New energy automobile battery health state monitoring system and implementation method

Abstract

The invention discloses a new energy automobile battery health state monitoring system and an implementation method, and relates to the field of new energy automobile intelligent monitoring. According to the system, in order to solve the problems that new energy automobile battery health state monitoring precision is insufficient, abnormal response lags behind and multi-dimensional data fusion efficiency is low, full-type battery compatibility and multi-physical-quantity accurate sensing are achieved through a battery data acquisition module; the data processing and analysis module is used for completing multi-modal data feature decoupling and depth modeling; by means of a battery state monitoring module, a display and interaction module and a communication module, millisecond-level abnormity early warning, health state visualization and full-link data intercommunication are realized; and finally, under a complex working condition, the intelligence and precision level of new energy automobile battery health state monitoring is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent monitoring of new energy vehicles, and more particularly to a new energy vehicle battery health state monitoring system and implementation method. BACKGROUND

[0002] As the core carrier of low-carbon development in the transportation field, the battery health state as a core indicator representing the capacity attenuation and internal resistance evolution of the battery directly affects the vehicle's endurance, safety performance and life cycle economy. In the prior art: In terms of sensor compatibility, traditional contact sensing technology is limited by the diversity of physical packaging forms, making it difficult to achieve full coverage monitoring of new batteries. Non-conformal sensors can easily cause interface contact impedance to rise, causing signal distortion. In addition, existing commercial sensors mostly use single physical quantity detection mechanisms, lacking the ability to simultaneously capture the characteristics of electrochemical-thermal-mechanical multi-field coupling, limiting the construction of high-precision SOH estimation models.

[0003] At the data processing algorithm level, traditional methods have the dual limitations of single feature extraction dimension and insufficient noise suppression capability. Mechanism-driven methods represented by the equivalent circuit model ECM have difficulty in accurately describing the nonlinear aging dynamics behavior of lithium-ion batteries, especially under high-rate charging and discharging conditions, with significantly increased parameter identification error.

[0004] In terms of anomaly detection and early warning mechanism, existing research is mostly based on fixed threshold or statistical process control SPC methods, which cannot adapt to the time-varying characteristics of battery aging parameters. Traditional methods have insufficient sensitivity to early micro-defects, with a false alarm rate as high as 15%-20%.

[0005] To address the above challenges, the present application proposes a multi-modal fusion driven battery health state monitoring system. SUMMARY

[0006] To address the existing deficiencies, the present application discloses a new energy vehicle battery health state monitoring system and implementation method, which can effectively improve the intelligence and reliability of battery management, meeting user needs.

[0007] The present application adopts the following scheme: a new energy vehicle battery health state monitoring system, wherein the system comprises: a battery data acquisition module comprising a conformal sensor array, the sensor array including a tunneling field effect transistor voltage current sensor, a terahertz thermal imaging sensor, and an atomic force electrochemical cycle counter; A battery state monitoring module based on real-time data stream analysis and dynamic threshold modeling, which analyzes collected data through a multi-scale feature extraction algorithm to achieve millisecond-level dynamic monitoring of the charging and discharging process; an abnormality detection engine driven by reinforcement learning is built into the battery state monitoring module to adaptively learn the battery operation mode; combined with battery historical data and aging models, a generative adversarial network is used to generate accurate maintenance prompts; A data processing and analysis module, which includes a data preprocessing submodule, a feature extraction submodule, a health state evaluation submodule, and an abnormality diagnosis submodule; The data preprocessing submodule uses a fusion algorithm of variational mode decomposition (VMD) and adaptive median filtering, introduces empirical wavelet transform (EWT) and fractal dimension calculation to extract 40+ dimensional key features from time domain, frequency domain, time-frequency domain, and complex system characteristics; the health state evaluation submodule constructs a battery health factor prediction model based on graph convolutional neural network (GCN); the abnormality diagnosis submodule develops an abnormal manifold modeling of generative adversarial network (GAN), which simulates normal state distribution through a generator and detects data deviation in real time through a discriminator; The output end of the data preprocessing submodule is connected to the input end of the feature extraction submodule, the output end of the feature extraction submodule is connected to the input end of the health state evaluation submodule, and the output end of the health state evaluation submodule is connected to the input end of the abnormality diagnosis submodule; A display and interaction module designed with an intuitive graphical interface; A communication module designed with a standardized communication protocol; The output end of the battery data collection module is connected to the input end of the data processing and analysis module, the output end of the data processing and analysis module is connected to the input end of the battery state monitoring module and the display and interaction module, and the communication module is bidirectionally connected to the data processing and analysis module. As a further technical solution of the present application, the battery data collection module includes a terahertz super surface intelligent sensing array module, a bionic electronic skin compatible architecture module, a neuromorphic data preprocessing engine module, a microwave photon anti-interference transmission link module, and a self-calibration and health management module; The terahertz super surface intelligent sensing array module uses laser direct writing to prepare a conformal sensor array with a thickness of ≤0.1 mm on a polyimide substrate; it uses terahertz time-domain spectroscopy to detect the internal temperature distribution of the battery with a temperature resolution of 0.01℃; it combines carbon nanotube cantilever beams with electrochemical impedance spectroscopy to achieve a charging and discharging count error of ≤0.1 times; the array includes a flexible solar thin film and a vibration energy harvester to achieve self-power supply with an average power consumption of ≤100μW; The biomimetic electronic skin compatible architecture module is based on liquid metal and shape memory alloy; it has a built-in multi-protocol gateway to realize the conversion of vehicle and industrial protocol data with a latency of ≤5μs; The neuromorphic data preprocessing engine module uses a spiking neural network to spatiotemporally encode the data. The microwave photonic anti-interference transmission link module is based on 60GHz millimeter-wave directional communication; The self-calibration and health management module automatically calibrates every hour; The output of the terahertz metasurface intelligent sensing array module is connected to the input of the biomimetic electronic skin compatible architecture module; the output of the biomimetic electronic skin compatible architecture module is connected to the input of the neuromorphic data preprocessing engine module; the output of the neuromorphic data preprocessing engine module is connected to the input of the microwave photonic anti-interference transmission link module; the outputs of the intelligent sensing array module and the terahertz metasurface intelligent sensing array module are connected to the input of the self-calibration and health management module.

[0008] As a further technical solution of the present invention, the battery status monitoring module includes: a multi-dimensional real-time monitoring module, a reinforcement learning anomaly detection module, and a generative maintenance decision module; The multi-dimensional real-time monitoring module performs multi-scale analysis and real-time monitoring of voltage, current, and temperature through wavelet packet decomposition and spatiotemporal grid model. The reinforcement learning anomaly detection module dynamically adjusts the anomaly threshold based on the deep deterministic policy gradient algorithm and achieves three-level early warning through Mahalanobis distance; The generative maintenance decision module generates accurate maintenance suggestions by combining battery status and user habits through generative adversarial networks and Pareto optimization. The output of the multi-dimensional real-time monitoring module is connected to the input of the reinforcement learning anomaly detection module; the output of the reinforcement learning anomaly detection module is connected to the input of the generative maintenance decision module.

[0009] As a further technical solution of the present invention, the data preprocessing submodule includes a variational mode decomposition (VMD) signal reconstruction module, an adaptive median filtering optimization module, and a data cleaning and invalid value processing module. The variational mode decomposition (VMD) signal reconstruction module employs multi-component adaptive decomposition to decompose the original voltage and current signals into six intrinsic mode functions, and avoids over-decomposition through the Bayesian information criterion. The adaptive median filtering optimization module dynamically adjusts the filtering window according to local data fluctuations. When the standard deviation is below 50% of the global standard deviation, a 3×3 window is used to suppress Gaussian noise. When the standard deviation is between 50% and 150%, a 5×5 window is used. When the standard deviation exceeds 150%, a 7×7 window is used to suppress impulse interference. The data cleaning and invalid value processing module uses the local outlier factor algorithm, with 5 nearest neighbors and a threshold of 2.0, to identify outliers with voltage jumps exceeding 80mV and current deviations exceeding 15% of the rated value; it also uses Kalman filtering combined with the expectation-maximization algorithm to optimize interpolation. The output of the variational mode decomposition (VMD) signal reconstruction module is connected to the input of the adaptive median filtering optimization module; the output of the adaptive median filtering optimization module is connected to the input of the invalid value processing module.

[0010] As a further technical solution of the present invention, the feature extraction submodule includes a time-frequency feature extraction module, a fractal dimension feature calculation module, a multi-domain feature fusion module, and a feature engineering toolchain module; The time-frequency feature extraction module performs adaptive frequency band segmentation on voltage and current signals through empirical wavelet transform. The adaptive frequency band consists of 6-8 frequency bands, covering the entire frequency range from 0.1Hz to 10kHz. The fractal dimension feature calculation module calculates the fractal dimension of the temperature field using the box dimension algorithm; The multi-domain feature fusion module includes time domain, frequency domain, time-frequency domain and fractal dimension features, forming a 40+ dimension health feature vector; The feature engineering toolchain module calculates feature importance in real time through an automated platform; The output of the time-frequency feature extraction module is connected to the input of the fractal dimension feature calculation module; the output of the fractal dimension feature calculation module is connected to the input of the multi-domain feature fusion module.

[0011] As a further technical solution of the present invention, the health status assessment submodule includes a battery pack topology modeling module, a graph convolutional neural network architecture module, a multi-scale health factor prediction module, a spatiotemporal evolution model module, and a model training and optimization module. The battery pack topology modeling module constructs a three-dimensional topology graph; The graph convolutional neural network architecture module extracts local, module-level and global features layer by layer through a three-layer GCN combined with a graph attention mechanism. The multi-scale health factor prediction module covers SOH, remaining lifespan, and failure probability through a multi-task prediction system. The spatiotemporal evolution model module predicts fault propagation in space and locks in the risk of module thermal runaway within 15 minutes. The model training and optimization module uses a hybrid loss function to enhance early aging learning, federated learning to protect privacy, and online fine-tuning every 50 cycles. The output of the battery pack topology modeling module is connected to the input of the graph convolutional neural network architecture module; the output of the graph convolutional neural network architecture module is connected to the multi-scale health factor prediction module; the output of the multi-scale health factor prediction module is connected to the spatiotemporal evolution model module; the output of the spatiotemporal evolution model module is connected to the model training and optimization module; and the output of the model training and optimization module is connected to the inputs of the graph convolutional neural network architecture module and the multi-scale health factor prediction module.

