possible implementation manner, step S130 includes:

[0059] Step S131, through the convolutional temporal encoding layer in the battery state monitoring model, capture local patterns of the standardized feature matrix to generate a primary spatio-temporal feature map.

[0060] In this embodiment, when the standardized feature matrix is input into the convolutional temporal coding layer in the battery state monitoring model, the convolutional temporal coding layer aims to capture local patterns of the standardized feature matrix to generate a primary spatio-temporal feature map. During the actual operation of an electric vehicle, the standardized feature matrix contains a lot of information related to the battery state, such as data on the extreme values of voltage fluctuations, current charge and discharge efficiency, etc. at different times and driving states. The convolutional kernels in the convolutional temporal coding layer are like specific filters that slide on the matrix at a certain stride and analyze different local regions. Taking the example of an electric vehicle during a long-distance drive, the driving process includes different road conditions such as flat roads, uphill slopes, downhill slopes, and different driving behaviors such as constant-speed driving, acceleration, and deceleration. During this process, various characteristics of the battery will exhibit specific local patterns at different times and operating conditions. The convolutional temporal coding layer can identify the correlation patterns between characteristics such as voltage, current, and temperature under specific operating conditions, for example, the specific combination patterns of voltage, current, temperature, etc. when driving at high speed and the battery is discharging at high power. In this way, local spatio-temporal features related to the battery state are extracted from the standardized feature matrix, and then a primary spatio-temporal feature map is generated, which can initially reflect the combination of the battery's state characteristics at different spatio-temporal positions.

[0061] Step S132: Through the attention weight assignment layer in the battery state monitoring model, strengthen the key time points of the primary spatio-temporal feature map to generate a weighted spatio-temporal feature map.

[0062] During the entire operation cycle of an electric vehicle, there are some key time points that are of great significance for judging the battery state. For example, during the deep discharge stage of the battery, the changes in characteristics such as the battery's voltage and temperature have a greater impact on the battery's health state and remaining life; or during the fast charging stage, the inflow of a large current may cause the chemical reactions inside the battery to intensify. The characteristics corresponding to these time points are crucial for accurately judging the battery state. The attention weight assignment layer strengthens the weights of the characteristics corresponding to these key time points in the primary spatio-temporal feature map according to the weights learned in the pre-trained model. For example, during the deep discharge stage, the weights of the voltage, current, temperature, etc. characteristics corresponding to this period in the primary spatio-temporal feature map will be increased, making these characteristics more prominent in subsequent analyses, thereby generating a weighted spatio-temporal feature map. Such a weighted spatio-temporal feature map can focus more on the feature information that is of great significance for judging the battery state.

[0063] Step S133: Through the bidirectional recurrent prediction layer in the battery state monitoring model, perform forward cumulative degradation analysis and backward life residual fitting on the weighted spatio-temporal feature map to generate a degradation cumulative amount and a life residual coefficient.

[0064] For example, the weighted spatio-temporal feature map is segmented by time steps. Taking a complete charge-discharge cycle of an electric vehicle as an example, assuming the time step is set to 5 seconds, the forward input sequence is the feature sequence in chronological order from the start of charging to the end of discharging, and the reverse input sequence is the feature sequence in reverse from the end of discharging to the start of charging. The hidden state of the forward input sequence is recursively calculated through the forward recurrent neural network unit. Within each time step, the degradation increment is calculated based on the various characteristics of the battery. For example, during the normal driving and discharging process of the vehicle, as time goes by, the capacity of the battery gradually decreases. The capacity reduction within each 5-second time step is calculated as the degradation increment based on features such as voltage and current in the weighted spatio-temporal feature map, and then these degradation increments are accumulated to obtain the degradation cumulative amount. At the same time, the hidden state of the reverse input sequence is traced back through the reverse recurrent neural network unit, and the life residual component is calculated within each time step. For example, during the reverse analysis of the charging process, the life residual component corresponding to each time step is calculated based on the characteristic changes of the battery during charging, and these life residual components are weighted and aggregated according to a certain weight to obtain the life residual coefficient. In addition, the final hidden state of the forward recurrent neural network unit is cross-connected with the initial hidden state of the reverse recurrent neural network unit to generate a joint hidden state vector, and the joint hidden state vector is non-linearly transformed through a fully connected layer to output the correction factors of the degradation cumulative amount and the life residual coefficient to further improve the accuracy of these two parameters.

[0065] Step S134: Through the multi-task output layer in the battery state monitoring model, jointly map the degradation cumulative amount and the life residual coefficient, and synchronously output the degradation state parameter, the health state parameter, and the remaining life prediction parameter.

[0066] For example, in a possible implementation, step S134 includes:

[0067] Step S1341: Linearly weight and splice the degradation cumulative amount and the life residual coefficient by time step to generate a degradation-residual joint feature vector, and perform cross-channel feature scaling on the degradation-residual joint feature vector to generate a normalized degradation feature and a normalized residual feature.

[0068] In this embodiment, during the operation of an electric vehicle, the time step is an important time interval unit for analyzing the battery state. For example, taking every 5 minutes as a time step, at different time steps, the cumulative degradation reflects the degree of performance degradation of the battery during long-term use, and the remaining life coefficient reflects the remaining characteristics related to the predicted life. By linearly weighting and splicing the cumulative degradation and the remaining life coefficient at each time step according to specific weights, a degradation-residual joint feature vector is formed, and this degradation-residual joint feature vector contains comprehensive information related to battery degradation and life. However, since the dimensions and numerical ranges of different features may be different, in order to make these features comparable in subsequent analyses, cross-channel feature scaling is required. Taking the voltage-related features and temperature-related features of the battery as an example, they may differ in numerical range and the weight of their impact on the battery state. Cross-channel feature scaling can adjust these different features to a suitable numerical range, thereby generating standardized degradation features and standardized residual features, and these two standardized features can more accurately reflect the state information of the battery in terms of degradation and life, laying a foundation for subsequent analyses.

[0069] Step S1342: Input the standardized degradation feature into the first fully connected branch, generate initial degradation state parameters through a degradation state activation function, input the standardized residual feature into the second fully connected branch, and generate initial life prediction parameters through a life residual activation function.

[0070] In the actual use of electric vehicles, the initial degradation state parameters of the battery reflect the performance change of the battery from the initial state to the current state. For example, when an electric vehicle has traveled a certain mileage and experienced multiple charge and discharge cycles, the chemical substances inside the battery will be depleted, and the electrode structure may change. These factors will all lead to a decline in battery performance. The standardized degradation features contain this degradation-related information and are processed through the degradation state activation function in the first fully connected branch. This activation function performs a non-linear transformation on the input standardized degradation features according to a preset mathematical model to generate the initial degradation state parameters. This initial degradation state parameter can intuitively represent the degradation degree of the battery in the current state relative to the initial state. For example, a numerical value can be used to represent the attenuation ratio of the battery capacity or the increase degree of the battery internal resistance, etc. At the same time, the standardized residual features are input into the second fully connected branch, and the initial life prediction parameters are generated through the life residual activation function. During the use of an electric vehicle, the remaining life of the battery is a key indicator. The standardized residual features contain information related to the remaining life of the battery, such as the remaining power of the battery under different working conditions and the change of the charge and discharge efficiency of the battery. Through the processing of the life residual activation function in the second fully connected branch, this activation function calculates the standardized residual features according to the pre-learned model to generate the initial life prediction parameters. This initial life prediction parameter can initially estimate the remaining time or the remaining number of charge and discharge cycles that the battery can still be used normally in the current state.

