A health status detection system and method for new energy vehicle battery packs

Abstract

This invention belongs to the field of battery health detection technology and discloses a health status detection system and method for new energy vehicle battery packs. It acquires key parameters such as voltage, current, temperature, and high-frequency impedance with nanosecond-level precision through a multi-channel synchronous acquisition unit of the BMS. During the vehicle's idle period after engine shutdown, a low-disturbance DC pulse excitation is injected to complete in-situ benchmark calibration. Simultaneously, it pre-detects the consistency of individual cells and marks abnormal cells, eliminating measurement errors caused by acquisition system drift, environmental disturbances, and initial cell inconsistencies. For three operating conditions—static idle, low-rate stable operation, and high-rate dynamic operation—an adaptive matching denoising algorithm is used. Combined with the mutual information entropy algorithm, core health features are extracted and redundant information is removed to ensure a high signal-to-noise ratio. In the cell estimation stage, a single-particle electrochemical model and a second-order equivalent circuit model are fused, and a lightweight CNN-LSTM network based on an attention mechanism is used for pre-training.

Description

Technical Field

[0001] This invention belongs to the field of battery health detection technology, specifically a health status detection system and method for new energy vehicle battery packs. Background Technology

[0002] The health status of a power lithium-ion battery pack directly determines the vehicle's driving range, charging and discharging performance, and driving safety. However, existing battery pack SOH detection technologies have the following technical problems: While offline testing technologies such as capacity calibration and electrochemical impedance spectroscopy offer high accuracy, they require disassembling the battery pack and removing it from the vehicle's operating conditions. This makes it impossible to achieve real-time online testing throughout the entire lifecycle and fails to meet the actual usage requirements of the vehicle. Model-based online detection technology, with equivalent circuit model and simplified electrochemical model as the core, relies on algorithms such as Kalman filtering to achieve estimation. However, under the complex dynamic conditions of the whole vehicle, wide temperature range environment, and late stage of battery aging, the deviation between model parameters and actual battery characteristics increases sharply. Furthermore, it does not consider system-level factors such as inconsistency of individual cells in the battery pack and circuit impedance loss, resulting in serious errors in the overall SOH estimation. Most existing technologies focus only on the single estimation of SOH values, without linking abnormal SOH decay with early battery failures and failing to provide early warnings of potential thermal runaway risks such as micro-short circuits, loose connection circuits, and abnormal shedding of active materials. At the same time, most detection models are offline and cannot adapt to the aging characteristics changes throughout the battery's entire life cycle, resulting in a significant decrease in detection accuracy in the later stages of battery use. Summary of the Invention

[0003] The purpose of this invention is to provide a health status detection system and method for new energy vehicle battery packs to solve one or more problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting the health status of a new energy vehicle battery pack, comprising the following specific steps: Furthermore, the acquisition and calibration stage uses the multi-channel synchronous acquisition unit of the battery management system (BMS) to acquire the total voltage of the battery pack, the charging and discharging circuit current, the terminal voltage of each individual battery cell, the surface temperature of each individual cell, the bus connection point temperature, and the high-frequency impedance signal of the charging and discharging circuit in real time with nanosecond-level synchronization accuracy. The nanosecond-level synchronization accuracy is achieved by the high-precision constant-temperature clock chip and multi-channel parallel synchronous trigger circuit built into the acquisition unit. All voltage, current, temperature and impedance acquisition channels share the same set of clock reference signals. The synchronous trigger command is sent to each acquisition channel at the same time, ensuring that all parameters are synchronously acquired and converted at exactly the same time node, thus completely eliminating data errors caused by multi-channel timing offset and sampling asynchrony from the hardware level.

[0005] During the resting window after the vehicle is turned off, a low-disturbance DC pulse excitation is injected into the battery pack, the pulse response signal is collected, and the reference ohmic impedance of the current battery pack and the zero-point drift parameters of the acquisition system are calibrated through an adaptive zero-point drift calibration algorithm to complete the in-situ reference calibration of the detection system. Simultaneously perform electrical parameter consistency pre-detection, compare the reference impedance and open circuit voltage deviation of each individual cell, mark cells with initial consistency abnormalities, and eliminate initial measurement errors caused by acquisition system drift, environmental disturbances, and initial cell inconsistencies.

[0006] Furthermore, the denoising and enhancement stage identifies the operating conditions of the original electrical signals acquired during the acquisition and calibration stage. Based on three core dimensions—charge / discharge rate, battery SOC range, and ambient temperature—the real-time operating conditions are divided into three categories—static stationary operating conditions, low-rate stable operating conditions, and high-rate dynamic operating conditions—using a density clustering algorithm. For each type of operating condition, an adaptive denoising algorithm is matched. For the static stationary operating conditions, unscented Kalman filtering is used to remove baseline drift noise; for the low-rate stable operating conditions, wavelet thresholding is used for denoising; and for the high-rate dynamic operating conditions, variational mode decomposition and wavelet thresholding are used for joint denoising. The joint denoising algorithm is executed step-by-step in a fixed order: variational mode decomposition followed by wavelet threshold denoising. First, variational mode decomposition adaptively decomposes the complex electrical signal with violent fluctuations under high-rate dynamic operating conditions into multiple stable mode components, separating strong interference noise caused by sudden changes in operating conditions and load impacts from effective health feature components. Then, wavelet threshold denoising is performed on the retained effective components to further filter out residual high-frequency small noise. Under the premise of completely preserving the core features related to battery health, various interference noises under high-rate dynamic operating conditions are eliminated to the greatest extent.

[0007] For the denoised signal, the core electrical and coupling features related to battery health are extracted, including peak features of capacity increment IC curve, impulse response relaxation features, ohmic impedance features, polarization impedance features, single-cell voltage gradient features, temperature-internal resistance coupling features, loop impedance change rate, and single-cell voltage fluctuation coefficient. The health feature enhancement and redundant feature removal are completed by mutual information entropy algorithm, and features strongly correlated with SOH decay and electrical faults are retained.

