Battery abnormal state online detection system based on pattern recognition

Abstract

The invention discloses a battery abnormal state online detection system based on pattern recognition, and the system comprises a data collection and preprocessing module which is used for collecting multi-dimensional parameter data and generating a key time sequence; the successive variational modal decomposition module is used for extracting an instantaneous amplitude and an instantaneous phase and generating modal features; the improved CoAtNet pattern recognition module is used for performing multi-dimensional modeling to obtain high-dimensional pattern features; the state judgment module is used for generating a classification result and an abnormal risk index and carrying out state division; and the early warning and linkage control module is used for executing online early warning and issuing a control instruction to the battery management platform. According to the invention, through successive variational mode decomposition and improvement of the CoAtNet network, advanced accurate online detection and linkage control of the abnormal state of the battery are realized.

Description

Technical Field

[0001] This invention relates to the field of battery state monitoring technology, and in particular to an online detection system for abnormal battery states based on pattern recognition. Background Technology

[0002] With the rapid development of electric vehicles, energy storage power stations, and uninterruptible power supplies (UPS), batteries are operating for extended periods under high-rate charging and discharging, wide temperature ranges, and complex conditions, making safe operation and health management increasingly important. Most existing battery management systems (BMS) set fixed safety thresholds for a few parameters such as voltage, current, and temperature, triggering protection actions through limit exceedance judgments and simple logic rules. Some systems introduce empirical curves or lookup table methods for state estimation and alarms. Current monitoring methods based on single or a few parameter thresholds lack sufficient multi-dimensional data mining during battery operation, failing to reflect the dynamic characteristics of batteries under different operating conditions, aging evolution, and environmental disturbances. They are slow to respond to early, subtle anomalies and latent faults, often only generating alarms when parameters severely exceed limits. This results in untimely and inaccurate anomaly identification, failing to meet the needs of new energy vehicles and large-scale energy storage scenarios for online detection of battery anomalies, highly reliable early warning, and refined linkage control.

[0003] In the current context, signal decomposition and intelligent recognition methods are gradually being introduced into the fields of battery state estimation and fault diagnosis. Some studies utilize empirical mode decomposition, variational mode decomposition, and wavelet analysis to perform time-frequency decomposition of voltage and current signals, constructing health indicators from component energy or spectral features. Other schemes employ convolutional neural networks, recurrent neural networks, and Transformer deep learning models for pattern recognition of battery operating data. However, existing mode decomposition methods generally suffer from problems such as mode aliasing, mode redundancy, and difficulty in effectively constraining residual signals in engineering applications. They lack orthogonalization processing and mode merging and renormalization mechanisms for battery signals, resulting in unclear decomposition results that are difficult to directly support identification. Most existing deep networks simply treat battery data as a one-dimensional time series or a regular two-dimensional matrix input, failing to fully utilize the structural relationships between the time dimension, physical feature dimension, and modal dimension. They lack dedicated network structures for phase and amplitude separation modeling, axial decoupling modeling, and modal relationship modeling for battery anomaly identification. The linkage between online detection results and battery management platforms is mostly limited to simple alarms, and an integrated online monitoring and early warning linkage control scheme for battery anomaly states based on multi-dimensional data and pattern recognition algorithms has not yet been formed.

[0004] Therefore, how to provide an online detection system for abnormal battery conditions based on pattern recognition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] An object of the present invention is to propose an online detection system for battery abnormal states based on pattern recognition. The present invention establishes a complete technical process around the acquisition, processing, and online warning of multi-dimensional battery operation data. Through successive variational mode decomposition, a modal orthogonality constraint and a modal merging and reorganization mechanism are introduced during the decomposition process to obtain intrinsic mode components with clear structures. In the pattern recognition stage, an improved CoAtNet network is constructed, adopting a phase-amplitude dual-channel structure, a three-axis decoupling operation structure of the time axis, the feature axis, and the modal axis, and a modal graph feature enhancement structure to jointly model the time dimension, the physical feature dimension, and the modal dimension, extract high-dimensional pattern features representing the battery operation state, complete state classification and abnormal risk assessment, and drive the battery management platform to execute linkage control. Compared with the existing monitoring method that relies on a single threshold judgment, the present invention can identify abnormalities in advance during the stage when parameters do not exceed the limit, improve the accuracy and timeliness of battery abnormal detection, and enhance the robustness and engineering applicability under complex working conditions.

[0006] An online detection system for battery abnormal states based on pattern recognition according to an embodiment of the present invention includes the following modules: A data acquisition and preprocessing module, configured to acquire multi-dimensional parameter data during the operation of the battery and generate a key time series; A successive variational mode decomposition module, configured to extract instantaneous amplitude and instantaneous phase according to the key time series and generate modal features; An improved CoAtNet pattern recognition module, configured to perform multi-dimensional modeling on the modal features based on a phase-amplitude dual-channel and a three-axis decoupling operation and a modal graph structure to obtain high-dimensional pattern features; A state determination module, configured to generate a classification result and an abnormal risk index of the current operation state of the battery according to the high-dimensional pattern features and perform state division; An early warning and linkage control module, configured to execute online warning according to the classification result and the abnormal risk index and issue a control instruction to the battery management platform.

[0007] Optionally, the modules are implemented by the following method: Acquire multi-dimensional parameter data of the battery during operation, arrange them in chronological order, and perform preprocessing to obtain a key time series; Perform successive variational mode decomposition on the key time series, solve a single modal component from the current residual signal in each round of decomposition, perform modal orthogonality processing, and perform modal merging and reorganization according to the center frequency difference and the spectral overlap degree to obtain intrinsic mode components; Construct a modal feature set according to the intrinsic mode components, extract an instantaneous amplitude sequence and an instantaneous phase sequence, combine the time dimension, the physical feature dimension, and the modal index dimension to construct a pseudo-two-dimensional feature map, and perform statistical operations to obtain modal features; An improved CoAtNet network is constructed, a phase-amplitude dual-path is set to extract modal features, and dimensional modeling is performed through three-axis decoupling operation. In the middle and high-level stages, a modal graph is set to construct modal graphs for modal-level features, and graph feature updates are performed. The updated modal-level features are mapped back to the corresponding modal channels to obtain high-dimensional mode features. Based on the high-dimensional pattern features, the classification results and abnormal risk indicators of the current operating status of the battery are generated, and the battery operating status is divided. The system provides online early warning and coordinated control based on classification results and abnormal risk indicators, and issues control commands to the battery management platform.