[0071] Step S1343: Perform cross-attention interaction on the standardized degradation features and the standardized residual features to generate a health state interaction weight matrix, and perform dynamic weighted aggregation on the standardized degradation features according to the health state interaction weight matrix to generate a health state intermediate vector.

[0072] Specifically, the health state of the battery is a comprehensive concept and is affected by multiple factors. There is a complex mutual relationship between the standardized degradation features and the standardized residual features. Through the cross-attention interaction mechanism, the mutual influence relationship between these features can be captured, thereby generating a health state interaction weight matrix. For example, in some cases, the degradation degree of the battery may have a greater impact on the remaining life, so a higher weight corresponds to it in the health state interaction weight matrix. According to this health state interaction weight matrix, dynamic weighted aggregation is performed on the standardized degradation features. Taking the voltage fluctuation and current change of the battery under different charge and discharge states as an example, the degradation-related features are weighted and summed according to the weights in the weight matrix to generate a health state intermediate vector. This health state intermediate vector comprehensively considers the impact of the features related to battery degradation and remaining life on the health state and can more comprehensively reflect the health state information of the battery.

[0073] Step S1344: Input the intermediate health state vector into the third fully connected branch, generate the initial health state parameters through the health state activation function, perform temporal alignment compensation on the initial degradation state parameters and the initial health state parameters to generate the compensated degradation state parameters and the compensated health state parameters, and perform residual coupling on the compensated degradation state parameters and the initial life prediction parameters to generate the coupled life prediction parameters.

[0074] In the actual usage scenario of an electric vehicle, this initial health state parameter can comprehensively reflect the health status of multiple aspects of the battery. For example, the comprehensive state of aspects such as the internal resistance of the battery, the available effective capacity, and the charge-discharge efficiency. The health state activation function processes the intermediate health state vector according to a pre-determined mathematical model and converts it into the initial health state parameters. The initial health state parameters can be represented by a numerical value or a set of numerical values to indicate the health degree of the battery. For example, a health state scoring system can be set, and according to this initial health state parameter, the health state of the battery is divided into different levels. Then, perform temporal alignment compensation on the initial degradation state parameters and the initial health state parameters to generate the compensated degradation state parameters and the compensated health state parameters. During the operation of an electric vehicle, since there is a certain correlation and change law between the degradation and the health state of the battery in terms of time, it is necessary to ensure the consistency of these two parameters in the time dimension. For example, at different driving mileage and charge-discharge cycle times, the changes in the degradation state and the health state of the battery may have a certain time delay or advance. Through the temporal alignment compensation mechanism, according to the actual operation data of the battery and the pre-learned time relationship model, adjust the initial degradation state parameters and the initial health state parameters to generate the compensated degradation state parameters and the compensated health state parameters. These compensated parameters can more accurately reflect the degradation and health state of the battery in the same time dimension.

[0075] Step S1345: Perform non-linear interpolation fusion on the compensated health state parameters and the coupled life prediction parameters to generate the interpolated health-life correlation parameters, and input the compensated degradation state parameters into the degradation state correction module to adjust the output corrected degradation state parameters through the degradation correction coefficient.

[0076] There is an inherent relationship between the degradation state of the battery and its remaining life. The compensated degradation state parameter reflects the actual degradation of the battery, while the initial life prediction parameter is a preliminary estimate of the battery's remaining life. Through the residual coupling mechanism, these two parameters are organically combined. For example, when the degradation degree of the battery is relatively high, it may have a greater impact on the remaining life. By considering this impact through a specific mathematical model, a coupled life prediction parameter is generated. This parameter can more accurately reflect the remaining life of the battery considering the degradation state. Then, the compensated state of health parameter and the coupled life prediction parameter are non-linearly interpolated and fused to generate an interpolated state of health-life correlation parameter. In the actual operation of an electric vehicle, the relationship between the state of health of the battery and its remaining life is not a simple linear relationship, but a complex non-linear relationship. Through the non-linear interpolation fusion mechanism, according to the pre-learned non-linear model, the compensated state of health parameter and the coupled life prediction parameter are fused to generate an interpolated state of health-life correlation parameter. This parameter can more comprehensively reflect the correlation between the state of health of the battery and its remaining life.

[0077] Step S1346: Input the interpolated state of health-life correlation parameter into the state of health correction module, adjust and output the corrected state of health parameter through the health correction coefficient, and input the coupled life prediction parameter into the life prediction correction module, adjust and output the corrected remaining life prediction parameter through the life correction coefficient.

[0078] Due to the influence of various factors, such as measurement errors and environmental changes, there may be certain deviations in the initially calculated degradation state parameter. The degradation state correction module adjusts the compensated degradation state parameter according to the pre-determined degradation correction coefficient. For example, when it is found that the degradation rate of the battery under certain specific working conditions deviates from the model prediction, the compensated degradation state parameter can be corrected by adjusting the degradation correction coefficient to obtain a more accurate corrected degradation state parameter. Similarly, input the interpolated state of health-life correlation parameter into the state of health correction module, adjust and output the corrected state of health parameter through the health correction coefficient, and input the coupled life prediction parameter into the life prediction correction module, adjust and output the corrected remaining life prediction parameter through the life correction coefficient. These corrected parameters can further improve the accurate description of the battery state.

[0079] Step S1347: Perform output consistency constraints on the corrected degradation state parameter, the corrected state of health parameter, and the corrected remaining life prediction parameter to generate the degradation state parameter, the state of health parameter, and the remaining life prediction parameter.

[0080] Among them, the output consistency constraints include:

[0081] Perform reverse correlation verification on the corrected health status parameters and the corrected remaining life prediction parameters. If the health status parameters are lower than the health threshold and the remaining life prediction parameters are synchronously lower than the life threshold, it is determined that the consistency is passed.

[0082] If the threshold status of the health status parameters and the remaining life prediction parameters is inconsistent, perform dynamic proportional scaling on the corrected health status parameters and the corrected remaining life prediction parameters until the reverse correlation condition is satisfied.

[0083] Perform trend matching verification on the corrected degradation status parameters and the corrected health status parameters. If the increasing rate of the degradation status parameters exceeds the decreasing rate of the health status parameters, apply a rate limiting factor to the corrected degradation status parameters to forcefully match the decreasing trend of the health status parameters.

[0084] Take the parameter combination that has passed the reverse correlation verification and the trend matching verification as the finally output degradation status parameters, health status parameters, and remaining life prediction parameters.

[0085] Specifically, in the battery management of an electric vehicle, if the health status parameters are lower than the health threshold and the remaining life prediction parameters are synchronously lower than the life threshold. For example, if the health status parameter indicates that the effective capacity of the battery has dropped below the set lower limit of the healthy capacity, and at the same time the remaining life prediction parameter indicates that the remaining charge-discharge cycles of the battery are also lower than the set lower limit of the life, then it is determined that the consistency is passed. However, if the threshold status of the health status parameters and the remaining life prediction parameters is inconsistent, for example, the health status parameter shows that the battery is still in a good state, but the remaining life prediction parameter shows that the remaining life is extremely low. In this case, perform dynamic proportional scaling on the corrected health status parameters and the corrected remaining life prediction parameters. According to the pre-set scaling rules and the actual operation data of the battery, adjust these two parameters until the reverse correlation condition is satisfied. At the same time, perform trend matching verification on the corrected degradation status parameters and the corrected health status parameters. During the use of an electric vehicle, as the battery is used, the degradation status parameters should show a certain trend relationship with the health status parameters. If the increasing rate of the degradation status parameters exceeds the decreasing rate of the health status parameters, for example, the battery degrades too fast and the decrease rate of the health status parameter does not keep up, in this case, apply a rate limiting factor to the corrected degradation status parameters to forcefully match the decreasing trend of the health status parameters. Through these output consistency constraint operations, take the parameter combination that has passed the reverse correlation verification and the trend matching verification as the finally output degradation status parameters, health status parameters, and remaining life prediction parameters. These final parameters can accurately reflect the actual state of the power battery of the electric vehicle, providing a reliable basis for battery management, maintenance, and decision-making.