[0008] Furthermore, the single-particle estimation stage, based on the high signal-to-noise ratio health features extracted in the denoising and enhancement stage, constructs a dual-drive estimation model that integrates electrochemical mechanisms and lightweight data-driven approaches. It adopts a design that combines parallel fusion with dynamic mechanism constraints to construct a simplified single-particle electrochemical model, coupled with a second-order equivalent circuit model, to simulate the core decay mechanisms of battery SEI film growth, active material shedding, and active lithium-ion loss. It sets a decay rate correlation factor to correlate the battery's real-time electrical parameters with the decay mechanism, generating a virtual sample dataset that includes the battery's entire life cycle, all operating conditions, and all temperature ranges. This completes the pre-training of the lightweight neural network model. The lightweight neural network adopts an attention mechanism-CNN-LSTM hybrid architecture, retaining only the core feature extraction layer and the fully connected layer. The simplified single-particle electrochemical model focuses on the three core processes inside the battery: lithium-ion diffusion, charge transfer, and active material decay, while ignoring non-critical electrochemical details, thus reducing computational complexity while ensuring simulation accuracy.

[0009] The second-order equivalent circuit model consists of ohmic internal resistance, polarization internal resistance, and polarization capacitance. It can simulate the voltage response and polarization characteristics during the charging and discharging process of a battery, and fully reflect the electrical performance of the internal electrochemical process of the battery.

[0010] The single-cell health features extracted during the denoising and enhancement stage are input into the pre-trained model. The initial single-cell SOH value output by the model is corrected in real time using unscented Kalman filtering. A dynamic mechanism constraint module is set up, which dynamically adjusts the mechanism constraint strength according to the real-time electrical state of the battery. The output results are physically constrained by the electrochemical mechanism model to eliminate abnormal estimates. At the same time, combined with the initial consistency abnormal cells marked in the collection and calibration stage, the SOH estimate is further calibrated to obtain the SOH baseline estimate value of each single cell battery.

[0011] For individual cells marked with initial consistency anomalies during the data acquisition and calibration phase, the system automatically retrieves historical data such as the cell's reference impedance, open-circuit voltage deviation, and initial anomaly level. Based on the conventional SOH estimation process, a dedicated anomaly compensation factor is added to directionally correct the estimation results according to the severity of the anomaly, eliminating the continuous impact of initial consistency deviation on the individual cell's SOH estimation. This ensures that the SOH reference estimation value of the anomaly cell always closely matches the true health status, avoiding distortion of the individual cell health assessment due to initial anomalies.

[0012] Furthermore, in the system correction phase, a battery pack system-level coupling attenuation model is constructed, and three core correction parameters are set: single-cell consistency weighting coefficient, loop impedance loss coefficient, and temperature gradient correction coefficient. The calculation mechanism and dynamic coupling logic of each parameter are clarified. Based on the SOH baseline estimate values ​​of each individual cell obtained in the individual cell estimation stage, and combined with the capacity, internal resistance, and voltage deviation of each individual cell, the individual cell consistency weight coefficient is calculated by the entropy weight method. Individual cells with poor consistency are given higher weights. At the same time, the initial consistency test results of the individual cells marked in the calibration stage are collected and the weight allocation is dynamically adjusted. Based on the loop impedance and bus temperature rise data collected during the calibration phase, and combined with the charge / discharge rate, the loop impedance loss coefficient is calculated using a linear regression algorithm to correct the capacity loss caused by the system loop contact resistance and wire loss. Based on the temperature distribution within the battery pack and the internal resistance temperature coefficient of each cell, a temperature gradient correction coefficient is calculated to correct the cell attenuation deviation caused by uneven temperature. A three-parameter dynamic coupling algorithm is set up to adjust the weights of the three correction parameters according to the real-time operating conditions of the battery. The SOH benchmark value of each cell is weighted and corrected through the three-parameter dynamic coupling algorithm to finally obtain the overall SOH estimate of the battery pack and generate the battery pack consistency health coefficient simultaneously.

[0013] Furthermore, in the fault tracing stage, a correlation mapping database is constructed between SOH decay anomaly and battery electrical faults. The correlation mapping database associates SOH decay rate with electrical fault characteristics and includes six major fault modes: abnormal SEI film growth, active material shedding, lithium ion deposition, single cell micro short circuit, loose connection circuit, and abnormal heat dissipation system. The electrical characteristic thresholds of each type of fault are clearly defined. The database is built based on a large amount of battery aging accelerated test data, real vehicle operation fault collection data, and after-sales fault disassembly verification data. It includes electrical characteristic samples of six typical faults under different temperature ranges, different charge and discharge rates, and different aging stages. The database has automatic incremental update capability. After each fault source is traced and verified to be effective, the characteristic data, attenuation law, and location results of this fault will be automatically entered into the database, and the fault characteristic threshold and matching rules will be optimized simultaneously.

[0014] Based on the single-cell SOH baseline value obtained in the single-cell estimation stage and the overall SOH estimate value obtained in the system correction stage, the SOH degradation rate of the battery pack is calculated. When the degradation rate exceeds the preset normal degradation threshold, the abnormal source is triggered. The core health features extracted in the noise reduction and enhancement stage are matched with the association mapping database. Combined with the dynamic changes of real-time electrical parameters of the battery, the cosine similarity algorithm is used to locate the root cause of abnormal degradation. Based on the fault type and the degree of abnormal degradation, battery health risks are divided into four levels: normal, attention, warning, and high risk, and corresponding fault warning signals and electrical fault location reports are generated.

[0015] Furthermore, the model iteration phase, based on the battery health status and fault characteristic data confirmed in the fault tracing phase, collects complete charging data during each complete constant current and constant voltage charging process of the vehicle, and calculates the actual usable capacity of the battery pack using the ampere-hour integration method, which serves as the true value for model accuracy verification; when the error between the overall SOH estimate obtained in the system correction phase and the SOH true value calculated from the actual capacity exceeds a preset threshold, the incremental learning update of the model is triggered. Based on the valid electrical data collected during this charging process, only the parameters of the fully connected layer of the lightweight neural network in the dual-drive model are incrementally fine-tuned; an aging state adaptive adjustment module is set up to dynamically adjust the learning rate of the neural network and the constraint strength of the mechanism constraint module according to the current aging degree of the battery, and at the same time, the iterative model parameters are fed back to the pre-training stage of the single-cell estimation stage.