[0012] As a further technical solution of the present invention, the anomaly diagnosis submodule includes a GAN anomaly manifold modeling module, a multi-scale anomaly detection module, a fault type classification module, a fault tracing and root cause analysis module, and a real-time early warning and response module. The GAN abnormal manifold modeling module uses a deep convolutional generative adversarial network (DCGAN). The generator outputs time-series pseudo samples of voltage, current, and temperature, and the discriminator outputs anomaly scores of 0-1. After 1000 rounds of training, the difference in distribution (FID) between the generated samples and the real data is reduced to below 15, accurately simulating the normal state of the battery. The multi-scale anomaly detection module includes point anomaly detection to identify sudden single-cell failures through reconstruction errors; context anomaly detection to analyze sequence continuity using LSTM to identify progressive aging; and collective anomaly detection to combine topological graphs and graph attention mechanisms to provide early warning of regional potential thermal runaway risks up to 2 hours in advance. The fault type classification module is based on the attention ResNet classifier and automatically weights the key features of temperature to output the probability of 8 types of faults. The fault tracing and root cause analysis module constructs a causal graph model with 30+ variables; The real-time early warning and response module is equipped with three levels of early warning. The output of the GAN anomaly manifold modeling module is connected to the input of the multi-scale anomaly detection module; the output of the multi-scale anomaly detection module is connected to the input of the fault type classification module; the output of the fault type classification module is connected to the input of the fault tracing and root cause analysis module; and the output of the fault tracing and root cause analysis module is connected to the input of the real-time early warning and response module.

[0013] As a further technical solution of the present invention, the method for monitoring the health status of new energy vehicle batteries is as follows: 1. System hardware construction and data acquisition; A battery data acquisition module employing a terahertz metasurface conformal sensor array and biomimetic electronic skin is used. S2, Data Processing and Analysis Flow; The data preprocessing submodule suppresses noise and preserves weak features through a fusion algorithm of variational mode decomposition and adaptive median filtering; S3, Battery Status Monitoring and Closed-Loop Management; The battery status monitoring module achieves millisecond-level monitoring of the charging and discharging process based on real-time data streams and dynamic thresholds; S4, System Interaction and Data Communication; The display and interaction module presents real-time data, historical records, and alert information through a graphical interface, and supports user-defined thresholds. S5. System closed-loop and optimization; The system forms a closed loop across the entire chain of "collection-analysis-monitoring-maintenance-feedback".

[0014] As a further technical solution of the present invention, the implementation method of the feature extraction submodule is: 1) Empirical Wavelet Transform (EWT); Empirical wavelet transform the original signal Decomposed into N adaptive frequency bands, introducing multiphysics and dynamic weighting mechanisms, the mathematical expression is: In formula (1), For time series functions, The t-transform is used to dynamically adjust the components of each frequency band based on the energy distribution and characteristics of the signal at different times. Contribution to the original signal; For multiphysics modulation operators, For real-time state of charge, Kelvin temperature scale data collected by a temperature sensor. This represents the effective value of the current within the sliding window. The capacity loss temperature sensitivity coefficient is calculated based on the Arrhenius equation; The k-th frequency band component The calculation method is as follows: In formula (2), The improved empirical wavelet filter not only depends on the frequency w, but also introduces time-varying parameters. ; Based on the local frequency characteristics of the signal, the shape and bandwidth of the filter are adaptively adjusted, making the filter more flexible in matching the characteristics of different frequency components and improving the accuracy of decomposition. Original signal The Fourier transform converts a time-domain signal to the frequency domain, revealing the frequency components and energy distribution of the signal. It is a frequency domain weighting function; 2) Time-frequency feature extraction; Root mean square value In formula (3), It represents the time-varying root mean square value at time t, dynamically reflecting the change in the energy intensity of the signal at different times; is the i-th sample value in the signal sequence at time t, representing the instantaneous amplitude of the signal in the time dimension; N is the number of sampling points, representing the length of the signal data involved in the calculation; The introduced time-varying weighting function is represented by L; L is the electrochemical and thermodynamic joint modulation function. SEI film growth rate coefficient calibrated based on aging tests; It is the temperature gradient tensor; The average current density within the sliding event window; This is a two-domain mixed processing function; kurtosis: In formula (4), It is a time-varying kurtosis used to characterize the amplitude distribution of a signal at time t, and to detect the impulse components in the signal and their changes. The introduced morphological adjustment function adjusts the contribution of different sampling points in the kurtosis calculation according to the characteristics of the signal morphology. As a driver of health; Commonly used in the battery field, it represents the state of charge, indicating the health or performance parameters of the system at time t; This is the expansion factor; The effect of current on the system varies with time; peak ripple factor: In formula (5), The time-varying crest factor is used to measure the relative magnitude of the peak value and the effective value of a signal at time t, and to assess the impact of the signal. The peak value of the weighted signal at time t. It is an adaptive enhancement function; This is the local gradient of the signal at point i, reflecting the rate of change; The gradient response function has the following parameters. Modulate sensitivity to mutations; This is an attenuation factor that is dynamically adjusted based on historical peak wave factors to prevent background fluctuation interference. 3) Box dimension calculation; In formula (6), D is the target fractal dimension; Based on the basic scale parameter, This indicates a gradual refinement of the analytical scale from macro to micro. Let represent the dynamic scale parameter set for the i-th local region and the j-th time window of the dataset; Ni,j(ϵ): the number of boxes required to cover the i-th local region and the j-th time window at the dynamic scale ϵi,j; M is the number of local regions divided in the spatial dimension; T is the number of windows divided in the temporal dimension. It is a spatiotemporal dynamic weighting function that is dynamically adjusted based on the information entropy of the local region and the importance of the time window; The variance of the number of boxes filled in the i-th region and the j-th time window; This is the variance influence factor, which controls the moderating effect of variance on weights. Let be the correlation coefficient between the i-th region and its neighboring regions in the j-th time window. Correlation influencing factors; 4) Multifractal analysis; In formula (7), Here, q is the generalized scaling exponent, and q is the order moment parameter. Focus on high-probability areas. Focus on low-probability, rare events. Corresponding to box dimension calculation; The basic scale parameter represents the initial box size used when covering the system. This indicates a refined analytical process from macroscopic to microscopic levels; In scale The total number of boxes required for the under-coverage system; Let be the probability measure of the i-th box. For dynamic feature interaction weights, quantize the order moment q and scale of the i-th box and its neighboring box j. The interaction strength under these conditions; The distance between boxes i and j in the feature space. This refers to the spatial interaction strength parameter; This is a dynamic scaling adjustment factor; The multi-scale interaction index quantifies the coupling effect of the i-th box across different scales. For multi-scale coupling strength parameters; 5) Feature fusion and weight adjustment; Time-domain feature weight formula: In formula (8), Here, the time-domain feature weights are defined for the nth iteration, where n is the current iteration number, ranging from 1 to n to N, and gradually increasing with each iteration; N is the total iteration number, defining the time span for weight adjustment and determining the iteration progress factor. The range of variation; The dynamic attenuation coefficient, This is the decay rate parameter; Adaptive rate of change coefficient; Dynamic offset coefficient; The correlation sensitivity coefficient; The real-time correlation coefficient between the time-domain and frequency-domain characteristics; when At that time, the center of weight change shifts towards the earlier cycle; when In time, the system shifts to a later, cyclical manner to achieve feature-driven weight adjustment; This is a cross-domain coupling factor that quantifies the influence of frequency domain features on time domain weights, with a value range of [−1, 1]. When, positive correlation enhances the time domain weights; when At that time, negative correlation suppresses time-domain weights; Variance regularization factor, which suppresses abnormal fluctuations caused by noise; The real-time variance of the time-domain features; Fractal feature weight formula: In formula (9), The fractal feature weights are defined in the nth iteration; n is the current iteration number, reflecting the data processing progress, with a value range of 1 ≤ n ≤ N; N is the total number of iterations. This is the dynamic fractal weight growth coefficient. The larger the value, the higher the upper limit of the fractal feature weight; The adaptive fractal weight growth rate coefficient replaces the fixed w. The larger the value, the faster the weight increases; This refers to the dynamic fractal weight growth offset coefficient. This is a fractal complexity enhancement factor, which strengthens the influence of fractal dimension on weights. The dimension of the fractal feature at the nth iteration; Frequency domain feature weight formula: ;In formula (10), The time-domain feature weights for the current loop are calculated using formula (8); The fractal feature weights for the current loop are calculated using formula (9); The time-domain weighting influence coefficient has a value range of (0,1). The larger the value, the stronger the reduction effect of the time-domain feature weights on the frequency-domain feature weights; This is the fractal weighting influence coefficient, with a value range of (0,1). The larger the value, the stronger the reduction effect of the fractal feature weights on the frequency domain feature weights. This has positive and beneficial effects. The new energy vehicle battery health status monitoring system offers numerous positive benefits. In terms of compatibility and data acquisition, the conformal sensor array is compatible with 99% of existing battery models, and multiple types of sensors accurately collect key voltage and current data in real time, providing a comprehensive information foundation for monitoring. For data processing and analysis, the system integrates algorithmic noise reduction, multi-dimensional feature extraction, and neural network modeling, significantly improving the accuracy of battery health assessment and the timeliness of anomaly diagnosis. Regarding dynamic monitoring and maintenance, millisecond-level monitoring combined with reinforcement learning and generative adversarial networks can adaptively identify battery operating modes and generate precise maintenance prompts, effectively preventing faults. In terms of interaction and communication, the graphical interface allows users to intuitively obtain information, and standardized communication protocols enable remote interaction and system upgrades, supporting intelligent management throughout the entire battery lifecycle. The collaborative operation of all system modules significantly improves battery management efficiency and safety, reduces maintenance costs, extends battery life, and powerfully promotes the safe and efficient development of the new energy vehicle industry. Attached Figure Description

[0015] To more clearly illustrate the embodiments of the present invention and existing solutions, the accompanying drawings used in the embodiments and existing descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is a flowchart illustrating the system of the present invention; Figure 2 This is a flowchart illustrating the health status assessment submodule of the present invention; Figure 3 This is a flowchart illustrating the abnormal diagnosis submodule of the present invention; Figure 4 This is a schematic diagram illustrating the implementation steps of the method of the present invention. Detailed Implementation

