Electric vehicle battery health evaluation method and system based on multi-source data fusion

Abstract

The application discloses a battery health evaluation method and system for an electric vehicle based on multi-source data fusion, and relates to the technical field of new energy vehicle battery applications. The method comprises the following steps: collecting Nyquist impedance curves, terminal voltage relaxation time series data and equalization current time series data based on constant current stage cutoff events of a battery pack of an electric vehicle; extracting impedance pattern embedding vectors corresponding to the Nyquist impedance curves based on a self-encoder; performing wavelet packet decomposition on the terminal voltage relaxation time series data to obtain corresponding multi-scale voltage relaxation energy feature vectors; calculating a dynamic equalization imbalance index based on the terminal voltage relaxation time series data and the equalization current time series data; and inputting the impedance pattern embedding vectors, the multi-scale voltage relaxation energy feature vectors and the dynamic equalization imbalance index into a battery SOH prediction model to output an SOH estimation value of the battery pack. Thus, the accuracy of the SOH evaluation result is improved by jointly modeling the coupling characteristics of the multiple physical processes of the battery.

Description

Technical Field

[0001] This application relates to the field of new energy vehicle battery application technology, and in particular to a method and system for assessing the health of electric vehicle batteries based on multi-source data fusion. Background Technology

[0002] As a crucial component of new energy transportation, the safety, reliability, and lifespan characteristics of electric vehicle power battery systems have become a key focus of current research and industry. State of Health (SOH) assessment of power batteries is one of the key technologies for realizing the functions of intelligent Battery Management Systems (BMS). Accurate SOH assessment helps optimize charge and discharge control strategies, extend service life, and reduce operating and maintenance costs.

[0003] Currently, mainstream battery health assessment methods typically rely on a single data source or indicator for modeling and judgment. For example, some methods construct criteria based on frequency domain measurement information, reflecting the internal state of the battery by acquiring its response characteristics under specific operating conditions; other methods utilize voltage and current change curves during charging and discharging to perform static or dynamic modeling to infer capacity decay and performance degradation.

[0004] However, current battery health assessment methods primarily focus on a single physical process or signal dimension. The extracted features are often limited to single-valued quantities at fixed frequencies or are based on parameter estimations from simplified fitting models, making it difficult to comprehensively reflect the complex time-frequency coupled evolution process within the battery. Furthermore, with increasing usage time, there is often a degradation in consistency between different cells within the battery module. The modeling process fails to effectively consider the dynamic differences and regulatory behaviors between cells, resulting in decreased accuracy of SOH assessment results when facing early imbalance states. Summary of the Invention

[0005] This application provides a method, system, storage medium, computer program product, and electronic device for assessing the health of electric vehicle batteries based on multi-source data fusion. This is intended to at least address the problem in current related technologies that rely on a single signal dimension and cannot accurately reveal the evolution mechanism of the mutual coupling of multiple physical fields during battery aging, leading to biases in SOH assessment results.

[0006] In a first aspect, embodiments of this application provide a method for assessing the health of electric vehicle batteries based on multi-source data fusion, comprising: simultaneously triggering electrochemical impedance spectroscopy scanning, terminal voltage relaxation recording, and battery pack equalization operations based on the constant current phase cutoff event of the battery pack for an electric vehicle, to obtain corresponding Nyquist impedance curves, terminal voltage relaxation time-series data, and equalization current time-series data, respectively; the terminal voltage relaxation time-series data includes multiple sampling times and corresponding terminal voltage values, and the equalization current time-series data includes multiple sampling times and corresponding equalization current values; extracting the impedance morphology embedding vector corresponding to the Nyquist impedance curve based on an autoencoder; performing wavelet packet decomposition on the terminal voltage relaxation time-series data to obtain corresponding multi-scale voltage relaxation energy feature vectors; and calculating the dynamic equalization misalignment index based on the terminal voltage relaxation time-series data and the equalization current time-series data.

[0007]

[0008] In the formula, M bal (t) is the dynamic equilibrium imbalance index; Δt represents the sampling time interval, σ V (t) represents the standard deviation of the terminal voltages of all individual cells at sampling time t. Let ΔI be the regularization constant. bal (t) represents the equalization current increment, indicating the difference in equalization current between two adjacent sampling times t-Δt and t.

[0009] The impedance morphology embedding vector, the multi-scale voltage relaxation energy characteristic vector, and the dynamic equilibrium misalignment index are input into the battery SOH prediction model to output the SOH estimate of the battery pack.

[0010] Secondly, embodiments of this application provide an electric vehicle battery health assessment system based on multi-source data fusion, comprising: a multi-source data acquisition unit, used to simultaneously trigger electrochemical impedance spectroscopy scanning, terminal voltage relaxation recording, and battery pack equalization operation based on the constant current stage cutoff event of the battery pack for an electric vehicle, to obtain corresponding Nyquist impedance curves, terminal voltage relaxation time-series data, and equalization current time-series data respectively; the terminal voltage relaxation time-series data includes multiple sampling times and corresponding terminal voltage values, and the equalization current time-series data includes multiple sampling times and corresponding equalization current values; an impedance morphology encoding unit, used to extract the impedance morphology embedding vector corresponding to the Nyquist impedance curve based on an autoencoder; a voltage relaxation multi-scale extraction unit, used to perform wavelet packet decomposition on the terminal voltage relaxation time-series data to obtain corresponding multi-scale voltage relaxation energy feature vectors; and a dynamic equalization misalignment index calculation unit, used to calculate the dynamic equalization misalignment index based on the terminal voltage relaxation time-series data and the equalization current time-series data.