[0016] Furthermore, the output control phase, based on the dual-drive estimation model and detection parameters optimized in the model iteration phase, constructs a multi-port hierarchical output architecture. It outputs corresponding detection results for the functional safety requirements of different application scenarios, with a focus on strengthening outputs related to electrical parameters. This includes outputting corrected overall SOH values, locations of consistency bottleneck cells, fault risk levels, charge / discharge power threshold correction parameters, loop impedance, and cell voltage consistency to the Battery Management System (BMS); outputting range correction values ​​and power output limit parameters corresponding to SOH to the Vehicle Controller Unit (VCU); and outputting battery health status, remaining usable life, and maintenance recommendations that users can understand to the onboard terminal. The reference impedance, SOH estimate, fault characteristic data, and model iteration parameters obtained from this test are fed back to the in-situ reference calibration stage in the data acquisition and calibration phase and the model pre-training stage in the individual unit estimation phase.

[0017] The present invention also provides a health status detection system for new energy vehicle battery packs, which, based on the above method, includes the following modules; The acquisition and calibration module integrates a multi-channel synchronous acquisition unit of the BMS to acquire key electrical parameters such as the total voltage of the battery pack and the charging and discharging circuit current in real time; it has a built-in low-disturbance pulse excitation unit and an adaptive zero-point drift calibration unit to complete the in-situ benchmark calibration after the battery is statically balanced, and simultaneously perform pre-detection of the consistency of individual electrical parameters and mark abnormal cells; The noise reduction and enhancement module has a built-in working condition recognition unit and an adaptive noise reduction unit. It divides three types of real-time working conditions and matches the corresponding noise reduction methods. It also integrates feature extraction and redundancy removal units to extract core features related to battery health. The single-cell estimation module includes a mechanism- and data-driven estimation model and a lightweight neural network unit to simulate the battery degradation mechanism and complete the model pre-training; it also has a built-in dynamic mechanism constraint and anomaly calibration unit to correct and calibrate the initial value of single-cell SOH and output the baseline estimated value of SOH for each single cell. The system correction module constructs a battery pack system-level coupled attenuation model, sets up three core correction parameter calculation units and a dynamic three-parameter dynamic coupling algorithm, and combines the data from the acquisition and calibration module to weighted correct the individual cell SOH benchmark value, outputting the overall SOH estimate and consistency health coefficient of the battery pack. The fault tracing module has a built-in database that maps SOH attenuation anomalies to battery electrical faults. It can calculate the SOH attenuation rate and trigger anomaly tracing, locate the root cause of the fault through feature matching, classify health risk levels, and generate fault warnings and location reports. The model iteration module integrates charging data acquisition and capacity calculation units, which can verify model accuracy and trigger incremental updates; it also has a built-in aging state adaptive adjustment unit that dynamically optimizes model parameters and feeds back the iteration parameters to the individual unit estimation module. The output control module constructs a multi-port hierarchical output architecture, outputting hierarchical output corresponding detection results and control parameters to the BMS, VCU and vehicle terminal; and feeding back the detection data to the acquisition and calibration module and the individual unit estimation module.

[0018] The beneficial effects of this invention are as follows: 1. This invention acquires key parameters such as voltage, current, temperature, and high-frequency impedance with nanosecond-level precision through a multi-channel synchronous acquisition unit of the BMS. During the vehicle's shutdown and resting window, a low-disturbance DC pulse excitation is injected to complete in-situ reference calibration. Simultaneously, the consistency of individual cells is pre-detected and abnormal cells are marked, eliminating measurement errors caused by acquisition system drift, environmental disturbances, and initial cell inconsistencies. For three types of operating conditions—static resting, low-magnification stable, and high-magnification dynamic—an adaptive matching denoising algorithm is used, combined with the mutual information entropy algorithm to extract healthy core features and remove redundant information, ensuring a high signal-to-noise ratio.

[0019] 2. This invention integrates a single-particle electrochemical model and a second-order equivalent circuit model in the single-cell estimation stage, and completes pre-training with an attention mechanism-CNN-LSTM lightweight network. It eliminates abnormal estimated values ​​through dynamic mechanism constraints to obtain the single-cell SOH baseline value. In the system correction stage, three types of correction coefficients are introduced: single-cell consistency, loop impedance loss, and temperature gradient. The entropy weight method is used to focus on the single-cell with consistency shortcomings, and the dynamic coupling algorithm is used for weighted correction to compensate for the estimation deviation caused by single-cell inconsistency, loop contact loss, and temperature unevenness, thereby improving the overall SOH estimation accuracy of the battery pack.

[0020] 3. This invention constructs a database that correlates SOH attenuation with electrical faults, including six typical fault modes. It locates the root cause of the fault by monitoring the attenuation rate and matching features, classifies the health risks into four levels and generates early warning and location reports, providing early warning of potential risks such as micro-short circuits and loose circuits. It calculates the actual capacity based on constant current and constant voltage charging data as the true value to verify accuracy. When the error exceeds the standard, it triggers incremental updates of the model, dynamically adjusting the learning rate and the strength of the mechanism constraints, so that the model can continuously adapt to the battery aging state. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the overall health status detection process of the battery pack according to the present invention. Figure 2 This is a flowchart of the SOH estimation and fault tracing process of this invention; Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figures 1 to 2 As shown in the figure, this invention provides a method for detecting the health status of a new energy vehicle battery pack, including the following specific steps: In this embodiment of the invention, the acquisition and calibration stage uses the multi-channel synchronous acquisition unit of the battery management system (BMS) to acquire the total voltage of the battery pack, the charging and discharging circuit current, the terminal voltage of each individual battery cell, the surface temperature of each individual cell, the bus connection point temperature, and the high-frequency impedance signal of the charging and discharging circuit in real time with nanosecond-level synchronization accuracy. The frequency range of the high-frequency impedance signal is 1kHz-1MHz. The high-frequency impedance signal is acquired using an independent acquisition channel, with the sampling rate set to a high-frequency sampling rate that meets the signal acquisition requirements. The acquisition channel is isolated from the voltage and current acquisition channels to avoid signal crosstalk.

[0024] The busbar connection temperature is collected at the solder joints connecting the positive and negative busbars of the battery pack to the individual cells. These solder joints are high-risk locations for circuit losses and poor contact. The collected temperature data is used to determine whether the connection status is abnormal.

[0025] The pulse excitation is uniformly injected into the main circuit of the battery pack, and no individual excitation is applied to any single cell throughout the process, ensuring that the internal electrical state of the battery pack is not affected by local disturbances, and the excitation signal is uniformly applied to the entire battery pack system. The pulse response signal is synchronously and in parallel acquired through the voltage acquisition channel and the high-frequency impedance acquisition channel of the battery management system. The two acquisition channels maintain nanosecond-level synchronization accuracy to avoid timing deviations that cause distortion of response data. After the dual-channel acquisition is completed, the data is fused and processed, and the reference impedance is calibrated by combining the voltage response and impedance response characteristics.