[0016] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0017] according to Figures 1-4 A new energy vehicle battery health status monitoring system, the system comprising: The battery data acquisition module includes a conformal sensor array, compatible with 99% of existing battery models, including 4680 batteries and blade batteries. The sensor array includes a tunneling field-effect transistor voltage and current sensor, a terahertz thermal imaging sensor, and an atomic force electrochemical cycle counter, which collects voltage, current, temperature, and charge / discharge cycle data in real time. The collected data is then transmitted to the data processing and analysis module. The battery status monitoring module, based on real-time data stream analysis and dynamic threshold modeling, parses the collected data through a multi-scale feature extraction algorithm to achieve millisecond-level dynamic monitoring of the charging and discharging process; the battery status monitoring module has a built-in reinforcement learning-driven anomaly detection engine that adaptively learns the battery operating mode; and combines historical battery data and aging models to generate accurate maintenance prompts using generative adversarial networks. The data processing and analysis module receives data transmitted from the battery data acquisition module, performs in-depth processing and analysis on the data, and calculates the battery's health status indicators through a constructed algorithm model. The data processing and analysis module includes a data preprocessing submodule, a feature extraction submodule, a health status assessment submodule, and an anomaly diagnosis submodule. The data preprocessing submodule employs a fusion algorithm of Variational Mode Decomposition (VMD) and Adaptive Median Filtering to suppress Gaussian noise and impulse interference while retaining weak characteristic signals such as voltage ripple and temperature abrupt changes. The feature extraction submodule innovatively introduces Empirical Wavelet Transform (EWT) and fractal dimension calculation to extract 40+ key features from the time domain, frequency domain, time-frequency domain, and complex system features. The health status assessment submodule constructs a battery health factor prediction model based on Graph Convolutional Neural Network (GCN) and integrates topological correlation information between battery cells. The anomaly diagnosis submodule develops Generative Adversarial Network (GAN) for abnormal manifold modeling, simulating normal state distribution through a generator and detecting data deviation in real time through a discriminator. The output of the data preprocessing submodule is connected to the input of the feature extraction submodule, the output of the feature extraction submodule is connected to the input of the health status assessment submodule, and the output of the health status assessment submodule is connected to the input of the abnormal diagnosis submodule. The display and interaction module features an intuitive graphical interface that presents battery health status data, monitoring results, and warning information to users. It also provides human-computer interaction functions, allowing users to query historical data and set warning thresholds. The communication module is designed with a standardized communication protocol to enable data transmission and interaction between the monitoring system and the new energy vehicle control system and cloud server, supporting remote data access and system upgrades. The output of the battery data acquisition module is connected to the input of the data processing and analysis module. The output of the data processing and analysis module is connected to the input of the battery status monitoring module and the display and interaction module, respectively. The communication module is bidirectionally connected to the data processing and analysis module. Furthermore, the battery data acquisition module includes a terahertz metasurface intelligent sensing array module, a biomimetic electronic skin compatible architecture module, a neuromorphic data preprocessing engine module, a microwave photonic anti-interference transmission link module, and a self-calibration and health management module. The terahertz metasurface intelligent sensing array module employs laser direct writing to fabricate a conformal sensor array with a thickness ≤0.1mm on a polyimide substrate. It includes tunneling field-effect transistors for high-precision measurement of 1pA current and 1μV voltage; terahertz time-domain spectroscopy for detecting the internal temperature distribution of the battery with a temperature resolution of 0.01℃; and a combination of carbon nanotube cantilever beams and electrochemical impedance spectroscopy to achieve a charge / discharge count error ≤0.1 times. The array includes a flexible solar film and a vibration energy harvester, enabling self-powered operation with an average power consumption ≤100μW. The biomimetic electronic skin compatible architecture module is based on liquid metal and shape memory alloy. The sensor array adaptively fits various batteries with a contact thermal resistance of ≤0.3K / W. The magnetic modular interface supports quick replacement within 10 seconds and is compatible with 99% of battery models. It has a built-in multi-protocol gateway to realize the conversion of vehicle and industrial protocol data with a latency of ≤5μs. The neuromorphic data preprocessing engine module uses a spiking neural network to spatiotemporally encode the data and detects anomalies based on contrastive learning; spatiotemporal sparse coding compresses the data volume by more than 75% and pre-extracts 12 key features; The microwave photonic anti-interference transmission link module is based on 60GHz millimeter-wave directional communication, which improves anti-interference capability by 50dB compared to traditional WiFi; it uses chaotic signals for physical layer encryption with a key generation rate of 5Gbit / s to ensure data security. The self-calibration and health management module automatically calibrates every hour, with voltage drift ≤5μV / day and temperature drift ≤0.003℃ / day; it predicts sensor lifespan through a physical-data fusion twin model with an error ≤3% and provides early warning of failures 7 days in advance. The output of the terahertz metasurface intelligent sensing array module is connected to the input of the biomimetic electronic skin compatible architecture module; the output of the biomimetic electronic skin compatible architecture module is connected to the input of the neuromorphic data preprocessing engine module; the output of the neuromorphic data preprocessing engine module is connected to the input of the microwave photonic anti-interference transmission link module; the outputs of the intelligent sensing array module and the terahertz metasurface intelligent sensing array module are connected to the input of the self-calibration and health management module.

[0018] Furthermore, the battery status monitoring module includes: a multi-dimensional real-time monitoring module, a reinforcement learning anomaly detection module, and a generative maintenance decision module; The multi-dimensional real-time monitoring module uses wavelet packet decomposition and a spatiotemporal grid model to perform multi-scale analysis and real-time monitoring of voltage, current, and temperature, capturing charging and discharging trends as low as 0.1Hz and voltage jumps above 50mV, reversing the internal heat flow distribution of the battery, and monitoring temperature gradients exceeding 2℃ / cm; combined with efficient circuits and neural networks to simulate battery status, it dynamically marks abnormal stages, with a monitoring delay of no more than 500ms and extracts features of more than 50 dimensions. The reinforcement learning-based anomaly detection module dynamically adjusts the anomaly threshold based on a deep deterministic policy gradient algorithm and achieves three-level early warning through Mahalanobis distance. It utilizes a convolutional long short-term memory network to identify six types of anomaly patterns and combines fault tree analysis to trace the root cause. The anomaly detection latency is within 1 second, with a low false alarm rate and a false negative rate of <0.1%. The generative maintenance decision module generates accurate maintenance suggestions by combining battery status and user habits with generative adversarial networks and Pareto optimization, with the error controlled within 5 cycles; it generates a 3D visual work order and displays it through AR, including effect backtracking and optimization strategies. The output of the multi-dimensional real-time monitoring module is connected to the input of the reinforcement learning anomaly detection module; the output of the reinforcement learning anomaly detection module is connected to the input of the generative maintenance decision module.

[0019] In a specific embodiment, the battery status monitoring module consists of three main functional modules. The multi-dimensional real-time monitoring module utilizes wavelet packet decomposition and a spatiotemporal grid model to achieve multi-scale real-time monitoring and feature extraction of battery electrical and thermal parameters. The reinforcement learning anomaly detection module leverages deep deterministic policy gradient algorithms and convolutional long short-term memory networks to dynamically adjust anomaly thresholds and accurately identify anomaly patterns. The generative maintenance decision module uses generative adversarial networks and Pareto optimization to generate maintenance suggestions based on battery status and user habits. This achieves highly sensitive real-time monitoring of battery electrical and thermal parameters, rapid and accurate anomaly detection, and an extremely low false alarm / false negative rate. The generated maintenance suggestions are tailored to actual needs, have minimal error, and are easily implemented through 3D visual work orders and AR display. Overall, this significantly improves the timeliness and accuracy of battery status monitoring and the scientific nature of maintenance decisions, effectively ensuring the safe and stable operation of the battery and reducing the risk of failure.

[0020] Furthermore, the data preprocessing submodule includes a variational mode decomposition (VMD) signal reconstruction module, an adaptive median filtering optimization module, and a data cleaning and invalid value processing module; The Variational Mode Decomposition (VMD) signal reconstruction module employs multi-component adaptive decomposition to decompose the original voltage and current signals into six intrinsic mode functions. Over-decomposition is avoided through the Bayesian information criterion. The sample entropy of each component is calculated, and modes with an entropy higher than 0.8 are identified as noise and removed. An improved wavelet threshold function is used to optimize the effective modes, increasing the signal-to-noise ratio of high-frequency features above 200Hz from 25dB to 40dB, while fully preserving weak voltage anomaly signals at the 50μV level. The adaptive median filtering optimization module dynamically adjusts the filtering window based on local data fluctuations. When the standard deviation is below 50% of the global standard deviation, a 3×3 window is used to suppress Gaussian noise. When the standard deviation is between 50% and 150%, a 5×5 window is used. When the standard deviation exceeds 150%, a 7×7 window is used to suppress impulse interference. For temperature data, a gradient protection threshold of 1℃ / ms is set. When a sudden temperature rise signal before thermal runaway is detected, linear interpolation is used. The data cleaning and invalid value processing module utilizes a local outlier factor algorithm with 5 nearest neighbors and a threshold of 2.0 to identify outliers such as voltage jumps exceeding 80mV and current deviations exceeding 15% of the rated value. It employs Kalman filtering combined with the expectation-maximization algorithm to optimize interpolation, controlling the interpolation error of missing values ​​in charge and discharge data to within 0.3%. Z-score standardization is applied to voltage, current, and temperature data, and nonlinear scaling is performed on data within the 20%-80% battery state of charge range, improving the sensitivity of subsequent models to these range features by 30%. The output of the variational mode decomposition (VMD) signal reconstruction module is connected to the input of the adaptive median filtering optimization module; the output of the adaptive median filtering optimization module is connected to the input of the invalid value processing module.

[0021] In a specific embodiment, the VMD signal reconstruction module decomposes the original voltage and current signals into six intrinsic mode functions, calculates sample entropy to remove noise components, and performs wavelet threshold optimization on the effective modes. Secondly, the adaptive median filtering module dynamically adjusts the filtering window based on the local data standard deviation and enables linear interpolation for signals with sudden temperature increases. Finally, the data cleaning module identifies and processes outliers, performs Kalman filtering interpolation on missing values, and completes standardization and interval scaling. These three modules are processed sequentially to form a complete data preprocessing workflow. This submodule significantly improves the quality of battery monitoring data, increasing the signal-to-noise ratio of high-frequency features above 200Hz from 25dB to 40dB, while fully preserving weak voltage anomaly signals at the 50μV level. The dynamic filtering strategy effectively suppresses different types of noise, improving the response speed of thermal runaway early warning. Data cleaning and standardization increase the sensitivity of subsequent models to features in the 20%-80% state of charge range by 30%, with the overall preprocessing error controlled within 0.3%, providing a solid data foundation for battery state estimation and fault diagnosis.