[0011]

[0012] In the formula, M bal (t) is the dynamic equilibrium imbalance index; Δt represents the sampling time interval, σ V (t) represents the standard deviation of the terminal voltages of all individual cells at sampling time t. Let ΔI be the regularization constant. bal (t) represents the equalization current increment, indicating the difference in equalization current between two adjacent sampling times t-Δt and t.

[0013] The battery pack SOH prediction unit is used to input the impedance morphology embedding vector, the multi-scale voltage relaxation energy characteristic vector, and the dynamic equilibrium misalignment index into the battery SOH prediction model to output the SOH estimate of the battery pack.

[0014] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the electric vehicle battery health assessment method based on multi-source data fusion according to any embodiment of this application.

[0015] Fourthly, embodiments of this application provide a storage medium storing a computer program thereon, characterized in that, when the program is executed by a processor, it implements the steps of the electric vehicle battery health assessment method based on multi-source data fusion according to any embodiment of this application.

[0016] Fifthly, embodiments of this application provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the electric vehicle battery health assessment method based on multi-source data fusion according to any embodiment of this application.

[0017] The electric vehicle battery health assessment method and system based on multi-source data fusion provided in this application can achieve at least the following technical effects:

[0018] (1) By simultaneously triggering electrochemical impedance spectroscopy scanning, terminal voltage relaxation, and equalization current measurement, the internal state of the battery was comprehensively characterized from three dimensions: time-frequency coupling, energy distribution, and cell consistency. On the one hand, the Nyquist impedance morphology embedding vector obtained by the autoencoder can capture the subtle changes in the charge transfer and diffusion processes at the battery interface, making the characterization of the internal electrochemical processes more discriminative. On the other hand, the multi-scale voltage relaxation energy features extracted by wavelet packet decomposition can reflect the polarization and recovery kinetics of the battery at different time scales, improving the model's sensitivity to signals in the early stages of capacity decay. Combined with the dynamic equalization misalignment index for real-time quantification of voltage imbalance and equalization behavior between cells, the early imbalance risk caused by consistency degradation can be accurately identified. The fusion of multi-source features enables the SOH prediction model to maintain high accuracy and robustness at different degradation stages, significantly improving the early degradation diagnosis capability and the assessment accuracy under imbalance conditions, thus providing a more reliable decision-making basis for intelligent BMS.

[0019] (2) By introducing a dynamic balance imbalance index, the standard deviation of the terminal voltage of all cells in the battery pack and the increment of the equalization current are quantified in real time, which quantitatively describes the difference in the cells during the equalization process. It can effectively identify the consistency degradation phenomenon inside the module and achieve accurate identification of the imbalance state and equalization behavior between cells. In particular, it can identify and adjust the imbalance state in the early stage of battery pack use, further improving the accuracy and reliability of the overall SOH assessment.

[0020] (3) The three types of time-frequency coupling, multi-energy distribution and single-cell consistency assessment features, namely impedance morphology embedding vector, multi-scale voltage relaxation energy characteristics and dynamic equilibrium imbalance index, are fused into the SOH prediction model to make the assessment results maintain high accuracy and robustness in each degradation stage and improve the accuracy of SOH assessment in the unbalanced state of the battery pack.

[0021] This technical solution achieves joint modeling of the coupling characteristics of multiple physical processes in the battery by simultaneously collecting electrochemical impedance spectroscopy, terminal voltage relaxation data, and battery pack equalization current data, and by integrating frequency domain response characteristics, time domain relaxation characteristics, and module consistency status. This enables a more comprehensive and accurate characterization of the battery's health status. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1A flowchart illustrating an example of an electric vehicle battery health assessment method based on multi-source data fusion according to an embodiment of this application is shown.

[0024] Figure 2 A schematic diagram showing the experimental comparison simulation results of an example of dynamic SOH estimation and static SOH estimation based on dynamic equilibrium misalignment index according to an embodiment of this application is provided.

[0025] Figure 3 This document illustrates an example of an operation flowchart for synchronously triggering multi-source data acquisition based on a constant current phase cutoff event, according to an embodiment of this application.

[0026] Figure 4 A schematic diagram showing an example of the encoder structure of a self-encoder according to an embodiment of this application is provided.