[0026] During the resting window after the vehicle is turned off, a low-disturbance DC pulse excitation with an amplitude ≤50mV and an adjustable frequency is injected into the battery pack. The adjustable frequency range is 10Hz-1kHz. The pulse response signal is collected, and the reference ohmic impedance of the current battery pack and the zero-point drift parameters of the acquisition system are calibrated through an adaptive zero-point drift calibration algorithm to complete the in-situ reference calibration of the detection system.

[0027] The adaptive zero-point drift calibration algorithm first acquires the baseline signal of the acquisition channel under no-excitation state, calculates the average offset and fluctuation range of the baseline signal, and then performs difference calculation by combining it with the reference signal of the pulse excitation response to automatically correct the zero-point offset error of the acquisition channel. During the calibration process, the influence coefficient of ambient temperature on acquisition drift is recorded simultaneously, and the drift compensation value is automatically updated in different temperature ranges to eliminate the hardware drift error of the acquisition system caused by long-term use.

[0028] The resting window period is determined by real-time monitoring of the voltage fluctuation amplitude and temperature change rate of individual cells by the battery management system. When the voltage fluctuation of individual cells is less than the set threshold and the internal temperature difference of the battery pack remains stable for 30 consecutive minutes, the battery is confirmed to have entered the electrochemical equilibrium state. The pulse excitation is injected into the main circuit of the battery pack through the excitation unit built into the battery management system. The injection process does not interrupt the low-voltage power supply of the vehicle and does not trigger the battery protection mechanism. The duration of the excitation signal is adaptively adjusted according to the battery pack capacity. When collecting the response signal, interference signals from other electrical components of the vehicle are simultaneously shielded.

[0029] Simultaneously perform electrical parameter consistency pre-test, compare the reference impedance and open circuit voltage deviation of each individual cell, mark the cells with initial consistency abnormalities, complete the in-situ reference calibration of the test system, and eliminate the initial measurement errors caused by data acquisition system drift, environmental disturbances and initial cell inconsistencies.

[0030] In this embodiment of the invention, the denoising and enhancement stage performs operating condition identification on the original electrical signals acquired in the acquisition and calibration stage. Based on three core dimensions—charge / discharge rate, battery SOC range, and ambient temperature—the real-time operating conditions are divided into three categories—static stationary operating conditions, low-rate stable operating conditions, and high-rate dynamic operating conditions—using a density clustering algorithm. For each type of operating condition, an adaptive denoising algorithm is matched. For the static stationary operating conditions, unscented Kalman filtering is used to remove baseline drift noise; for the low-rate stable operating conditions, wavelet thresholding is used for denoising; and for the high-rate dynamic operating conditions, variational mode decomposition and wavelet thresholding are used for joint denoising. The moment the density clustering algorithm completes the condition type determination, the corresponding denoising algorithm will be automatically triggered and started immediately. No manual intervention or manual switching of algorithm modes is required throughout the process. The denoising processing operation rhythm is completely synchronized with the real-time changes in battery condition. When the battery operating state switches rapidly between the three types of conditions, the algorithm will dynamically switch in real time according to the condition determination result, without delay, lag, or algorithm conflict. At the same time, the transition state of the condition boundary is smoothed to avoid signal processing distortion caused by frequent algorithm switching.

[0031] The density clustering algorithm sets clear thresholds based on actual operating data. When the charging and discharging current is close to zero and the SOC does not change significantly, it is judged as a static resting condition. When the charging and discharging rate is in the low range, the current fluctuation amplitude is small and the duration is stable, it is judged as a low-rate stable condition. When the charging and discharging rate is in the high range, the current changes frequently and the load changes rapidly, it is judged as a high-rate dynamic condition. The algorithm updates the condition clustering center in real time to adapt to the changes in operating conditions caused by different driving habits and road conditions.

[0032] For the denoised signal, the core electrical and coupling features related to battery health are extracted, including peak features of capacity increment IC curve, impulse response relaxation features, ohmic impedance features, polarization impedance features, single-cell voltage gradient features, temperature-internal resistance coupling features, loop impedance change rate, and single-cell voltage fluctuation coefficient. The health feature enhancement and redundant feature removal are completed by mutual information entropy algorithm, and features strongly correlated with SOH decay and electrical faults are retained.

[0033] The mutual information entropy algorithm calculates the correlation between each electrical feature and SOH attenuation and electrical faults, quantifies the effective contribution of features, eliminates redundant features with correlation below a set threshold, and amplifies the weight of highly correlated features to enhance feature expression. The algorithm automatically eliminates invalid features caused by operating condition interference and retains only the core features that can stably reflect battery aging and faults.

[0034] In this embodiment of the invention, the single-particle estimation stage is based on the high signal-to-noise ratio health features extracted in the denoising and enhancement stage. It constructs a dual-drive estimation model that integrates electrochemical mechanisms and lightweight data-driven approaches. It adopts a design that combines parallel fusion with dynamic mechanism constraints to construct a simplified single-particle electrochemical model, coupled with a second-order equivalent circuit model, to simulate the core decay mechanisms of battery SEI film growth, active material shedding, and active lithium-ion loss. It sets a decay rate correlation factor to correlate real-time battery electrical parameters such as impedance and voltage change rate with the decay mechanism, and generates a virtual sample dataset that includes the entire battery life cycle, all operating conditions, and the entire temperature range. It completes the pre-training of the lightweight neural network model. The lightweight neural network adopts an attention mechanism-CNN-LSTM hybrid architecture, retaining only the core feature extraction layer and the fully connected layer. Through model pruning and quantization optimization, the model is adapted to the vehicle edge computing scenario.

[0035] Pre-training follows a standardized process of data preprocessing, model initialization, batch training, error backpropagation, and verification convergence. The specific steps are as follows: Voltage, impedance, temperature, and multiplier data in the virtual sample set are subjected to min-max normalization to uniformly map them to the [0, 1] numerical range, eliminating dimensional differences; the batch size is set to 32, and a mini-batch gradient descent strategy is used to improve training stability; the initial learning rate is 1×10⁻⁶. -4 The Adam adaptive optimizer was selected, and the weight decay factor was set to 1×10. -5 To avoid parameter overfitting, the mean square error between the estimated SOH value of a single battery cell and the true label was used as the loss function, and the maximum number of training iterations was set to 500. A validation set test was performed every 20 iterations, and the model was considered to have converged when the loss function value was consistently below 0.001 and showed no significant decrease for 50 consecutive iterations. The dataset was divided into training and validation sets in an 8:2 ratio, including a temperature range of -20℃ to 60℃, a charge / discharge rate of 0 to 2C, and a SOC range of 0% to 100%, to ensure that the pre-trained model's generalization ability met the standards.