[0086] In a possible implementation, step S140 includes:

[0087] Step S141: Compare the degradation state parameter with a preset degradation threshold range. If the degradation state parameter exceeds the degradation threshold range, generate a first anomaly flag.

[0088] For example, in the battery management system of an electric vehicle, the preset degradation threshold range is determined based on the performance degradation data of a large number of batteries of the same type under normal usage conditions. Suppose for a certain model of electric vehicle power battery, the preset degradation threshold range is [0, 0.3]. This range represents the acceptable degradation degree range of the battery during normal use, such as after a certain mileage of driving and charge-discharge cycles. When the degradation state parameter of the battery is 0.35, it exceeds this set threshold range, indicating that the performance degradation of the battery has exceeded the normal expectation. This kind of exceeding may be caused by reasons such as excessive loss of internal chemical substances in the battery, severe damage to the electrode structure, or a malfunction of the battery management system. At this time, generate a first anomaly flag, which serves as a signal for the abnormal degradation situation of the battery and provides a basis for subsequent maintenance decisions.

[0089] Step S142: Calculate the trend deviation degree between the health state parameter and the historical health state baseline. If the trend deviation degree exceeds the dynamic deviation threshold, generate a second anomaly flag.

[0090] Throughout the entire life cycle of an electric vehicle, the health state of the battery is a key indicator. The historical health state baseline is obtained through statistical analysis of a large amount of health state data of batteries of the same model under normal usage conditions. For example, in the early stage of normal battery use, the health state parameter may remain at a relatively stable level. As the mileage of use and the number of charge-discharge cycles increase, the health state will gradually decline, but this decline process has a certain pattern. Suppose the historical health state baseline is a numerical range or a changing trend curve of the health state determined according to different usage stages. When calculating the trend deviation degree between the current battery's health state parameter and the historical health state baseline, multiple factors need to be considered, such as the change trends of the battery's effective capacity, charge-discharge efficiency, internal resistance, etc. For example, set the dynamic deviation threshold to 0.2. If the calculated trend deviation degree is 0.25, it means that there is a relatively large deviation between the health state of the battery and the historical health state under normal conditions. This deviation may be caused by factors such as abnormal temperature effects, overcharge and overdischarge during battery use, or imbalance between battery cells. Once the trend deviation degree exceeds the dynamic deviation threshold, generate a second anomaly flag, indicating that there is an abnormal change trend in the health state of the battery and further attention and handling are required.

[0091] Step S143: Conduct a relative position analysis of the remaining life prediction parameter and the median of the life distribution of batteries in the same batch. If the remaining life prediction parameter is lower than the life warning quantile, generate a third anomaly flag.

[0092] Specifically, batteries in the same batch have similar performance and life characteristics because they follow the same manufacturing process and quality standards during production. In the battery management of electric vehicles, the median of the life distribution can be obtained through life tests and statistical analysis of a large number of samples of batteries in the same batch. For example, for a certain batch of power batteries of electric vehicles, through the actual use and life monitoring of numerous battery samples, it is found that the median of the life distribution is 5 years, indicating that under normal use conditions, half of the batteries in this batch will have a life exceeding 5 years, and the other half will be less than 5 years. At the same time, the life warning quantile is set to 0.3, which means that when the remaining life prediction parameter of the battery is lower than 30% of the median of the life distribution of batteries in the same batch, a warning needs to be issued. Suppose the remaining life prediction parameter of the current battery is 1.5 years, and the median of the life distribution of batteries in the same batch is 5 years. 1.5 years is lower than 30% of 5 years (i.e., 1.5 years < 5 × 0.3 = 1.5 years). In this case, the system will generate a third anomaly flag, indicating that the remaining life of the battery has reached a relatively low level and may be approaching the service life limit, and timely measures need to be taken.

[0093] Step S144: According to the combination pattern of the first anomaly flag, the second anomaly flag, and the third anomaly flag, match the preset maintenance strategy knowledge base to generate the maintenance decision recommendation including the maintenance priority, the type of maintenance operation, and the maintenance time window.

[0094] Specifically, the preset maintenance strategy knowledge base is formulated according to different combinations of abnormal conditions, as well as the actual usage requirements and safety requirements of the battery. For example, if only the first abnormal mark appears, indicating that the battery is mainly abnormal in the degradation state, the possible maintenance decision suggestion is a medium maintenance priority. The type of maintenance operation can be to perform a deep charge and discharge cycle to attempt to restore the battery performance, and the maintenance time window can be set to be carried out within the next week. If both the first abnormal mark and the third abnormal mark appear simultaneously, this indicates that the battery not only has a degradation abnormality but also has a relatively low remaining life. At this time, the maintenance decision suggestion may be a high maintenance priority, and the type of maintenance operation may be to replace the battery module or conduct a comprehensive inspection and repair of the battery. The maintenance time window needs to be more urgent, such as within the next 24 hours. If all three abnormal marks appear, it means that the battery state is very bad, and it may be necessary to immediately stop using the vehicle and perform emergency maintenance. The maintenance priority is the highest, and the type of maintenance operation may be to directly replace the entire battery pack, and the maintenance time window is to be executed immediately. By matching the maintenance strategy knowledge base according to the combination pattern of abnormal marks in this way, reasonable and effective maintenance decision suggestions can be provided for the power battery of electric vehicles, ensuring the safety, reliability, and service life of the battery.

[0095] In a possible implementation manner, step S150 includes:

[0096] Step S151, according to the abnormal mark type in the abnormal detection result, select the corresponding weight update rule from the model parameter adjustment strategy library.

[0097] Specifically, the model parameter adjustment strategy library is established in advance and contains weight update strategies for different abnormal mark types. For example, when only the first abnormal mark appears in the abnormal detection result (that is, the degradation state parameter exceeds the preset degradation threshold interval), this indicates that the degradation of the battery is abnormal. The weight update rule selected from the model parameter adjustment strategy library may focus on adjusting the weights of the model parameters related to the degradation state. In the battery state monitoring model, the convolutional temporal encoding layer is responsible for capturing local patterns in the battery state data and plays an important role in detecting the degradation characteristics of the battery. Therefore, the weight update rule for this abnormal mark type may increase the adjustment amplitude of the weights of the convolutional kernels related to the degradation characteristics in the convolutional temporal encoding layer. For example, if a certain convolutional kernel is mainly used to detect the relationship between the battery voltage fluctuation and the degradation degree, when the first abnormal mark appears, the adjustment amplitude of the weight of this convolutional kernel may be larger than normal, so that the model can more accurately capture the degradation characteristics of the battery.

[0098] Step S152, according to the maintenance priority in the maintenance decision suggestion, determine the learning rate adjustment amplitude and regularization strength of the battery state monitoring model.

[0099] For the case where the first anomaly marker and the third anomaly marker appear simultaneously (i.e., the degradation state is abnormal and the remaining life prediction parameter is lower than the life warning quantile), the weight update rule in the model parameter adjustment strategy library may be more comprehensive. It will not only increase the weight adjustment related to degradation in the convolutional temporal encoding layer, but also adjust the parameter weights related to remaining life prediction in the bidirectional recurrent prediction layer. Because this anomaly combination indicates that the battery not only has the current degradation problem, but also involves a serious shortening of the remaining life, it is necessary to simultaneously adjust the model parameter weights related to these two aspects to improve the model's prediction ability for such complex situations.