[0008] Optionally, the multidimensional parameter data includes individual cell voltage, battery pack voltage, current, temperature, state of charge, and state of health.

[0009] Optionally, obtaining the key time series includes: The individual cell voltage, battery pack voltage, current, temperature, state of charge and health status of the battery during operation are sampled synchronously. Multidimensional raw sampling data with timestamps are obtained according to the sampling period, and the multidimensional raw sampling data is sorted according to the timestamps to form a time-continuous multidimensional raw sampling data sequence. Denoising, missing value imputation, and outlier correction are performed on each parameter channel of the time-continuous multidimensional raw sampled data sequence. Resampling and time alignment are performed on the time axis to obtain the multidimensional preprocessed data sequence. Based on charging, discharging, and resting conditions, the multidimensional preprocessed data sequence is identified and divided into intervals. The continuous time intervals of different conditions are divided into multiple data segments. Amplitude normalization and dimension rearrangement are performed on the multidimensional preprocessed data sequence in each data segment. Key time series are constructed with battery running time as the vertical index and multidimensional parameters as the horizontal index.

[0010] Optionally, obtaining the intrinsic mode components includes: The key time series is recorded as the original signal. The maximum number of modes, residual energy threshold, variational mode decomposition penalty factor and iteration termination threshold are set. The initial residual signal is initialized. The value of the initial residual signal at each time index is equal to the value of the original signal at the current time index. In each round of variational mode decomposition, the residual signal updated in the previous round is used as the input signal for the current round of decomposition. The single narrowband mode component and its corresponding center frequency are solved by variational mode decomposition, and the single narrowband mode component obtained in the current round is added to the mode component set. Orthogonalization is performed on the current set of modal components, projecting each modal component onto orthogonal directions to obtain an orthogonal set of modal components. The original signal is then linearly reconstructed using the orthogonal set of modal components to generate a reconstructed signal. The residual signal is updated by subtracting the value of the reconstructed signal from the value of the original signal. In the set of orthogonal modal components, calculate the center frequency difference, spectral overlap and energy ratio of any two orthogonal modal components. Based on the threshold, divide the orthogonal modal components with similar center frequencies or large spectral overlap into the same mode cluster. Perform linear merging on the orthogonal modal components within each mode cluster and perform local variational solution to generate merged renormalized modal components. All merged and renormalized mode components are collected to form an intrinsic mode component set. The successive variational mode decomposition process ends when the residual signal energy is lower than the residual energy threshold, or the number of intrinsic mode components reaches the maximum number of modes, and the difference between the center frequency of the latest merged and renormalized mode component and the center frequency of any intrinsic mode component in the intrinsic mode component set is lower than the center frequency difference threshold.

[0011] Optionally, the step of performing statistical operations to obtain modal features includes: An analytical signal is constructed for each intrinsic mode component. The amplitude of the analytical signal is read at each time index as an instantaneous amplitude sequence, and the phase angle of the analytical signal is read at each time index as an instantaneous phase sequence. All instantaneous amplitude sequences and instantaneous phase sequences are stored in the modal feature set. Using time index as the vertical index and the index of the types of physical parameters and intrinsic modal components collected during battery operation as the horizontal index, the corresponding positions are filled according to the value order of instantaneous amplitude sequence and instantaneous phase sequence to construct a pseudo two-dimensional feature map, which is then incorporated into the modal feature set. On the pseudo-two-dimensional feature map, as well as the instantaneous amplitude sequence and instantaneous phase sequence, mean operation, variance operation, energy operation and extreme value operation are performed according to the time window and preset modal range to generate corresponding statistical features. The statistical features are then matched with the pseudo-two-dimensional feature map index to form modal features.

[0012] Optionally, obtaining the high-dimensional pattern features includes: An improved CoAtNet network is constructed, including a phase-amplitude dual-path feature extraction layer, a three-axis decoupling operation layer, a modality graph feature enhancement layer, and an output layer; The phase-amplitude dual-path feature extraction layer divides modal features into amplitude feature sets and phase feature sets. The amplitude feature sets are input into the convolution branch, and amplitude patterns are extracted through multiple convolution blocks and nonlinear transformations. The phase feature sets are input into the attention branch, and phase patterns are extracted through multiple self-attention blocks and a feedforward network. At the end of the phase-amplitude dual-path, the first fused feature map is generated through feature concatenation and linear mapping. The three-axis decoupling operation layer takes the first fused feature map as input, sets the time index as the time axis, the physical parameter type as the feature axis, and the modality index as the modality axis. It performs one-dimensional convolution operation in the time axis direction through the time axis convolution unit to extract local temporal features, performs fully connected mapping through the feature axis mixing unit to complete feature axis mixing, and performs self-attention operation on all modal channels under each time index through the modality axis attention unit to complete modality axis related modeling and generate the second fused feature map. The modality graph feature enhancement layer constructs a modality-level feature set based on the second fused feature map, calculates the similarity between any two modalities using the modality-level feature set, connects similar modalities according to the similarity threshold to form a modality graph, performs attention operation on the modality graph to update the modality-level features, and obtains the modality-level feature set. The output layer maps the modal-level feature set back to the modal channel corresponding to the second fusion feature map according to the modal index, and then fuses it with the second fusion feature map through channel concatenation to obtain high-dimensional modal features.

[0013] Optionally, the step of generating classification results and abnormal risk indicators of the current battery operating state based on high-dimensional pattern features, and classifying the battery operating state, includes: High-dimensional pattern features are classified through a fully connected layer and an output unit. The output unit is set with output nodes for normal state, mildly abnormal state, and severely abnormal state, and the score values ​​for the three states are calculated respectively. Normalization is performed on the three state score values. After normalization, the three score values ​​are used as the probability of normal state, the probability of slightly abnormal state, and the probability of severely abnormal state, respectively. The current operating state classification result of the battery is determined based on the largest value. An abnormal risk index is calculated based on the probabilities of mild and severe abnormal states. At each sampling time, the probability of mild abnormal state is multiplied by a first weighting coefficient, and the probability of severe abnormal state is multiplied by a second weighting coefficient. The products of the two are added together to obtain the time series of the abnormal risk index. The battery operating status risk level is classified according to the relationship between the time series of the abnormal risk index and the risk threshold.

