A method and system for battery pack health prediction

Abstract

The present invention relates to the technical field of electrical performance detection, and particularly to a method and system for battery pack health prediction. The method includes the following steps: performing multi-modal sensing data acquisition on the battery pack to obtain voltage data, current data, temperature data, and acoustic vibration data as the original monitoring data set; preprocessing the original monitoring data set and extracting time-frequency domain features, and constructing a battery pack operation feature matrix; obtaining the historical operation data of the battery pack; and constructing a battery pack health status intelligent diagnosis model by using the historical operation data of the battery pack and the battery pack operation feature matrix. The present invention accurately monitors the health status of the battery pack, distinguishes the aging trend from the short-term fluctuation through multi-modal data acquisition, intelligent diagnosis model and time series decomposition, and combines health warning and pre-maintenance plan design, thereby improving the intelligent level and reliability of battery pack management.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical performance detection, and particularly to a method and system for battery pack health prediction. Background Art

[0002] With the rapid development of the new energy industry, lithium-ion batteries have been widely used in electric vehicles, energy storage systems, and consumer electronics fields. However, during long-term use, battery packs are affected by multiple factors such as cyclic aging, environmental impact, and usage patterns, resulting in gradual performance degradation and even safety accidents in severe cases. Early battery health prediction methods mainly relied on physical models, such as those based on equivalent circuit models (ECMs) and mechanism-based models (PBMs), which described the battery health state by establishing internal dynamics equations of the battery. However, such methods have a relatively complex description of the internal chemical reactions of the battery and require a large amount of experimental data for parameter calibration, making it difficult to adapt to battery packs of different types and working environments. Subsequently, data-driven methods have gradually become the mainstream of research, such as health prediction methods based on statistical analysis, machine learning, and deep learning. Statistical analysis methods analyze the changing trend of battery health state through historical data, but cannot effectively capture complex non-linear characteristics; traditional machine learning methods have improved the prediction accuracy, but rely on manual feature extraction and are difficult to fully mine the deep information in multi-modal data. Although certain progress has been made in the prior art, there are still deficiencies. Traditional methods rely on a single data source for modeling the battery health state and cannot fully utilize the information of multi-modal sensing data; although machine learning and deep learning methods have improved the prediction accuracy, most methods lack in-depth analysis of the process of battery health state change and cannot effectively distinguish short-term fluctuations from long-term aging trends; existing prediction methods still face great challenges in abnormal state detection and health warning and are difficult to provide accurate pre-maintenance plans. Summary of the Invention

[0003] Based on this, it is necessary for the present invention to provide a method and system for battery pack health prediction to solve at least one of the above technical problems.

[0004] To achieve the above object, a method for battery pack health prediction includes the following steps:

[0005] Step S1: Collect multi-modal sensing data of the battery pack to obtain voltage data, current data, temperature data, and acoustic vibration data as the original monitoring data set;

[0006] Step S2: Preprocess the original monitoring data set, extract time-frequency domain features, and construct an operating feature matrix of the battery pack;

[0007] Step S3: Obtain the historical operation data of the battery pack; construct an intelligent diagnosis model for the health status of the battery pack by using the historical operation data of the battery pack and the operation feature matrix of the battery pack, wherein constructing the intelligent diagnosis model for the health status of the battery pack includes training a generative adversarial network and training a multi-input deep neural network;

[0008] Step S4: Based on the intelligent diagnosis model for the health status of the battery pack, evaluate the state of the battery pack to obtain the health status evaluation data of the battery pack;

[0009] Step S5: Perform time series decomposition on the health status evaluation data of the battery pack to separate the aging trend and short-term fluctuations, and obtain the life evolution characteristics of the battery pack; perform an abnormal level evaluation on the health status evaluation data of the battery pack to obtain the abnormal state level of the battery pack;

[0010] Step S6: Based on the life evolution characteristics of the battery pack and the abnormal state level of the battery pack, design a health warning and pre-maintenance plan to obtain the health prediction result of the battery pack.

[0011] Through multi-modal sensing data acquisition, the present invention improves the comprehensiveness of the health status monitoring of the battery pack, can obtain multi-dimensional data such as voltage, current, temperature and acoustic vibration, and overcomes the limitation of traditional methods relying on a single data source. By preprocessing the original monitoring data set and extracting time-frequency domain features, the noise in the data is removed, the effectiveness of the feature data is improved, and an operation feature matrix of the battery pack is constructed, making the expression form of the data more in line with the requirements of health status analysis. By combining the historical operation data of the battery pack with the operation feature matrix of the battery pack, an intelligent diagnosis model for the health status is constructed, in which a generative adversarial network and a multi-input deep neural network are combined to achieve high-precision health status recognition. Compared with traditional statistical analysis and machine learning methods, it can more effectively capture the non-linear characteristics of the battery pack and improve the accuracy of health status prediction. Based on the state evaluation of the intelligent diagnosis model, detailed health status evaluation data can be provided, enabling the operation state of the battery pack to be accurately quantified, laying a data foundation for subsequent life prediction and abnormal analysis. Through the time series decomposition method, the aging trend and short-term fluctuations in the health status data are effectively distinguished, making the life evolution characteristics of the battery pack clearer, and combined with the abnormal level evaluation, the refined recognition of the abnormal state of the battery pack is realized, overcoming the problem that existing methods are difficult to distinguish short-term fluctuations from long-term aging trends. Further combining the life evolution characteristics and the abnormal state level, a health warning and pre-maintenance plan are designed, so that the health prediction is not only limited to the evaluation of the current state, but also can predict in advance the fault trend of the battery pack, provide accurate health warnings, and formulate targeted pre-maintenance strategies, solving the deficiencies of existing methods in abnormal detection and pre-maintenance, thereby improving the intelligent level of battery pack management, extending the service life of the battery pack, and enhancing the safety and reliability of the system.

[0012] Preferably, the present invention further provides a system for battery pack health prediction, which is used to execute the above-mentioned method for battery pack health prediction. The system for battery pack health prediction includes:

[0013] A multimodal data acquisition module, which is used to collect multimodal sensing data of the battery pack to obtain voltage data, current data, temperature data, and acoustic vibration data as the original monitoring data set;

[0014] A data preprocessing and feature extraction module, which is used to preprocess the original monitoring data set, extract time-frequency domain features, and construct a battery pack operation feature matrix;

[0015] An intelligent diagnosis model construction module, which is used to obtain the historical operation data of the battery pack; use the historical operation data of the battery pack and the battery pack operation feature matrix to construct an intelligent diagnosis model for the battery pack health situation, where constructing the intelligent diagnosis model for the battery pack health situation includes training a generative adversarial network and training a multi-input deep neural network;

[0016] A health status evaluation module, which is used to evaluate the status of the battery pack based on the intelligent diagnosis model of the battery pack health situation to obtain battery pack health status evaluation data;

[0017] A life evolution and anomaly detection module, which is used to perform time series decomposition on the battery pack health status evaluation data to separate the aging trend and short-term fluctuations to obtain the battery pack life evolution characteristics; perform an anomaly level evaluation on the battery pack health status evaluation data to obtain the battery pack anomaly status level;

[0018] A health prediction and pre-maintenance module, which is used to perform health warning and pre-maintenance plan design based on the battery pack life evolution characteristics and the battery pack anomaly status level to obtain the battery pack health prediction result.

[0019] The present invention can comprehensively obtain various operation data of the battery pack, providing rich original data support for subsequent health assessment. By cleaning, feature extraction, and constructing an operation feature matrix, the quality and analyzability of the data are ensured, providing an effective information basis for the training of the model. Using historical data and operation features, a powerful intelligent diagnosis model for the battery pack health situation is constructed, which can accurately identify the health status of the battery pack, and through the training of the generative adversarial network and the multi-input deep neural network, the diagnostic accuracy and robustness are improved. Accurately evaluate the status of the battery pack, generate detailed health status data, providing a key reference for subsequent life prediction and anomaly detection. Through time series decomposition and anomaly level evaluation, the aging trend and short-term fluctuations of the battery pack are revealed, which helps to accurately identify potential risks and improve the ability of fault prediction. Based on the health assessment results, design health warning and pre-maintenance plans, providing a scientific basis for the maintenance strategy of the battery pack to ensure the reliability and service life of the battery pack. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non - restrictive embodiments read in conjunction with the accompanying drawings:

[0021] Figure 1 It is a schematic flow chart of the steps of a method for predicting the health of a battery pack according to the present invention;

[0022] Figure 2 is Figure 1 a detailed schematic flow chart of step S1 in

[0023] Figure 3 is Figure 1 a detailed schematic flow chart of step S2 in DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] The technical method of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0025] In addition, the drawings are only schematic diagrams of the present invention and are not necessarily drawn to scale. The same reference numerals in the drawings represent the same or similar parts, and thus their repeated description will be omitted. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor methods and / or microcontroller methods.

[0026] It should be understood that although terms such as "first", "second", etc. may be used here to describe each unit, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, the first unit can be called the second unit, and similarly the second unit can be called the first unit. The term "and / or" used here includes any and all combinations of one or more of the listed associated items.

[0027] To achieve the above object, please refer to Figures 1 to 3 The present invention provides a method for predicting the health of a battery pack, and the method includes the following steps:

[0028] In an embodiment of the present invention, referring to Figure 1As shown in the figure, it is a schematic diagram of the step flow of a method for battery pack health prediction according to the present invention. In this example, the method for battery pack health prediction includes the following steps:

[0029] Step S1: Collect multi-modal sensing data of the battery pack to obtain voltage data, current data, temperature data, and acoustic vibration data as the original monitoring data set.