[0100] Specifically, the maintenance priority reflects the urgency and importance of the battery state. If the maintenance priority in the maintenance decision recommendation is high, this means that the battery state is already very urgent and the model needs to be adjusted as soon as possible to improve the prediction accuracy. At this time, to accelerate the model's convergence speed, the learning rate will be increased significantly. For example, if the original learning rate is 0.001, in the case of high maintenance priority, the learning rate may be increased to 0.005. At the same time, to prevent the model from overfitting during the rapid adjustment process, the regularization strength will also be appropriately enhanced. Assuming the original regularization strength is 0.1, it may be increased to 0.2. A higher regularization strength can limit the excessive growth of model parameters and ensure that the model can still maintain good generalization ability while quickly learning new data features.

[0101] If the maintenance priority is medium, the adjustment range of the learning rate will be relatively small, for example, adjusted from 0.001 to 0.002, and the regularization strength will only be moderately adjusted, such as from 0.1 to 0.15. This adjustment range can enable the model to be optimized to a certain extent according to new data information and avoid unstable model performance caused by excessive adjustment.

[0102] Step S153, construct a virtual feedback data set according to the maintenance operation type and the maintenance time window, and perform hybrid sampling on the virtual feedback data set and the historical training data.

[0103] Different maintenance operation types and maintenance time windows will have different impacts on the battery state. For example, when the maintenance operation type is deep charge and discharge repair and the maintenance time window is 24 hours, based on the understanding of the characteristics and expected changes of the battery during the deep charge and discharge repair process, a virtual feedback dataset is constructed. This dataset contains the expected change values of various state parameters of the battery during the deep charge and discharge repair, such as the change trends of parameters such as battery voltage, current, and temperature during the repair process. At the same time, considering that the maintenance time window is 24 hours, the time series data in the dataset will be constructed according to this time range to reflect the state changes of the battery under this specific maintenance operation and time range.

[0104] Suppose the mixing ratio is 1:3. The purpose of doing this is to introduce the new information contained in the virtual feedback dataset on the basis of using historical training data, so that the model can learn the change rules of the battery state under specific maintenance operations and time windows. For example, the historical training data contains a large amount of state data of the battery during normal use, while the virtual feedback dataset supplements the state data under special maintenance conditions. Through mixed sampling, the model can learn and optimize on a more comprehensive data basis.

[0105] Step S154, based on the dataset after mixed sampling, the updated learning rate, and the regularization strength, perform backpropagation optimization on the parameter weights of the convolutional time series encoding layer, the attention weight allocation layer, and the bidirectional recurrent prediction layer of the battery state monitoring model.

[0106] In the electric vehicle battery state monitoring model, backpropagation optimization is a process of gradually adjusting the model parameter weights according to the sample errors in the dataset after mixed sampling. For the convolutional time series encoding layer, taking one of the convolutional kernels as an example, during the backpropagation process, according to the error between the battery state data in the mixed sampling dataset and the model prediction result, calculate the contribution degree of this convolutional kernel to the error, and then adjust the weight of this convolutional kernel according to the updated learning rate. For example, if in the mixed sampling dataset, there is a large error between the actual degradation state of a certain battery sample and the predicted degradation state of the model, and this error is related to the features extracted by a certain convolutional kernel in the convolutional time series encoding layer, then the weight of this convolutional kernel will be adjusted according to the direction and magnitude of the error to reduce the prediction error.

[0107] Step S155, freeze the parameter weights of the multi-task output layer in the battery state monitoring model, and only perform iterative updates on the parameter weights of the convolutional time series encoding layer, the attention weight allocation layer, and the bidirectional recurrent prediction layer until the prediction error converges.

[0108] For the attention weight allocation layer, during the backpropagation optimization process, the attention weights are adjusted according to the importance of battery state data at different time points in the mixed sampling dataset and the difference between the model's attention to this data and the actual situation. For example, if in the battery state change after a certain maintenance operation, the battery state data at a specific time point (such as the voltage change at the initial stage of charging) is more important for the overall battery state judgment than the model previously thought, then during the backpropagation optimization, the attention weight corresponding to this time point will be increased, so that the model can more accurately focus on the state information of these key time points in subsequent predictions.

[0109] For the bidirectional recurrent prediction layer, during the backpropagation optimization, the weights in the bidirectional recurrent neural network units are adjusted according to the error between the forward and backward state change data of the battery in the mixed sampling dataset and the model prediction results. For example, during the charge and discharge cycle of the battery, if the prediction error of the remaining battery life by the model is large, and this error is related to the weights when the bidirectional recurrent prediction layer calculates the degradation accumulation and the life residual coefficient, then these weights will be adjusted according to the error situation to improve the prediction accuracy of the remaining battery life by the model.

[0110] During the entire backpropagation optimization process, the parameter weights of the multi-task output layer in the battery state monitoring model are frozen, and only the parameter weights of the convolutional temporal encoding layer, the attention weight allocation layer, and the bidirectional recurrent prediction layer are iteratively updated until the prediction error converges. The parameter weights of the multi-task output layer are frozen because this layer is mainly responsible for jointly mapping the results processed by the previous layers and outputting the final degradation state parameters, health state parameters, and remaining life prediction parameters. Its parameter weights are relatively stable in the overall structure of the model and do not need to be adjusted frequently. The parameter weights of the convolutional temporal encoding layer, the attention weight allocation layer, and the bidirectional recurrent prediction layer are related to core operations such as feature extraction of battery state data, key time point enhancement, and cumulative degradation analysis. By continuously iteratively updating these weights, the model can better adapt to the changes in the battery state and improve the prediction accuracy. During the iterative update process, the prediction error of the model is continuously monitored. When the prediction error gradually decreases and reaches the convergence criterion (such as the change in the prediction error is less than a certain set threshold), the iterative update is stopped. At this time, the prediction accuracy of the model is updated and improved, and it can more accurately monitor and predict the state of the electric vehicle power battery.

[0111] In a possible implementation manner, before step S110, the method further includes:

[0112] Step S111, configuring a distributed sensing node network connected to the target battery, where the distributed sensing node network includes a voltage acquisition node, a current acquisition node, a temperature acquisition node, and a time synchronization node.

[0113] To accurately monitor the state of an electric vehicle's power battery during charge-discharge cycles, it is first necessary to configure a distributed sensing node network connected to the target battery. This distributed sensing node network is a system specifically constructed to obtain comprehensive battery state information, which includes voltage acquisition nodes, current acquisition nodes, temperature acquisition nodes, and time synchronization nodes.

[0114] Step S112: Capture the voltage sequence through the voltage acquisition nodes at a first sampling frequency and capture the current sequence through the current acquisition nodes at a second sampling frequency.

[0115] The voltage acquisition nodes play an important role in obtaining the battery voltage sequence throughout the network. It captures the battery voltage at a first sampling frequency. During the operation of an electric vehicle, the battery voltage fluctuates continuously with the vehicle's driving state, charging state, and the battery's own performance. For example, when the electric vehicle is accelerating, the battery needs to provide a larger current, and at this time, the chemical reactions inside the battery intensify, and the battery voltage may drop briefly; while when the vehicle is driving at a constant speed or during charging, the battery voltage shows different change trends. The voltage acquisition nodes capture the voltage value at a relatively high sampling frequency, such as once every 5 milliseconds, to ensure that these voltage changes can be accurately captured. Through continuous acquisition, a voltage sequence reflecting the change of battery voltage over time is formed, and this voltage sequence contains rich battery state information, such as the state of charge of the battery and the degree of polarization inside the battery.