[0036] The hybrid architecture employs a topology of serial cascading and feature residual connections, built in a fixed order: input feature layer, single-head self-attention mechanism layer, 2D convolutional neural network layer, long short-term memory recurrent layer, global pooling layer, two-layer fully connected mapping layer, and SOH output layer. The attention mechanism layer is used to extract electrical features strongly correlated with battery health using weighted methods. The convolutional layer uses two 3×3 convolutional kernels, combined with ReLU nonlinear activation function and batch normalization layer to complete local feature extraction. The LSTM recurrent layer uses 64-dimensional hidden state units to capture the temporal decay features of battery electrical signals. The fully connected layer uses two linear transformation layers to map features to SOH values. Between layers, a Dropout layer randomly deactivates 20% of neurons to prevent overfitting. The model removes redundant neurons through structured pruning and compresses parameters using 8-bit fixed-point quantization.

[0037] The single-cell health features extracted during the denoising and enhancement stage are input into the pre-trained model. The initial single-cell SOH value output by the model is corrected in real time using unscented Kalman filtering. A dynamic mechanism constraint module is set up. The dynamic mechanism constraint module dynamically adjusts the mechanism constraint strength according to the real-time electrical state such as battery impedance change and voltage stability. The output results are physically constrained by the electrochemical mechanism model to eliminate abnormal estimates. At the same time, combined with the initial consistency abnormal cells marked in the collection and calibration stage, the SOH estimate is further calibrated to obtain the SOH baseline estimate value of each single cell battery.

[0038] The abnormal estimated values ​​include two typical cases: first, the estimated SOH value exceeds the normal theoretical range of batteries of the same model, capacity, and aging stage, which does not conform to the basic physical law of battery electrochemical degradation; second, the estimated SOH values ​​sampled at adjacent time points show a large jump, which does not conform to the actual characteristics of battery steady degradation. The dynamic mechanism constraint module will perform double verification on each set of SOH estimation results in real time, identify the above abnormal values, and automatically replace the abnormal values ​​with reasonable correction values ​​that conform to physical laws.

[0039] The dynamic mechanism constraint module sets physical boundary constraints based on the basic electrochemical characteristics of the battery. When the estimation result exceeds the normal electrical characteristics range of the battery, the constraint correction is automatically triggered. The constraint strength gradually increases as the battery impedance increases and the voltage fluctuation intensifies. The constraint process does not change the core structure of the model, but only corrects abnormal estimation values ​​to ensure compliance.

[0040] In this embodiment of the invention, the system correction stage constructs a battery pack system-level coupling attenuation model, sets three core correction parameters: single-cell consistency weighting coefficient, loop impedance loss coefficient, and temperature gradient correction coefficient, and clarifies the calculation mechanism and dynamic coupling logic of each parameter. Based on the baseline estimated SOH values ​​of each individual cell obtained in the individual cell estimation stage, and combined with the capacity, internal resistance, and voltage deviation of each individual cell, the individual cell consistency weight coefficient is calculated by the entropy weight method. The individual cells with the lowest SOH and the highest internal resistance are given higher weights. At the same time, the initial consistency test results of the individual cells marked in the calibration stage are collected and the weight allocation is dynamically adjusted. The entropy weighting method objectively assigns weights based on the electrical differences between individual battery pack cells. The input dimensions are three core features: the state of equilibrium (SOH) value of each cell, DC internal resistance, and terminal voltage deviation, avoiding subjective bias from manual weighting. The specific calculation process is as follows: a feature judgment matrix is ​​constructed, and the feature values ​​of each cell are standardized; the information entropy and difference coefficient of each feature are calculated, with higher weights for larger differences; initial weights are generated based on the difference coefficients, and a compensation mechanism is implemented to automatically increase the weight by 20% for cells with poor consistency (lowest SOH, highest internal resistance, and largest voltage deviation), strengthening the impact of these shortcomings on the overall health of the battery pack; all cell weights are normalized to ensure that the sum of the weights is strictly 1, with unbiased weighting correction; the weight results are synchronously associated with the initial abnormal cell markers collected during the calibration phase, dynamically adapting to changes in cell consistency.

[0041] Based on the loop impedance and bus temperature rise data collected during the calibration phase, and combined with the charge / discharge rate, the loop impedance loss coefficient is calculated using a linear regression algorithm to correct the capacity loss caused by the system loop contact resistance and wire loss. The circuit impedance loss coefficient is specifically used to correct the overall capacity and overall state of harmonics (SOH) of the battery pack. It mainly compensates for the capacity reduction caused by system-level losses such as poor busbar contact, main circuit wire resistance loss, and loose connection points causing overheating. The circuit impedance loss coefficient is dynamically adjusted according to the real-time circuit impedance and charge / discharge rate to compensate for the estimation deviation caused by circuit loss and to truly reflect the usable capacity and health status of the cells inside the battery pack.

[0042] Based on the temperature distribution within the battery pack and the temperature coefficient of the internal resistance of each cell, a temperature gradient correction coefficient is calculated to correct the cell degradation deviation caused by uneven temperature. A three-parameter dynamic coupling algorithm is set up to dynamically adjust the coupling weight of the three correction parameters according to real-time operating conditions such as battery charge / discharge state and ambient temperature. The three-parameter dynamic coupling algorithm is used to weight and correct the baseline value of the cell SOH, and finally obtain the overall SOH estimate of the battery pack. At the same time, a battery pack consistency health coefficient is generated, which improves the accuracy of system-level SOH detection. The consistency health coefficient is used to quantify the degree of balance of electrical parameters of each cell in the battery pack. The higher the coefficient, the better the consistency of the cells, and the lower the coefficient, the greater the difference between cells. The coefficient is calculated in combination with the cell SOH, internal resistance, and voltage deviation to directly reflect the overall balance state and service life potential of the battery pack.