[0014] Optionally, the online early warning and linkage control, which involves sending control commands to the battery management platform, includes: At each sampling moment, the current operating status classification result of the battery and the corresponding abnormal risk indicators are read. The normal state, mild abnormal state and severe abnormal state in the classification result, together with the abnormal risk indicators, are used as input information for online early warning and linkage control. When the classification result is a slightly abnormal state or the abnormal risk indicator is in the warning risk range, a warning message is generated and prompted through the monitoring interface and alarm channel. At the same time, control instructions to limit the charging and discharging current, limit the charging and discharging power, and adjust the operating conditions are sent to the battery management platform. When the classification result is a severe abnormal state or the abnormal risk indicator is in the dangerous risk range, an emergency alarm message is generated, and a high-priority alarm is issued through the monitoring interface and alarm channel. Control commands are also sent to the battery management platform to cut off the charging and discharging circuit, activate the heat dissipation device, and lock the battery pack's operating status.

[0015] The beneficial effects of this invention are: By introducing successive variational mode decomposition and an improved CoAtNet network, deep structural modeling and refined pattern recognition of multidimensional battery operating signals are achieved. Utilizing modal orthogonality constraints and mode merging and renormalization mechanisms, time series such as voltage, current, and temperature are successively decomposed to obtain eigenmode components with clear structures and well-defined frequency band divisions, effectively suppressing mode aliasing and redundant decomposition problems. A phase-amplitude dual-path structure, a three-axis decoupled operation structure (time axis, feature axis, and modal axis), and a modal graph feature enhancement structure are constructed to jointly model instantaneous amplitude, instantaneous phase, and multimodal features. This enables the identification of abnormal evolution patterns before parameters exceed limits, improving the sensitivity, accuracy, and robustness to complex operating conditions and high-noise environments in battery anomaly detection.

[0016] This integrated process, encompassing key time series construction, modal feature extraction, high-dimensional pattern feature recognition, state classification and risk assessment, and coordinated control, achieves closed-loop online monitoring of battery anomalies from data acquisition and analysis to control. Based on the state classification results and anomaly risk indicators generated from high-dimensional pattern features, the battery management platform can automatically execute control strategies such as current limiting, power limiting, circuit disconnection, and heat dissipation activation according to risk level, completing the automatic transition from anomaly detection to anomaly handling. Compared to traditional solutions relying on single thresholds or empirical rules, this technical approach significantly improves the timeliness and predictability of anomaly identification, overcoming bottlenecks such as low data utilization, coarse pattern characterization, and single-mode coordinated control. This is beneficial for the safe operation and intelligent maintenance of battery systems in electric vehicles and energy storage power stations. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0018] Figure 1 This is a flowchart of an online battery abnormality detection system based on pattern recognition proposed in this invention; Figure 2This is a structural block diagram of an online battery abnormality detection method based on pattern recognition proposed in this invention. Figure 3 This is a functional diagram of the improved CoAtNet network, which is a pattern recognition-based online detection method for abnormal battery states proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 An online detection system for abnormal battery conditions based on pattern recognition includes the following modules: The data acquisition and preprocessing module is used to collect multi-dimensional parameter data during battery operation and generate key time series. The successive variational mode decomposition module is used to extract instantaneous amplitude and instantaneous phase from key time series and generate modal features; An improved CoAtNet pattern recognition module is used to perform multidimensional modeling of modal features based on phase-amplitude dual-path and triaxial decoupling operations and modal graph structure to obtain high-dimensional pattern features. The state determination module is used to generate classification results and abnormal risk indicators of the current operating state of the battery based on high-dimensional pattern features, and to classify the state. The early warning and linkage control module is used to issue online early warnings based on classification results and abnormal risk indicators, and to send control commands to the battery management platform.

[0021] refer to Figure 2 and Figure 3 A method for online detection of abnormal battery states based on pattern recognition, comprising: Collect multidimensional parameter data of the battery during operation, arrange them in chronological order, and preprocess them to obtain key time series; Successive variational mode decomposition is performed on the key time series. In each round of decomposition, a single mode component is solved from the current residual signal. Mode orthogonalization is performed. Mode merging and renormalization are carried out based on the center frequency difference and spectral overlap to obtain the intrinsic mode components. Modal feature sets are constructed based on intrinsic modal components, instantaneous amplitude sequences and instantaneous phase sequences are extracted, and pseudo-two-dimensional feature maps are constructed by combining time dimension, physical feature dimension and modal index dimension. Statistical operations are then performed to obtain modal features. An improved CoAtNet network is constructed, a phase-amplitude dual-path is set to extract modal features, and dimensional modeling is performed through three-axis decoupling operation. In the middle and high-level stages, a modal graph is set to construct modal graphs for modal-level features, and graph feature updates are performed. The updated modal-level features are mapped back to the corresponding modal channels to obtain high-dimensional mode features. Based on the high-dimensional pattern features, the classification results and abnormal risk indicators of the current operating status of the battery are generated, and the battery operating status is divided. The system provides online early warning and coordinated control based on classification results and abnormal risk indicators, and issues control commands to the battery management platform.

[0022] In this embodiment, the multidimensional parameter data includes individual cell voltage, battery pack voltage, current, temperature, state of charge, and state of health.