[0116] At the same time, the current acquisition nodes capture the current sequence at a second sampling frequency. During the operation of an electric vehicle, the change of current is closely related to the vehicle's power demand. When the vehicle starts, climbs a slope, or drives at high speed, it requires a large amount of power, and the battery will output a large current; while when the vehicle decelerates or coasts, the battery may be in a charging state (through the regenerative braking system), and at this time, the direction of the current will change and flow into the battery. The current acquisition nodes record these current changes at a specific sampling frequency, such as once every 3 milliseconds. Such frequent sampling can accurately reflect the charge-discharge conditions of the battery under different working conditions, and the formed current sequence can be used to analyze important parameters such as the charge-discharge efficiency of the battery and the internal resistance of the battery.

[0117] Step S113: Capture the temperature sequence through the temperature acquisition nodes at a third sampling frequency and perform microsecond-level alignment of the sampling moments of the voltage acquisition nodes, current acquisition nodes, and temperature acquisition nodes through the time synchronization nodes.

[0118] The temperature acquisition node captures the temperature sequence at the third sampling frequency. The battery temperature is a crucial state indicator because the battery generates heat during charging and discharging, and excessive temperature may affect the battery's performance, lifespan, and even safety. During the actual operation of an electric vehicle, for example, after rapid charging or long - time high - speed driving, the battery temperature will increase significantly. The temperature acquisition node collects the temperatures at different positions of the battery at an appropriate sampling frequency, such as once every 10 milliseconds. Since the temperature distribution inside the battery may be uneven, multiple temperature acquisition points can obtain the overall temperature information of the battery, and the changes in these collected temperature values over time constitute the temperature sequence. This temperature sequence helps monitor the battery's thermal management and detect potential local overheating problems.

[0119] In this distributed sensor node network, the time synchronization node plays a crucial role. It aligns the sampling moments of the voltage acquisition node, current acquisition node, and temperature acquisition node at the microsecond level. In the complex operating environment of an electric vehicle, precise time synchronization is the key to ensuring the validity and relevance of data. For example, when the vehicle accelerates suddenly at a certain moment, the voltage, current, and temperature of the battery will all change simultaneously. If the sampling moments of these data are not precisely aligned, then when analyzing the battery state subsequently, it is impossible to accurately determine the relationships between these parameters. The time synchronization node ensures that the voltage values, current values, and temperature values collected at the same moment can accurately reflect the true state of the battery at that moment through microsecond - level alignment.

[0120] Step S114, send the voltage sequence, current sequence, temperature sequence, and timestamp sequence to the edge computing gateway through an encrypted transmission protocol for temporary caching and integrity verification.

[0121] The use of the encrypted transmission protocol is to ensure the security and confidentiality of data, preventing data from being tampered with or stolen during transmission. In the environment of an electric vehicle, there may be various electromagnetic interferences and potential network security threats, and encrypted transmission can effectively protect the integrity of the battery state data. The edge computing gateway, as a data transfer station, receives data from each acquisition node and performs temporary caching. This is because during the data acquisition process, there may be fluctuations in data transmission or rate mismatches between the acquisition devices and the subsequent analysis system, and temporary caching can ensure that data is not lost. At the same time, the edge computing gateway will perform integrity verification on the received data. Integrity verification is to calculate the data through specific algorithms, such as checksum algorithms or hash algorithms, and compare the calculated value with the pre - calculated verification value at the sending end. If the data has not been corrupted or tampered with during transmission, then the verification values should match.

[0122] Step S115, in response to the passing of the integrity check, batch read the real-time operation data from the edge computing gateway.

[0123] Only when the integrity check is successful can it be ensured that the acquired data is accurate and reliable. Batch reading data from the edge computing gateway can improve the efficiency of data acquisition, obtaining data at multiple time points at once, which is convenient for subsequent comprehensive analysis of the battery state. These batch-read real-time operation data, including voltage sequences, current sequences, temperature sequences, and timestamp sequences, will serve as the basis for further analyzing the battery state, such as performing multi-dimensional feature extraction, inputting into the battery state monitoring model, etc., so as to achieve precise monitoring and management of the power battery state of electric vehicles.

[0124] In a possible implementation manner, step S121 includes:

[0125] Step S1211, perform extreme point detection on the voltage sequence, identify all local maximum points and local minimum points, and calculate the absolute difference between adjacent maximum and minimum values, taking the maximum absolute difference as the voltage fluctuation extreme value.

[0126] During a complete driving and charging cycle of an electric vehicle, the battery voltage is constantly changing. For example, when the vehicle starts, the battery needs to instantaneously output a large current, and the voltage will have an obvious drop. Subsequently, during the stable driving process of the vehicle, the voltage will fluctuate within a certain range. When the vehicle performs regenerative braking (charging process), the voltage will rise again. During this process, through the extreme point detection algorithm, all local maximum points and local minimum points can be identified. For example, during a certain driving process, the detected local maximum point is 4.1 volts, and the immediately following local minimum point is 3.9 volts. The absolute difference between the adjacent maximum and minimum values is calculated to be 0.2 volts. In the entire voltage sequence, such calculations are continuously performed, and finally the maximum absolute difference is taken as the voltage fluctuation extreme value. Suppose that during the entire driving and charging cycle, the maximum absolute difference between adjacent maximum and minimum values is 0.3 volts. This 0.3 volts is the voltage fluctuation extreme value of the battery during this cycle. This extreme value can reflect the maximum amplitude of voltage fluctuation of the battery under different working conditions and is an important indicator of battery stability. If the voltage fluctuation extreme value is too large, it may mean that the chemical reaction inside the battery is unstable or there is a problem with the battery management system.

[0127] Step S1212, perform a first-order difference operation on the voltage sequence to obtain a differential voltage sequence, and calculate the arithmetic mean of the differential voltage sequence within a sliding time window as the average voltage change rate.

[0128] The rate of change of the battery voltage can reflect the charge and discharge state as well as the health status of the battery. When the vehicle is in the acceleration stage, the battery discharges rapidly and the voltage drops relatively fast; while when driving at a constant speed, the voltage drops relatively slowly. By performing a first-order difference operation on the voltage sequence, a difference voltage sequence is obtained. For example, within a certain 10-second time interval, the values of the voltage sequence are 4.0 volts, 3.95 volts, 3.9 volts, etc. After performing the first-order difference operation, the difference voltage sequence obtained may be -0.05 volts, -0.05 volts, etc. Then, calculate the arithmetic mean of this difference voltage sequence within a sliding time window as the mean value of the voltage change rate. Assume that the sliding time window is set to 5 seconds, and the values of the difference voltage sequence within this window are -0.03 volts, -0.04 volts, -0.03 volts, etc. Their arithmetic mean is -0.033 volts, and this -0.033 volts is the mean value of the voltage change rate within this 5-second sliding time window. By calculating the mean values of the voltage change rate within different sliding time windows, the voltage change trend of the battery at different time periods can be comprehensively understood, so as to evaluate the performance and state of the battery.

[0129] Step S1213: Detect the plateau interval of the voltage sequence, identify the maximum continuous time period during which the rate of change of the voltage continuously remains lower than the plateau threshold, and use the length of the maximum continuous time period as the duration of the voltage plateau.