[0043] The temperature gradient correction coefficient is calculated based on the real-time temperature distribution differences of each cell within the battery pack. For cells in high-temperature regions with faster aging and degradation rates, the SOH estimation weight is appropriately lowered, while for cells in low-temperature regions with slower aging and degradation rates, the SOH estimation weight is appropriately increased. By dynamically balancing the cell degradation differences caused by uneven temperature distribution through the coefficient, the overall health level and usable capacity of the battery pack under the current actual temperature environment are truly reflected.

[0044] The three-parameter dynamic coupling algorithm allocates coupling weights according to the priority of real-time operating conditions. During low-temperature, high-rate charging and discharging, the weights of the temperature gradient correction coefficient and the loop impedance loss coefficient are increased. Under static conditions, the weight of the single-cell consistency coefficient is increased. The weight adjustment process is automated without human intervention and is based entirely on the real-time electrical state of the battery. After the three parameters are coupled, a unified correction factor is formed to smoothly weight the single-cell SOH benchmark value, avoiding jumps in the estimation results caused by sudden weight changes.

[0045] In this embodiment of the invention, the fault tracing stage constructs a correlation mapping database between SOH decay anomaly and battery electrical faults. The correlation mapping database associates SOH decay rate with electrical fault characteristics, such as micro-short circuits corresponding to current fluctuations, loose circuits corresponding to impedance changes, and abnormal heat dissipation corresponding to temperature-internal resistance coupling anomalies. The database includes six major fault modes: abnormal SEI film growth, active material shedding, lithium ion deposition, single-cell micro-short circuits, loose connection circuits, and abnormal heat dissipation systems. The electrical characteristic thresholds for each type of fault are clearly defined. Based on the single-cell SOH baseline value obtained in the single-cell estimation stage and the overall SOH estimate value obtained in the system correction stage, the SOH degradation rate of the battery pack is calculated. When the degradation rate exceeds the preset normal degradation threshold, the abnormal source is triggered. The core health features extracted in the noise reduction and enhancement stage, such as the loop impedance change rate and single-cell voltage fluctuation coefficient, are matched with the associated mapping database. Combined with the dynamic changes of real-time electrical parameters such as battery voltage, current, and impedance, the cosine similarity algorithm is used to locate the root cause of abnormal degradation. The cosine similarity algorithm is customized for battery fault tracing scenarios. The input data consists of an 8-dimensional health feature vector extracted in real time and standard feature vectors of 6 typical faults in the database. The 8-dimensional health feature vector includes ohmic impedance, polarization impedance, single-cell voltage gradient, temperature-internal resistance coupling value, loop impedance change rate, voltage fluctuation coefficient, IC curve peak value, and pulse relaxation time. The algorithm measures the feature matching degree by calculating the cosine value of the angle between the two vectors. The calculation formula is cos(θ) = A·B / (||A||×||B||). The output result is the similarity value and the matched fault type. The judgment rule is: when the similarity is ≥0.85, it is judged as a highly matched fault; 0.70~0.85 is a suspected fault that needs continuous monitoring; and <0.70 is a fault with no matching fault features. After matching, it directly associates six major fault modes: abnormal growth of SEI film, active material shedding, lithium ion deposition, single-cell micro-short circuit, loose connection circuit, and abnormal heat dissipation system, thus completing the root cause location.

[0046] Based on the fault type and the degree of abnormal degradation, battery health risks are divided into four levels: normal, attention, warning, and high risk. Corresponding fault warning signals and electrical fault location reports are generated, realizing the transformation from SOH numerical estimation to electrical fault diagnosis and safety risk warning.

[0047] The four health risk levels correspond to tiered control actions. At the normal level, no warnings are triggered, and the system operates stably in its standard detection mode. At the monitoring level, high-frequency real-time monitoring is activated, continuously recording abnormal data and uploading it to the backend terminal to remind users to pay attention to changes in battery status. At the warning level, charging and discharging power and power output are automatically limited, and continuous reminders are issued through the vehicle terminal, advising users to perform battery checks and maintenance as soon as possible. At the high-risk level, the vehicle's safety protection mechanism is immediately activated, cutting off high-current charging and discharging output, issuing emergency audible and visual warnings, and preventing safety risks such as thermal runaway.

[0048] The fault warning signals are divided into two forms: audible and visual prompts and data commands. Visual warning prompts are output to the vehicle terminal, and fault control commands are output to the battery management system. Different risk levels correspond to different warning frequencies and prompting methods, with high-risk warnings being triggered first and locked for display.

[0049] Health risk levels are classified according to the SOH decay rate, fault characteristic intensity, and safety impact. A normal level is defined as a normal SOH decay rate with no fault characteristics; a level of concern is defined as a SOH decay rate slightly higher than the threshold with minor abnormal characteristics; a level of warning is defined as a SOH decay rate significantly exceeding the standard with clear fault characteristics; and a level of high risk is defined as a SOH decay rate rapidly exceeding the standard with thermal runaway-related fault characteristics. Different levels correspond to different warning methods and maintenance recommendations, and the level determination results are synchronized to the vehicle control system in real time.

[0050] In this embodiment of the invention, the model iteration stage is based on the battery health status and fault characteristic data confirmed in the fault tracing stage. During each complete constant current and constant voltage charging process of the vehicle, complete charging data is collected, and the actual usable capacity of the battery pack is calculated by the ampere-hour integration method as the true value for model accuracy verification. When the error between the overall SOH estimate obtained in the system correction stage and the SOH true value calculated by the actual capacity exceeds 3%, the incremental learning update of the model is triggered. The ampere-hour integration method is implemented during the complete constant current and constant voltage charging process. It continuously integrates the product of charging current and time from the start of charging to the end of charging, and simultaneously corrects the influence of temperature and current ratio on the integration result, eliminating the power consumed by the vehicle's low-voltage power consumption during the charging process. The effective capacity value is calculated only after the battery is fully charged and has reached equilibrium, avoiding calculation errors caused by incomplete charging or unstable voltage.

[0051] Based on the valid electrical data collected during this charging process, only the parameters of the fully connected layer of the lightweight neural network in the dual-drive model are incrementally fine-tuned; an aging state adaptive adjustment module is set up, which dynamically adjusts the learning rate of the neural network and the constraint strength of the mechanism constraint module according to the current aging degree of the battery, which is determined by the SOH benchmark value and the impedance change rate. At the same time, the iterative model parameters are fed back to the pre-training stage of the single-cell estimation stage.