[0030] In the embodiment of the present invention, the multi-modal sensing data collection of the battery pack is realized by arranging multi-type sensor arrays on the surface and periphery of the battery pack. The sensing array consists of 16 voltage acquisition points, 8 current acquisition points, 24 temperature sensor nodes, and 12 acoustic vibration sensors. The voltage data is collected using a high-precision digital multimeter with a sampling accuracy of 0.01V and a sampling frequency set to 10Hz. The current data is collected using a Hall current sensor with a range of 0-500A and an accuracy of 0.5% of the full scale, and the sampling frequency is set to 20Hz. The temperature data is collected using a PT100 platinum resistance temperature sensor array with a measurement range of -20°C to 120°C, a resolution of 0.1°C, and a sampling frequency of 1Hz. The sensors are arranged according to the hot spot coverage principle and are arranged in a 3×8 grid on the surface of the battery pack. The acoustic vibration data is collected using a piezoelectric acceleration sensor with a frequency response range of 20Hz-20kHz, a sensitivity of 100mV / g, and a sampling frequency set to 40kHz. The sensors are respectively fixed at the four corners and the central position of the battery pack. All sensor signals are digitized by a 16-bit ADC converter and then transmitted to the data acquisition unit. The multi-modal data collection sets the sampling duration to the entire process of the standard charge and discharge cycle, and records the cumulative cycle number, ambient temperature, and load type of the battery pack at the beginning of each sampling. The acquisition unit aligns and merges the four types of sensing data according to the unified time stamp into a structured data set, forming an original monitoring data set containing a voltage vector, a current vector, a temperature matrix, and an acoustic spectrum matrix. The data set is stored and numbered separately for each charge and discharge cycle.

[0031] Step S2: Preprocess the original monitoring data set, extract time-frequency domain features, and construct a battery pack operation feature matrix.

[0032] In the embodiments of the present invention, data cleaning is performed on the voltage data, current data, temperature data, and acoustic vibration data obtained in step S1, including removing outliers, removing pulse interference in the voltage data and current data through the median filtering algorithm, eliminating random fluctuations in the temperature data by using the moving average method, and processing high-frequency noise in the acoustic vibration data by using the wavelet threshold denoising method; then, time-domain feature extraction is performed on the cleaned data, calculating the mean, variance, peak factor, kurtosis, and skewness of the voltage data, calculating the mean, standard deviation, maximum value, minimum value, and volatility of the current data, calculating the highest temperature, lowest temperature, temperature gradient, and temperature change rate of the temperature data, and calculating the root mean square value, peak value, waveform factor, and pulse factor of the acoustic vibration data; then, frequency-domain feature extraction is performed on the cleaned data, applying the fast Fourier transform to the voltage data and current data to extract the spectral energy distribution, main frequency component, and spectral centroid, applying the wavelet packet transform to the acoustic vibration data to extract the energy ratio of different frequency bands and the frequency band power spectral density; finally, all the extracted time-domain features and frequency-domain features are organized into a feature vector in the order of acquisition time, and the feature vectors obtained from multiple samplings are combined to construct a battery pack operation feature matrix. Each row of this matrix represents all the features at a time point, and each column represents the values of a feature at different time points. The matrix dimension is n×m, where n is the number of samplings and m is the total number of features.

[0033] Step S3: Obtain the historical operation data of the battery pack; construct an intelligent diagnosis model for the health status of the battery pack by using the historical operation data of the battery pack and the battery pack operation feature matrix, where constructing the intelligent diagnosis model for the health status of the battery pack includes training a generative adversarial network and training a multi-input deep neural network;

[0034] In the embodiments of the present invention, historical operation data of the battery pack under different charge and discharge cycles and environmental conditions is extracted from the battery management system database, including historical voltage fluctuation records, historical charge and discharge current curves, historical temperature distribution data, historical acoustic vibration characteristics, and the corresponding battery pack capacity attenuation rate, internal resistance growth rate, and cycle life health status label. Then, the historical operation data is subjected to time alignment and normalization processing with the battery pack operation characteristic matrix constructed in step S2, and the data is divided into a training set accounting for 70% and a validation set accounting for 30%. Then, a generative adversarial network is constructed. The generator adopts a four-layer convolutional neural network structure, with an input of a 100-dimensional random noise vector. The number of neurons in the hidden layers is 256, 512, 256, and 128 respectively. The activation function is LeakyReLU, and the output layer uses the Tanh activation function. The discriminator adopts a five-layer fully connected network, with an input of the battery pack operation characteristic matrix. The number of neurons in the hidden layers is 128, 256, 512, and 256 respectively. The activation function is ReLU, and the output layer uses the Sigmoid activation function. Alternating training is performed 500 rounds with a learning rate of 0.0002 through the Adam optimizer until the generator can generate synthetic feature data close to the real data distribution. Subsequently, the original feature data and the generated synthetic feature data are used to jointly train a multi-input deep neural network. The network includes four parallel input branches that respectively process voltage characteristics, current characteristics, temperature characteristics, and acoustic vibration characteristics. Each branch consists of three layers of fully connected layers, with the number of neurons being 64, 32, and 16 respectively. The outputs of the four branches are merged through a connection layer and then connected to three layers of fully connected layers, with the number of neurons being 128, 64, and 32 respectively. The number of neurons in the final output layer is 3, corresponding to the capacity attenuation rate, internal resistance growth rate, and remaining life prediction value respectively. The mean square error is used as the loss function, and training is performed 1000 rounds with a learning rate of 0.001 through the stochastic gradient descent algorithm. Finally, when the prediction error on the validation set is lower than 3%, it is confirmed that the training of the battery pack health status intelligent diagnosis model is completed, and the model parameters are saved.

[0035] Step S4: Based on the battery pack health status intelligent diagnosis model, the state of the battery pack is evaluated to obtain battery pack health status evaluation data;

[0036] The embodiments of the present invention load an intelligent diagnosis model for the health status of a battery pack, including the parameters of a generative adversarial network model and the parameters of a multi-input deep neural network model; then, the raw data collected during the current monitoring period is preprocessed and feature-extracted according to the processing method of step S2 to form the operation feature vector of the battery pack at the current moment; then, the operation feature vector of the battery pack is input into the multi-input deep neural network, and the predicted values of three key health indicators, namely the capacity attenuation rate, the internal resistance growth rate, and the remaining life, are obtained through forward propagation calculation; subsequently, the actual measured terminal voltage response curve of the battery pack is compared with the ideal terminal voltage response curve predicted by the multi-input deep neural network, and the root mean square error, the correlation coefficient, and the consistency index are calculated to quantitatively evaluate the deviation degree of the current operation state of the battery pack from the ideal state; at the same time, the discriminator of the generative adversarial network is used to calculate the K-L divergence between the probability distribution of the current operation characteristics of the battery pack and the probability distribution of historical normal samples. When the K-L divergence exceeds the preset threshold of 0.5, it is marked as an abnormal state; further, according to the voltage consistency, the temperature distribution uniformity, and the acoustic vibration signal correlation of each single battery inside the battery pack, the weights of each index are determined by combining the analytic hierarchy process, and the comprehensive health score is calculated. The health score ranges from 0 to 100, and the higher the score, the better the health state of the battery pack; finally, the capacity attenuation rate, the internal resistance growth rate, the predicted value of the remaining life, the terminal voltage response deviation, the K-L divergence, and the comprehensive health score index are combined to form the battery pack health state evaluation data, which is stored in the database according to the time granularity of hours, days, weeks, and months.

[0037] Step S5: Perform time series decomposition on the battery pack health state evaluation data to separate the aging trend and short-term fluctuations, and obtain the life evolution characteristics of the battery pack; perform an abnormal level evaluation on the battery pack health state evaluation data to obtain the abnormal state level of the battery pack.

[0038] Extract the three key indicators of capacity attenuation rate, internal resistance growth rate, and comprehensive health score from the battery pack health status evaluation data obtained in step S4 of the embodiment of the present invention to construct a time series data set; then, use the seasonal decomposition method to process the time series data, and decompose the data into three parts: trend term, seasonal term, and residual term, where the trend term reflects the long-term aging trend of the battery pack, the seasonal term reflects the periodic changes of charge and discharge cycles, and the residual term reflects random fluctuations; then, apply the polynomial fitting method to the trend term for processing to determine the optimal fitting curve, the fitting polynomial order is 3, and the fitting accuracy requirement is that the root mean square error is less than 0.02. Calculate the aging rate and aging acceleration of the battery pack through the first derivative and second derivative of the fitting curve; subsequently, combine the typical aging curves in the battery pack historical aging database, and calculate the similarity between the current aging trend and the typical aging mode through the dynamic time warping algorithm to identify the aging type of the battery pack, which is divided into three categories: uniform aging type, accelerated aging type, and sudden aging type; at the same time, perform an abnormal level evaluation on the battery pack health status evaluation data, and based on three indicators: K-L divergence value, voltage deviation value, and temperature uniformity, use the fuzzy comprehensive evaluation method to construct an abnormal level evaluation matrix, and divide the abnormal state into five levels, namely normal (level 1, K-L divergence < 0.3), slightly abnormal (level 2, 0.3 ≤ K-L divergence < 0.5), moderately abnormal (level 3, 0.5 ≤ K-L divergence < 0.7), severely abnormal (level 4, 0.7 ≤ K-L divergence < 0.9), and dangerous state (level 5, K-L divergence ≥ 0.9); finally, integrate the battery pack aging trend characteristics, aging rate, aging acceleration, aging type, and abnormal level information into the battery pack life evolution characteristics and the battery pack abnormal state level, and store them in the database.

[0039] Step S6: Design a health warning and pre-maintenance plan based on the battery pack life evolution characteristics and the battery pack abnormal state level to obtain the battery pack health prediction result.

[0040] Embodiments of the present invention extract the life evolution characteristics of a battery pack, including the aging trend curve, aging rate, aging acceleration, and aging type. The exponential smoothing prediction method is used to predict the capacity decay trajectories for the next 30 days, 90 days, and 180 days, with the smoothing coefficient set to 0.85 and the prediction confidence level at 95%. Then, in combination with the rated capacity and design life of the battery pack, the expected value and upper and lower bounds of the remaining service life of the battery pack are calculated. When the predicted capacity decays to 80% of the rated capacity, it is determined as the end-of-life point. Then, based on the abnormal state level of the battery pack obtained in step S5, a four-level health warning mechanism is established, corresponding to different warning colors: levels 1-2 correspond to a green warning, indicating normal operation; level 3 corresponds to a yellow warning, indicating that the monitoring frequency needs to be increased; level 4 corresponds to an orange warning, indicating that the charge and discharge current needs to be restricted; level 5 corresponds to a red warning, indicating that use should be stopped immediately and maintenance should be carried out. Subsequently, pre-maintenance plans are formulated for different warning levels and aging types: for a battery pack with a yellow warning and a uniform aging type, an equalization charging strategy is formulated, restricting the charging cut-off voltage to 95% of the rated value; for a battery pack with an orange warning and an accelerated aging type, a current limiting measure is formulated, restricting the maximum discharge rate to 0.5C, and a periodic equalization charging plan is formulated; for a battery pack with a red warning and a sudden aging type, an emergency shutdown procedure is triggered, and a single cell detection task is generated, marking the location of potential faulty single cells. Finally, the predicted remaining life of the battery pack, the prediction confidence interval, the health warning level, the pre-maintenance plan, and the recommended replacement time are integrated into the health prediction result of the battery pack, which is transmitted to the battery management system through a data interface, and a prediction report is generated, including a capacity decay trend graph, an internal resistance growth trend graph, a time axis of key abnormal events, and a list of pre-maintenance suggestions, providing data support for battery pack management decisions.