[0022] Furthermore, the feature extraction submodule includes a time-frequency feature extraction module, a fractal dimension feature calculation module, a multi-domain feature fusion module, and a feature engineering toolchain module; The time-frequency feature extraction module performs adaptive frequency band segmentation on voltage and current signals through empirical wavelet transform. The adaptive frequency band is 6-8 bands, covering the entire frequency band from 0.1Hz to 10kHz. It extracts 12 feature parameters, including root mean square value and kurtosis, and generates a 72-dimensional time-frequency feature vector. After correlation screening, 40 key features are retained. The fractal dimension feature calculation module calculates the fractal dimension of the temperature field using the box dimension algorithm. The fractal dimension decreases by 0.15 3 hours before thermal runaway, and the sensitivity is improved by 30%. It performs multifractal analysis on the voltage sequence and provides early warning of aging 50 cycles in advance. It also calculates the fractal dimension of the state of charge-voltage hysteresis loop to predict the growth of battery internal resistance. The multi-domain feature fusion module includes time domain, frequency domain, time-frequency domain, and fractal dimension features, forming a 40+ dimension health feature vector; among them, the time domain features reflect the electrode contact impedance, the frequency domain features are related to the electrolyte concentration, the time-frequency domain features provide early warning of thermal runaway, and the fractal dimension features monitor early aging. The feature engineering toolchain module calculates feature importance in real time through an automated platform, marks abnormal feature points, and dynamically adjusts feature weights. In the early stage of the loop, it focuses on the time domain and in the later stage, it strengthens fractal and frequency domain characteristics. The output of the time-frequency feature extraction module is connected to the input of the fractal dimension feature calculation module; the output of the fractal dimension feature calculation module is connected to the input of the multi-domain feature fusion module.

[0023] In a specific embodiment, the battery feature extraction submodule operates collaboratively through four functional modules. The time-frequency feature extraction module uses empirical wavelet transform to adaptively segment and filter the frequency bands of electrical signals; the fractal dimension feature calculation module uses box-dimensional algorithms to mine potential features of battery aging and thermal runaway from temperature and voltage data; the multi-domain feature fusion module integrates time-domain, frequency-domain, time-frequency-domain, and fractal dimension features to construct a multi-dimensional health feature vector; and the feature engineering toolchain module dynamically adjusts feature weights based on an automated platform. Each module has a clear division of labor, achieving comprehensive and accurate extraction of battery features. First, the time-frequency feature extraction module segments the frequency bands of voltage and current signals and extracts and filters features; then, the fractal dimension feature calculation module performs fractal analysis on temperature and voltage data; next, the multi-domain feature fusion module integrates multi-dimensional features; finally, the feature engineering toolchain module calculates feature importance in real time and dynamically adjusts weights, forming a complete battery feature extraction process. This submodule achieves efficient extraction of battery electrical signal features across the entire frequency band, can provide early warning of battery aging up to 50 cycles in advance, captures fractal dimension changes 3 hours before thermal runaway, and improves sensitivity by 30%. By fusing multi-domain features and dynamically adjusting weights, a multi-dimensional feature vector that comprehensively reflects the battery health status is constructed, providing accurate data support for battery status assessment and fault prediction, and effectively improving the reliability and foresight of the battery management system.

[0024] Furthermore, the health status assessment submodule includes a battery pack topology modeling module, a graph convolutional neural network architecture module, a multi-scale health factor prediction module, a spatiotemporal evolution model module, and a model training and optimization module; The battery pack topology modeling module constructs a three-dimensional topology graph, with the battery cells as nodes, containing 12-dimensional state parameters; the edge weights are dynamically set according to electrical connections, thermal coupling, and electrochemical interactions: the edge weight of series-connected cells is 0.6, the thermal conduction weight of close-range cells is 0.2, and the electrochemical similarity determines the weight of 0.2, presenting the internal interaction characteristics of the battery pack; The graph convolutional neural network architecture module extracts local, module-level, and global features layer by layer through a three-layer GCN combined with a graph attention mechanism, and outputs a 48-dimensional health vector; the mechanism automatically enhances the weight of cells with high temperature and abnormal voltage to highlight key information. The multi-scale health factor prediction module covers SOH, remaining lifespan, and failure probability through a multi-task prediction system; the SOH prediction focuses on aging characteristics with an error of ≤3%; the remaining lifespan prediction combines historical data with an error of <100 cycles; and the failure probability classification F1 score is ≥0.92, achieving comprehensive assessment. The spatiotemporal evolution model module predicts fault propagation in space and locks in the risk of module thermal runaway within 15 minutes; in time, it dynamically adjusts parameters according to the charging and discharging strategy, predicts the frequency up to 10Hz during fast charging, and tracks health changes in real time. The model training and optimization modules employ a hybrid loss function to enhance early aging learning, while federated learning protects privacy, with online fine-tuning every 50 iterations. After 2000 iterations, the model prediction error increases by only 1.2%. The output of the battery pack topology modeling module is connected to the input of the graph convolutional neural network architecture module; the output of the graph convolutional neural network architecture module is connected to the multi-scale health factor prediction module; the output of the multi-scale health factor prediction module is connected to the spatiotemporal evolution model module; the output of the spatiotemporal evolution model module is connected to the model training and optimization module; and the output of the model training and optimization module is connected to the inputs of the graph convolutional neural network architecture module and the multi-scale health factor prediction module.

[0025] The battery pack topology modeling module constructs a 3D topology graph reflecting the electrical, thermal, and electrochemical interactions of the cells; the graph convolutional neural network architecture module uses a three-layer GCN combined with a graph attention mechanism to extract multi-scale health features; the multi-scale health factor prediction module predicts SOH, remaining lifetime, and failure probability in parallel through a multi-task system; the spatiotemporal evolution model module realizes fault propagation space prediction and dynamic adjustment of charge and discharge parameters; the model training and optimization module uses a hybrid loss function and federated learning to ensure model accuracy and privacy. First, the battery pack topology modeling module constructs a 3D topology graph containing 12-dimensional state parameters; second, the graph convolutional neural network architecture module extracts a 48-dimensional health vector through a three-layer GCN combined with an attention mechanism; next, the multi-scale health factor prediction module outputs SOH, remaining lifetime, and failure probability in parallel; subsequently, the spatiotemporal evolution model module performs fault propagation space prediction and dynamic adjustment of charge and discharge parameters; finally, the model training and optimization module fine-tunes the model online through a hybrid loss function and federated learning, and feeds the optimized parameters back to the front-end module.

[0026] Furthermore, the anomaly diagnosis submodule includes a GAN anomaly manifold modeling module, a multi-scale anomaly detection module, a fault type classification module, a fault tracing and root cause analysis module, and a real-time early warning and response module; The GAN abnormal manifold modeling module uses a deep convolutional generative adversarial network (DCGAN). The generator outputs time-series pseudo samples of voltage, current, and temperature, and the discriminator outputs anomaly scores of 0-1. After 1000 rounds of training, the difference in distribution (FID) between the generated samples and the real data is reduced to below 15, accurately simulating the normal state of the battery. The multi-scale anomaly detection module includes point anomaly detection to identify sudden single-cell failures through reconstruction errors; context anomaly detection to analyze sequence continuity using LSTM to identify progressive aging; and collective anomaly detection to combine topological graphs and graph attention mechanisms to provide early warning of regional potential thermal runaway risks up to 2 hours in advance. The fault type classification module is based on the attention ResNet classifier, automatically weights the key features of temperature, and outputs the probability of 8 types of faults, with a rare fault recall rate of 85%. It introduces uncertainty assessment, triggering secondary verification when the entropy value exceeds 0.6, thereby improving diagnostic reliability. The fault tracing and root cause analysis module constructs a causal graph model with 30+ variables to analyze the causal chain of "fast charging-temperature-aging". It quantifies the impact of faults through intervention calculation and counterfactual reasoning to assist in root cause location and prevention decisions. The real-time early warning and response module has three levels of early warning: green warning triggers BMS self-learning; yellow warning prompts maintenance and recommends service stations; red warning limits power to 50% and remotely alarms; response time <200ms, thermal runaway warning 30 minutes in advance, ensuring emergency response time; The output of the GAN anomaly manifold modeling module is connected to the input of the multi-scale anomaly detection module; the output of the multi-scale anomaly detection module is connected to the input of the fault type classification module; the output of the fault type classification module is connected to the input of the fault tracing and root cause analysis module; and the output of the fault tracing and root cause analysis module is connected to the input of the real-time early warning and response module.

[0027] In a specific embodiment, the battery anomaly diagnosis submodule implements full lifecycle fault management through a five-layer architecture. The GAN anomaly manifold modeling module uses DCGAN to generate high-fidelity normal state pseudo-samples to construct a baseline manifold; the multi-scale anomaly detection module combines point anomaly, contextual anomaly, and collective anomaly detection to achieve multi-dimensional fault capture; the fault type classification module uses an attention-based ResNet classifier to accurately identify eight types of faults and introduces uncertainty assessment to improve reliability; the fault tracing and root cause analysis module constructs a causal graph model and quantifies the impact of the fault chain through intervention calculation; the real-time early warning and response module establishes a three-level response mechanism to ensure rapid handling. First, the GAN anomaly manifold modeling module trains DCGAN to generate normal state samples, and the discriminator outputs anomaly scores; second, the multi-scale anomaly detection module identifies different types of anomalies through reconstruction error, LSTM sequence analysis, and graph attention mechanism; next, the fault type classification module outputs the fault probability distribution and performs uncertainty assessment based on attention mechanism weighted key features; subsequently, the fault tracing and root cause analysis module constructs a causal graph and locates the root cause through intervention calculation; finally, the real-time early warning and response module triggers a three-level response according to the level of the anomaly, with a response time of <200ms. This submodule enables accurate diagnosis and rapid response to battery faults, reducing the difference in distribution between generated samples and real data (FID) to below 15, achieving a rare fault recall rate of 85%, and providing a 30-minute advance warning of thermal runaway.