[0052] After the parameters of the lightweight neural network fully connected layer are incrementally fine-tuned, the new parameters will take effect immediately in real time, directly replacing the original parameters of the model and permanently storing them in the non-volatile storage unit of the battery management system. The parameter update process adopts a hot loading mode, which does not require interruption of the detection system operation or restart of the vehicle control system. The next round of SOH estimation will automatically call the updated parameters, and the system will automatically back up the previous version of valid parameters. If an estimation error occurs after the update, it can be quickly rolled back.

[0053] The model iteration phase employs a full lifecycle adaptive incremental update mechanism. The learning rate and constraint strength are finely set in segments according to the battery aging degree. When the state of health (SOH) is 80%~100% in a brand new state, the neural network learning rate is set to 1×10. -4 ① The mechanistic constraint strength is set to medium to preserve the model's generalization ability; ② When the SOH is in a moderately aged state of 60%~80%, the learning rate is reduced to 5×10. -5 The mechanistic constraint strength is increased to medium-high, strengthening the electrochemical and physical constraints; when the state of steam hydration (SOH) is below 60% in the deep aging state, the learning rate is further reduced to 1×10. -5 The mechanistic constraint strength is set to the highest level to match the degradation characteristics after battery aging; incremental updates only fine-tune the parameters of the two fully connected layers of the lightweight neural network, and the single weight update amplitude is controlled within 5% to not change the main structure of the model; when the overall SOH estimation error exceeds 3%, an update is triggered, and the updated parameters are fed back to the pre-training stage of the individual estimation module in real time.

[0054] In this embodiment of the invention, the output control stage is based on the dual-drive estimation model and detection parameters optimized in the model iteration stage, and constructs a multi-port hierarchical output architecture. It outputs corresponding detection results for the functional safety requirements of different application scenarios, with a focus on strengthening the output related to electrical parameters. The BMS outputs corrected overall SOH values, locations of consistency bottleneck cells, fault risk levels, charge / discharge power threshold correction parameters, loop impedance, and cell voltage consistency, among other core electrical parameters. The VCU outputs the range correction value and power output limit parameters corresponding to the SOH to the vehicle controller. The vehicle terminal outputs user-understandable battery health status, remaining usable life, and maintenance recommendations. The multi-port hierarchical output architecture divides dedicated output ports according to the receiving object. Each port uses an independent communication protocol to transmit data, transmitting high-frequency core electrical and control parameters to the battery management system, periodic power control parameters to the vehicle controller, and visualized health data to the vehicle terminal. The output data is encrypted and transmitted according to the security level, and fault warning signals adopt a priority transmission mechanism to ensure that critical safety information is delivered to the corresponding control unit as soon as possible.

[0055] The reference impedance, SOH estimate, fault characteristic data, and model iteration parameters obtained from this test are fed back to the in-situ reference calibration stage in the data acquisition and calibration phase and the model pre-training stage in the individual unit estimation phase.

[0056] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0057] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting the health status of a new energy vehicle battery pack, characterized in that, The specific steps include the following: During the data acquisition and calibration phase, the key electrical parameters of the battery pack and each cell are collected in real time through the synchronous acquisition device of the battery management system. After the vehicle is turned off and the battery reaches electrochemical equilibrium, a low-disturbance DC pulse excitation is injected and the response signal is collected to complete the in-situ benchmark calibration of the detection system. At the same time, the consistency of the individual cell electrical parameters is pre-detected, and the initial abnormal cells are marked. In the noise reduction and enhancement stage, based on the collected original electrical signals and reference parameters, the real-time operating conditions of the battery are identified and an appropriate noise reduction method is matched. Core features related to battery health status and electrical faults are extracted and redundant information is removed. In the single-cell estimation stage, based on the denoised health characteristics, a dual-drive estimation model integrating electrochemical mechanism and lightweight data-driven approach is constructed. Combined with dynamic mechanism constraints and abnormal single-cell-specific calibration, the SOH benchmark estimation results of each single-cell battery are obtained. During the system correction phase, a battery pack system-level coupled attenuation model is constructed. Three core correction parameters are set: individual cell consistency weighting coefficient, loop impedance loss coefficient, and temperature gradient correction coefficient. Through a three-parameter dynamic coupling algorithm, the individual cell SOH is weighted and corrected by combining the individual cell consistency weighting coefficient, loop impedance loss coefficient, and temperature gradient correction coefficient to obtain the overall SOH estimate and consistency health coefficient of the battery pack. During the fault tracing phase, based on the overall SOH estimation results of the battery pack and individual cells, the SOH decay rate is calculated and the anomaly tracing is triggered. Through the SOH decay anomaly and electrical fault association mapping database, the cosine similarity algorithm is used for matching to locate the root cause of the abnormal decay, classify the battery health risk level, and generate fault warning signals and location reports. During the model iteration phase, the actual usable capacity of the battery is calculated based on the data collected during the constant current and constant voltage charging process as the accuracy verification benchmark. When the estimation error exceeds the standard, the model incremental update is triggered, and the model parameters are dynamically adjusted in combination with the battery aging status. In the output control phase, a multi-port hierarchical output architecture is constructed to output corresponding detection results and control parameters to the battery management system, vehicle controller and vehicle terminal, while the detection data is fed back to the acquisition and calibration phase and the pre-training phase of the single-unit estimation model.

2. The method for detecting the health status of a new energy vehicle battery pack according to claim 1, characterized in that, During the data acquisition and calibration phase, the battery management system (BMS) uses a multi-channel synchronous acquisition unit to acquire the total voltage of the battery pack, the charging and discharging circuit current, the terminal voltage of each individual battery cell, the temperature of each individual cell and busbar, and the high-frequency impedance signal of the circuit with nanosecond-level synchronous accuracy. During the resting window after the vehicle is turned off, a low-disturbance DC pulse excitation is injected. Through an adaptive zero-point drift calibration algorithm, the reference ohmic impedance of the battery pack and the zero-point drift parameters of the acquisition system are calibrated to complete the in-situ reference calibration. Simultaneously, the reference impedance and open-circuit voltage deviation of each cell are compared to complete the consistency pre-detection and mark the initial abnormal cells, thereby eliminating the measurement errors caused by the drift of the acquisition system, environmental disturbances and the inconsistency of the initial cells.