[0041] The present invention improves the comprehensiveness of the health state monitoring of the battery pack through multi-modal sensing data acquisition, can obtain multi-dimensional data of voltage, current, temperature and acoustic vibration, and overcomes the limitation of traditional methods relying on a single data source. By preprocessing the original monitoring data set and extracting time-frequency domain features, the noise in the data is removed, the effectiveness of the feature data is improved, and a battery pack operation feature matrix is constructed, making the expression form of the data more in line with the requirements of health state analysis. Combining the historical operation data of the battery pack with the battery pack operation feature matrix, a health situation intelligent diagnosis model is constructed, in which a generative adversarial network and a multi-input deep neural network are combined to achieve high-precision health state recognition. Compared with traditional statistical analysis and machine learning methods, it can more effectively capture the non-linear features of the battery pack and improve the accuracy of health state prediction. Based on the state evaluation of the intelligent diagnosis model, detailed health state evaluation data can be provided, enabling the operation state of the battery pack to be accurately quantified, laying a data foundation for subsequent life prediction and anomaly analysis. Through the time series decomposition method, the aging trend and short-term fluctuations in the health state data are effectively distinguished, making the life evolution characteristics of the battery pack clearer, and combined with the anomaly level evaluation, a refined identification of the abnormal state of the battery pack is realized, overcoming the problem that existing methods are difficult to distinguish short-term fluctuations from long-term aging trends. Further combining the life evolution characteristics and the abnormal state level, health early warning and pre-maintenance plan design are carried out, so that health prediction is not only limited to the evaluation of the current state, but also can predict in advance the fault trend of the battery pack, provide accurate health early warning, and formulate targeted pre-maintenance strategies, solving the deficiencies of existing methods in anomaly detection and pre-maintenance, thereby improving the intelligent level of battery pack management, extending the service life of the battery pack, and enhancing the safety and reliability of the system.

[0042] Preferably, step S1 includes the following steps:

[0043] Step S11: Perform high-frequency sampling on the terminal voltage of the battery pack to obtain voltage time series data;

[0044] Step S12: Monitor the charge and discharge current of the battery pack in real time to obtain current time series data;

[0045] Step S13: Perform multi-point measurement of the surface temperature of the battery pack based on a distributed temperature sensor array to obtain temperature distribution data;

[0046] Step S14: Collect acoustic vibration signals through piezoelectric sensors attached to the surface of the battery pack to obtain vibration spectrum data;

[0047] Step S15: Perform time synchronization and integration on the voltage time series data, current time series data, temperature distribution data and vibration spectrum data to obtain a multi-modal original monitoring data set.

[0048] In the embodiments of the present invention, a precision voltage acquisition module LTC6804-2 is connected to the positive and negative wiring terminals of the battery pack. The sampling resolution is set to 16 bits, the sampling frequency is set to 10 kHz, the voltage signal is adjusted to the range of 0-5V through a pre-stage voltage divider circuit, an RC low-pass filter is used to set the cut-off frequency to 4 kHz to eliminate high-frequency interference, the sampling duration is 60 seconds per cycle, the sampling cycle is triggered once per hour, and the acquired voltage data is transmitted to the data processing unit through the CAN bus. The specific implementation method for real-time monitoring of the charge and discharge current of the battery pack is as follows: a Hall current sensor ACS758 is installed in series in the main circuit of the battery pack, with a measurement range of -500A to +500A and an accuracy of ±0.5% of the full scale. The sampling frequency is set to 5 kHz, and digital conversion is performed through a 16-bit ADC. The current waveform and mutation events are recorded in real time, and the average value is calculated every 100 ms to form current time-series data. The specific implementation method for multi-point measurement of the surface temperature of the battery pack based on a distributed temperature sensor array is as follows: a total of 30 NTC thermistor sensors are installed on the surface of the battery pack according to a 5×6 grid layout. The temperature measurement range of the sensor is -20°C to +80°C, the measurement accuracy is ±0.5°C, the sensor spacing is 10 cm, the sampling frequency is set to 1 Hz, and data acquisition is performed through a 24-bit delta-sigma ADC. The discrete temperature point data acquired is reconstructed into a complete temperature distribution heat map through a bilinear interpolation algorithm. The specific implementation method for acquiring acoustic vibration signals through piezoelectric sensors attached to the surface of the battery pack is as follows: 5 high-sensitivity piezoelectric acceleration sensors are installed at the four corners and the center of the battery pack housing, with a sensitivity of 100 mV / g, a measurement range of ±50g, a frequency response range of 10 Hz to 20 kHz, the sampling frequency is set to 44.1 kHz, and data acquisition is performed through a 24-bit ADC. The short-time Fourier transform is used to convert the time-domain signal into a frequency-domain signal, and the analysis frequency range is 10 Hz to 5 kHz to form vibration spectrum data. The specific implementation method for time synchronization and integration of the voltage time-series data, current time-series data, temperature distribution data, and vibration spectrum data is as follows: all acquired data is marked based on a unified microsecond-level timestamp, and an FPGA is used to achieve precise trigger synchronization at the hardware level, with a trigger error less than 5 μs. The data acquired at different frequencies is unified to a sampling rate of 100 Hz through a data compensation algorithm. The voltage and current data are downsampled by taking the average value, the vibration data is downsampled after low-pass filtering, and the temperature data is upsampled through linear interpolation. Finally, the four types of data are organized and integrated along the same time axis into a multi-modal raw monitoring dataset in MATLAB mat format. The dataset contains five main fields: timestamp, voltage matrix, current vector, temperature distribution matrix, and vibration spectrum matrix.

[0049] By performing high-frequency sampling on the terminal voltage of the battery pack, the present invention can accurately capture the subtle voltage changes during the operation of the battery, reflect the internal chemical reactions and aging status, and improve the sensitivity of state-of-health monitoring. By monitoring the charge and discharge current in real time, the current response characteristics of the battery pack under different working conditions can be quantified, providing data support for evaluating the load capacity and internal impedance changes of the battery pack. Based on a distributed temperature sensor array for multi-point measurement, the temperature distribution on the surface of the battery pack can be comprehensively obtained, local overheating phenomena can be accurately identified, the effectiveness of thermal management can be improved, and battery performance degradation or safety hazards caused by local overheating can be avoided. By collecting acoustic vibration signals through piezoelectric sensors attached to the surface of the battery pack, non-destructive detection of the internal structure state of the battery can be realized, battery anomalies can be identified, and the comprehensiveness of the state-of-health assessment of the battery pack can be improved. By synchronizing and integrating various types of data in time, a multi-modal original monitoring data set is formed to ensure the temporal consistency of different types of sensing data, improve the quality of data fusion, make the subsequent state-of-health assessment more accurate, overcome the problem of insufficient information in a single data source, and provide comprehensive and reliable input data for the intelligent diagnosis model of the health situation.

[0050] Preferably, step S2 includes the following steps:

[0051] Step S21: Perform data cleaning and outlier detection on the original monitoring data set to obtain a preliminary processed data set with noise points and outliers removed;

[0052] Step S22: Suppress the noise of the preliminary processed data set to obtain a denoised data set;

[0053] Step S23: Perform standardization and normalization processing on the denoised data set to obtain a standardized data set;

[0054] Step S24: Perform continuous wavelet transform on the standardized data set to obtain the time-frequency domain decomposition result;

[0055] Step S25: Extract the energy distribution, scale coefficient, and frequency characteristics from the time-frequency domain decomposition result to obtain a multi-dimensional feature vector;

[0056] Step S26: Perform dimensionality reduction and feature selection on the multi-dimensional feature vector to obtain the operation feature matrix of the battery pack.