[0023] In this embodiment, obtaining the key time series includes: The individual cell voltage, battery pack voltage, current, temperature, state of charge and health status of the battery during operation are sampled synchronously. Multidimensional raw sampling data with timestamps are obtained according to the sampling period, and the multidimensional raw sampling data is sorted according to the timestamps to form a time-continuous multidimensional raw sampling data sequence. Denoising, missing value imputation, and outlier correction are performed on each parameter channel of the time-continuous multidimensional raw sampled data sequence. Resampling and time alignment are performed on the time axis to obtain the multidimensional preprocessed data sequence. Based on charging, discharging, and resting conditions, the multidimensional preprocessed data sequence is subjected to condition identification and interval division. The continuous time intervals of different conditions are divided into multiple data segments. Amplitude normalization and dimensional rearrangement are performed on the multidimensional preprocessed data sequence in each data segment. A key time series is constructed using battery running time as the vertical index and multidimensional parameters as the horizontal index. Specifically, the condition identification and interval division are as follows: Read the current, voltage, power, relay status and state of charge change trends at each sampling moment. When the current is positive and the absolute value of the current is not less than one-tenth of the rated current, the charging power is not less than one-tenth of the rated power and the state of charge continues to rise, it is determined to be the charging condition. When the current is negative and the absolute value of the current is not less than one-tenth of the rated current, the discharging power is not less than one-tenth of the rated power and the state of charge continues to fall, it is determined to be the discharging condition. When the absolute value of the current is less than one-twentieth of the rated current, the power is less than one-twentieth of the rated power, and the relay state and the state of charge are basically unchanged, it is determined to be a stationary working condition. The sampling points that are continuously judged in chronological order are merged, and the continuous time period with a duration of not less than 30 seconds is divided into working condition intervals. The time period with a duration of less than 30 seconds is merged into the previous or next adjacent working condition interval, resulting in a data segment composed of charging interval, discharging interval and stationary interval.

[0024] In this embodiment, obtaining the intrinsic mode components includes: The key time series is denoted as the original signal. A maximum number of modes, a residual energy threshold, a variational mode decomposition penalty factor, and an iteration termination threshold are set. An initial residual signal is initialized, with the value of the initial residual signal at each time index equal to the value of the original signal at the current time index. Specifically, setting the maximum number of modes, the residual energy threshold, the variational mode decomposition penalty factor, and the iteration termination threshold involves: The maximum number of modes is fixed at six based on the spectral distribution of the battery signal. The residual energy threshold is set at two percent of the total energy of the original signal, and the decomposition ends when the residual signal energy is less than two percent of the total energy of the original signal. The variational mode decomposition penalty factor is determined by grid search of multiple sets of typical battery operating data during the offline calibration stage. A fixed value of 3,500 is uniformly taken in electric vehicle and energy storage scenarios to constrain the mode bandwidth and suppress high-frequency noise. The iteration termination threshold is set by the ratio of the difference between the mode components obtained in two adjacent iterations to the amplitude of the current mode component. In each round of variational mode decomposition, the residual signal updated in the previous round is used as the input signal for the current round of decomposition. The single narrowband mode component and its corresponding center frequency are solved through variational mode decomposition. The single narrowband mode component obtained in this round is then added to the mode component set. Specifically, solving for the single narrowband mode component and its corresponding center frequency through variational mode decomposition involves: Perform a Discrete Fourier Transform on the current residual signal to obtain the complex spectrum values ​​of each frequency point in frequency order. Calculate the energy of each frequency point as the square of the amplitude of the current point. Calculate the initial center frequency of the current mode by weighting the frequency with the energy of each frequency point. Only retain the non-negative frequency points and double their energy to form an analytical spectrum. Calculate the bandwidth penalty for each frequency point based on its distance from the current center frequency. The bandwidth penalty is equal to the product of the penalty factor and the square of the frequency distance. Use the bandwidth penalty to attenuate the spectrum amplitude, weakening components far from the center frequency and retaining components close to the center frequency, thus obtaining a single narrowband mode component. The inverse discrete Fourier transform is performed on the analyzed spectrum after bandwidth contraction to obtain the updated time-domain modal signal. Then, the energy-weighted average frequency is recalculated based on the spectral energy distribution of the updated modal signal to obtain the center frequency. Orthogonalization is performed on the current set of modal components, projecting each modal component onto orthogonal directions to obtain an orthogonal set of modal components. The original signal is then linearly reconstructed using this orthogonal set of modal components to generate a reconstructed signal. The residual signal is updated by subtracting the values ​​of the reconstructed signal from the values ​​of the original signal, where: The orthogonalization process is as follows: each modal component in the current modal component set is treated as a vector on the time axis, and processed one by one in a fixed order. The sum of the products of the modal component and the corresponding sampling points of each orthogonal modal component over the entire time range is calculated, and then divided by the sum of squares of the corresponding sampling points of the orthogonal modal component to obtain the projection coefficient. The projection coefficient is multiplied by the orthogonal modal component and subtracted point by point on the time axis. The projection of all orthogonal modal components is subtracted in turn to obtain the new modal component. The linear reconstruction is specifically performed as follows: taking the time index as a reference, at each time index, the values ​​of all orthogonal mode components under the time index are read, the values ​​are added together as the reconstructed signal value at the time index, and the same point-by-point summation operation is performed along the time axis for all time indices to obtain a reconstructed signal sequence covering the entire time range. In the set of orthogonal modal components, calculate the center frequency difference, spectral overlap, and energy ratio of any two orthogonal modal components. Based on a threshold, orthogonal modal components with similar center frequencies or large spectral overlap are grouped into the same mode cluster. Linear merging is performed on the orthogonal modal components within each mode cluster, and local variational solutions are obtained to generate merged renormalized modal components, where: The calculation of the center frequency difference, spectral overlap, and energy ratio of any two orthogonal modal components is specifically as follows: Perform a discrete Fourier transform on each orthogonal mode component to obtain the spectral amplitude of each frequency point in frequency order. Take the square of the amplitude of each frequency point as the energy of the frequency point. Then multiply the frequency value of each frequency point by the energy of the point and accumulate them. Divide by the sum of the energy of the whole frequency band to obtain the center frequency. The difference between the center frequencies of the two modes is the absolute value of the difference between the center frequencies of the two modes. Spectral overlap is determined by comparing the energy of two modes at the same frequency point point by point. When both energies are not less than one-tenth of their respective maximum energy values, they are determined to be overlapping frequency points. The sum of the energies of all overlapping frequency points is divided by the total energy of the mode with smaller energy to obtain the spectral overlap. The energy ratio is obtained by dividing the total energy of the mode with smaller energy by the total energy of the mode with larger energy. The thresholds are specifically: the thresholds include the center frequency difference threshold, the spectral overlap threshold, and the energy ratio threshold. The center frequency difference threshold is set at one-tenth of the total width of the analysis bandwidth, the spectral overlap threshold is fixed at 0.5, and the energy ratio threshold is fixed at 0.2. All merged and renormalized mode components are collected to form an intrinsic mode component set. The successive variational mode decomposition process ends when the residual signal energy is lower than the residual energy threshold, or the number of intrinsic mode components reaches the maximum number of modes, and the difference between the center frequency of the latest merged and renormalized mode component and the center frequency of any intrinsic mode component in the intrinsic mode component set is lower than the center frequency difference threshold.