[0028] Furthermore, the method for monitoring the health status of new energy vehicle batteries is as follows: S1. System Hardware Setup and Data Acquisition The battery data acquisition module employs a terahertz metasurface conformal sensor array and biomimetic electronic skin. It is compatible with 99% of existing battery models through liquid metal bonding and magnetic interface. It uses tunneling field-effect transistors and terahertz thermal imaging sensors to collect voltage, current, temperature and charge / discharge cycles in real time and transmits them to the data processing module with a delay of <50ms. S2, Data Processing and Analysis Workflow The data preprocessing submodule suppresses noise and preserves weak features through a fusion algorithm of variational mode decomposition and adaptive median filtering; the feature extraction submodule uses empirical wavelet transform and fractal dimension calculation to extract 40+ key features from multiple domains; the health status assessment submodule uses graph convolutional neural network to fuse cell topology information to achieve multi-task prediction of SOH and remaining life; the anomaly diagnosis submodule uses generative adversarial network to model the normal state manifold, detects anomalies, and traces the root cause. S3, Battery Status Monitoring and Closed-Loop Management The battery status monitoring module achieves millisecond-level monitoring of the charging and discharging process based on real-time data streams and dynamic thresholds. It has a built-in reinforcement learning anomaly detection engine for adaptive early warning and combines generative adversarial networks to generate maintenance suggestions. Through 3D work orders and AR interaction, it realizes a closed-loop end-to-end system from monitoring to maintenance, reducing maintenance time by 30%. S4, System Interaction and Data Communication The display and interaction module presents real-time data, historical records, and warning information through a graphical interface, and supports user-defined thresholds. The communication module realizes data interaction with the vehicle system and the cloud through a standardized protocol, supports remote access and upgrades, and has a communication latency of <100ms. Cloud data is used to optimize global maintenance strategies. S5, System Closed Loop and Optimization The system forms a closed loop of "collection-analysis-monitoring-maintenance-feedback". The strategy library is optimized weekly through a maintenance effect backtracking mechanism, and online model upgrades are achieved using federated learning to adapt to individual battery differences. The specific implementation method is as follows: I. Solving the problems of battery data acquisition adaptability and real-time performance The battery data acquisition module employs a terahertz metasurface conformal sensor array and a biomimetic electronic skin architecture. Through a liquid metal and magnetic interface design, it can quickly adapt to 99% of existing battery models within 10 seconds, achieving a contact thermal resistance as low as 0.3 K / W. Utilizing tunneling field-effect transistors, it achieves high-precision measurements of 1 pA current and 1 μV voltage. Combined with terahertz time-domain spectroscopy, it can detect the internal temperature distribution of the battery in real time with a temperature resolution of 0.01℃. Data is transmitted to the data processing module via a microwave photonic anti-interference transmission link based on 60 GHz millimeter-wave directional communication with a delay of <50 ms, ensuring the accuracy and real-time transmission of the acquired data and providing a solid foundation for subsequent monitoring.

[0029] II. Solving the Challenges of Battery Data Processing and Feature Extraction The data preprocessing submodule employs a fusion algorithm of variational mode decomposition and adaptive median filtering to decompose the original voltage and current signals into six intrinsic mode functions. By using the Bayesian information criterion to avoid over-decomposition, the signal-to-noise ratio of high-frequency features above 200Hz is increased from 25dB to 40dB, fully preserving weak voltage anomaly signals at the 50μV level. The feature extraction submodule uses empirical wavelet transform to perform 6-8 adaptive frequency band segmentation on the electrical signal. Combined with fractal dimension calculation, it extracts 40+ key features from multiple domains including the time domain, frequency domain, time-frequency domain, and fractal dimension. Core information is retained through correlation screening, providing accurate data support for battery health assessment.

[0030] III. Solving problems related to accurate assessment of battery health status and diagnosis of anomalies The health status assessment submodule, based on a graph convolutional neural network architecture, constructs a 3D topological graph of the battery pack containing 12-dimensional state parameters. Through a three-layer GCN combined with a graph attention mechanism, it outputs a 48-dimensional health vector, achieving multi-task prediction of SOH and remaining lifetime. The SOH prediction error is ≤3%, and the remaining lifetime prediction error is <100 iterations. The anomaly diagnosis submodule utilizes a deep convolutional generative adversarial network (DCGAN) to model the normal state manifold. After 1000 rounds of training, the difference in distribution (FID) between generated samples and real data is reduced to below 15. Combined with multi-scale anomaly detection and a fault-based causal graph model, it provides an early warning of thermal runaway risk up to 2 hours in advance, achieving an 85% recall rate for rare faults and accurately locating the root cause of the fault.

[0031] IV. Solving the Problems of Closed-Loop Management and Interaction in Battery Status Monitoring The battery status monitoring module, based on real-time data streams and a reinforcement learning anomaly detection engine, achieves millisecond-level monitoring of the charging and discharging process. It generates precise maintenance suggestions through generative adversarial networks and combines 3D visual work orders with AR interactive technology, reducing maintenance time by 30%. The display and interaction module presents data via a graphical interface and supports user-defined thresholds. The communication module interacts with the vehicle and cloud via standardized protocols, with communication latency <100ms. Cloud data is used to optimize maintenance strategies, achieving closed-loop management across the entire chain from monitoring to maintenance.

[0032] V. Continuous System Optimization and Resolution of Individual Adaptation Issues The system constructs a closed-loop system encompassing data collection, analysis, monitoring, maintenance, and feedback, optimizing the strategy library weekly through a maintenance effectiveness feedback mechanism. Utilizing federated learning technology, it achieves online model upgrades to adapt to individual battery differences, with prediction errors increasing by only 1.2% after 2000 iterations. The system continuously optimizes diagnostic accuracy and robustness, ensuring the health management and safe operation of new energy vehicle batteries throughout their entire lifecycle, effectively improving battery efficiency and vehicle reliability.

[0033] Furthermore, the implementation method of the feature extraction submodule is as follows: 1) Empirical Wavelet Transform (EWT); Empirical wavelet transform the original signal Decomposed into N adaptive frequency bands, introducing multiphysics and dynamic weighting mechanisms, the mathematical expression is: In formula (1), For time series functions, The t-transform is used to dynamically adjust the components of each frequency band based on the energy distribution and characteristics of the signal at different times. The contribution of the signal to the original signal, thereby more accurately capturing the time-varying characteristics of the signal; For multiphysics modulation operators, For real-time state of charge, Kelvin temperature scale data collected by a temperature sensor. This represents the effective value of the current within the sliding window. The capacity loss temperature sensitivity coefficient is calculated based on the Arrhenius equation; The k-th frequency band component The calculation method is as follows: In formula (2), The improved empirical wavelet filter not only depends on the frequency w, but also introduces time-varying parameters. ; Based on the local frequency characteristics of the signal, the shape and bandwidth of the filter are adaptively adjusted, making the filter more flexible in matching the characteristics of different frequency components and improving the accuracy of decomposition. Original signal The Fourier transform converts a time-domain signal to the frequency domain, revealing the frequency components and energy distribution of the signal. This is a frequency domain weighting function that assigns different weights to different frequency ranges in the frequency domain based on the signal's distribution in the frequency domain, further optimizing the extraction of frequency band components, highlighting key frequency components, and suppressing noise and interference. In a specific embodiment, the Empirical Wavelet Transform (EWT) achieves accurate decomposition of the original signal through a multi-physics-coupled dynamic weighting mechanism and adaptive parameter optimization. It utilizes a time-series function to dynamically adjust the contribution of each frequency band component to the original signal based on the energy distribution of the signal at different times. A multi-physics modulation operator is introduced, fusing parameters such as the battery's real-time state of charge, temperature, and effective current value to quantify the influence of physical factors on the signal. The improved EWT filter, combined with time-varying parameters, can adaptively adjust filter characteristics, while the frequency domain weighting function further optimizes frequency band component extraction. These multiple mechanisms work together to improve signal decomposition accuracy. First, the time-series function is calculated to determine the dynamic weights of each frequency band. Second, the multi-physics modulation operator is calculated based on the battery's state of charge, temperature, and current parameters. Then, a Fourier transform is performed on the original signal, and the improved EWT filter and time-varying frequency domain weighting function are used to calculate each frequency band component. Finally, the original signal is reconstructed after processing each frequency band component. EWT achieves adaptive frequency band decomposition of the original signal, accurately capturing the signal's time-varying characteristics and effectively suppressing noise. By fusing multi-physics parameters, the impact of battery operating status on signals is clarified, enhancing the correlation between signal decomposition and battery characteristics. Adaptive filters and frequency domain weighting functions highlight the extraction of key frequency components, providing high-precision signal processing results for battery health status monitoring applications and significantly improving the accuracy and reliability of signal analysis. Table 1 compares the core performance of the traditional empirical wavelet transform and the improved method in signal processing by introducing a multi-physics field fusion dynamic weighting mechanism. In key frequency component extraction, the improved method improves accuracy by 15.1% through multi-physics field modulation operators and frequency domain weighting functions, more accurately identifying key frequencies in power signals; noise suppression ratio is improved by 30.5%, effectively reducing signal interference. The time-varying feature capture delay is significantly reduced by 62.5%, enabling faster response to signal changes. For power equipment capacity loss prediction, the improved method combines physical parameters such as state of charge, temperature, and current, reducing prediction error by 63.6%. Furthermore, the adaptive frequency band adjustment speed is improved by 212.5%, enabling the system to adapt to signal frequency changes more quickly and flexibly, resulting in overall performance significantly superior to traditional methods.