[0057] In the embodiments of the present invention, an outlier detection method based on the interquartile range is adopted. For each type of data, Q1 (the 25th percentile), Q3 (the 75th percentile), and IQR (the interquartile range, equal to Q3 - Q1) are calculated, and the data points outside the range of [Q1 - 1.5×IQR, Q3 + 1.5×IQR] are marked as outliers; for voltage data, the sliding window median method is used to detect and mark mutation points, the window size is set to 51 data points, and when the voltage difference between adjacent data points exceeds 0.1V, it is determined as a mutation point; for current data, the 3σ criterion is applied, the mean μ and the standard deviation σ are calculated, and the points deviating from the mean by more than 3 standard deviations are marked as outliers; for temperature data, a spatial consistency test is adopted, and when the temperature difference between a certain point and the average temperature of the adjacent 8 points exceeds 5°C, it is marked as an outlier point; for vibration spectrum data, the peak clipping method is used, and the isolated peaks with amplitudes exceeding 80% of the global maximum value are marked as interference points; the specific implementation manner of noise suppression for the preliminary processed data set is as follows: for voltage data, a Butterworth low-pass filter is applied, the cut-off frequency is set to 500Hz, and the filter order is 4, to eliminate high-frequency noise; for current data, an adaptive median filtering algorithm is applied, and the window size is dynamically adjusted between 3 and 15 to remove pulse interference; for temperature data, Gaussian filtering is applied, the kernel size is set to 5×5, and the standard deviation is set to 1.2, to smooth the temperature distribution map; for vibration spectrum data, wavelet threshold denoising is applied, the db4 wavelet basis is adopted, the decomposition level is 5 layers, and the soft threshold function is used, with the threshold coefficient being 3.5; the specific implementation manner of standardization and normalization processing for the denoised data set is as follows: for voltage data, the Min-Max normalization method is adopted to map the voltage values to the interval [0, 1]; for current data, the Z-score standardization method is adopted to make the data mean 0 and the standard deviation 1; for temperature data, the maximum temperature normalization is adopted to divide all temperature points by the highest temperature value of this batch; for vibration spectrum data, the band energy normalization is adopted to divide the energy of each band by the total energy; the specific implementation manner of continuous wavelet transform for the standardized data set is as follows: for voltage and current data, the Mexican hat wavelet (Ricker wavelet) is used for transformation, the scale range is set to 1 to 128, and the scale step is 1, to obtain the time-frequency energy spectrum; for temperature data, two-dimensional continuous wavelet transform is adopted, the Haar wavelet basis is used, and the decomposition scale is set to 5 to obtain the multi-scale features of the temperature field; for vibration spectrum data, the Morlet wavelet transform is adopted, and the central frequency is set to 0.8125 Hz, with a scale range from 1 to 64, to obtain a time-frequency diagram with high time-frequency resolution; the specific implementation of extracting the energy distribution, scale coefficient, and frequency characteristics from the time-frequency domain decomposition result is as follows: extract 20 features including the maximum energy ratio, energy entropy, dominant scale value, and wavelet variance from the voltage wavelet coefficients; extract 25 features including the band energy distribution, wavelet entropy, scale correlation coefficient, and duration distribution from the current wavelet coefficients; extract 15 features including the temperature gradient feature, spatial distribution entropy, local extreme point density, and hot spot migration rate from the temperature wavelet coefficients; extract 30 features including the characteristic frequency energy ratio, ridge line length, ridge line curvature, and instantaneous frequency distribution from the vibration wavelet coefficients; the specific implementation of dimensionality reduction and feature selection for the multi-dimensional feature vector is as follows: first, use the Pearson correlation coefficient analysis method to screen out redundant features with a correlation higher than 0.85; then use the principal component analysis method for dimensionality reduction, set the cumulative contribution rate threshold to 95%, and arrange the dimensionality-reduced feature vector into an m×n-dimensional matrix, where m is the number of sampling times and n is the number of features after dimensionality reduction, usually n is between 20 and 30; finally, combine the feature vectors of multiple sampling periods into a battery pack operation feature matrix according to the time series, where each row of the matrix represents all the features at a time point, and each column represents the change of a certain feature at different time points.

[0058] The present invention can effectively eliminate invalid data and measurement errors, reduce the interference of abnormal points on health assessment, and improve the reliability of data by performing data cleaning and outlier detection on the original monitoring data set. Perform noise suppression on the preliminarily processed data set to further reduce environmental interference and measurement noise, and ensure the stability and accuracy of the data. Perform standardization and normalization processing on the denoised data set to make data with different dimensions have a consistent scale, improve the stability of model training, and enhance the comparability between different features. Perform continuous wavelet transform on the standardized data set to achieve time-frequency domain decomposition, so that the battery pack state characteristics can be accurately characterized in both the time and frequency dimensions, and improve the analytical ability for non-stationary signals. Extract the energy distribution, scale coefficient, and frequency characteristics from the time-frequency domain decomposition result, construct a multi-dimensional feature vector, make the description of the battery operation state more comprehensive, and enhance the identification ability for battery aging and abnormal modes. Perform dimensionality reduction and feature selection on the multi-dimensional feature vector, remove redundant information, improve the calculation efficiency, and construct a battery pack operation feature matrix to provide high-quality input data for the subsequent health trend intelligent diagnosis model, thereby improving the accuracy of health assessment and life prediction.

[0059] Preferably, the training of the preset generative adversarial network using historical data in step S3 includes:

[0060] Screen and classify the historical data to obtain a normal behavior pattern data set, where the signal-to-noise ratio is greater than or equal to 25 dB;

[0061] Construct a conditional generative adversarial network based on the normal behavior pattern dataset to obtain an initial network structure;

[0062] The conditional input dimension in the conditional generative adversarial network is 4 - 8 dimensions, the number of hidden layers is 3 - 5 layers, and the number of neurons in each layer is 64 - 256;

[0063] Initialize the parameters of the generator for the initial network structure to obtain a generator model, where the learning rate is 0.0001 - 0.001;

[0064] Conduct discriminator adversarial training based on the normal behavior pattern dataset and the generator model to obtain an initial discriminator, where the proportion of real samples is 40% - 60%, the proportion of generated samples is 40% - 60%, and the update frequency of the initial discriminator is every 1 - 3 generator updates;

[0065] Iteratively optimize the generator model and the initial discriminator to obtain a generative adversarial network model;

[0066] Extract the discriminator part based on the generative adversarial network model to obtain a discriminator model.

[0067] In the embodiments of the present invention, the long-term monitoring data in the historical battery pack operation database is screened to extract data segments with a signal-to-noise ratio greater than or equal to 25 dB. The data is classified through an expert annotation system to distinguish the normal behavior pattern data set, and data samples with high voltage stability, good temperature uniformity, and gentle internal resistance change are retained. The data is stratified and sampled according to the battery pack operation ambient temperature, charge-discharge rate, and number of cycles to ensure the representativeness of the data set. A conditional generative adversarial network is constructed based on the normal behavior pattern data set. The network structure is designed with four-dimensional conditional input, including temperature, current, internal resistance, and remaining capacity. The number of hidden layers is set to 4, and the number of neurons in each layer is 128. The generator adopts a cascaded architecture of transposed convolutional layers and upsampling layers, and the discriminator adopts a multi-scale progressive discriminant structure. The generator parameters of the initial network structure are initialized. The initialization method selects the Xavier algorithm, the learning rate is set to 0.0005, the momentum coefficient is 0.9, and the weight decay coefficient is 1e-5. The weight matrix is randomly initialized using a truncated normal distribution, and the bias term is initialized to 0 to obtain the generator model. Based on the normal behavior pattern data set and the generator model, adversarial training of the discriminator is carried out. The proportion of real samples in the training data set is set to 50%, the proportion of generated samples is set to 50%, and the discriminator update frequency is once every 2 generator updates. The loss function uses the Wasserstein distance, and a gradient penalty term is introduced to prevent mode collapse. The generator model and the initial discriminator are iteratively optimized. The maximum number of iteration rounds is set to 500 rounds. When the Fréchet inception distance between the generated data distribution and the real data distribution is less than 0.1, the training stops. Overfitting is prevented through a dynamic learning rate adjustment strategy and an early stopping mechanism to obtain the generative adversarial network model. Based on the generative adversarial network model, the discriminator part is extracted, all parameters of the discriminator except the last layer are frozen, the discriminator feature extraction network and the output layer are retained, and the extracted features are mapped to a 4D health state space to obtain the final discriminator model. This model can extract key features and perform state discrimination in the intelligent diagnosis of the battery pack health situation.

[0068] The present invention screens and classifies historical data to ensure that the signal-to-noise ratio of the normal behavior pattern data set used for training is greater than or equal to 25 dB, improves data quality, reduces noise interference, and enables the model to learn more accurate health state features. A conditional generative adversarial network is constructed based on the normal behavior pattern data set to form an initial network structure, enabling the generator to generate samples conforming to the battery health state features under known condition constraints, and improving the model's fitting ability for complex health states. The conditional input dimension is set to 4 - 8 dimensions, the number of hidden layers is 3 - 5 layers, and the number of neurons in each layer is 64 - 256, ensuring that the network has a sufficiently powerful expression ability to adapt to the health state modeling requirements of different types of battery packs. The generator parameters of the initial network structure are initialized, and the learning rate is set between 0.0001 and 0.001, enabling the generator model to converge stably during training and improving the quality of the generated samples. Discriminator adversarial training is performed based on the normal behavior pattern data set and the generator model. By reasonably allocating the ratio of real samples to generated samples (40% - 60%) and setting the discriminator update frequency to every 1 - 3 generator updates, the balance of adversarial training is ensured, mode collapse is avoided, and the model's recognition ability for the real battery health state is improved. The generator model and the initial discriminator are iteratively optimized to gradually converge the generative adversarial network, enhancing the authenticity and diversity of the generated samples and making them closer to the real health state distribution. Finally, the discriminator part is extracted based on the trained generative adversarial network model to obtain a discriminator model, enabling it to be independently used for health state assessment, improving the accuracy of battery anomaly detection and health assessment, and providing a highly reliable discrimination ability for subsequent intelligent diagnosis of health trends.

[0069] Preferably, in step S3, training the preset multi-input deep neural network model with the battery pack operation feature matrix as the input includes:

[0070] Grouping the battery pack operation feature matrix according to data types to obtain a voltage feature sub-matrix, a current feature sub-matrix, a temperature feature sub-matrix, and an acoustic feature sub-matrix;

[0071] The voltage feature dimension is 0 - 20 dimensions, the current feature dimension is 8 - 16 dimensions, the temperature feature dimension is 12 - 24 dimensions, and the acoustic feature dimension is 20 - 40 dimensions;

[0072] Constructing a voltage-current process based on the voltage feature sub-matrix and the current feature sub-matrix to obtain a time-series electrical characteristic neural network functional unit, where the number of hidden layers is 2 - 3 layers, the number of hidden units is 64 - 128, and the sequence length is 20 - 50 time steps;

[0073] Construct temperature data processing based on the temperature feature sub - matrix to obtain the temperature distribution spatial feature neural network functional unit, where the number of convolutional layers is 2 - 4 layers, the number of convolutional kernels is 16 - 64, and the size of the convolutional kernel is 3×3;

[0074] Construct acoustic vibration data processing based on the acoustic feature sub - matrix to obtain the vibration spectrum feature neural network functional unit, where the number of frequency segments is 8 - 16 segments, the number of attention heads is 4 - 8, and the output feature dimension is 32 - 64 dimensions;

[0075] Design a multi - head attention mechanism based on the time - series electrical characteristic neural network functional unit, the temperature distribution spatial feature neural network functional unit, and the vibration spectrum feature neural network functional unit to obtain a weight calculation model, where the number of attention heads is 4 - 8;

[0076] Fuse the outputs of the time - series electrical characteristic neural network functional unit, the temperature distribution spatial feature neural network functional unit, and the vibration spectrum feature neural network functional unit through the weight calculation model to obtain a multi - modal fusion feature vector;

[0077] Construct a fully - connected layer based on the multi - modal fusion feature vector to obtain a battery pack health state classifier and a degradation degree regressor;

[0078] The number of layers of the fully - connected layer is 2 - 3 layers, where the number of hidden neurons in the first layer is 128 - 256, halving layer by layer, and the dropout rate is 0.2 - 0.5;

[0079] Jointly train the battery pack health state classifier and the degradation degree regressor, and optimize the parameters based on the joint training results to obtain a battery pack state recognition model.