[0025] In this embodiment, the step of performing statistical operations to obtain modal features includes: An analytical signal is constructed for each intrinsic mode component. The amplitude of the analytical signal is read at each time index as an instantaneous amplitude sequence, and the phase angle of the analytical signal is read at each time index as an instantaneous phase sequence. All instantaneous amplitude and phase sequences are stored in the modal feature set. The construction of the analytical signal specifically involves: For each intrinsic mode component, the discrete sampled values ​​on the time axis are arranged in time order to form a real value sequence. The real value sequence is processed by Hilbert transform to obtain an imaginary value sequence with the same length as the original sequence. The original real value sequence is used as the real part of the analytic signal, and the imaginary value sequence obtained by Hilbert transform is used as the imaginary part of the analytic signal. At each sampling time, the real part and the imaginary part are paired to form a pair of values, which are arranged in time order to form the analytic signal of the intrinsic mode component. Using time index as the vertical index and the index of physical parameter types and intrinsic mode components collected during battery operation as the horizontal index, the corresponding positions are filled according to the value order of instantaneous amplitude sequence and instantaneous phase sequence to construct a pseudo two-dimensional feature map. The pseudo two-dimensional feature map is then incorporated into the modal feature set. The physical parameter types include single cell voltage, battery pack voltage, current, temperature, state of charge, and state of health. On the pseudo-two-dimensional feature map, the instantaneous amplitude sequence, and the instantaneous phase sequence, mean calculation, variance calculation, energy calculation, and extreme value calculation are performed according to the time window and the preset modal range to generate corresponding statistical features. The statistical features are then matched with the pseudo-two-dimensional feature map index to form modal features. Specifically, the mean calculation, variance calculation, energy calculation, and extreme value calculation are performed according to the time window and the preset modal range as follows: The preset modal range is segmented from low to high based on the modal center frequency. The first two low-frequency modes are defined as the low-frequency range, the middle two modes are defined as the mid-frequency range, and the last two modes are defined as the high-frequency range. Instantaneous amplitude and phase are extracted at corresponding locations within each time window and modal range. The mean is calculated for each set of data. The values ​​at all sampling times within the window are summed and then divided by the number of sampling points to obtain the average value of the current window modal range. The variance is calculated by calculating the difference between the current value and the corresponding mean at each point within the window, squaring the differences, summing them, and then dividing the sum of squares by the number of sampling points to obtain the dispersion of the values ​​around the mean. The energy is calculated by squaring the values ​​at each point within the window and summing them. The sum of squares is used as the energy index of the window modal range. The extremum is calculated by traversing all sampling points within the window and recording the maximum and minimum values ​​respectively.

[0026] In this embodiment, obtaining the high-dimensional pattern features includes: An improved CoAtNet network is constructed, comprising a phase-amplitude dual-path feature extraction layer, a three-axis decoupling operation layer, a modality graph feature enhancement layer, and an output layer, wherein: A phase-amplitude dual-path feature extraction layer is set on the input side. Amplitude convolution branch and phase attention branch are used to replace single-path feature extraction. Amplitude features and phase features are encoded separately and then fused at the end of the path. A three-axis decoupling operation layer is introduced in the middle structure. The processing method that only relies on two-dimensional convolution or global attention is replaced by time axis convolution unit, feature axis mixing unit and modal axis attention unit. The time dimension, physical feature dimension and modal dimension are modeled separately and jointly fused. After the three-axis decoupling operation layer, a modal graph feature enhancement layer is connected in series. Modal graphs and modal-level feature vectors constructed based on modal similarity are introduced to participate in feature update and enhancement. An output layer is set to replace the simple pooling classification head. The features after multi-layer fusion are aggregated and mapped to form a high-dimensional modal feature output for battery status classification and abnormal risk assessment. The phase-amplitude dual-path feature extraction layer divides modal features into amplitude feature sets and phase feature sets. The amplitude feature sets are input into a convolutional branch, where amplitude patterns are extracted through multiple convolutional blocks and nonlinear transformations. The phase feature sets are input into an attention branch, where phase patterns are extracted through multiple self-attention blocks and a feedforward network. At the end of the phase-amplitude dual-path, a first fused feature map is generated through feature concatenation and linear mapping. Specifically, dividing the modal features into amplitude and phase feature sets involves: According to the time index, extract the instantaneous amplitude sequence and instantaneous phase sequence for each intrinsic mode component. Write the instantaneous amplitude values ​​corresponding to the same time index, the same physical parameter, and the same intrinsic mode component into the reserved amplitude positions in the pseudo-two-dimensional feature map according to the row and column positions. Write the corresponding instantaneous phase values ​​into the phase positions in the pseudo-two-dimensional feature map. Copy all channels or columns storing instantaneous amplitude values ​​in the pseudo-two-dimensional feature map in chronological order to form an amplitude feature set. Copy all channels or columns storing instantaneous phase values ​​in the pseudo-two-dimensional feature map in chronological order to form a phase feature set. The three-axis decoupling operation layer takes the first fused feature map as input, sets the time index as the time axis, the physical parameter type as the feature axis, and the modality index as the modality axis. It performs one-dimensional convolution operation in the time axis direction through the time axis convolution unit to extract local temporal features, performs fully connected mapping through the feature axis mixing unit to complete feature axis mixing, and performs self-attention operation on all modal channels under each time index through the modality axis attention unit to complete modality axis related modeling and generate the second fused feature map. The modality graph feature enhancement layer constructs a modality-level feature set based on the second fused feature map. It then calculates the similarity between any two modalities using this feature set. Based on a similarity threshold, it connects similar modalities to form a modality graph. Attention operations are performed on the modality graph to update the modality-level features, resulting in the modality-level feature set: The construction of the modal-level feature set based on the second fused feature map specifically involves: using the modal index as the modal dimension identifier in the second fused feature map; performing averaging and maximizing operations in each modal channel; arranging the average, maximum, energy, and variance of the same modal channel in a fixed order into a modal feature vector; repeating the operation for all modal channels; and collecting all modal feature vectors in the order of modal index to form a modal-level feature set. The calculation of the similarity between any two modalities is specifically as follows: select any two modal feature vectors from the modal-level feature set, multiply the values ​​at corresponding positions in each modal feature vector one by one and sum them up to obtain the inner product of the two vectors, and at the same time calculate the square root of the sum of squares of each element of each modal feature vector as the length of the current vector, and then divide the inner product value by the product of the lengths of the two vectors in turn to obtain the similarity of the modal feature vectors. The process of connecting similar modalities to form a modal graph based on a similarity threshold is as follows: each modal feature vector in the modal-level feature set is assigned a modal node number; modal pairs with a similarity not lower than a preset similarity threshold are considered similar modal pairs. The similarity threshold is fixed at 0.8; an undirected edge is added between the corresponding two modal nodes; the similarity is used as the weight value of the current edge; all modal node pairs are traversed to complete the edge connection operation that satisfies the similarity threshold condition, thus forming a modal graph structure composed of modal nodes and weighted edges. The output layer maps the modal-level feature set back to the corresponding modal channel of the second fused feature map according to the modal index, and then fuses it with the second fused feature map through channel concatenation to obtain high-dimensional modal features. Specifically, the mapping back to the corresponding modal channel of the second fused feature map according to the modal index is as follows: In the second fused feature map, each modal channel is numbered sequentially. In the modal-level feature set, each modal feature vector is numbered in the same order to make them correspond. For any modal number, the corresponding modal feature vector is read and expanded into a feature map with the same size as the current modal channel by performing linear transformations in the time direction and feature direction. The expanded feature map is written into the modal channel position with the corresponding number in the second fused feature map and concatenated with the original channel features in the channel dimension to complete the mapping and fusion of modal-level feature vectors to modal channels.