[0027] Figure 5 A schematic diagram of the structural connections of an example battery SOH prediction model according to an embodiment of this application is shown;

[0028] Figure 6 A structural block diagram of an example of an electric vehicle battery health assessment system based on multi-source data fusion according to an embodiment of this application is shown;

[0029] Figure 7 This is a schematic diagram of the structure of an embodiment of the electronic device of this application. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0031] It should be noted that current battery health assessment methods mainly include analytical methods based on electrochemical impedance spectroscopy (EIS) and assessment methods based on static fitting of charge-discharge curves. In EIS-based methods, the internal electrochemical state of the battery is reflected by measuring its AC impedance at different frequencies. Typically, the impedance magnitude or phase angle at a specific frequency point of 1Hz or 10Hz is selected as a characteristic indicator for SOH estimation. However, this method lacks comprehensive consideration of the overall morphological changes of the spectral curve and struggles to distinguish subtle curve changes caused by different degradation mechanisms. Furthermore, single-point characteristics exhibit poor stability in the presence of measurement noise, easily leading to assessment bias.

[0032] In charge-discharge curve analysis-based methods, the voltage relaxation curve after charge-discharge cutoff is fitted, and the time constant is extracted using an exponential or double exponential function for SOH estimation. However, the voltage relaxation process is essentially a coupling effect of multiple physical processes (such as charge transfer, concentration polarization, and diffusion), and its decay characteristics exhibit significant multi-timescale features. Currently used simple function models cannot effectively decouple the decay components at different timescales, resulting in insufficient sensitivity for detecting early, minor degradation. Furthermore, static fitting methods struggle to capture the nonlinear characteristics of dynamic voltage changes, limiting their adaptability under complex operating conditions.

[0033] Furthermore, battery modules are composed of multiple cells connected in series and parallel. During use, due to manufacturing differences, inconsistent aging rates, and other reasons, voltage inconsistencies often occur between cells. BMS systems typically adjust cell voltages through active or passive balancing strategies, but most existing health assessment methods ignore the dynamic information of balancing current and cell voltage difference, thus failing to identify early signs of internal consistency degradation in a timely manner, affecting the sensitivity and accuracy of SOH assessment.

[0034] It should be understood that the above description of the relevant technologies is intended only to help the public better understand the inventive spirit and motivation of this application, and is not intended to limit this application. Furthermore, the technical solutions described in the above-mentioned relevant technologies are not prior art, and may also be undisclosed technical solutions, such as those under research or in the laboratory stage.

[0035] The technical solutions in this application, including the collection, storage, use, processing, transmission, provision, and disclosure of users' personal information, comply with relevant laws and regulations and do not violate public order and good morals.

[0036] Figure 1 A flowchart illustrating an example of an electric vehicle battery health assessment method based on multi-source data fusion according to an embodiment of this application is shown.

[0037] It should be noted that static characteristics only reflect the overall state of the battery pack at the end of a single charge-discharge cycle, and cannot capture the details of instantaneous voltage differences between individual cells and the lag in the equalization current response during the cycle. When the capacity or internal resistance of a few cells begins to deviate due to manufacturing differences or localized damage, this deviation will be partially smoothed out by the equalization circuit after the cycle ends, making the static characteristics appear to be within the normal range, but masking the risks that have already begun to accumulate internally. Furthermore, as the number of cycles increases, the differences in capacity and internal resistance among the cells within the battery pack will gradually accumulate, and the real-time compensation speed of the equalization circuit varies. This dynamic imbalance is a precursor to potential early failures, but it is often overlooked.

[0038] Regarding the execution subject of the method in this application embodiment, it can be any controller or processor with computing or processing capabilities. Specifically, it can be implemented by a BMS controller or a SOH evaluation component. It adopts a data-driven and physical feature fusion approach for feature extraction and modeling. For the first time, it proposes to incorporate the dynamic balance imbalance index into the SOH evaluation, thereby including the inconsistency of individual cells within the group and their real-time balance response in the model. The constructed health evaluation model has good generalization ability and anti-interference ability, and is applicable to power battery systems of different models and under different operating conditions.

[0039] In some examples, it can be integrated into an electronic device or terminal through software, hardware, or a combination of both, and the type of terminal or electronic device can be diverse, such as mobile phones, tablets, or desktop computers, etc.

[0040] like Figure 1 As shown, in step S110, electrochemical impedance spectroscopy scanning, terminal voltage relaxation recording, and battery pack equalization operation are synchronously triggered based on the constant current stage cutoff event of the battery pack for electric vehicles, so as to obtain the corresponding Nyquist impedance curve, terminal voltage relaxation time series data, and equalization current time series data, respectively.

[0041] Here, the terminal voltage relaxation timing data includes multiple sampling times and corresponding terminal voltage values, and the equalization current timing data includes multiple sampling times and corresponding equalization current values. For example, the terminal voltage value and equalization current value are synchronously triggered based on the same sampling timestamp.

[0042] The constant current phase cutoff event refers to a state transition event triggered by the system during the constant current charging and discharging process of the battery pack when a preset cutoff condition is reached (such as the individual cell voltage reaching the upper limit threshold, the total charging capacity reaching the threshold, or the cumulative time exceeding the limit). This event marks the termination of the constant current phase and triggers subsequent operations (such as switching to the constant voltage phase, starting data acquisition, etc.).