[0080] In the embodiments of the present invention, first, the operation characteristic matrix of the battery pack is accurately grouped according to the data type, obtaining a 16-dimensional voltage characteristic sub-matrix, a 12-dimensional current characteristic sub-matrix, a 20-dimensional temperature characteristic sub-matrix, and a 36-dimensional acoustic characteristic sub-matrix; a time-series electrical characteristic neural network functional unit is constructed according to the voltage characteristic sub-matrix and the current characteristic sub-matrix, adopting a gated recurrent unit (GRU) structure, with the hidden layer set to 3 layers, 96 hidden units in each layer, the sequence length being 40 time steps, and the thresholds of the input gate and the forget gate being dynamically adjusted; a temperature distribution spatial characteristic neural network functional unit is constructed according to the temperature characteristic sub-matrix, adopting a spatial residual convolutional neural network, with 3 convolutional layers, 32 convolutional kernels in the first layer, 48 in the second layer, and 64 in the third layer, the size of the convolutional kernels all being 3×3, and batch normalization and a spatial attention mechanism being introduced; a vibration spectrum characteristic neural network functional unit is constructed according to the acoustic characteristic sub-matrix, dividing the spectrum data into 12 frequency bands, designing a multi-head self-attention mechanism, with 6 attention heads, the output feature dimension being 48 dimensions, and enhancing the feature extraction ability through residual connection; a weight calculation model is designed based on the above three functional units, adopting a multi-head attention mechanism, with the number of attention heads set to 6, and the weight calculation method combining dot-product attention and additive attention; the outputs of the three functional units are subjected to feature fusion through the weight calculation model to obtain a multi-modal fusion feature vector, and residual connection and layer normalization are introduced in the fusion process; a fully connected layer is constructed based on the multi-modal fusion feature vector, with 192 hidden neurons in the first layer, 96 in the second layer, the dropout rate being set to 0.3, and the activation function being selected as LeakyReLU, constructing a battery pack health state classifier (4 output categories) and a degradation degree regressor (continuous value output); a multi-task joint training method is adopted for the battery pack health state classifier and the degradation degree regressor, the loss function combining cross-entropy loss and mean square error loss, introducing a weight balance coefficient, training through the Adam optimizer, with an initial learning rate of 0.001, adaptively decaying every 10 training epochs, and performing parameter regularization and early stopping based on the performance on the validation set, finally obtaining a battery pack state recognition model.

[0081] The present invention groups the battery pack operation feature matrix according to data types to form a voltage feature sub-matrix, a current feature sub-matrix, a temperature feature sub-matrix, and an acoustic feature sub-matrix, making the data structure clearer and improving the pertinence of feature extraction. The dimension ranges of different feature sub-matrices are set to reasonably allocate various feature dimensions, ensuring information integrity and improving the expressive ability of the model. A time-series electrical characteristic neural network functional unit is constructed based on voltage and current features, which can extract the electrical dynamic features during the operation of the battery and improve the sensitivity to changes in the battery health state. A temperature distribution spatial feature neural network functional unit is constructed based on temperature features, and the battery temperature distribution pattern is extracted through multi-layer convolution operations to improve the recognition ability of thermal runaway and abnormal heating conditions. A vibration spectrum feature neural network functional unit is constructed based on acoustic features, and the signal analysis ability is enhanced through frequency segmentation and attention mechanism to improve the detection accuracy of abnormal vibrations inside the battery. A multi-head attention mechanism is designed for weight calculation, enabling the features output by each functional unit to dynamically allocate weights and improving the rationality of multi-modal data fusion. A weight calculation model is used for feature fusion to generate a multi-modal fusion feature vector, fully combining electrical, temperature, and acoustic information to improve the accuracy of battery health state assessment. A fully connected layer is constructed based on the multi-modal fusion feature vector to achieve health state classification and degradation degree regression, improving the prediction ability of the model. By jointly training the battery pack health state classifier and the degradation degree regressor and optimizing the parameters, the model can accurately identify the battery health state and accurately predict the battery degradation trend, providing efficient support for subsequent health assessment and life prediction.

[0082] Preferably, step S4 includes the following steps:

[0083] Step S41: Input the battery pack operation feature matrix into the discriminator model, calculate the deviation degree from the normal battery pack behavior pattern, and obtain a preliminary anomaly metric;

[0084] Step S42: Input the battery pack operation feature matrix into the battery pack state recognition model, obtain the error vector between the predicted values and the actual values of each parameter of the battery pack, and obtain a parameter performance degradation index;

[0085] Step S43: Fuse the preliminary anomaly metric and the parameter performance degradation index to construct a battery pack health index. The range of the battery pack health index is 0-100, where 100 represents a brand-new state and 0 represents complete failure;

[0086] Step S44: Calculate the abnormal probability distribution map of each component of the battery pack based on the parameter performance degradation index;

[0087] Step S45: Integrate the battery pack health indicators and the abnormal probability distribution map to generate a battery pack health status assessment report, where the battery pack health status assessment report includes a health score, a heat map of abnormal areas, and a deviation degree of key parameters.

[0088] In the embodiment of the present invention, first, the battery pack operation characteristic matrix is input into the discriminator model, and the Euclidean distance calculation method is used to measure the deviation degree between the current battery pack operation state and the standard normal battery pack behavior pattern, obtaining a preliminary abnormal metric index. The specific calculation formula is deviation value = sqrt(sum((current feature - standard feature)^2)); subsequently, the battery pack operation characteristic matrix is input into the battery pack state recognition model. By comparing the predicted value with the actual value, the root mean square error algorithm is used to calculate the performance decay index of each parameter, obtaining a parameter prediction error vector; then, according to the weight fusion method, the preliminary abnormal metric index and the parameter performance decay index are linearly combined to construct a battery pack health indicator in the range of 0 - 100, where the weight distribution is set according to the historical data experience coefficient; based on the parameter performance decay index, the probability density estimation method is used to draw the abnormal probability distribution map of each component of the battery pack. Specifically, the kernel density estimation algorithm is used to calculate the abnormal occurrence probability of each component; finally, the battery pack health indicator and the abnormal probability distribution map are integrated to generate a battery pack health status assessment report including a health score, a heat map of abnormal areas, and a deviation degree of key parameters. The health score directly corresponds to the health indicator value of 0 - 100. The heat map of abnormal areas is based on the color mapping of the probability distribution map, and the deviation degree of key parameters is obtained by calculating the relative deviation percentage of each parameter from the standard value.

[0089] The present invention inputs the battery pack operation characteristic matrix into the discriminator model and calculates the deviation degree from the normal battery pack behavior pattern, enabling the system to quickly identify abnormal situations in the battery pack operation state and improving the accuracy of abnormal detection. By inputting the battery pack operation characteristic matrix into the battery pack state recognition model to obtain the error vector between the predicted value and the actual value of each parameter, the attenuation of each battery operation parameter can be quantified, providing a key basis for battery performance evaluation. By fusing the preliminary abnormal metric index and the parameter performance decay index to construct a health indicator and adopting a scoring mechanism of 0 - 100, the assessment of the battery pack health status is made more intuitive and convenient for comparative analysis under different operation states. Calculating the abnormal probability distribution map of each part of the battery pack based on the parameter performance decay index can locate the specific abnormal occurrence location and improve the accuracy of fault diagnosis. Finally, integrating the battery pack health indicator and the abnormal probability distribution map to generate a battery pack health status assessment report makes the health assessment result more systematic. The report includes a health score, a heat map of abnormal areas, and a deviation degree of key parameters, providing a comprehensive health analysis reference for operation and maintenance personnel and improving the intelligent level of the battery management system.

[0090] Particularly importantly, inputting the battery pack operation characteristic matrix into the discriminator model and calculating the deviation degree from the normal battery pack behavior pattern includes:

[0091] Perform data normalization processing on the battery pack operation characteristic matrix to obtain a normalized characteristic matrix;

[0092] Perform feature matching on the discriminator model based on the normalized characteristic matrix to obtain the deviation degree of the current operation characteristics of the battery pack;

[0093] Perform statistical analysis on the deviation degree of the current operation characteristics of the battery pack to obtain the deviation amplitude of each characteristic variable;

[0094] Perform cluster analysis on the overall operation mode of the battery pack based on the deviation amplitude of each characteristic variable to obtain the classification result of the battery pack operation mode;

[0095] Perform comparative analysis on the classification result of the battery pack operation mode to obtain the abnormal deviation value of the battery pack operation mode;

[0096] Calculate the abnormal probability distribution of the battery pack based on the abnormal deviation value of the battery pack operation mode to obtain a preliminary abnormal probability vector;

[0097] Perform feature weighting calculation on the preliminary abnormal probability vector to obtain a preliminary abnormal metric index of the battery pack.

[0098] In the embodiment of the present invention, first, perform min-max normalization processing on the battery pack operation characteristic matrix to map each characteristic variable to the interval [0,1]. The normalization formula is X_norm = (X - X_min) / (X_max - X_min) to ensure that features with different dimensions can be directly compared. Subsequently, adopt a feature matching algorithm to calculate the Euclidean distance between the normalized characteristic matrix and the standard normal battery pack characteristic matrix to obtain the deviation degree of the current operation characteristics of the battery pack. The specific distance calculation method is sqrt(sum((standard feature - current feature)^2)). Then, perform statistical analysis on the deviation degree of the current operation characteristics of the battery pack, and use descriptive statistical methods to calculate the deviation amplitude of each characteristic variable, including the absolute deviation value and the relative deviation percentage. Based on the deviation amplitude of each characteristic variable, perform K-means cluster analysis to divide the battery pack operation mode into four categories: healthy, sub-healthy, warning, and dangerous. Perform comparative analysis on the classification result of the battery pack operation mode, and calculate the distance between categories and the within-class variance to obtain the abnormal deviation value of the battery pack operation mode. According to the abnormal deviation value of the operation mode, use the probability density estimation method to calculate the abnormal probability distribution of the battery pack and generate a preliminary abnormal probability vector. Finally, perform feature weighting calculation on the preliminary abnormal probability vector, and the weight coefficient is determined according to historical data and expert experience to obtain a preliminary abnormal metric index of the battery pack.