[0130] During the charging process of the battery, especially when transitioning from the constant current charging stage to the constant voltage charging stage, a voltage plateau phenomenon will occur. At this time, the rate of change of the voltage continuously remains lower than the plateau threshold. For example, the plateau threshold is set to 0.01 volts / second. During the charging process, the interval during which the rate of change of the voltage continuously remains lower than this threshold is identified through a detection algorithm. Assume that during a certain stage of charging, from the 100th second to the 150th second, the rate of change of the voltage is always lower than 0.01 volts / second. This 50-second maximum continuous time period is the duration of the voltage plateau. The duration of the voltage plateau is of great significance for evaluating the charging characteristics and health status of the battery. If the duration of the voltage plateau is abnormal, it may indicate that the state of the chemical substances inside the battery has changed, or there is a problem with the matching between the charging device and the battery.

[0131] In a possible implementation manner, step S129 includes:

[0132] Step S1291: Perform a first-order seasonal difference operation on each feature dimension in the dynamic feature subset to eliminate the trend component and retain the periodic fluctuation component.

[0133] During the operation of an electric vehicle, many characteristics related to the battery state have certain trends and periodicities. Taking the dynamic characteristic of the battery charge and discharge efficiency as an example, as the number of battery usage times increases and the battery ages, its charge and discharge efficiency generally shows a downward trend, but there are also periodic fluctuations under different seasons or driving conditions. Through first-order seasonal differencing operations, for example, for the charge and discharge efficiency data represented by a time series, calculating the differences between adjacent time points can eliminate this long-term downward trend and only retain the periodic fluctuation components. For other dynamic characteristic dimensions, such as the temperature change rate of the battery under different working conditions, the same first-order seasonal differencing operation is also used, so that each characteristic dimension removes the trend component and only leaves the information related to the periodic changes of the battery state.

[0134] Step S1292: Perform a fast Fourier transform on the differenced characteristic dimensions, extract the dominant frequency components, and construct a reference phase signal based on the dominant frequency components.

[0135] Various state characteristics of the battery have different manifestations in the frequency domain. For example, some periodic behaviors of the battery may be more obvious at specific frequencies. Taking the dynamic characteristic of battery temperature change as an example, by performing a fast Fourier transform on the differenced temperature change rate data, it can be transformed from the time domain to the frequency domain. In the frequency domain, the dominant frequency component with the highest energy or the greatest influence on the characteristic can be found. Suppose in the frequency domain analysis of this temperature change rate, it is found that the frequency component of 0.1 Hz has the highest energy, and this 0.1 Hz is the dominant frequency component. A reference phase signal is constructed based on this dominant frequency component, and this reference phase signal can represent a reference phase state of the characteristic in the frequency domain. For other dynamic characteristic dimensions, the fast Fourier transform is also performed in the same way to find their respective dominant frequency components and construct the corresponding reference phase signals.

[0136] Step S1293: Calculate the cyclic phase difference between the phase components of each characteristic dimension and the reference phase signal to generate a set of phase offsets.

[0137] There may be phase differences between different dynamic characteristic dimensions. Taking the two dynamic characteristics of the battery charge and discharge current and the battery temperature as an example, their phases may be inconsistent due to various factors such as chemical reactions and heat conduction inside the battery. By calculating the cyclic phase difference between the phase components of each characteristic dimension and the reference phase signal, a set of phase offsets can be obtained. For example, for the charge and discharge current characteristic dimension, the calculated phase offset from the reference phase signal is 0.2 radians; for the battery temperature characteristic dimension, the calculated phase offset is -0.1 radians, etc. This set of phase offsets can quantify the degree of phase deviation of each characteristic dimension relative to the reference phase signal.

[0138] Step S1294, perform time-domain translation compensation on each feature dimension according to the set of phase offsets, so that the phase difference of the dominant frequency components of all feature dimensions is less than a preset phase tolerance.

[0139] To more accurately analyze the relationship between different feature dimensions, it is necessary to align the phases. For example, the preset phase tolerance is 0.1 radian. For the charge and discharge current feature dimension, since its phase offset is 0.2 radians, it is necessary to perform time-domain translation compensation on it so that the phase difference between its dominant frequency component and the phase of the reference phase signal is less than 0.1 radian. Similarly, perform such time-domain translation compensation operations on other dynamic feature dimensions such as battery temperature. In this way, it is ensured that the dominant frequency components of all dynamic feature dimensions are as close as possible in phase for subsequent analysis and processing.

[0140] Step S1295, perform amplitude normalization and timestamp resampling on the feature dimensions after phase compensation to generate the dynamically aligned feature subset.

[0141] Finally, after completing the phase compensation, since the amplitudes of different feature dimensions may vary greatly, this will affect the subsequent comprehensive analysis of the battery state. For example, the amplitude of the charge and discharge current of the battery may be in the tens of amperes, while the amplitude of the battery temperature may be in the tens of degrees Celsius. Through amplitude normalization, these different amplitude ranges are adjusted to a unified standard range. At the same time, since the time series of the data may have been adjusted during the previous processing, it is necessary to perform timestamp resampling to ensure the consistency of each feature dimension in time. The feature dimensions after amplitude normalization and timestamp resampling constitute the dynamically aligned feature subset. This dynamically aligned feature subset has better consistency in both time and frequency, can more accurately reflect the characteristic information of the battery in different states, and provides a more reliable data basis for subsequent battery state monitoring and analysis.

[0142] In a possible implementation manner, step S133 includes:

[0143] Step S1331, divide the weighted spatio-temporal feature map into a forward input sequence and a backward input sequence according to the time step.

[0144] In this embodiment, during the actual driving, charging, and discharging processes of an electric vehicle, the time step is an important time interval unit for analyzing the battery state. For example, with a time step of every 5 minutes, the weighted spatio-temporal feature map contains the feature information of the battery at different times and spaces (where the space here can be understood as different physical positions of the battery or different state parameter spaces). After segmentation according to this time step, the forward input sequence can be regarded as starting from the initial state of the battery (such as the state of a new vehicle or the initial state after a full charge), and as time progresses forward, it contains the feature information of the battery at each time step. This sequence reflects the trend of the battery state change during normal use. For example, as the driving mileage and the number of charging and discharging cycles increase, the changes in features such as battery voltage, current, and temperature. The reverse input sequence starts from a later time point (such as the current time or the time point after the last deep discharge) and traces back to an earlier state of the battery. This sequence contains the state change information in the opposite direction to the forward sequence. For example, when the battery traces back from the current state to a previous state, the feature information such as the charging recovery process and voltage recovery of the battery at different time steps.

[0145] Step S1332: Recursively calculate the hidden state of the forward input sequence through the forward recurrent neural network unit, calculate the degradation increment at each time step, and accumulate to obtain the degradation cumulative amount.

[0146] During each driving, charging, and discharging cycle of an electric vehicle, the battery will experience a certain degree of degradation. When using the forward recurrent neural network unit to process the forward input sequence, taking the capacity degradation of the battery as an example, within each time step, according to the battery voltage, current, temperature, and other relevant feature information in the weighted spatio-temporal feature map, through the neuron calculation mechanism inside the forward recurrent neural network unit, calculate the capacity degradation increment of the battery at this time step. For example, within a 5-minute time step, based on factors such as the average discharge current, voltage drop amplitude, and battery temperature of the battery during this period, through the calculation methods such as weights and activation functions in the pre-trained forward recurrent neural network unit, the capacity degradation increment of the battery is obtained as 0.01 (assuming a certain capacity unit is used here, such as ampere-hour). As time goes by, such calculations are performed for each time step, and then these degradation increments are accumulated in sequence. Suppose after 10 time steps, the degradation increments calculated for each time step are 0.01, 0.015, 0.02, etc. The 0.18 obtained by accumulating these values is the degradation cumulative amount within these 10 time steps. This degradation cumulative amount reflects the overall degradation degree of the battery during this period, comprehensively considering the performance degradation of the battery affected by various factors at different time steps.