[0043] For example, the BMS system issues a constant current phase end trigger signal when any of the following preset conditions are met:

[0044] Termination voltage condition: The battery pack terminal voltage rises to or falls to the preset charge / discharge termination voltage;

[0045] Capacity threshold condition: The cumulative charge and discharge capacity reaches the set fixed capacity threshold;

[0046] Time constraint: The duration of constant current charging and discharging reaches the pre-defined maximum time;

[0047] SOC / SOH safety threshold conditions: The state of charge (SOC) or state of harm (SOH) of the battery pack reaches the safety threshold that requires switching the control strategy.

[0048] When any of the above conditions are met, the BMS system immediately interrupts the constant current control mode, switches to the terminal voltage relaxation measurement and equalization operation mode, and simultaneously triggers an electrochemical impedance spectroscopy scan to ensure the timing alignment of various signals.

[0049] It should be noted that after the constant current phase ends, the charge distribution inside the battery has reached the specified bias level; then it enters a constant voltage or open-circuit relaxation state, at which point the system's dynamic response slows down significantly, and the electrochemical process becomes more controllable. Using this boundary point as the triggering time, it is possible to obtain multi-source signals that contain both the characteristics of the end of charge and discharge (polarization effect under high current) and the intrinsic recovery kinetics in the subsequent static relaxation phase.

[0050] More specifically, electrochemical impedance spectroscopy (EIS) measurements must be performed under steady-state conditions or with a known bias current to accurately reflect the battery interface and diffusion processes. Around the time the constant current cycle ends, the bias current has already reached the set constant current value. Applying a scanning perturbation voltage immediately after triggering ensures a clear and consistent initial reference for EIS measurements, avoiding data distortion caused by activating EIS during dynamic charge-discharge processes.

[0051] The voltage relaxation curve reflects the recovery kinetics from constant current cutoff to electrochemical equilibrium. Starting from the constant current cutoff stage, the entire relaxation process from the maximum polarization value can be systematically recorded, and the energy distribution characteristics under different time constants can be extracted through subsequent multi-scale analysis, which greatly improves the sensitivity to early capacity decay and polarization evolution.

[0052] Furthermore, BMS systems typically perform equalization compensation on each cell after the constant current charging / discharging cycle ends or after entering constant voltage mode. Switching to the equalization circuit at this time best reflects the differences in the state of the cells within the group. Triggering the equalization current measurement based on this event allows for real-time quantification of the BMS system's response to imbalance conditions, providing a direct comparison with the degree of voltage imbalance and generating a highly correlated dynamic misalignment index.

[0053] In some implementations, the battery pack current state is monitored in real time within the BMS system. When the constant current charging / discharging process is detected to be nearing termination (e.g., reaching a set termination voltage condition or capacity threshold condition), or via a constant current mode switching interrupt signal (e.g., generated by the charge / discharge controller), this moment is identified as a "constant current phase cutoff event," and the corresponding absolute timestamp t0 is recorded. Furthermore, at the end of the constant current charging / discharging phase of the electric vehicle battery pack, a trigger command is synchronously issued by the BMS control unit: on the one hand, the electrochemical impedance spectroscopy (EIS) scanner is activated, applying a small sinusoidal AC perturbation voltage and recording the current response; on the other hand, at the same moment, the terminal voltage relaxation measurement begins, and simultaneously, the equalization operation sampling of the battery pack equalization circuit is initiated. To ensure the timing alignment of the three signals, the BMS controller can pre-calibrate the clocks of each measurement subsystem and issue a "synchronous sampling" command via hardware interrupt or a dedicated bus (such as CAN or LIN).

[0054] Specifically, the EIS scanner applies a ±10mV peak-to-peak sinusoidal AC voltage, covering frequencies of 10mHz, 100mHz, 1Hz, 10Hz, 100Hz, and 1kHz. The sampling time for each frequency is approximately 100ms, with a total scanning time of <1s. Specifically, the scanning hardware is integrated into the BMS daughterboard and can run in parallel with subsequent relaxation recordings. The scanning results are recorded in real-time as t0+Δt. i (Δt i (Time stamp for startup delay at each frequency point)

[0055] The voltage relaxation measurement module records the terminal voltage of all cells according to the sampling frequency. Specifically, starting from time t0, it records the terminal voltage sequence V(t) at a sampling rate of 100Hz. The default recording duration is 300s, which can be adjusted according to the diffusion time constant of the battery material.

[0056] The equalization current measurement unit applies a constant voltage bypass or bidirectional equalization circuit by triggering the equalization module switch, and records the equalization current waveform through a shunt resistor and a high-precision ADC with the same sampling timing as the relaxation measurement.

[0057] Thus, multidimensional data is collected under a unified event-driven approach, with high time alignment, enabling a cross-domain synchronous data collection mechanism for the three types of data, which provides a foundation for accurate feature fusion and modeling.

[0058] In step S120, the impedance morphology embedding vector corresponding to the Nyquist impedance curve is extracted based on the autoencoder.

[0059] Here, the impedance morphology embedding vector can be a low-dimensional numerical representation of a fixed length, which can be compressed from the "shape information" of a Nyquist impedance curve and used by subsequent models. The impedance morphology embedding vector reflects the electrochemical impedance characteristics of the battery pack in its current state.