[0099] The present invention performs data normalization on the operation feature matrix of the battery pack, which can eliminate the dimensional differences of different feature data, improve the comparability of the data, and ensure the stability of model calculation. Feature matching is performed on the discriminator model based on the normalized feature matrix, enabling the system to accurately identify the differences between the current operation features of the battery pack and the normal mode, and improving the sensitivity of anomaly detection. By statistically analyzing the deviation degree of the current operation features of the battery pack, the deviation amplitude of each feature variable can be quantified, providing fine-grained anomaly features for further pattern recognition. Cluster analysis is performed on the overall operation mode of the battery pack based on the deviation amplitude of each feature variable, making the division of the health state of the battery pack clearer and improving the resolution of anomaly detection. Comparative analysis of the classification results of the operation modes of the battery pack can further extract the mode anomaly deviation value and accurately quantify the anomaly degree of the operation state of the battery pack. Calculating the anomaly probability distribution based on the anomaly deviation value of the battery pack operation mode helps to construct a more interpretable anomaly probability vector and improve the reliability of health assessment. Finally, feature weighting calculation is performed on the preliminary anomaly probability vector, comprehensively considering the importance of each feature, to obtain the preliminary anomaly metric index of the battery pack, providing a key basis for subsequent health state assessment and early warning.

[0100] Preferably, the time series decomposition of the battery pack health state assessment data in step S5 to separate the aging trend and short-term fluctuations includes:

[0101] Performing time series reconstruction on the battery pack health state assessment data to obtain the battery pack health state time series data set;

[0102] Performing decomposition of the trend component, periodic component, and random component on the battery pack health state time series data set to obtain the multi-dimensional decomposition features of the battery pack health state, where the number of IMF components is limited to within 3-8;

[0103] Performing significance screening on the multi-dimensional decomposition features of the battery pack health state to obtain the key features of the battery pack health state;

[0104] Performing evolutionary pattern recognition on the key features of the battery pack health state to obtain the evolutionary law of the battery pack health state;

[0105] Performing degradation curve fitting on the key features of the battery pack health state based on the evolutionary law of the battery pack health state to obtain the battery pack life degradation model;

[0106] Performing change point detection on the battery pack life degradation model to obtain the division of the key stages of the battery pack life;

[0107] Extracting the degradation characteristics of each stage based on the division of the key stages of the battery pack life to obtain the stage-by-stage life characteristics of the battery pack;

[0108] Perform feature fusion and normalization on the stage life characteristics of battery components to obtain the comprehensive battery life feature set;

[0109] Verify and optimize the comprehensive battery life feature set based on historical data to obtain the battery life evolution characteristics.

[0110] In the embodiments of the present invention, first, time series reconstruction is performed on the battery health state assessment data. The sliding window method is used to convert discrete data points into a continuous time series. The reconstruction window length is set to 24 data points, and the step size is 6 data points. Subsequently, the empirical mode decomposition method is used to decompose the reconstructed time series into trend components, periodic components, and random components. The number of intrinsic mode function (IMF) components is strictly controlled within the range of 3-8. The instantaneous frequency and amplitude of each IMF component are extracted through the Hilbert-Huang transform algorithm. Then, significance screening is performed on the multi-dimensional decomposition features of the battery health state. Using the variance contribution rate criterion, the feature components with an explained variance contribution rate exceeding 85% are retained. The evolution pattern of the screened key features is identified. The clustering analysis method is used to divide the features into three typical evolution stages: initial, stable, and decline. Based on the battery health state evolution law, the nonlinear least squares method is used to fit the degradation curve of the key features. The specific curve model is the two-parameter Weibull distribution function. The change point detection is performed on the battery life degradation model. The PELT (Pruned Exact Linear Time) algorithm is applied to identify the key turning points of the life. The degradation characteristics of each stage are extracted according to the key stage division of the life. The feature statistics of each stage are calculated, including mean, variance, skewness, and kurtosis. The Z-score normalization process is performed on the stage life characteristics of the battery components to eliminate the influence of dimensions. Finally, cross-validation is performed on the comprehensive battery life feature set based on the labeled data in the historical database. The prediction accuracy of the feature set is evaluated through the repeated sampling method to obtain the battery life evolution characteristics.

[0111] Through time series reconstruction, the present invention can effectively retain the dynamic change information of the health state of the battery pack, providing a complete data basis for subsequent analysis. The decomposition of trend, cycle and random components helps to accurately extract the long-term degradation trend and short-term fluctuation characteristics of the health state of the battery pack, improving the resolution ability of state assessment. Significance screening can reduce redundant information, extract key features, improve analysis efficiency and enhance the generalization ability of the model. Evolution pattern recognition can reveal the change law of the health state of the battery pack, making the life prediction more accurate and reliable. Degradation curve fitting ensures the rationality of the life model, enabling it to accurately describe the degradation process of the battery. Change point detection can identify the key stages of battery life, providing a basis for health assessment and maintenance strategies. Feature extraction in different life stages makes the characteristics of different life stages clearer, providing support for differential maintenance. Feature fusion and normalization processing improve the consistency and comparability of data, making the prediction ability of the model more stable and reliable. The verification and optimization of historical data enhance the accuracy of life evolution characteristics and improve the adaptability and practicability of the prediction model.

[0112] Preferably, the abnormal level assessment of the battery pack health state assessment data in step S5 includes:

[0113] Conduct historical statistical analysis on the battery pack health state assessment data to obtain the battery pack health state benchmark threshold;

[0114] Based on the battery pack health state benchmark threshold, calculate the deviation of the battery pack health state assessment data to obtain the battery pack health state deviation index;

[0115] Conduct pattern recognition and classification on the battery pack health state deviation index to obtain the battery pack health state abnormal pattern characteristics;

[0116] Based on the battery pack health state abnormal pattern characteristics, conduct quantitative assessment of the abnormal severity to obtain the battery pack abnormal degree index;

[0117] The quantization level of the abnormal severity quantitative assessment is 5 levels, where the slight abnormal threshold is 5%-10% deviation, the warning abnormal threshold is 10%-15% deviation, the moderate abnormal threshold is 15%-20% deviation, the severe abnormal threshold is 20%-25% deviation, and the critical abnormal threshold is greater than 25% deviation;

[0118] Conduct a fault impact assessment on the battery pack abnormal degree index to obtain the battery pack system risk level, where the weight of the impact factor for safety is 0.4-0.5, for performance is 0.2-0.3, and for life is 0.2-0.3, and the risk score range is 0-100;

[0119] Based on the battery pack system risk level, conduct multi-dimensional abnormal index analysis to obtain the battery pack abnormal association rules;

[0120] Hierarchically classify the abnormal association rules of the battery pack to obtain the hierarchical abnormal evaluation criteria for the battery pack;

[0121] Based on the hierarchical abnormal evaluation criteria of the battery pack, map the abnormal level of the current battery pack state to obtain the real-time abnormal level result of the battery pack;

[0122] Verify and confirm the real-time abnormal level result of the battery pack through domain knowledge rules to obtain the abnormal state level of the battery pack.

[0123] In the embodiment of the present invention, first, perform descriptive statistical analysis on the historical operation data of the battery pack, calculate the median and interquartile range of the key parameters of voltage, current, and temperature, and establish a benchmark threshold for the health state of the battery pack, where the benchmark threshold is based on the 95% confidence interval of the normal operation range; then use the Euclidean distance calculation method to calculate the deviation between the battery pack health state evaluation data and the benchmark threshold to obtain the battery pack health state deviation index, and the specific calculation formula is sqrt(sum((current data - benchmark threshold)^2)); then use the clustering analysis method to perform pattern recognition and classification on the battery pack health state deviation index, and divide the abnormal patterns into three typical characteristics: clustering abnormality, mutational abnormality, and progressive abnormality; based on the abnormal pattern characteristics, evaluate according to the predefined 5-level abnormal severity quantification standard, where minor abnormality corresponds to 5%-10% deviation, warning abnormality corresponds to 10%-15% deviation, moderate abnormality corresponds to 15%-20% deviation, severe abnormality corresponds to 20%-25% deviation, and critical abnormality corresponds to more than 25% deviation; perform a fault impact assessment on the abnormal degree index, construct a risk scoring model, with a weight ratio of safety factor 0.45, performance factor 0.25, and life factor 0.30, and calculate the system risk level of 0-100 points; perform multi-dimensional abnormal index analysis on the battery pack system risk level through the association rule mining algorithm, extract abnormal association rules; use the analytic hierarchy process to classify the abnormal association rules and construct a multi-level abnormal evaluation standard; based on the constructed hierarchical abnormal evaluation standard, map the current battery pack state to the corresponding abnormal level; finally, verify and confirm the abnormal level through the domain expert knowledge rule base to obtain the final abnormal state level of the battery pack.

[0124] Through historical statistical analysis, the present invention determines the health status benchmark threshold, which can provide an objective reference standard and improve the accuracy of anomaly detection. Deviation calculation can quantify the degree of deviation of the battery pack status, making anomaly identification more intuitive. Pattern recognition and classification help to accurately distinguish different types of anomaly patterns and provide a basis for subsequent anomaly handling. The quantitative assessment of anomaly severity enables a more hierarchical division of anomaly levels through a grading standard, facilitating the adoption of targeted maintenance measures. The assessment of fault impact can comprehensively consider safety, performance, and lifespan factors, improving the comprehensiveness of system risk assessment. The analysis of multi-dimensional anomaly indicators can uncover the correlation relationships between anomalies and enhance the intelligence level of anomaly detection. The hierarchical anomaly level division ensures the systematicness and operability of anomaly assessment, making the assessment results more scientific and reasonable. Anomaly level mapping can achieve real-time judgment of the current status and improve the timeliness of battery pack anomaly warning. The verification of domain knowledge rules enhances the reliability of the assessment results, reduces misjudgment and missed judgment, and improves the accuracy of anomaly status levels.