[0027] In this embodiment, the step of generating classification results and abnormal risk indicators of the current battery operating state based on high-dimensional pattern features, and classifying the battery operating state, includes: High-dimensional pattern features are classified using a fully connected layer and an output unit. The output unit is configured with output nodes for normal state, slightly abnormal state, and severely abnormal state, and scores are calculated for each of the three states. The calculation of the scores for the three states is as follows: The high-dimensional pattern features are expanded into a feature vector in a fixed order and input into a fully connected unit. In the fully connected unit, a set of weight coefficients and bias coefficients are set for the normal state, the mild abnormal state and the severe abnormal state respectively. For each set of weight coefficients, each component of the feature vector is multiplied by the corresponding weight coefficient and then added together. The current set of bias coefficients is then added to obtain a real number output for the corresponding state. These three real number outputs are recorded as the score values ​​for the normal state, the mild abnormal state and the severe abnormal state respectively. Normalization is performed on the three state score values. After normalization, the three score values ​​are used as the probability of normal state, the probability of slightly abnormal state, and the probability of severely abnormal state, respectively. The current operating state classification result of the battery is determined based on the largest value. An anomaly risk index is calculated based on the probabilities of mild and severe anomalies. At each sampling time, the probability of a mild anomaly is multiplied by a first weighting coefficient, and the probability of a severe anomaly is multiplied by a second weighting coefficient. The products are then added together to obtain the time series of the anomaly risk index. The battery operating status risk level is classified according to the relationship between the anomaly risk index time series and the risk threshold, where: The first weighting coefficient and the second weighting coefficient are set to 0.4 and 0.6, respectively. The method for classifying battery operating status risk levels is as follows: battery operating status risk levels are classified according to fixed thresholds. Sampling times with abnormal risk indicators less than 0.3 are classified as normal risk levels, sampling times with abnormal risk indicators greater than or equal to 0.3 and less than 0.6 are classified as early warning risk levels, and sampling times with abnormal risk indicators greater than or equal to 0.6 are classified as dangerous risk levels.

[0028] In this embodiment, the online early warning and linkage control, which involves sending control commands to the battery management platform, includes: At each sampling moment, the current operating status classification result of the battery and the corresponding abnormal risk indicators are read. The normal state, mild abnormal state and severe abnormal state in the classification result, together with the abnormal risk indicators, are used as input information for online early warning and linkage control. When the classification result is a slightly abnormal state or the abnormal risk indicator is in the warning risk range, a warning message is generated and prompted through the monitoring interface and alarm channel. At the same time, control instructions to limit the charging and discharging current, limit the charging and discharging power, and adjust the operating conditions are sent to the battery management platform. When the classification result is a severe abnormal state or the abnormal risk indicator is in the dangerous risk range, an emergency alarm message is generated, and a high-priority alarm is issued through the monitoring interface and alarm channel. Control commands are also sent to the battery management platform to cut off the charging and discharging circuit, activate the heat dissipation device, and lock the battery pack's operating status.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to a charging station of an electric bus operator, equipped with 24 sets of lithium iron phosphate battery cabinets, each with a rated capacity of 240 kWh, connecting a total of 120 pure electric buses. The original battery management system mainly relied on three fixed thresholds for protection control: upper limit of single cell voltage, upper limit of temperature, and upper limit of charging and discharging current. The operator reported that there were 11 abnormal temperature rises during charging between January and March 2025. Three of these alarms were triggered when the single cell temperature approached the safety limit, and two cases showed rapid rise in single cell voltage without exceeding the limit, but accelerated capacity decay was detected afterward. This indicates a problem of delayed anomaly identification and untimely detection of potential hazards.

[0030] The pattern recognition-based online battery anomaly detection system of this invention is deployed as software on a station-level monitoring server. It receives multi-dimensional parameters of each battery cabinet in real time, including individual cell voltage, battery pack voltage, current, temperature, state of charge, and health status, via existing communication interfaces, with a sampling frequency set to 10 Hz. The server preprocesses the collected multi-dimensional data, constructs key time series, and uses successive variational mode decomposition to perform mode decomposition on the voltage, current, and temperature signals of typical charge and discharge stages, generating eigenmode components and corresponding instantaneous amplitude and phase characteristics. These are then input into an improved CoAtNet network structure for pattern recognition, providing online battery status classification results and anomaly risk indicators. When the anomaly risk reaches the warning level, the station control layer issues current-limiting and power-limiting commands to the existing battery management platform; when the risk reaches the danger level, the charging circuit is disconnected and forced cooling is initiated, achieving a closed loop from identification to control.