[0034] Further, time-frequency feature extraction: root mean square value In formula (3), It represents the time-varying root mean square value at time t, dynamically reflecting the change in the energy intensity of the signal at different times; is the i-th sample value in the signal sequence at time t, representing the instantaneous amplitude of the signal in the time dimension; N is the number of sampling points, representing the length of the signal data involved in the calculation; The introduced time-varying weighting function assigns weights to the sampled values ​​at different times based on the local characteristics of the signal. For example, it increases the weight in regions of signal abrupt change to highlight key features, and decreases the weight in regions of concentrated noise to enhance noise immunity. L is the electrochemical and thermodynamic joint modulation function. SEI film growth rate coefficient calibrated based on aging tests; It is the temperature gradient tensor; The average current density within the sliding event window; This is a two-domain mixed processing function; kurtosis: In formula (4), It is a time-varying kurtosis used to characterize the amplitude distribution of a signal at time t, and to detect the impulse components in the signal and their changes. The introduced morphological adjustment function adjusts the contribution of different sampling points in the kurtosis calculation according to the characteristics of the signal morphology. As a driver of health; Commonly used in the battery field, it represents the state of charge, indicating the health or performance parameters of the system at time t; This is the expansion factor; The effect of current on the system varies with time; peak ripple factor: In formula (5), The time-varying crest factor is used to measure the relative magnitude of the peak value and the effective value of a signal at time t, and to assess the impact of the signal. The peak value of the weighted signal at time t. As an adaptive enhancement function, it amplifies the signal characteristics at the signal peak while suppressing other interferences, so that the peak factor more accurately reflects the signal's impact characteristics. This is the local gradient of the signal at point i, reflecting the rate of change; The gradient response function has the following parameters. Modulate sensitivity to mutations; To prevent background fluctuation interference, an attenuation factor dynamically adjusted based on historical peak saturation factors is used. In a specific embodiment, this time-frequency feature extraction method constructs a high-precision signal feature analysis system through multi-physics coupling and an adaptive weighting mechanism. In the root mean square (RMS) value calculation, a time-varying weighting function and an electrochemical-thermodynamic joint modulation function are introduced. Combined with parameters of state of charge, temperature gradient, and current density, the sampling weights are dynamically adjusted to accurately capture signal energy changes. Kurtosis calculation utilizes a morphological adjustment function and a health driving factor, integrating multiple physical quantities to characterize the signal's impact characteristics. The peak saturation factor uses an adaptive enhancement function and a gradient response function to highlight the signal's peak characteristics, and a dynamic attenuation factor suppresses background interference. These three elements work together to achieve multi-dimensional, adaptive analysis of the signal's time-frequency characteristics. First, the RMS value is calculated by determining the joint modulation function based on the physical parameters of state of charge and temperature gradient, and weighting the sampling points using a time-varying weighting function to obtain the time-varying RMS value reflecting the signal energy intensity. Second, kurtosis is calculated by adjusting the contribution of sampling points through a morphological adjustment function, integrating parameters of state of charge and health driving factors, and calculating the signal amplitude distribution morphology index. Finally, the peak-to-peak factor is calculated, and the signal peak value is amplified using an adaptive enhancement function. The gradient response function is then used to locate abrupt change points, and the result is divided by the root mean square (RMS) value corrected by a dynamic attenuation factor to assess the signal impact severity. This method significantly improves the accuracy and anti-interference capability of time-frequency feature extraction. The RMS value improves the sensitivity of capturing abrupt signals by 40% and the noise suppression ratio by 12 dB; kurtosis calculation increases the accuracy of signal impact component detection from 78% to 93%; the peak-to-peak factor reduces the response delay to transient impacts to 20 ms and reduces background fluctuation interference by 38%. In battery health monitoring scenarios, the capacity decay prediction error is controlled within 3.2%, a 58% improvement over traditional methods, providing strong technical support for power equipment condition monitoring and fault early warning. Table 2 compares the key performance differences between traditional time-frequency feature extraction methods and improved methods that incorporate multi-physics coupling and adaptive weighting mechanisms. The improved method significantly enhances the analytical capability of signal features through dynamic weighting and multi-physics parameter modulation. Furthermore, the box dimension is calculated: In formula (6), D is the target fractal dimension. By integrating spatiotemporal dynamic characteristics and adaptive weighting mechanism, the complexity and self-similar structural characteristics of the dataset are more accurately quantified, reflecting the geometric shape change law of the dataset under spatiotemporal scale. The base scale parameter represents the global initial box size used when covering the dataset. This indicates a gradual refinement of the analytical scale from macro to micro. The dynamic scale parameter is set for the i-th local region and j-th time window of the dataset, and is adaptively adjusted in combination with spatiotemporal characteristics; Ni,j(ϵ): the number of boxes required to cover the i-th local region and j-th time window under the dynamic scale ϵi,j, capturing the local features of the data distribution in more detail through spatiotemporal decomposition; M is the number of local regions divided in the spatial dimension, which is adaptively adjusted according to the spatial distribution characteristics and complexity of the data, and is used for multi-resolution analysis in space; T is the number of windows divided in the time dimension, used to process time series data or dynamic systems, and to realize time-varying analysis of fractal features; It is a spatiotemporal dynamic weighting function that is dynamically adjusted based on the information entropy of the local region and the importance of the time window; The variance of the number of boxes filled in the i-th region and the j-th time window reflects the uniformity of the local data distribution and is used to suppress the influence of noisy regions. The variance influence factor controls the degree to which variance modulates the weights, and the optimal value is determined through cross-validation. Let be the correlation coefficient between the i-th region and its neighboring regions in the j-th time window, capturing the spatial dependence between regions. The correlation factor controls the impact of inter-region correlation on fractal dimension calculation. In a specific embodiment, this box dimension calculation method constructs a high-precision fractal feature quantification model through spatiotemporal dynamic weighting and a multi-scale adaptive mechanism. A spatiotemporal dynamic weighting function is introduced, dynamically adjusting weights based on local information entropy and the importance of time windows to highlight the feature contributions of key regions and time periods. The dataset is spatiotemporally decomposed using the dynamic scale parameter ϵ(i,j) to achieve multi-resolution analysis from macro to micro. The variance factor λ suppresses the influence of noisy regions, and the correlation factor α captures spatial dependence, making the fractal dimension more accurately reflect the complexity and self-similar structure of the data, overcoming the limitations of traditional box dimension calculations at a static single scale. First, the dataset is divided into M local regions in the spatial dimension and T windows in the temporal dimension, initializing the global basic scale parameter ϵ. Second, for each local region i and time window j, the dynamic scale parameter ϵ(i,j) is calculated, and the number of boxes required to cover the data in that region is counted. Then, the spatiotemporal dynamic weights are calculated. The weight values ​​are adjusted by combining the regional information entropy and window importance; at the same time, the variance of the number of boxes filled is calculated. Correlation coefficient between regions The influence of factors λ and α on the weights is adjusted. Finally, the above parameters are substituted into the formula to calculate the fractal dimension D that integrates spatiotemporal characteristics, quantifying the geometric shape and time-varying laws of the data. This method significantly improves the accuracy and robustness of fractal dimension calculation. In complex power signal analysis, the ability to suppress noise regions is improved by 55%, and the feature extraction error is reduced to 0.08 dimensions. The spatiotemporal dynamic weights improve the accuracy of fractal feature recognition in key areas from 79% to 94%, especially in non-stationary scenarios such as grid load fluctuations and partial discharge of equipment. The multi-scale analysis mechanism improves the computational efficiency by 40%, adapting to real-time monitoring requirements. In the application of battery thermal failure early warning, the early fault detection accuracy based on this fractal dimension reaches 96.2%, which is 27 percentage points higher than the traditional method, providing an effective tool for the nonlinear feature analysis of complex power system data. Table 3 compares the core performance of the traditional box-dimensionality algorithm and the improved method that introduces spatiotemporal dynamic weights and adaptive mechanisms in fractal feature quantization. The improved method significantly enhances the analytical capabilities of complex data through spatiotemporal decomposition, dynamic weights, and noise suppression mechanisms. 4) Multifractal analysis; In formula (7), The generalized scaling index, by integrating dynamic feature interaction and multi-scale coupling effect, more accurately describes the scaling characteristics of the system under different moments q. Its variation curve reveals the evolution law of the system from micro to macro multifractal structure; q is the moment parameter, which takes values ​​in the real number range and controls the sensitivity to different probability measure regions. Focus on high-probability areas. Focus on low-probability, rare events. Corresponding to box dimension calculation; The basic scale parameter represents the initial box size used when covering the system. This indicates a refined analytical process from macroscopic to microscopic levels; In scale The total number of boxes required for the lower coverage system reflects the degree of spatial discretization of the system at that scale; The probability measure for the i-th box is defined as the proportion of the data volume within the box to the total data volume of the system, characterizing the relative importance of a local region. For dynamic feature interaction weights, quantize the order moment q and scale of the i-th box and its neighboring box j. The interaction strength under these conditions; The distance between boxes i and j in the feature space. The spatial interaction strength parameter controls the decay rate of the influence between adjacent boxes and is adaptively determined through cross-validation. This is a dynamic scaling factor that adaptively adjusts the analysis scale based on local complexity. The multi-scale interaction index quantifies the coupling effect of the i-th box across different scales. This is a multi-scale coupling strength parameter, which controls the degree of influence of interactions between different scales on the results; Table 4 compares the key performance differences between traditional multifractal methods and improved methods that incorporate dynamic feature interactions and multi-scale coupling mechanisms. The improved method significantly enhances the analytical capabilities of complex systems through dynamic adjustment of the moment parameter, spatial effect attenuation correction, and adaptive scale adjustment. Furthermore, weight adjustment; the formula for time-domain feature weights is: In formula (8), The temporal feature weights at the nth iteration are dynamically adjusted based on iteration progress, feature correlation, variance fluctuation, and adaptive parameters to determine the contribution of temporal features in the fusion process. n is the current iteration number, reflecting the data processing progress, with a value ranging from 1 ≤ n ≤ N, gradually increasing with iteration. N is the total iteration number, defining the time span for weight adjustment and determining the iteration progress factor. The range of variation; The dynamic attenuation coefficient, This is the decay rate parameter; Adaptive rate of change coefficient; Dynamic offset coefficient; The correlation sensitivity coefficient; The real-time correlation coefficient between the time-domain and frequency-domain characteristics; when At that time, the center of weight change shifts towards the earlier cycle; when In time, the system shifts to a later, cyclical manner to achieve feature-driven weight adjustment; This is a cross-domain coupling factor that quantifies the influence of frequency domain features on time domain weights, with a value range of [−1, 1]. When, positive correlation enhances the time domain weights; when At that time, negative correlation suppresses time-domain weights; Variance regularization factor, which suppresses abnormal fluctuations caused by noise; The real-time variance of the time-domain features; fractal feature weight formula: In formula (9), The fractal feature weights at the nth iteration are dynamically adjusted based on the iteration progress, fractal feature complexity, inter-feature correlation, and adaptive parameters, to determine the contribution ratio of fractal features in the fusion process. n is the current iteration number, reflecting the data processing progress, with a value range of 1 ≤ n ≤ N, which gradually increases with algorithm iterations to measure the time progress of weight adjustment. N is the total iteration number, determining the complete time span of weight adjustment and the variation range of the iteration progress factor Nn, providing a benchmark for dynamic weight changes. The dynamic fractal weight growth coefficient replaces the traditional fixed one. It employs a dynamic mechanism that varies with the fractal dimension; The larger the value, the higher the upper limit of the fractal feature weight; The adaptive fractal weight growth rate coefficient replaces the fixed w. The larger the value, the faster the weight increases; This refers to the dynamic fractal weight growth offset coefficient. This is a fractal complexity enhancement factor, which strengthens the influence of fractal dimension on weights. The dimension of the fractal feature at the nth iteration; Frequency domain feature weight formula: ;In formula (10), The time-domain feature weights for the current loop are calculated using formula (8); The fractal feature weights for the current loop are calculated using formula (9); The time-domain weight influence coefficient, with a value range of (0,1), is used to adjust the degree of influence of the time-domain feature weights on the frequency-domain feature weights. The larger the value, the stronger the reduction effect of the time-domain feature weights on the frequency-domain feature weights; This is the fractal weight influence coefficient, with a value range of (0,1), used to adjust the degree of influence of fractal feature weights on frequency domain feature weights. The larger the value, the stronger the reduction effect of fractal feature weights on frequency domain feature weights. In a specific embodiment, this feature weight calculation method achieves precise weight allocation of time-domain, fractal, and frequency-domain features during the fusion process through a multi-dimensional dynamic adaptive mechanism. The time-domain feature weights are dynamically adjusted based on the interaction relationship between features, taking into account the cycle progress, time-frequency domain correlation, variance fluctuation, and adaptive parameters, thus strengthening the feature contribution during key periods. The fractal feature weights introduce dynamic coefficients that change with the fractal dimension, combined with the cycle progress and complexity factor, highlighting the importance of complex structural features. The frequency-domain feature weights are calculated inversely through a linear combination of the time-domain and fractal weights, achieving a synergistic balance among the three weights. Overall, a feature interaction-driven, dynamically adaptive weight adjustment system is formed, avoiding the mismatch or omission of feature information caused by traditional fixed weights. First, determine the total number of loops N. In each loop n, calculate the real-time variance and time-frequency correlation coefficient of the time-domain features. Combine the adaptive parameters of dynamic decay and rate of change, and calculate the time-domain feature weights using formula (8). Second, obtain the fractal feature dimension values. Based on the dynamic growth coefficient, offset coefficient, and complexity enhancement factor, calculate the fractal feature weights using formula (9). Finally, based on the preset time-domain and fractal weight influence coefficients λ1 and λ2, and combined with the calculation results of the first two, derive the frequency-domain feature weights using formula (10), thus completing the dynamic adjustment of the weights for one loop. Repeat the above process until the Nth loop ends to achieve iterative optimization of the feature weights. This method significantly improves the accuracy and reliability of feature fusion. In power equipment fault diagnosis, the time-domain feature weights increase the sensitivity of capturing abrupt signals by 43% and reduce the false positive rate of abnormal fluctuations by 62%. Fractal feature weights increase the accuracy of identifying complex fault modes from 78% to 95%, and are particularly suitable for nonlinear signal analysis of partial discharge. The synergistic adjustment of frequency-domain feature weights with the previous two methods improves the overall fault warning accuracy by 38% and increases the dynamic pre-warning quantity by 50%. Compared with the traditional fixed-weight method, this dynamic adaptive mechanism exhibits stronger robustness in non-stationary and strong-interference scenarios, providing more reliable data analysis support for power system condition monitoring. Table 5 compares the core performance differences between the traditional fixed-weight method and the dynamic adaptive-weight method in feature fusion. The improved method significantly enhances the analytical capability of complex signals through a dynamic weighting coordination mechanism of time-domain, fractal, and frequency-domain features.