[0125] Preferably, step S6 includes the following steps:

[0126] Step S61: Perform time series prediction on the battery pack life evolution characteristics to obtain an estimated remaining service life;

[0127] Step S62: Based on the battery pack anomaly status level, conduct classification and recognition of the battery pack anomaly patterns to obtain the fault type diagnosis result;

[0128] Step S63: Perform multi-threshold division on the estimated remaining service life to obtain the health warning level division standard;

[0129] Step S64: Based on the fault type diagnosis result, optimize the allocation of maintenance resources to obtain a differentiated maintenance strategy library;

[0130] Step S65: Obtain the real-time status data of the battery pack; compare and analyze the real-time status data of the battery pack with the health warning level division standard to obtain health warning information;

[0131] Step S66: Based on the differentiated maintenance strategy library, optimize and determine the battery pack maintenance timing to obtain a pre-maintenance execution plan;

[0132] Step S67: Integrate and make a decision on the health warning information and the pre-maintenance execution plan to obtain the battery pack health prediction result.

[0133] In the embodiments of the present invention, first, an autoregressive integrated moving average (ARIMA) time series prediction method is adopted. Based on the historical operation data and life evolution characteristics of the battery pack, a prediction model is constructed. The model parameters are estimated by the least squares method to predict the remaining useful life of the battery pack, and the confidence level of the prediction interval is set to 95%. Subsequently, according to the abnormal state level of the battery pack, cluster analysis in a multi-dimensional feature space is performed, and the battery pack failure modes are divided into four typical failure types: insulation failure, electrode material degradation, electrolyte degradation, and increased internal resistance. Multiple threshold divisions are implemented on the estimated value of the remaining useful life to construct a four-level health warning level standard, specifically including: safe use period (>80% of the life), warning period (60%-80% of the life), decline period (40%-60% of the life), and critical period (<40% of the life). Based on the diagnosis results of the failure types, a differential maintenance strategy library is constructed through an expert knowledge base and historical maintenance data, including specific maintenance plans for detection, replacement, and repair for different failure types. The operation data of the battery pack, including key parameters such as voltage, current, and temperature, are obtained in real time. The real-time state data is quantitatively compared with the health warning level division standard to calculate the deviation degree and abnormal index, and detailed health warning information is generated. According to the differential maintenance strategy library, combined with the current health state, remaining useful life, and failure type of the battery pack, the optimal maintenance timing is determined through the decision tree algorithm optimization to form a pre-maintenance execution plan. Finally, multi-criteria comprehensive evaluation is performed on the health warning information and the pre-maintenance execution plan. The comprehensive score is calculated by the weighted summation method to generate the final battery pack health prediction result, and a complete diagnosis report including the health state, warning level, and maintenance suggestions is output.

[0134] The time series prediction of the present invention can capture the battery life evolution trend and improve the accuracy of the remaining life estimation. The abnormal mode classification and recognition can accurately distinguish different failure types and provide a reliable basis for fault diagnosis and maintenance decision-making. The multiple threshold division ensures the rationality of the health warning level and makes the warning information clearer and more intuitive. The optimized allocation of maintenance resources improves the scientificity and economy of resource allocation and reduces unnecessary maintenance costs. The acquisition and comparison analysis of real-time state data can dynamically monitor the battery health status and improve the timeliness of health warning. The optimized maintenance timing ensures that the maintenance measures are executed at the best time, extends the service life of the battery pack, and reduces the failure risk. The integrated decision combines multi-dimensional information, improves the comprehensiveness and reliability of health prediction, and provides more accurate support for battery management.

[0135] Especially importantly, the comparison and analysis of the real-time state data of the battery pack with the health warning level division standard includes:

[0136] Perform time series alignment processing on the real-time state data of the battery pack to obtain a standardized real-time operation feature matrix;

[0137] Perform matching calculations on the health warning level division criteria based on the standardized real-time operation feature matrix to obtain a health warning matching degree vector;

[0138] Perform an optimal classification mapping on the health warning matching degree vector to obtain the current health warning level of the battery pack;

[0139] Perform a correlation analysis on the historical health status data based on the current health warning level of the battery pack to obtain the health status change trend;

[0140] Perform an anomaly analysis on the health status change trend to obtain health status anomaly warning indicators;

[0141] Generate health warning information based on the health status anomaly warning indicators, where the health warning information includes the current health warning level, health status trend, and anomaly warning information.

[0142] In the embodiment of the present invention, first, perform a time series alignment process on the real-time status data of the battery pack, use the linear interpolation method to unify the sampling frequency, resample the multi-dimensional data of voltage, current, and temperature to a unified time interval, and use the maximum-minimum normalization method to map each feature variable to the [0, 1] interval to generate a standardized real-time operation feature matrix; subsequently, based on the standardized real-time operation feature matrix, use the Euclidean distance calculation method to perform matching calculations on the health warning level division criteria, and the specific calculation formula is sqrt(sum((real-time feature - standard feature)^2)) to obtain a health warning matching degree vector; then, use the maximum a posteriori probability criterion to perform an optimal classification mapping on the health warning matching degree vector, and divide the current health status of the battery pack into four levels: safe, warning, decline, and dangerous; based on the current health warning level of the battery pack, perform a correlation analysis on the historical health status data of the battery pack, use the sliding window method to calculate the slope and acceleration of the health status change within a continuous time period to obtain the health status change trend; perform an anomaly analysis on the health status change trend, identify abnormal fluctuation characteristics by calculating the variance of the trend line, the number of inflection points, and the extreme value frequency to generate health status anomaly warning indicators; finally, based on the health status anomaly warning indicators, generate a comprehensive health warning information including the current health warning level, a description of the health status trend, and anomaly warning information, and quantitatively and qualitatively evaluate the abnormal characteristics.

[0143] The present invention ensures the synchronization of real-time data, enabling the real-time operating feature matrix to match the standardized warning criteria and improving the comparability of data. Through the health warning matching degree vector generated by matching calculations, the determination of the warning level is made more accurate. The optimal classification mapping can accurately delimit the current health warning level of the battery pack, providing a clear basis for subsequent decision-making. The correlation analysis of historical health status data reveals the long-term change trend of the battery pack's health status, helping to identify potential problems in advance. The anomaly analysis further extracts anomaly warning indicators, improving the sensitivity and accuracy of anomaly detection. The finally generated health warning information comprehensively reflects the changes in the health status of the battery pack, provides warnings in a timely manner, and can support the optimization of decision-making and maintenance plans.

[0144] Preferably, the present invention further provides a system for battery pack health prediction for performing the above-mentioned method for battery pack health prediction. The system for battery pack health prediction includes:

[0145] A multi-modal data acquisition module for performing multi-modal sensing data acquisition on the battery pack to obtain voltage data, current data, temperature data, and acoustic vibration data as the original monitoring data set;

[0146] A data preprocessing and feature extraction module for preprocessing the original monitoring data set, performing time-frequency domain feature extraction, and constructing an operating feature matrix of the battery pack;

[0147] An intelligent diagnosis model construction module for obtaining the historical operating data of the battery pack; constructing an intelligent diagnosis model for the health situation of the battery pack using the historical operating data of the battery pack and the operating feature matrix of the battery pack, where constructing the intelligent diagnosis model for the health situation of the battery pack includes training a generative adversarial network and training a multi-input deep neural network;

[0148] A health status evaluation module for evaluating the status of the battery pack based on the intelligent diagnosis model for the health situation of the battery pack to obtain battery pack health status evaluation data;

[0149] A life evolution and anomaly detection module for performing time series decomposition on the battery pack health status evaluation data to separate the aging trend and short-term fluctuations to obtain the battery pack life evolution characteristics; performing an anomaly level evaluation on the battery pack health status evaluation data to obtain the battery pack anomaly status level;

[0150] A health prediction and pre-maintenance module for performing health warning and pre-maintenance plan design based on the battery pack life evolution characteristics and the battery pack anomaly status level to obtain the battery pack health prediction result.

[0151] Therefore, from any perspective, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is not defined by the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the application documents are intended to be encompassed within the present invention.

[0152] The above are only specific embodiments of the present invention, enabling those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the broadest scope consistent with the principles and novel features invented herein.

Claims

1. A method for predicting the health of a battery pack, characterized in that, It includes the following steps: Step S1: Collect multi-modal sensing data of the battery pack to obtain voltage data, current data, temperature data, and acoustic vibration data as the original monitoring data set; Step S2: Preprocess the original monitoring data set, extract time-frequency domain features, and construct an operating feature matrix of the battery pack; Step S3: Obtain the historical operating data of the battery pack; use the historical operating data of the battery pack and the operating feature matrix of the battery pack to construct an intelligent diagnosis model for the health status of the battery pack. The construction of the intelligent diagnosis model for the health status of the battery pack includes training a generative adversarial network and training a multi-input deep neural network. Specifically: use the historical data to train a preset generative adversarial network to obtain a discriminator model; use the operating feature matrix of the battery pack as the input to train a preset multi-input deep neural network model to obtain a battery pack state recognition model; Step S4: Evaluate the state of the battery pack based on the intelligent diagnosis model for the health status of the battery pack to obtain the health status evaluation data of the battery pack. Step S4 includes the following steps: Step S41: Input the operating feature matrix of the battery pack into the discriminator model, calculate the deviation degree from the normal battery pack behavior pattern, and obtain a preliminary anomaly metric; Step S42: Input the operating feature matrix of the battery pack into the battery pack state recognition model, obtain the error vector between the predicted value and the actual value of each parameter of the battery pack, and obtain a parameter performance degradation index; Step S43: Integrate the preliminary anomaly metric and the parameter performance degradation index to construct a health index for the battery pack. The range of the health index of the battery pack is 0-100, where 100 represents a brand-new state and 0 represents complete failure; Step S44: Calculate the abnormal probability distribution map of each component of the battery pack based on the parameter performance degradation index; Step S45: Integrate the health index of the battery pack and the abnormal probability distribution map to generate a health status evaluation report for the battery pack. The health status evaluation report for the battery pack includes a health score, a heat map of the abnormal area, and a deviation degree of key parameters; Step S5: Perform time series decomposition on the health status evaluation data of the battery pack to separate the aging trend and short-term fluctuations, and obtain the life evolution characteristics of the battery pack; perform an abnormal level evaluation on the health status evaluation data of the battery pack to obtain the abnormal state level of the battery pack; Step S6: Based on the life evolution characteristics of the battery pack and the abnormal state level of the battery pack, design a health warning and pre-maintenance plan to obtain the health prediction result of the battery pack.