[0060] It should be noted that on the Nyquist plot, different degradation mechanisms will cause the curve to exhibit different features such as semicircular arcs, depressions, overlapping double semicircles, or the tail end of a Warburg line. We collectively refer to these geometric structures as "morphology." The encoder part of the autoencoder "converges" the morphological information scattered on the curve at different frequency bands, thereby obtaining the impedance morphology embedding vector. Here, each component of the impedance morphology embedding vector does not directly correspond to the impedance value at a certain frequency point, but rather integrates "geometric" features learned at different scales by multiple convolutions, such as inflection points, depression depths, semicircle diameters, and tail slopes, to more effectively provide morphological discriminative features.

[0061] Furthermore, the autoencoder comprises an encoder and a decoder. The encoder maps the Nyquist curve to a low-dimensional latent space (i.e., an impedance morphological embedding vector), while the decoder attempts to reconstruct the original curve to ensure that the extracted features have sufficient information fidelity. It should be noted that the decoder is used to optimize the encoder and is applied only during the training phase of the autoencoder; it can also be configured to have a mirror-symmetric structure with the encoder.

[0062] It should be noted that in traditional methods, using only single-point impedance values ​​or equivalent circuit parameters easily overlooks information about the overall morphological changes of the curve, resulting in insufficient ability to distinguish different degradation mechanisms. In contrast, in the embodiments of this application, during the process of compressing and reconstructing the impedance curve by the autoencoder, the key physical mechanisms such as interface charge transfer, electrolyte diffusion, and electrode aging are automatically extracted and comprehensively characterized in the frequency domain by morphological discrimination, enabling the embedded vector to have both denoising and high-dimensional discrimination capabilities.

[0063] In step S130, wavelet packet decomposition is performed on the terminal voltage relaxation time series data to obtain the corresponding multi-scale voltage relaxation energy feature vector.

[0064] The terminal voltage relaxation curve typically encompasses a dynamic process lasting from several seconds to several minutes from the end of constant current to the equilibrium stage. To capture the characteristics at different kinetic scales, wavelet packet decomposition is introduced for the terminal voltage relaxation time series data to extract multi-scale voltage relaxation energy feature vectors, which reflect the dynamic changes of the battery at different time scales.

[0065] In some implementations, the terminal voltage sequence sampled at a uniform interval can be used as the input signal to ensure that the sampling frequency covers both the rapidly changing and slowly changing voltage segments. Furthermore, based on the non-stationary characteristics of the voltage curve, a tightly supported wavelet basis (such as the Daubechies wavelet) is selected, and a 3-5 layer wavelet packet decomposition structure is set to achieve comprehensive capture of high-frequency and low-frequency information. By calculating the energy values ​​in each sub-band as feature vector elements, a multi-scale energy distribution map describing the voltage relaxation process is formed.

[0066] In this embodiment, by employing wavelet packet multi-scale decomposition, the energy distribution of both short-term polarization effects and long-term diffusion recovery processes can be simultaneously reflected. This allows for the separate quantification of dynamic characteristics under different time constants, enabling the model to perceive the evolutionary trajectories of three mechanisms: early charge accumulation, surface polarization, and deep diffusion. Compared to single time windows or Fourier transforms, the wavelet packet method exhibits superior time-frequency localization capabilities, significantly enhancing its ability to capture the differences in capacity decay curves at various degradation stages.

[0067] In step S140, the dynamic balance offset index is calculated based on the terminal voltage relaxation timing data and the equalization current timing data.

[0068] In this embodiment, a dynamic balance imbalance index is introduced as a measure of consistency degradation within the module, aiming to infer the trend of uneven distribution of health status from the differences between cells and balance behavior.

[0069] In some implementations, firstly, the standard deviation of multiple individual cell voltage values ​​is calculated to reflect the degree of imbalance within the battery pack. Then, the incremental change between two equalization current samplings is calculated to characterize the strength of the equalization action. Finally, the standard deviation and current increment are combined for ratio calculation, and regularization is added to suppress the influence of extreme values. Furthermore, a sliding window method is used for continuous updating, and a reasonable calculation period (e.g., 100ms) and time window width (e.g., 1s) are set in the controller. This allows the exponential curve to smoothly display the dynamic balance process of voltage dispersion and current adjustment. The calculated dynamic imbalance index can be fed back to the BMS decision module in real time to determine whether to trigger equalization-side compensation or an early warning.

[0070] More specifically, the dynamic equilibrium imbalance index is calculated using the following formula:

[0071]

[0072] In the formula, M bal (t) represents the dynamic equilibrium misalignment index, indicating the inconsistency index within the battery pack calculated at sampling time t. It characterizes the ratio of individual cell terminal voltage differences to equilibrium current fluctuations. Δt represents the sampling time interval, and σ... V(t) represents the standard deviation of the terminal voltages of all individual cells at sampling time t. Let ΔI be the regularization constant. bal (t) represents the equalization current increment, indicating the difference in equalization current between two adjacent sampling times t-Δt and t.