[0035] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Various omissions, substitutions, and changes can be made to the details of the methods and systems described above without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result using substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A new energy vehicle battery health status monitoring system, characterized in that: The system includes: The battery data acquisition module includes a conformal sensor array, which comprises a tunneling field-effect transistor voltage and current sensor, a terahertz thermal imaging sensor, and an atomic force electrochemical cycle counter. The battery status monitoring module, based on real-time data stream analysis and dynamic threshold modeling, parses the collected data through a multi-scale feature extraction algorithm to achieve millisecond-level dynamic monitoring of the charging and discharging process; the battery status monitoring module has a built-in reinforcement learning-driven anomaly detection engine that adaptively learns the battery operating mode; and combines historical battery data and aging models to generate accurate maintenance prompts using generative adversarial networks. The data processing and analysis module includes a data preprocessing submodule, a feature extraction submodule, a health status assessment submodule, and an anomaly diagnosis submodule. The data preprocessing submodule employs a fusion algorithm of variational mode decomposition (VMD) and adaptive median filtering, incorporating empirical wavelet transform (EWT) and fractal dimension calculation to extract 40+ key features from the time domain, frequency domain, time-frequency domain, and complex system features. The health status assessment submodule constructs a battery health factor prediction model based on graph convolutional neural network (GCN). The anomaly diagnosis submodule develops an adversarial generative network (GAN) for abnormal manifold modeling, simulating normal state distribution through a generator and detecting data deviation in real time through a discriminator. The output of the data preprocessing submodule is connected to the input of the feature extraction submodule, the output of the feature extraction submodule is connected to the input of the health status assessment submodule, and the output of the health status assessment submodule is connected to the input of the abnormal diagnosis submodule. The display and interaction module features an intuitive graphical interface. The communication module is designed with a standardized communication protocol. The output of the battery data acquisition module is connected to the input of the data processing and analysis module. The output of the data processing and analysis module is connected to the input of the battery status monitoring module and the display and interaction module, respectively. The communication module is bidirectionally connected to the data processing and analysis module.

2. The new energy vehicle battery health status monitoring system according to claim 1, characterized in that: The battery data acquisition module includes a terahertz metasurface intelligent sensing array module, a biomimetic electronic skin compatible architecture module, a neuromorphic data preprocessing engine module, a microwave photonic anti-interference transmission link module, and a self-calibration and health management module. The terahertz metasurface intelligent sensing array module uses laser direct writing to fabricate a conformal sensor array with a thickness ≤0.1mm on a polyimide substrate; Terahertz time-domain spectroscopy was used to detect the internal temperature distribution of the battery with a temperature resolution of 0.01℃; the charge-discharge count error was ≤0.1 times by combining carbon nanotube cantilever beams with electrochemical impedance spectroscopy; the array includes a flexible solar thin film and a vibration energy harvester to achieve self-powered operation with an average power consumption of ≤100μW. The biomimetic electronic skin compatible architecture module is based on liquid metal and shape memory alloy; Built-in multi-protocol gateway enables data conversion between automotive and industrial protocols with a latency of ≤5μs; The neuromorphic data preprocessing engine module uses a spiking neural network to spatiotemporally encode the data. The microwave photonic anti-interference transmission link module is based on 60GHz millimeter-wave directional communication; The self-calibration and health management module automatically calibrates every hour; The output of the terahertz metasurface intelligent sensing array module is connected to the input of the biomimetic electronic skin compatible architecture module; the output of the biomimetic electronic skin compatible architecture module is connected to the input of the neuromorphic data preprocessing engine module; the output of the neuromorphic data preprocessing engine module is connected to the input of the microwave photonic anti-interference transmission link module; the outputs of the intelligent sensing array module and the terahertz metasurface intelligent sensing array module are connected to the input of the self-calibration and health management module.

3. The new energy vehicle battery health status monitoring system according to claim 1, characterized in that: The battery status monitoring module includes: a multi-dimensional real-time monitoring module, a reinforcement learning anomaly detection module, and a generative maintenance decision module; The multi-dimensional real-time monitoring module performs multi-scale analysis and real-time monitoring of voltage, current, and temperature through wavelet packet decomposition and spatiotemporal grid model. The reinforcement learning anomaly detection module dynamically adjusts the anomaly threshold based on the deep deterministic policy gradient algorithm and achieves three-level early warning through Mahalanobis distance; The generative maintenance decision module generates accurate maintenance suggestions by combining battery status and user habits through generative adversarial networks and Pareto optimization. The output of the multi-dimensional real-time monitoring module is connected to the input of the reinforcement learning anomaly detection module; the output of the reinforcement learning anomaly detection module is connected to the input of the generative maintenance decision module.

4. The new energy vehicle battery health status monitoring system according to claim 1, characterized in that: The data preprocessing submodule includes a variational mode decomposition (VMD) signal reconstruction module, an adaptive median filtering optimization module, and a data cleaning and invalid value processing module. The variational mode decomposition (VMD) signal reconstruction module employs multi-component adaptive decomposition to decompose the original voltage and current signals into six intrinsic mode functions, and avoids over-decomposition through the Bayesian information criterion. The adaptive median filtering optimization module dynamically adjusts the filtering window according to local data fluctuations. When the standard deviation is below 50% of the global standard deviation, a 3×3 window is used to suppress Gaussian noise. When the standard deviation is between 50% and 150%, a 5×5 window is used. When the standard deviation exceeds 150%, a 7×7 window is used to suppress impulse interference. The data cleaning and invalid value processing module uses the local outlier factor algorithm, with 5 nearest neighbors and a threshold of 2.0, to identify outliers with voltage jumps exceeding 80mV and current deviations exceeding 15% of the rated value; it also uses Kalman filtering combined with the expectation-maximization algorithm to optimize interpolation. The output of the variational mode decomposition (VMD) signal reconstruction module is connected to the input of the adaptive median filtering optimization module; the output of the adaptive median filtering optimization module is connected to the input of the invalid value processing module.

5. The new energy vehicle battery health status monitoring system according to claim 1, characterized in that: The feature extraction submodule includes a time-frequency feature extraction module, a fractal dimension feature calculation module, a multi-domain feature fusion module, and a feature engineering toolchain module; The time-frequency feature extraction module performs adaptive frequency band segmentation on voltage and current signals through empirical wavelet transform. The adaptive frequency band consists of 6-8 frequency bands, covering the entire frequency range from 0.1Hz to 10kHz. The fractal dimension feature calculation module calculates the fractal dimension of the temperature field using the box dimension algorithm; The multi-domain feature fusion module includes time domain, frequency domain, time-frequency domain and fractal dimension features, forming a 40+ dimension health feature vector; The feature engineering toolchain module calculates feature importance in real time through an automated platform; The output of the time-frequency feature extraction module is connected to the input of the fractal dimension feature calculation module; the output of the fractal dimension feature calculation module is connected to the input of the multi-domain feature fusion module.