2. The method for predicting the health of a battery pack according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Perform high-frequency sampling on the terminal voltage of the battery pack to obtain voltage time series data; Step S12: Monitor the charge and discharge current of the battery pack in real time to obtain current time series data; Step S13: Based on a distributed temperature sensor array, perform multi-point measurement of the surface temperature of the battery pack to obtain temperature distribution data; Step S14: Collect acoustic vibration signals through piezoelectric sensors attached to the surface of the battery pack to obtain vibration spectrum data; Step S15: Perform time synchronization and integration on the voltage time series data, current time series data, temperature distribution data, and vibration spectrum data to obtain a multi-modal original monitoring data set.

3. The method for predicting the health of a battery pack according to claim 1, wherein, Step S2 includes the following steps: Step S21: Perform data cleaning and outlier detection on the original monitoring data set to obtain a preliminary processed data set with noise points and outliers removed; Step S22: Perform noise suppression on the preliminary processed data set to obtain a denoised data set; Step S23: Perform standardization and normalization processing on the denoised data set to obtain a standardized data set; Step S24: Perform continuous wavelet transform on the standardized data set to obtain a time-frequency domain decomposition result; Step S25: Extract energy distribution, scale coefficient, and frequency characteristics from the time-frequency domain decomposition result to obtain a multi-dimensional feature vector; Step S26: Perform dimensionality reduction and feature selection on the multi-dimensional feature vector to obtain a battery pack operation feature matrix.

4. The method for predicting the health of a battery pack according to claim 1, wherein, The training of the preset generative adversarial network using historical data in Step S3 includes: Screen the historical data, extract data segments with a signal-to-noise ratio greater than or equal to 25 dB, and classify the data through an expert annotation system to obtain a normal behavior pattern data set; Construct a conditional generative adversarial network based on the normal behavior pattern data set to obtain an initial network structure; The conditional input dimension in the conditional generative adversarial network is 4 - 8 dimensions, the number of hidden layers is 3 - 5 layers, and the number of neurons in each layer is 64 - 256; Initialize the generator parameters of the initial network structure to obtain a generator model, where the learning rate is 0.0001 - 0.001; Perform discriminator adversarial training based on the normal behavior pattern data set and the generator model to obtain an initial discriminator, where the proportion of real samples is 40% - 60%, the proportion of generated samples is 40% - 60%, and the update frequency of the initial discriminator is once every 1 - 3 generator updates; Iteratively optimize the generator model and the initial discriminator to obtain a generative adversarial network model; Extract the discriminator part based on the generative adversarial network model to obtain a discriminator model.

5. The method for predicting the health of a battery pack according to claim 1, characterized in that, The training of the preset multi-input deep neural network model with the battery pack operation feature matrix as the input in Step S3 includes: Group the battery pack operation feature matrix according to data types to obtain a voltage feature sub-matrix, a current feature sub-matrix, a temperature feature sub-matrix, and an acoustic feature sub-matrix; The voltage feature dimension is 0 - 20 dimensions, the current feature dimension is 8 - 16 dimensions, the temperature feature dimension is 12 - 24 dimensions, and the acoustic feature dimension is 20 - 40 dimensions; Construct a voltage-current processing based on the voltage feature sub-matrix and the current feature sub-matrix to obtain a time-series electrical characteristic neural network functional unit, where the number of hidden layers is 2 - 3 layers, the number of hidden units is 64 - 128, and the sequence length is 20 - 50 time steps; Construct a temperature data processing based on the temperature feature sub-matrix to obtain a temperature distribution spatial feature neural network functional unit, where the number of convolutional layers is 2 - 4 layers, the number of convolutional kernels is 16 - 64, and the size of the convolutional kernel is 3×3; Construct an acoustic vibration data processing based on the acoustic feature sub-matrix to obtain a vibration spectrum feature neural network functional unit, where the number of frequency segments is 8 - 16 segments, the number of attention heads is 4 - 8, and the output feature dimension is 32 - 64 dimensions; Design a multi-head attention mechanism based on the neural network functional unit of the temporal electrical characteristics, the neural network functional unit of the spatial characteristics of the temperature distribution, and the neural network functional unit of the vibration spectrum characteristics to obtain a weight calculation model, where the number of attention heads is 4-8; Fuse the outputs of the neural network functional unit of the temporal electrical characteristics, the neural network functional unit of the spatial characteristics of the temperature distribution, and the neural network functional unit of the vibration spectrum characteristics through the weight calculation model to obtain a multi-modal fusion feature vector; Construct a fully connected layer based on the multi-modal fusion feature vector to obtain a battery pack health state classifier and a degradation degree regressor; The number of layers of the fully connected layer is 2-3 layers, where the number of hidden neurons in the first layer is 128-256, halving layer by layer, and the dropout rate is 0.2-0.5; Jointly train the battery pack health state classifier and the degradation degree regressor, and optimize the parameters based on the joint training results to obtain a battery pack state recognition model.

6. The method for predicting the health of a battery pack according to claim 1, characterized in that, The time series decomposition of the battery pack health state evaluation data in step S5 to separate the aging trend and short-term fluctuations includes: Perform time series reconstruction on the battery pack health state evaluation data to obtain a battery pack health state time series data set; Decompose the battery pack health state time series data set into trend components, periodic components, and random components to obtain multi-dimensional decomposition features of the battery pack health state, where the number of IMF components is limited to within 3-8; Perform significance screening on the multi-dimensional decomposition features of the battery pack health state to obtain key features of the battery pack health state; Perform evolutionary pattern recognition on the key features of the battery pack health state to obtain the evolutionary law of the battery pack health state; Fit a degradation curve to the key features of the battery pack health state based on the evolutionary law of the battery pack health state to obtain a battery pack life degradation model; Perform change point detection on the battery pack life degradation model to obtain the key stage division of the battery pack life; Extract the degradation characteristics of each stage based on the key stage division of the battery pack life to obtain the stage-by-stage life characteristics of the battery pack; Perform feature fusion and normalization processing on the stage-by-stage life characteristics of the battery pack to obtain a comprehensive feature set of the battery pack life; Verify and optimize the comprehensive feature set of the battery pack life based on historical data to obtain the evolutionary characteristics of the battery pack life.

7. The method for predicting the health of a battery pack according to claim 1, wherein The abnormal level evaluation of the battery pack health state evaluation data in step S5 includes: Perform historical statistical analysis on the battery pack health state evaluation data to obtain the battery pack health state benchmark threshold; Calculate the deviation of the battery pack health state evaluation data based on the battery pack health state benchmark threshold to obtain the battery pack health state deviation index; Perform pattern recognition and classification on the battery pack health state deviation index to obtain the abnormal pattern features of the battery pack health state; Quantitatively evaluate the abnormal severity based on the abnormal pattern features of the battery pack health state to obtain the battery pack abnormal degree index; The quantization level of the abnormal severity quantitative evaluation is 5 levels, where the mild abnormal threshold is 5%-10% deviation, the warning abnormal threshold is 10%-15% deviation, the moderate abnormal threshold is 15%-20% deviation, the severe abnormal threshold is 20%-25% deviation, and the critical abnormal threshold is greater than 25% deviation; Evaluate the fault impact of the battery pack anomaly degree index to obtain the system risk level of the battery pack. Among them, the weight of the influencing factor for safety is 0.4 - 0.5, for performance is 0.2 - 0.3, and for lifespan is 0.2 - 0.

3. The risk score range is 0 - 100; Conduct multi-dimensional anomaly index analysis based on the system risk level of the battery pack to obtain the battery pack anomaly association rules; Perform hierarchical anomaly level division on the battery pack anomaly association rules to obtain the battery pack hierarchical anomaly evaluation criteria; Perform real-time anomaly level mapping of the current battery pack state based on the battery pack hierarchical anomaly evaluation criteria to obtain the battery pack real-time anomaly level result; Verify and confirm the battery pack real-time anomaly level result with domain knowledge rules to obtain the battery pack anomaly state level.

8. The method for battery pack health prediction according to claim 1, wherein Step S6 includes the following steps: Step S61: Conduct time series prediction on the battery pack lifespan evolution characteristics to obtain the estimated remaining service life; Step S62: Classify and identify the battery pack anomaly patterns based on the battery pack anomaly state level to obtain the fault type diagnosis result; Step S63: Perform multi-threshold division on the estimated remaining service life to obtain the health warning level division criteria; Step S64: Optimize the allocation of maintenance resources based on the fault type diagnosis result to obtain a differentiated maintenance strategy library; Step S65: Obtain the battery pack real-time status data; Compare and analyze the battery pack real-time status data with the health warning level division criteria to obtain the health warning information; Step S66: Optimize and determine the battery pack maintenance timing based on the differentiated maintenance strategy library to obtain a pre-maintenance execution plan; Step S67: Integrate and make a decision on the health warning information and the pre-maintenance execution plan to obtain the battery pack health prediction result.

9. A system for predicting the health of a battery pack, characterized in that, A system for performing the method for battery pack health prediction as described in claim 1, the system for battery pack health prediction includes: A multi-modal data acquisition module for collecting multi-modal sensing data of the battery pack to obtain voltage data, current data, temperature data, and acoustic vibration data as the original monitoring data set; A data preprocessing and feature extraction module for preprocessing the original monitoring data set and extracting time-frequency domain features, and constructing a battery pack operation feature matrix; An intelligent diagnosis model construction module for obtaining the battery pack historical operation data; Using the battery pack historical operation data and the battery pack operation feature matrix to construct an intelligent diagnosis model for the battery pack health situation. Among them, constructing the intelligent diagnosis model for the battery pack health situation includes training a generative adversarial network and training a multi-input deep neural network. Specifically: Training a preset generative adversarial network with historical data; Using the battery pack operation feature matrix as the input to train a preset multi-input deep neural network model; A health status evaluation module for evaluating the status of the battery pack based on the intelligent diagnosis model of the battery pack health situation to obtain battery pack health status evaluation data; A lifespan evolution and anomaly detection module for performing time series decomposition on the battery pack health status evaluation data to separate the aging trend and short-term fluctuations to obtain the battery pack lifespan evolution characteristics; Evaluating the anomaly level of the battery pack health status evaluation data to obtain the battery pack anomaly state level; A health prediction and pre-maintenance module is used to design a health warning and pre-maintenance plan based on the battery pack life evolution characteristics and the battery pack abnormal state level, and obtain the battery pack health prediction result.

Images