[0031] During the trial operation phase from April to June 2025, the system continuously monitored the entire operation of ten battery cabinets, recording over 350 complete charge-discharge cycles and approximately 1,500 hours of on-grid data. The system identified 21 minor abnormal operating conditions related to factors such as decreased cooling capacity, excessively high charging current settings, and excessively high ambient temperature, and provided corresponding time series of abnormal risks. For most conditions, the first risk escalation warning appeared within 10 to 40 seconds after the start of charging. The system also identified several operating segments with abnormal voltage fluctuations that did not reach conventional over-limit conditions, generating corresponding risk curves and operating condition labels for maintenance personnel to replay and analyze on the station-level monitoring terminal. Simultaneously, relevant data was automatically archived for model parameter optimization and maintenance strategy adjustments, providing on-site maintenance with traceable records of abnormal patterns and more detailed decision-making basis.

[0032] Table 1. Performance Comparison of Different Battery Anomaly Detection Methods

[0033] As shown in the table, in terms of detection performance, the detection accuracy of the method of this invention is 96.5%, which is 11.2% higher than the threshold method (85.3%), 6.8% higher than EMD+SVM (89.7%), and also higher than VMD+CNN (92.4%), one-dimensional CNN (90.8%), and Transformer (93.1%). Regarding anomaly recall, the method of this invention achieves 95.1%, which is 16.9% higher than the threshold method (78.2%), 6.4% higher than VMD+CNN (88.7%), and 4.7% higher than Transformer (90.4%). The corresponding false positive rate is only 2.7%, lower than the 4.1% to 6.5% of the comparative methods; the false negative rate is 3.5%, while the false negative rate of the comparative methods ranges from 7.4% to 15.3%. This invention achieves the best numerical results in both accuracy and recall.

[0034] In terms of timeliness, the average early warning time is 8.6 seconds for the threshold method, 14.3 seconds for EMD+SVM, 21.7 seconds for VMD+CNN, 19.8 seconds for one-dimensional CNN, and 23.5 seconds for Transformer. The method of this invention achieves 30.9 seconds, which is 22.3 seconds longer than the threshold method and 7.4 seconds longer than the Transformer, providing a greater margin for processing time. Regarding average detection latency, the threshold method is 7.5 milliseconds, EMD+SVM is 24.3 milliseconds, VMD+CNN is 32.8 milliseconds, one-dimensional CNN is 15.6 milliseconds, and Transformer is 27.9 milliseconds. The method of this invention is 19.4 milliseconds. Although slightly higher than the threshold method, it is significantly lower than VMD+CNN and Transformer, falling within the acceptable range for real-time online detection.

[0035] In terms of robustness and adaptability to different operating conditions, the noise immunity index is as follows: threshold method 0.62, EMD+SVM 0.70, VMD+CNN 0.76, one-dimensional CNN 0.73, Transformer 0.80, and this invention achieves 0.89, which is 0.09 higher than the second-best method. Regarding operating condition adaptability, the threshold method scores 6.0, EMD+SVM 7.1, VMD+CNN 8.0, one-dimensional CNN 7.5, Transformer 8.3, and this invention scores 9.2, demonstrating better stability and adaptability under various charging / discharging conditions, temperature environments, and operating condition switching.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A pattern recognition based online detection system for abnormal states of a battery, characterized in that, The method comprises the following steps: a data acquisition and preprocessing module is used to acquire multi-dimensional parameter data of the battery during operation and generate key time series; a successive variational modal decomposition module is used to extract instantaneous amplitude and instantaneous phase according to the key time series and generate modal features; an improved CoAtNet pattern recognition module is used to perform multi-dimensional modeling on the modal features based on a phase-amplitude dual channel, three-axis decoupling operation and modal graph structure, and obtain high-dimensional pattern features; a state determination module is used to generate classification results and abnormal risk indicators of the current operation state of the battery according to the high-dimensional pattern features, and perform state division; a warning and linkage control module is used to perform online warning according to the classification results and abnormal risk indicators, and issue control instructions to the battery management platform. 2.The online battery abnormal state detection method based on pattern recognition according to claim 1, applied to the online battery abnormal state detection system based on pattern recognition according to claim 1, characterized in that, The method comprises the following steps: acquire multi-dimensional parameter data of the battery during operation, arrange them in time sequence, and perform preprocessing to obtain key time series; perform successive variational modal decomposition on the key time series, solve a single modal component from the current residual signal in each round of decomposition, perform modal orthogonalization processing, and perform modal merging and reorganization according to the center frequency difference and spectral overlap degree to obtain an intrinsic modal component; construct a modal feature set according to the intrinsic modal component, extract an instantaneous amplitude sequence and an instantaneous phase sequence, combine time dimension, physical feature dimension and modal index dimension to construct a pseudo two-dimensional feature graph, and perform statistical operation to obtain modal features; construct an improved CoAtNet network, set up a phase-amplitude dual channel to extract patterns from the modal features, perform dimension modeling through three-axis decoupling operation, set up a modal graph at the medium and high level stage to construct a modal graph for the modal level features, and perform graph feature updating, map the updated modal level features back to the corresponding modal channel to obtain high-dimensional pattern features; generate classification results and abnormal risk indicators of the current operation state of the battery according to the high-dimensional pattern features, and divide the battery operation state; perform online warning and linkage control on the classification results and abnormal risk indicators, and issue control instructions to the battery management platform.

3. The method of claim 2, wherein the method comprises: The multi-dimensional parameter data includes single cell voltage, battery pack voltage, current, temperature, state of charge and state of health.

4. The method of claim 2, wherein the method comprises: The key time series is obtained by: synchronously sampling the single cell voltage, battery pack voltage, current, temperature, state of charge and state of health of the battery during operation, obtaining multi-dimensional original sampling data with timestamps according to the sampling period, and sorting the multi-dimensional original sampling data according to the timestamps to form a time-continuous multi-dimensional original sampling data sequence; performing denoising, missing value filling and abnormal value correction processing on the time-continuous multi-dimensional original sampling data sequence in each parameter channel, and performing resampling and time alignment processing on the time axis to obtain a multi-dimensional preprocessed data sequence; performing work condition identification and interval division on the multi-dimensional preprocessed data sequence according to the charging condition, discharging condition and standing condition, dividing the continuous time interval of different working conditions into multiple data segments, performing amplitude normalization and dimension rearrangement processing on the multi-dimensional preprocessed data sequence in each data segment, and constructing the key time series with battery operation time as the vertical index and multi-dimensional parameters as the horizontal index.