[0073] Here, the dynamic balance imbalance index integrates both information on the dispersion of individual cell voltage and the intensity of balance current regulation. It can reflect the degree of imbalance and measure the response efficiency of the BMS to the imbalance. In the early stage of battery consistency degradation, the index change is more discriminative than single voltage fluctuations or current changes, and can provide early warning of potential imbalance risks several cycles in advance.

[0074] Figure 2 A schematic diagram showing the experimental comparison simulation results of an example of dynamic SOH estimation and static SOH estimation based on the dynamic equilibrium misalignment index according to an embodiment of this application is illustrated.

[0075] like Figure 2 As shown, both the static and dynamic SOH estimation curves show a downward trend, but in the critical range (50-100 cycles), the dynamic curve significantly leads the static curve in reflecting the accelerated decline in health.

[0076] exist Figure 2 In the 50th cycle at point A, the dynamic curve has dropped to 95%, while the static curve remains at 98%. At this point, the dynamic imbalance index rises due to increased individual inconsistency, allowing the fusion model to capture weak degradation signals in early cycles, while the static method, which relies solely on impedance or relaxation characteristics, has not yet shown any change.

[0077] exist Figure 2 In the 100th cycle at point B, the dynamic SOH had further decreased to 90%, while the static curve only began to decline rapidly from 95%. The dynamic model consistently retained its sensitivity to inconsistency fluctuations, while the static model only made a significant response to degradation until the end of the cycle, because it only evaluated the "steady state" characteristics at the end of the cycle and could not reflect the accumulation of fluctuations during the cycle in a timely manner.

[0078] Furthermore, referring to Figure 2 In the shaded region (i.e., the 50th to 100th cycles), the difference between the SOH estimates of the two curves gradually widens, reaching a maximum of about 5%. This is precisely the result of the dynamic model continuously adjusting the SOH prediction slope through the combined effect of the real-time equilibrium misalignment index and static features. The dynamic SOH consistently leads within this range, providing a longer lead time for maintenance and equilibrium strategy optimization. Therefore, through the embodiments of this application, the dynamic equilibrium misalignment index is deeply integrated with static impedance and relaxation features, preserving both static macroscopic decay information and supplementing real-time feedback of dynamic consistency changes.

[0079] In some examples of embodiments of this application, the dynamic imbalance index also helps to optimize the balancing strategy. More specifically, when the index shows a significant upward trend, the compensation intensity is increased, and after the decline stabilizes, the balancing switching frequency can be reduced, thereby reducing system power consumption and balancing individual unit lifetimes.

[0080] For example, the dynamic imbalance index is compared with a preset imbalance threshold, and when the dynamic imbalance index exceeds the imbalance threshold, an alarm operation for battery cell inconsistency fault is triggered.

[0081] It should be noted that when the dynamic equilibrium imbalance index M bal (t) exceeds the preset threshold M th When an alarm is triggered, it indicates that the voltage difference between individual cells within the battery pack is too large, and the balancing current response rate is insufficient to quickly bridge this difference. Cell inconsistency faults can encompass various scenarios, such as uneven capacity degradation, abnormal increase in internal resistance, balancing circuit failure, or performance degradation. Upon alarm activation, the system simultaneously writes the fault cycle number, timestamp, and relevant exponential curve data to the log for subsequent fault location and analysis. Therefore, by using a comprehensive, quantitative dynamic balancing misalignment index compared to a strict threshold, an alarm can be issued when cell inconsistency just exceeds the normal fluctuation range, making it more sensitive and accurate than traditional methods relying solely on voltage or temperature difference limits.

[0082] In step S150, the impedance morphology embedding vector, the multi-scale voltage relaxation energy characteristic vector, and the dynamic equilibrium misalignment index are input into the battery SOH prediction model to output the SOH estimate of the battery pack.

[0083] In some implementations, a fusion-structured neural network model (such as a Transformer or LSTM-CNN hybrid structure) can be used to jointly learn static and dynamic features and output the current SOH estimate. Specifically, the three types of features mentioned above—impedance morphological embedding, multi-scale relaxation energy vector, and dynamic misalignment index—are concatenated in a temporal or parallel manner and then input into the SOH prediction model to comprehensively capture nonlinear and temporal dependencies.

[0084] Training the battery SOH prediction model can involve constructing an experimental sample library of battery packs with different cycle counts, temperatures, and load conditions, and labeling the SOH values ​​based on actual capacity decay or authoritative test data. Overfitting is avoided during training through cross-validation, hyperparameter search, and regularization techniques. During online inference, the model first standardizes the input features, then outputs the SOH value, and the BMS automatically adjusts the charging and discharging strategy or issues maintenance prompts based on the prediction results.

[0085] Here, by jointly learning multi-source features representing different physical mechanisms in the same model, a comprehensive perception of battery SOH can be achieved. The battery SOH prediction model can not only identify the aging state of electrodes and electrolytes based on morphological impedance characteristics, but also determine the capacity decay process by combining relaxation energy spectrum, and compensate for consistency deviations by relying on the offset index, reducing interference from single-signal noise, improving the model's generalization ability to complex degradation modes, and enabling SOH estimation to maintain high accuracy (error less than ±2%) and high robustness (significantly reduced standard deviation) under extreme operating conditions and imbalance conditions, providing reliable decision support for intelligent BMS.