[0147] Step S1333: Through the reverse recurrent neural network unit, perform hidden state backtracking on the reverse input sequence, calculate the lifetime residual component at each time step, and weighted aggregate to obtain the lifetime residual coefficient.

[0148] For example, the remaining lifetime of a battery is affected by various factors. Through reverse analysis, the influence relationship of these factors on the remaining lifetime can be better captured. When the reverse recurrent neural network unit processes the reverse input sequence, it is also based on various feature information in the weighted spatio-temporal feature map. For example, during the process of the battery backtracking from the current state to the previous state, within each time step, according to factors such as the battery's charging efficiency, voltage recovery speed, and temperature change, calculate the lifetime residual component related to the remaining lifetime. Suppose that within a certain time step, the calculated lifetime residual component based on these factors is 0.05 (here it is assumed to be a dimensionless value after normalization, representing a relative quantity related to the remaining lifetime). After performing such calculations for each time step, different weights need to be assigned according to the importance of the influence of different time steps on the remaining lifetime, and then these weighted lifetime residual components are weighted aggregated to obtain the lifetime residual coefficient. For example, for the earlier time steps, since their influence on the remaining lifetime is relatively small, the assigned weight is 0.1; while for the relatively recent time steps, the assigned weight is 0.3. The lifetime residual coefficient obtained through such a weighted aggregation method can comprehensively reflect the feature information related to the remaining lifetime of the battery during the reverse analysis process.

[0149] Step S1334: Cross-connect the final hidden state of the forward recurrent neural network unit and the initial hidden state of the reverse recurrent neural network unit to generate a joint hidden state vector.

[0150] The final hidden state obtained by the forward recurrent neural network unit after processing the forward input sequence contains the final state information of the battery in the forward time series, such as the state characteristics of the battery after a series of usage and degradation; while the initial hidden state of the backward recurrent neural network unit when processing the backward input sequence contains the initial state information of the battery at the start of the backward trace from the current state. Although these two state information come from different analysis directions, they are both of great significance for comprehensively understanding the battery state. By cross-layer connecting these two hidden states, they are combined into a joint hidden state vector. For example, the final hidden state of the forward recurrent neural network unit is a vector containing multiple elements (such as numerical values representing state-related information such as battery capacity, internal resistance, voltage, etc.), assumed to be [0.1, 0.2, 0.3]; the initial hidden state of the backward recurrent neural network unit is [0.4, 0.5, 0.6], and the joint hidden state vector obtained by connecting them is [0.1, 0.2, 0.3, 0.4, 0.5, 0.6]. This joint hidden state vector integrates the key state information of the battery in the forward and backward analyses, providing a rich information source for more accurately correcting the degradation accumulation amount and the life residual coefficient in the subsequent steps.

[0151] Step S1335, perform a non-linear transformation on the joint hidden state vector through a fully connected layer, and output the correction factors for the degradation accumulation amount and the life residual coefficient.

[0152] The fully connected layer has a powerful non - linear mapping ability. After the combined hidden state vector enters the fully connected layer, each neuron in the fully connected layer is connected to each element in the combined hidden state vector, and a non - linear transformation is performed through calculation mechanisms such as the weights and activation functions inside the neurons. For example, the neuron weights in the fully connected layer are obtained through pre - training. According to these weights, the elements in the combined hidden state vector are weighted and summed, and then a non - linear transformation is performed through an activation function (such as the ReLU function or other suitable non - linear activation functions). After such a calculation process, the output result is the correction factor for the degradation cumulative amount and the life residual coefficient. This correction factor can adjust the previously calculated degradation cumulative amount and life residual coefficient according to the comprehensive information in the combined hidden state vector. For example, if the correction factor is 1.2, then the previously calculated degradation cumulative amount and life residual coefficient will be multiplied by this correction factor respectively, so as to obtain more accurate degradation cumulative amount and life residual coefficient, in order to better reflect the actual degradation situation and remaining life situation of the battery. This operation mode of the bidirectional recurrent prediction layer comprehensively considers the state information of the battery in the forward and reverse time series, as well as the hidden state information in different analysis directions, and outputs the correction factor through non - linear transformation. Finally, it can generate the degradation cumulative amount and life residual coefficient more accurately, laying a solid foundation for the subsequent accurate evaluation of the battery state parameters (such as degradation state parameters, health state parameters, and remaining life prediction parameters).

[0153] Figure 2 The hardware structure diagram of the machine - learning - based battery state monitoring and analysis system 100 for implementing the above - mentioned machine - learning - based battery state monitoring and analysis method provided by the embodiment of the present invention is shown, as Figure 2 shown, the machine - learning - based battery state monitoring and analysis system 100 may include a processor 110, a machine - readable storage medium 120, a bus 130, and a communication unit 140.

[0154] The machine - readable storage medium 120 can store data and / or instructions. In some embodiments, the machine - readable storage medium 120 can store data obtained from an external terminal. In some embodiments, the machine - readable storage medium 120 can store the data and / or instructions used by the machine - learning - based battery state monitoring and analysis system 100 to execute or use to complete the exemplary methods described in the present invention.

[0155] In a specific implementation process, one or more processors 110 execute computer-executable instructions stored in a machine-readable storage medium 120, enabling the processors 110 to execute the machine learning-based battery state monitoring and analysis method as described in the above method embodiments. The processors 110, the machine-readable storage medium 120, and the communication unit 140 are connected through a bus 130, and the processors 110 can be used to control the transceiver actions of the communication unit 140.

[0156] For the specific implementation process of the processors 110, reference can be made to the various method embodiments executed by the above machine learning-based battery state monitoring and analysis system 100. Their implementation principles and technical effects are similar, and will not be elaborated here in this embodiment.

[0157] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the above machine learning-based battery state monitoring and analysis method is implemented.

[0158] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the previous description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. A battery status monitoring and analysis method based on machine learning, characterized in that: The method comprises: Acquire real-time operation data of the target battery in the charge and discharge cycle, wherein the real-time operation data includes a voltage sequence, a current sequence, a temperature sequence, and a timestamp sequence; Performing multi-dimensional feature extraction on the real-time operation data to generate an original feature set, and performing dynamic time series processing on the original feature set to obtain a standardized feature matrix; Inputting the standardized feature matrix into a pre-trained battery state monitoring model to generate degradation state parameters, health state parameters and remaining life prediction parameters of the target battery; Generate an abnormality detection result and maintenance decision suggestion for the target battery according to the degradation state parameter, health state parameter and remaining life prediction parameter; Based on the abnormality detection results and maintenance decision suggestions, adaptively adjust the parameter weights of the battery status monitoring model to update the prediction accuracy of the battery status monitoring model; The step of performing dynamic time series processing on the original feature set to obtain a standardized feature matrix includes: Perform missing value filling and noise filtering on each feature dimension in the original feature set to generate a denoised feature set; According to the statistical distribution characteristics of each feature dimension in the denoising feature set, the denoising feature set is divided into a static feature subset and a dynamic feature subset; Performing sliding window mean smoothing processing on the static feature subset to generate a static smooth feature subset; Performing time series difference transformation and phase alignment processing on the dynamic feature subset to generate a dynamic alignment feature subset; The static smooth feature subset and the dynamic alignment feature subset are cross-dimensionally fused according to a timestamp sequence to generate the standardized feature matrix.