3. The method for detecting the health status of a new energy vehicle battery pack according to claim 2, characterized in that, The denoising and enhancement stage is based on three core dimensions: charge / discharge rate, battery SOC range, and ambient temperature. It uses a density clustering algorithm to divide the real-time operating conditions into three categories: static idle condition, low-rate stable condition, and high-rate dynamic condition. The corresponding denoising algorithm is adaptively matched for each condition. The core electrical and coupling features related to battery health are extracted from the denoised signal. The mutual information entropy algorithm is used to enhance the health features and remove redundant features, retaining features that are strongly correlated with SOH decay and electrical faults.

4. The method for detecting the health status of a new energy vehicle battery pack according to claim 3, characterized in that, The single-cell estimation stage, based on the denoised high signal-to-noise ratio health characteristics, constructs a parallel dual-drive estimation model that integrates electrochemical mechanisms and lightweight data-driven approaches. By coupling a simplified single-particle electrochemical model with a second-order equivalent circuit model, it simulates the core degradation mechanism of the battery and generates a virtual sample dataset of the entire life cycle, completing the pre-training of the lightweight neural network. The single-cell health characteristics are input into the pre-trained model, and the initial SOH value is corrected in real time through unscented Kalman filtering. Anomaly estimates are eliminated by combining a dynamic mechanism constraint module, and the SOH estimates of the initial abnormal cells are specifically calibrated to finally obtain the baseline SOH estimates for each single cell.

5. The method for detecting the health status of a new energy vehicle battery pack according to claim 4, characterized in that, The system correction phase is based on the SOH baseline estimate of each individual cell. The consistency weight coefficient of each cell is calculated by the entropy weight method. The cells with poor consistency are given higher weight and dynamically adjusted in combination with the initial consistency test results. Based on the loop impedance, busbar temperature rise data and charge / discharge rate, the loop impedance loss coefficient is calculated to correct the capacity loss caused by the system loop. Based on the battery pack temperature distribution and the temperature coefficient of the internal resistance of each cell, a temperature gradient correction coefficient is calculated to correct the degradation deviation caused by temperature unevenness. Through a three-parameter dynamic coupling algorithm, the coupling weights of the three correction parameters are adjusted according to the real-time operating conditions to complete the weighted correction of the single cell SOH, and the overall SOH estimate and consistency health coefficient of the battery pack are obtained.

6. The method for detecting the health status of a new energy vehicle battery pack according to claim 5, characterized in that, The fault tracing stage pre-constructs a correlation mapping database between SOH attenuation anomalies and battery electrical faults. The database associates SOH attenuation rate with electrical fault characteristics, including core battery attenuation and electrical fault modes, and sets corresponding feature thresholds. When the battery SOH attenuation rate exceeds the preset normal threshold, anomaly tracing is triggered. The extracted core health features are matched with the database, and the cosine similarity algorithm is used to locate the root cause of the abnormal attenuation. Based on the fault type and the degree of abnormal degradation, battery health risks are divided into four levels, generating corresponding fault warning signals and electrical fault location reports.

7. The method for detecting the health status of a new energy vehicle battery pack according to claim 6, characterized in that, During the model iteration phase, the actual usable capacity of the battery pack under full charge is calculated using the ampere-hour integration method during each complete constant current and constant voltage charging process of the vehicle, and this value is used as the true value for verifying the model accuracy. When the error between the estimated SOH value of the battery pack and the true value exceeds a preset threshold, the model is triggered to perform incremental updates, and the parameters of the fully connected layer of the lightweight neural network are incrementally fine-tuned based on the effective data of this charging. An aging state adaptive adjustment module is set up to dynamically adjust the neural network learning rate and mechanism constraint strength according to the current aging degree of the battery, and simultaneously feed the iterated model parameters back to the model pre-training stage.

8. The method for detecting the health status of a new energy vehicle battery pack according to claim 7, characterized in that, In the output control phase, a multi-port hierarchical output architecture is constructed to meet the functional safety requirements of different application scenarios: the corrected overall SOH value, the location of the inconsistency bottleneck cell, the fault risk level, the charge and discharge power threshold correction parameters, and the core electrical parameters are output to the battery management system (BMS); the range correction value and power output limit parameters are output to the vehicle controller (VCU); the battery health status, remaining usable life, and maintenance suggestions are output to the vehicle terminal; and the benchmark data and model iteration parameters of this test are simultaneously fed back to the in-situ benchmark calibration and model pre-training phase.

9. A health status detection system for a new energy vehicle battery pack, based on the method described in any one of claims 1-8, characterized in that, Includes the following modules: The acquisition and calibration module integrates a multi-channel synchronous acquisition unit of the BMS to acquire key electrical parameters of the battery pack and individual cells in real time. It has a built-in low-disturbance pulse excitation unit and an adaptive zero-point drift calibration unit to complete the in-situ benchmark calibration after the battery is placed at rest, pre-detection of the consistency of individual electrical parameters, and initial abnormal individual cell marking. The noise reduction and enhancement module has a built-in working condition recognition unit and an adaptive noise reduction unit. It divides the real-time working condition of the battery and matches the corresponding noise reduction method. It integrates feature extraction and redundancy removal units to extract core features related to battery health and remove redundant information. The single-cell estimation module has a built-in dual-drive estimation model of electrochemical mechanism and data-driven, a lightweight neural network unit, simulates the battery degradation mechanism and completes model pre-training, sets dynamic mechanism constraints and anomaly calibration unit, corrects the initial value of single-cell SOH and outputs the baseline estimated value of SOH for each single-cell. The system correction module constructs a battery pack system-level coupled attenuation model, sets up three core correction parameter calculation units and a three-parameter dynamic coupling algorithm unit, weighted corrects the individual cell SOH benchmark value, and outputs the overall SOH estimate and consistency health coefficient of the battery pack. The fault tracing module has a built-in database that maps SOH attenuation anomalies to battery electrical faults. It calculates the SOH attenuation rate and triggers anomaly tracing, locates the root cause of the fault through feature matching, classifies health risk levels, and generates fault warnings and location reports. The model iteration module integrates charging data acquisition and capacity calculation units, verifies model accuracy and triggers incremental updates, and has a built-in aging state adaptive adjustment unit to dynamically optimize model parameters and feed them back to the individual unit estimation module. The output control module constructs a multi-port hierarchical output architecture, outputting detection results and control parameters to the corresponding vehicle control unit and terminal, and simultaneously feeding back the detection data to the acquisition and calibration module and the individual unit estimation module.

Images