[0086] Figure 3 A flowchart illustrating an example of multi-source data acquisition synchronously triggered based on a constant current phase cutoff event, according to an embodiment of this application, is shown.

[0087] To ensure that the data collected during subsequent EIS scanning, voltage relaxation, and equalization operations all reflect the same "battery pack completes constant current charging and discharging and enters open-circuit resting" state, a predetermined time point is needed to trigger it uniformly. This time point is not necessarily the sampling time itself, but can also be the event time confirmed by the current judgment logic.

[0088] like Figure 3 As shown, in step S310, the charging and discharging current is sampled in real time at the shunt resistor of the battery pack, and the sampled signal is processed by first-order low-pass filtering and moving average to obtain a smooth current curve.

[0089] Here, high-frequency noise caused by electromagnetic interference and current pulsation is effectively suppressed through first-order low-pass filtering and moving average processing.

[0090] In step S320, the current curve is smoothed. With preset threshold I thr When comparing, And the duration is not less than the dejitter time t deb When the current reaches a certain value, the constant current phase is considered to have ended.

[0091] In some implementations, I thr It is recommended to set it to 2% of the rated current or a fixed 0.5A to balance false alarms and response speed, and it can also be dynamically adjusted through production calibration or online learning. The dejitter time can be selected as 10-50ms, which filters out abnormal glitches of tens of milliseconds while ensuring that the roll-off trend within tens of milliseconds is captured. When all the above conditions are met, the physical state corresponding to this judgment period is identified as the "constant current stage cutoff and entry into open circuit rest" constant current stage cutoff event.

[0092] In step S330, if the constant current stage is determined to be cut off, a hardware trigger message is generated and broadcast through the vehicle communication bus to synchronously trigger electrochemical impedance spectroscopy scanning, terminal voltage relaxation recording and battery pack equalization operation. The hardware trigger message includes a cycle number and a timestamp of the sampling time.

[0093] Once the event is confirmed, a hardware trigger message is immediately generated and broadcast via the vehicle communication bus to notify downstream modules to initiate EIS scanning, terminal voltage relaxation timing data recording, and battery pack balancing operations in parallel. The "cycle number N" and "sampling time timestamp t0" carried in the hardware trigger message serve as key indexes to ensure that the three parallel data acquisition channels can be matched one-to-one at the backend according to the same cycle and the same point of state, eliminating data alignment ambiguity.

[0094] Here, by combining the hardware interrupt signal and the vehicle bus message broadcast dual-channel triggering mechanism, not only is the triggering time alignment error controlled to below the millisecond level, but the cyclic number and precise timestamp can also be synchronously transmitted to each acquisition module, ensuring that multi-source data is collected consistently under the same physical state.

[0095] The signal optimization and dual-channel triggering fusion scheme proposed in the embodiments of this application can significantly improve the accuracy and stability of the constant current stage cutoff event identification compared with the traditional method of directly determining the original current threshold.

[0096] Figure 4 A schematic diagram showing an example of the encoder structure of a self-encoder according to an embodiment of this application is provided.

[0097] It should be noted that in battery health assessment, Nyquist impedance spectroscopy contains multiple mechanistic information such as electrochemical interfaces, diffusion processes, and material internal resistance. However, traditional single-point impedance values ​​or equivalent circuit parameters are insufficient to fully characterize the overall morphological changes. Therefore, in this embodiment, an encoder structure with an autoencoder is introduced to map the geometric shape of the impedance curve to a low-dimensional embedding vector, providing high-dimensional and distinguishable morphological features for multi-source data fusion.

[0098] like Figure 4 As shown, the encoder structure 400 of the autoencoder includes a cascaded input preprocessing layer 410, a multi-scale dilated convolutional layer 420, a channel attention pooling layer 430, and a fully connected dimensionality reduction layer 440.

[0099] The input preprocessing layer 410 is used to perform the following operations:

[0100] The obtained discrete impedance spectrum point set The Nyquist curve is obtained through interpolation and arc length parameterization, with the arc length set as follows:

[0101]

[0102] Divide the Nyquist curve into a resampling point sequence P according to equal arc length intervals L / N. j =(x j ,y j ), where L is the total arc length of the Nyquist curve, N is the total number of resampling points, j = 1,...,N; and ω i This represents the frequency of the i-th sinusoidal excitation signal. and The measured excitation frequency ω i The real and imaginary parts of the electrochemical impedance of the battery pack are given, where K represents the total number of discrete frequency points collected from the Nyquist curve, τ is the curve parameter, and s(t) represents the geometric arc length from the curve start point to the parameter value t; P j Let x represent the Nyquist plane coordinates of the j-th resampled point. j Corresponding to the real part, y j Corresponding to the imaginary part.