2. The battery status monitoring and analysis method based on machine learning according to claim 1 is characterized in that: The performing multi-dimensional feature extraction on the real-time operation data to generate an original feature set includes: Extracting voltage fluctuation extreme value, voltage change rate mean and voltage platform duration from the voltage sequence; Extracting current charge and discharge efficiency, current peak interval period and current decay slope from the current sequence; Extracting temperature gradient distribution, temperature rise rate threshold and temperature balance index from the temperature sequence; Extracting the charge and discharge cycle length, the rest phase duration and the cycle interval consistency coefficient from the timestamp sequence; The voltage fluctuation extreme value, voltage change rate mean, voltage platform duration, current charge and discharge efficiency, current peak interval period, current decay slope, temperature gradient distribution, temperature rise rate threshold, temperature balance index, charge and discharge cycle length, static stage length and cycle interval consistency coefficient are normalized and spliced ​​to generate the original feature set.

3. The battery status monitoring and analysis method based on machine learning according to claim 1 is characterized in that: The step of inputting the standardized feature matrix into a pre-trained battery state monitoring model to generate degradation state parameters, health state parameters and remaining life prediction parameters of the target battery includes: The standardized feature matrix is ​​captured locally through a convolutional temporal coding layer in the battery state monitoring model to generate a primary spatiotemporal feature map; Through the attention weight allocation layer in the battery status monitoring model, the primary spatiotemporal feature graph is strengthened at key time points to generate a weighted spatiotemporal feature graph; Through the bidirectional cyclic prediction layer in the battery state monitoring model, forward cumulative degradation analysis and reverse life residual fitting are performed on the weighted spatiotemporal characteristic graph to generate degradation accumulation and life residual coefficient; The degradation accumulation and life residual coefficient are jointly mapped through the multi-task output layer in the battery state monitoring model, and the degradation state parameters, health state parameters and remaining life prediction parameters are synchronously output.

4. The battery status monitoring and analysis method based on machine learning according to claim 1 is characterized in that: The generating, according to the degradation state parameter, health state parameter and remaining life prediction parameter, an abnormality detection result and a maintenance decision suggestion for the target battery includes: Comparing the degradation state parameter with a preset degradation threshold interval, and generating a first abnormality mark if the degradation state parameter exceeds the degradation threshold interval; Calculate the trend deviation between the health status parameter and the historical health status baseline, and if the trend deviation exceeds a dynamic deviation threshold, generate a second abnormal mark; Performing relative position analysis on the remaining life prediction parameter and the median of the life distribution of batteries in the same batch, and generating a third abnormal mark if the remaining life prediction parameter is lower than the life warning quantile; According to the combination pattern of the first abnormal mark, the second abnormal mark and the third abnormal mark, a preset maintenance strategy knowledge base is matched to generate the maintenance decision suggestion including the maintenance priority, the maintenance operation type and the maintenance time window; The method of adaptively adjusting the parameter weights of the battery status monitoring model based on the abnormality detection results and the maintenance decision suggestions, and updating the prediction accuracy of the battery status monitoring model, includes: According to the abnormality mark type in the abnormality detection result, select a corresponding weight update rule from the model parameter adjustment strategy library; Determining a learning rate adjustment range and a regularization strength of the battery status monitoring model according to the maintenance priority in the maintenance decision suggestion; According to the maintenance operation type and the maintenance time window, a virtual feedback data set is constructed, and the virtual feedback data set is mixed and sampled with historical training data; Based on the mixed sampled data set, the updated learning rate and the regularization strength, back-propagation optimization is performed on the parameter weights of the convolutional temporal coding layer, the attention weight allocation layer and the bidirectional recurrent prediction layer of the battery state monitoring model; The parameter weights of the multi-task output layer in the battery state monitoring model are frozen, and only the parameter weights of the convolutional temporal coding layer, the attention weight allocation layer, and the bidirectional recurrent prediction layer are iteratively updated until the prediction error converges.

5. The battery status monitoring and analysis method based on machine learning according to claim 1 is characterized in that: Before acquiring the real-time operation data of the target battery in the charge and discharge cycle, the method further includes: Configuring a distributed sensor node network connected to the target battery, the distributed sensor node network comprising a voltage acquisition node, a current acquisition node, a temperature acquisition node and a time synchronization node; Capturing the voltage sequence at a first sampling frequency through the voltage acquisition node, and capturing the current sequence at a second sampling frequency through the current acquisition node; The temperature acquisition node captures the temperature sequence at a third sampling frequency, and the time synchronization node performs microsecond-level alignment on the sampling moments of the voltage acquisition node, the current acquisition node, and the temperature acquisition node; Sending the voltage sequence, current sequence, temperature sequence and timestamp sequence to the edge computing gateway through an encrypted transmission protocol for temporary caching and integrity verification; In response to the integrity check passing, the real-time operation data is read in batches from the edge computing gateway.

6. The battery status monitoring and analysis method based on machine learning according to claim 2 is characterized in that: The step of extracting the voltage fluctuation extreme value, the voltage change rate mean value and the voltage platform duration from the voltage sequence includes: Perform extreme point detection on the voltage sequence, identify all local maximum points and local minimum points, and calculate the absolute difference between adjacent maximum and minimum values, and use the maximum absolute difference as the voltage fluctuation extreme value; Performing a first-order difference operation on the voltage sequence to obtain a differential voltage sequence, and calculating an arithmetic mean value of the differential voltage sequence within a sliding time window as the voltage change rate mean value; A platform interval detection is performed on the voltage sequence to identify a maximum continuous time period during which the voltage change rate is continuously lower than a platform threshold, and the length of the maximum continuous time period is used as the voltage platform duration.

7. The battery status monitoring and analysis method based on machine learning according to claim 1 is characterized in that: The step of performing time series difference transformation and phase alignment processing on the dynamic feature subset to generate a dynamic alignment feature subset includes: Performing a first-order seasonal difference operation on each feature dimension in the dynamic feature subset to eliminate trend components and retain periodic fluctuation components; Performing a fast Fourier transform on the feature dimension after differentiation, extracting the dominant frequency component, and constructing a reference phase signal based on the dominant frequency component; Perform cyclic phase difference calculation on the phase component of each characteristic dimension and the reference phase signal to generate a phase offset set; Performing time domain translation compensation on each characteristic dimension according to the phase offset set, so that the phase difference of the dominant frequency components of all characteristic dimensions is less than a preset phase tolerance; The feature dimension after phase compensation is amplitude normalized and timestamp resampled to generate the dynamic alignment feature subset.

8. The battery status monitoring and analysis method based on machine learning according to claim 3 is characterized in that: The forward cumulative degradation analysis and reverse life residual fitting are performed on the weighted spatiotemporal feature graph through the bidirectional cyclic prediction layer in the battery state monitoring model to generate the degradation accumulation and life residual coefficient, including: Splitting the weighted spatiotemporal feature map into a forward input sequence and a reverse input sequence according to the time step; Performing hidden state recursion on the forward input sequence through a forward recurrent neural network unit, calculating the degradation increment of each time step, and accumulating to obtain the degradation accumulation; Perform hidden state backtracking on the reverse input sequence through a reverse recurrent neural network unit, calculate the life residual component of each time step, and perform weighted aggregation to obtain the life residual coefficient; Cross-layer connection is performed between the final hidden state of the forward recurrent neural network unit and the initial hidden state of the reverse recurrent neural network unit to generate a joint hidden state vector; The joint hidden state vector is nonlinearly transformed through a fully connected layer to output a correction factor of the degradation accumulation amount and the life residual coefficient.

9. A battery status monitoring and analysis system based on machine learning, characterized in that: The battery status monitoring and analysis system based on machine learning includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the battery status monitoring and analysis method based on machine learning as described in any one of claims 1 to 8.

Images