5. The method of claim 2, wherein the method is characterized by: The intrinsic modal component includes: Record the key time sequence as the original signal, set the maximum modal number, residual energy threshold, variational modal decomposition penalty factor and iteration termination threshold, initialize the initial residual signal, and the value of the initial residual signal at each time index is equal to the value of the original signal at the current time index; In each round of variational modal decomposition, the updated residual signal of the previous round is taken as the input signal of the current round, and a single narrowband modal component and a corresponding center frequency are solved by variational modal decomposition, and the single narrowband modal component solved in the current round is added to the modal component set; Orthogonalization processing is performed on the current modal component set, each modal component is projected onto the orthogonal direction to obtain an orthogonal modal component set, and the original signal is linearly reconstructed by the orthogonal modal component set to generate a reconstructed signal, and the residual signal is updated by subtracting the value of the reconstructed signal from the value of the original signal; The center frequency difference, spectral overlap and energy ratio of any two orthogonal modal components in the orthogonal modal component set are calculated, and the orthogonal modal components with close center frequencies or large spectral overlap are divided into the same modal cluster according to the threshold, linear combination is performed on the orthogonal modal components in each modal cluster, and local variational solving is performed to generate a combined and reconstructed modal component; All combined and reconstructed modal components are collected to form an intrinsic modal component set, and when the residual signal energy is lower than the residual energy threshold, or the number of intrinsic modal components reaches the maximum modal number, and the difference between the center frequency of the latest combined and reconstructed modal component and the center frequency of any intrinsic modal component in the intrinsic modal component set is lower than the center frequency difference threshold, the successive variational modal decomposition process is ended.

6. The method of claim 2, wherein the method is characterized by: The statistical operation includes: For each intrinsic modal component, an analytic signal is constructed, the amplitude value of the analytic signal at each time index is read as an instantaneous amplitude sequence, and the phase angle of the analytic signal at each time index is read as an instantaneous phase sequence, and all instantaneous amplitude sequences and instantaneous phase sequences are stored in a modal feature set; A pseudo two-dimensional feature map is constructed by taking the time index as the vertical index, taking the index of the intrinsic modal component as the horizontal index, and filling the corresponding position according to the value order of the instantaneous amplitude sequence and the instantaneous phase sequence, and the pseudo two-dimensional feature map is incorporated into the modal feature set; On the pseudo two-dimensional feature map and the instantaneous amplitude sequence and the instantaneous phase sequence, mean value operation, variance operation, energy operation and extreme value operation are performed according to the time window and the preset modal range to generate corresponding statistical features, and the statistical features are corresponded to the pseudo two-dimensional feature map index to form the modal features.

7. The method of claim 2, wherein the method is characterized by: The high-dimensional pattern feature includes: An improved CoAtNet network is constructed, including a phase-amplitude dual-channel feature extraction layer, a three-axis decoupling operation layer, a modal map feature enhancement layer and an output layer; The phase-amplitude dual-channel feature extraction layer divides the modal features into an amplitude feature set and a phase feature set, inputs the amplitude feature set into a convolution branch, extracts amplitude patterns through multiple convolution blocks and nonlinear transformations, inputs the phase feature set into an attention branch, extracts phase patterns through multiple self-attention blocks and a feedforward network, and generates a first fusion feature map at the end of the phase-amplitude dual-channel through feature splicing and linear mapping; The three-axis decoupling operation layer takes the first fusion feature map as input, sets the time index as the time axis, the physical parameter type as the feature axis, and the modal index as the modal axis, extracts time local features through one-dimensional convolution operation in the time axis direction by a time axis convolution unit, completes feature axis mixing through a feature axis mixing unit by performing full connection mapping, and completes modal axis related modeling by performing self-attention operation on all modal channels at each time index through a modal axis attention unit, to generate a second fusion feature map; The modal graph feature enhancement layer constructs a modal level feature set according to the second fusion feature map, calculates the similarity between any two modalities using the modal level feature set, connects similar modalities according to a similarity threshold to form a modal graph, and updates the modal level features by performing attention operation on the modal graph to obtain the modal level feature set; The output layer maps the modal level feature set back to the modal channels corresponding to the second fusion feature map according to the modal index, fuses the second fusion feature map through channel splicing, and obtains high-dimensional pattern features.

8. The method of claim 2, wherein the method is characterized by: The classification result and the abnormal risk index of the current running state of the battery are generated according to the high-dimensional pattern features, and the running state of the battery is divided, including: The high-dimensional pattern features are classified through a full connection and an output unit, the output unit sets normal state output nodes, mild abnormal state output nodes and severe abnormal state output nodes, and the scores of the three states are calculated respectively; The three state score values are normalized, and after normalization, the three score values are taken as the normal state probability, the mild abnormal state probability and the severe abnormal state probability respectively, and the classification result of the current running state of the battery is determined according to the largest value; The abnormal risk index is calculated according to the mild abnormal state probability and the severe abnormal state probability, the mild abnormal state probability is multiplied by a first weight coefficient at each sampling time, the severe abnormal state probability is multiplied by a second weight coefficient, and the two products are added to obtain an abnormal risk index time series, and the risk level of the running state of the battery is divided according to the size relationship between the abnormal risk index time series and the risk threshold.

9. The method of claim 2, wherein the method is characterized by: The online early warning and linkage control is performed, and control instructions are issued to the battery management platform, including: The current running state classification result and the corresponding abnormal risk index are read at each sampling time, and the normal state, the mild abnormal state and the severe abnormal state in the classification result and the abnormal risk index are taken as input information of the online early warning and linkage control together; When the classification result is a mild abnormal state or the abnormal risk index is in a pre-warning risk interval, pre-warning information is generated and prompted through a monitoring interface and an alarm channel, and control instructions of limiting charging and discharging current, limiting charging and discharging power and adjusting operating conditions are issued to a battery management platform; When the classification result is a severe abnormal state or the abnormal risk index is in a dangerous risk interval, emergency alarm information is generated, high-priority alarm is performed through the monitoring interface and the alarm channel, and control instructions of cutting off the charging and discharging loop, enabling the heat dissipation device and locking the operating state of the battery pack are issued to the battery management platform.

Images