6. The new energy vehicle battery health status monitoring system according to claim 1, characterized in that: The health status assessment submodule includes a battery pack topology modeling module, a graph convolutional neural network architecture module, a multi-scale health factor prediction module, a spatiotemporal evolution model module, and a model training and optimization module. The battery pack topology modeling module constructs a three-dimensional topology graph; The graph convolutional neural network architecture module extracts local, module-level and global features layer by layer through a three-layer GCN combined with a graph attention mechanism. The multi-scale health factor prediction module covers SOH, remaining lifespan, and failure probability through a multi-task prediction system. The spatiotemporal evolution model module predicts fault propagation in space and locks in the risk of module thermal runaway within 15 minutes. The model training and optimization module uses a hybrid loss function to enhance early aging learning, federated learning to protect privacy, and online fine-tuning every 50 cycles. The output of the battery pack topology modeling module is connected to the input of the graph convolutional neural network architecture module; the output of the graph convolutional neural network architecture module is connected to the multi-scale health factor prediction module; the output of the multi-scale health factor prediction module is connected to the spatiotemporal evolution model module; the output of the spatiotemporal evolution model module is connected to the model training and optimization module; and the output of the model training and optimization module is connected to the inputs of the graph convolutional neural network architecture module and the multi-scale health factor prediction module.

7. The new energy vehicle battery health status monitoring system according to claim 1, characterized in that: The anomaly diagnosis submodule includes a GAN anomaly manifold modeling module, a multi-scale anomaly detection module, a fault type classification module, a fault tracing and root cause analysis module, and a real-time early warning and response module. The GAN abnormal manifold modeling module uses a deep convolutional generative adversarial network (DCGAN). The generator outputs time-series pseudo samples of voltage, current, and temperature, and the discriminator outputs anomaly scores of 0-1. After 1000 rounds of training, the difference in distribution (FID) between the generated samples and the real data is reduced to below 15, accurately simulating the normal state of the battery. The multi-scale anomaly detection module includes point anomaly detection to identify sudden single-cell failures through reconstruction errors; context anomaly detection to analyze sequence continuity using LSTM to identify progressive aging; and collective anomaly detection to combine topological graphs and graph attention mechanisms to provide early warning of regional potential thermal runaway risks up to 2 hours in advance. The fault type classification module is based on the attention ResNet classifier and automatically weights the key features of temperature to output the probability of 8 types of faults. The fault tracing and root cause analysis module constructs a causal graph model with 30+ variables; The real-time early warning and response module is equipped with three levels of early warning. The output of the GAN anomaly manifold modeling module is connected to the input of the multi-scale anomaly detection module; the output of the multi-scale anomaly detection module is connected to the input of the fault type classification module; the output of the fault type classification module is connected to the input of the fault tracing and root cause analysis module; and the output of the fault tracing and root cause analysis module is connected to the input of the real-time early warning and response module.

8. The method for monitoring the health status of new energy vehicle batteries according to claim 1, characterized in that... The method for monitoring the health status of new energy vehicle batteries is as follows: S1. System hardware setup and data acquisition; A battery data acquisition module employing a terahertz metasurface conformal sensor array and biomimetic electronic skin is used. S2, Data Processing and Analysis Flow; The data preprocessing submodule suppresses noise and preserves weak features through a fusion algorithm of variational mode decomposition and adaptive median filtering; S3, Battery Status Monitoring and Closed-Loop Management; The battery status monitoring module achieves millisecond-level monitoring of the charging and discharging process based on real-time data streams and dynamic thresholds; S4, System Interaction and Data Communication; The display and interaction module presents real-time data, historical records, and alert information through a graphical interface, and supports user-defined thresholds. S5. System closed-loop and optimization; The system forms a closed loop across the entire chain of "collection-analysis-monitoring-maintenance-feedback".

9. The method for monitoring the health status of new energy vehicle batteries according to claim 1, characterized in that: The implementation method of the feature extraction submodule is as follows: 1) Empirical Wavelet Transform (EWT); Empirical wavelet transform the original signal Decomposed into N adaptive frequency bands, introducing multiphysics and dynamic weighting mechanisms, the mathematical expression is: In formula (1), For time series functions, The t-transform is used to dynamically adjust the components of each frequency band based on the energy distribution and characteristics of the signal at different times. Contribution to the original signal; For multiphysics modulation operators, For real-time state of charge, Kelvin temperature scale data collected by a temperature sensor. This represents the effective value of the current within the sliding window. The capacity loss temperature sensitivity coefficient is calculated based on the Arrhenius equation; The k-th frequency band component The calculation method is as follows: ; In formula (2), The improved empirical wavelet filter not only depends on the frequency w, but also introduces time-varying parameters. ; Based on the local frequency characteristics of the signal, the shape and bandwidth of the filter are adaptively adjusted, making the filter more flexible in matching the characteristics of different frequency components and improving the accuracy of decomposition. Original signal The Fourier transform converts a time-domain signal to the frequency domain, revealing the frequency components and energy distribution of the signal. It is a frequency domain weighting function; 2) Time-frequency feature extraction; Root mean square value ; In formula (3), It represents the time-varying root mean square value at time t, dynamically reflecting the change in the energy intensity of the signal at different times; is the i-th sample value in the signal sequence at time t, representing the instantaneous amplitude of the signal in the time dimension; N is the number of sampling points, representing the length of the signal data involved in the calculation; The introduced time-varying weighting function is represented by L; L is the electrochemical and thermodynamic joint modulation function. SEI film growth rate coefficient calibrated based on aging tests; It is the temperature gradient tensor; The average current density within the sliding event window; This is a two-domain mixed processing function; kurtosis: ; In formula (4), It is a time-varying kurtosis used to characterize the amplitude distribution of a signal at time t, and to detect the impulse components in the signal and their changes. The introduced morphological adjustment function adjusts the contribution of different sampling points in the kurtosis calculation according to the characteristics of the signal morphology. As a driver of health; Commonly used in the battery field, it represents the state of charge, indicating the health or performance parameters of the system at time t; This is the expansion factor; The effect of current on the system varies with time; peak ripple factor: In formula (5), The time-varying crest factor is used to measure the relative magnitude of the peak value and the effective value of a signal at time t, and to assess the impact of the signal. The peak value of the weighted signal at time t. It is an adaptive enhancement function; This represents the local gradient of the signal at point i, reflecting the rate of change. The gradient response function has the following parameters. Modulate sensitivity to mutations; This is an attenuation factor that is dynamically adjusted based on historical peak wave factors to prevent background fluctuation interference. 3) Box dimension calculation; In formula (6), D is the target fractal dimension; Based on the basic scale parameter, This indicates a gradual refinement of the analytical scale from macro to micro. Let represent the dynamic scale parameter set for the i-th local region and the j-th time window of the dataset; Ni,j(ϵ): the number of boxes required to cover the i-th local region and the j-th time window at the dynamic scale ϵi,j; M is the number of local regions divided in the spatial dimension; T is the number of windows divided in the temporal dimension. It is a spatiotemporal dynamic weighting function that is dynamically adjusted based on the information entropy of the local region and the importance of the time window; The variance of the number of boxes filled in the i-th region and the j-th time window; This is the variance influence factor, which controls the moderating effect of variance on weights. Let be the correlation coefficient between the i-th region and its neighboring regions in the j-th time window. Correlation influencing factors; 4) Multifractal analysis; In formula (7), Here, q is the generalized scaling exponent, and q is the order moment parameter. Focus on high-probability areas. Focus on low-probability, rare events. Corresponding to box dimension calculation; The basic scale parameter represents the initial box size used when covering the system. This indicates a refined analytical process from macroscopic to microscopic levels; In order to scale The total number of boxes required for the under-coverage system; Let be the probability measure of the i-th box. For dynamic feature interaction weights, quantize the order moment q and scale of the i-th box and its neighboring box j. The interaction strength under these conditions; The distance between boxes i and j in the feature space. This refers to the spatial interaction strength parameter; This is a dynamic scaling adjustment factor; The multi-scale interaction index quantifies the coupling effect of the i-th box across different scales. For multi-scale coupling strength parameters; 5) Feature fusion and weight adjustment; Time-domain feature weight formula: ; In formula (8), Here, the time-domain feature weights are defined for the nth iteration, where n is the current iteration number, ranging from 1 to n to N, and gradually increasing with each iteration; N is the total iteration number, defining the time span for weight adjustment and determining the iteration progress factor. The range of variation; The dynamic attenuation coefficient, This is the decay rate parameter; Adaptive rate of change coefficient; Dynamic offset coefficient; The correlation sensitivity coefficient; The real-time correlation coefficient between the time-domain and frequency-domain characteristics; when At that time, the center of weight change shifts towards the earlier cycle; when In time, the system shifts to a later, cyclical manner to achieve feature-driven weight adjustment; This is a cross-domain coupling factor that quantifies the influence of frequency domain features on time domain weights, with a value range of [−1, 1]. When, positive correlation enhances the time domain weight; when At that time, negative correlation suppresses time-domain weights; Variance regularization factor, which suppresses abnormal fluctuations caused by noise; The real-time variance of the time-domain features; Fractal feature weighting formula: ; In formula (9), The fractal feature weights are used in the nth iteration. n is the current loop number, reflecting the data processing progress, and its value range is 1≤n≤N; N is the total number of loops; This is the dynamic fractal weight growth coefficient. The larger the value, the higher the upper limit of the fractal feature weight; The adaptive fractal weight growth rate coefficient replaces the fixed w. The larger the value, the faster the weight increases; This is the dynamic fractal weight growth offset coefficient; This is a fractal complexity enhancement factor, which strengthens the influence of fractal dimension on weights. is the dimension value of the fractal feature in the nth iteration; Frequency domain feature weight formula: ;In formula (10), The time-domain feature weights for the current loop are calculated using formula (8); The fractal feature weights for the current loop are calculated using formula (9); The time-domain weighting influence coefficient has a value range of (0,1). The larger the value, the stronger the reduction effect of the time-domain feature weights on the frequency-domain feature weights; This is the fractal weighting influence coefficient, with a value range of (0,1). The larger the value, the stronger the reduction effect of the fractal feature weight on the frequency domain feature weight.

Images