[0103] It should be noted that the discrete impedance spectrum point set is obtained through EIS scanning. For example, at the i-th excitation frequency ω... i Next, a small voltage excitation is applied to the battery port, and the corresponding AC voltage or current response is measured. After lock-in amplification or fast Fourier transform, the real part of the impedance is obtained. and the virtual part A total of K such frequency pairs were collected, forming a discrete impedance spectrum point set.

[0104] The impedance spectrum points obtained from the original EIS scan are unevenly distributed on the frequency or impedance plane, and the curves are sparse due to the limitations of the number of measurement points and the scanning strategy. In the input preprocessing layer 410, the discrete points are padded into a geometrically equidistant sequence by curve arc length parameterization and equidistant resampling, and then normalized to eliminate the scaling differences of the original data.

[0105] After stacking the resampling points row by row, the input matrix H is obtained. (0) Each row of the matrix is ​​(x j ,y j ).

[0106] The multi-scale dilated convolutional layer 420 is used to perform the following operations:

[0107] Input matrix H (0) Input three layers of one-dimensional dilated convolutions, each layer is...

[0108]

[0109] Iteratively obtained

[0110] Among them, H (l) This represents an intermediate feature of the l-th layer. This indicates that the expansion rate d is 2. l One-dimensional convolution operation, b (l) Here, C represents the bias of the l-th layer, and C is the number of channels.

[0111] It should be noted that different mechanistic features on the Nyquist curve manifest as semicircles and depressions of varying scales. Single-scale convolution cannot simultaneously capture both minute depressions and large semicircles; therefore, dilated convolution is employed to create "hollow" receptive fields with different dilation rates, thus balancing local and global morphology. Specifically, the first layer has a dilation rate of 1 to capture the most minute single-point depressions, the second layer has a dilation rate of 2 to connect wider features between two sampling points, and the third layer has a dilation rate of 4 to expand to a larger area (such as covering the entire semicircular structure). This establishes a multi-scale receptive field, automatically learning microscopic and macroscopic morphological features.

[0112] Channel attention pooling layer 430 is used to perform the following operations:

[0113] For H (3) Each row vector Calculate attention weights

[0114]

[0115] In the formula, α j This represents the attention weight corresponding to the j-th resampling point. Represents the attention mapping matrix. This is the attention weight vector. The denominator is... This represents the normalization term, used to obtain α through softmax normalization. j .

[0116] It should be noted that different sampling locations contribute differently to health status—low-frequency semi-circular depressions are often better at distinguishing aging types, but ordinary global pooling cannot differentiate them. Based on a channel attention mechanism, the network learns location weights to focus on the most discriminative local features.

[0117] Calculate the global morphological feature vector:

[0118]

[0119] Therefore, by dynamically adjusting feature aggregation and suppressing irrelevant noise, it is robust to sudden measurement errors. By visualizing attention weights, it is possible to explain which frequency bands and which curve regions contribute the most to the SOH evaluation, making the fused vector z more discriminative.

[0120] Fully connected dimensionality reduction layer 440 is used to perform the following operations:

[0121] The global morphological feature vector z is mapped through a two-level fully connected layer to form an output impedance morphological embedding vector.

[0122]

[0123] In the formula, Represents the first-level dimensionality reduction mapping matrix, Represents the first layer bias vector, Represents the second-level dimensionality reduction mapping matrix, This represents the second-layer bias vector, where D is the dimension of the impedance shape embedding vector, and E represents the impedance shape embedding vector.

[0124] Here, the high-dimensional aggregated feature z is mapped to the preset embedding space D, which facilitates subsequent splicing and unified fusion with other modal features. The two-level mapping is used to realize feature fusion and dimensionality compression respectively, thereby retaining sufficient nonlinear expressive power and ensuring good differentiation of different degradation modes.

[0125] Through the embodiments of this application, uniform arc length resampling is performed on the Nyquist curve constructed from the discrete impedance spectrum point set, and multi-scale dilated convolution combined with attention pooling is used to enable the embedding vector to accurately capture the morphological differences of various degradation mechanisms on the Nyquist curve, thereby capturing more discriminative impedance morphological features.

[0126] It should be noted that the decoder structure in an autoencoder can be diverse and can be mirror-symmetric to the encoder structure. For example, the decoder structure consists of three one-dimensional transposed convolutional layers symmetric to the encoder, with the number of channels and kernel size of each layer mirroring the corresponding encoder layer. Furthermore, the activation function (ReLU) and batch normalization are used in the same way as the encoder, and the last layer linearly maps back to the original N×2 Nyquist sequence. During autoencoder training, the reconstruction error (MSE) plus curvature loss can be used as the overall loss to optimize encoding performance.

[0127] Regarding the explanation of battery voltage relaxation, it should be noted that battery voltage relaxation is the result of a superposition of multiple stages and physical processes. After discharge or charging, the voltage first undergoes rapid reconstruction of the double-layer capacitance (typically with time constants in the millisecond range), followed by slower processes such as liquid-phase diffusion (in the second range) and solid-phase (solid-state) diffusion (in the tens to hundreds of seconds range). Traditional single-exponential or double-exponential fitting models can only roughly separate the two time constants, failing to distinguish more mechanisms and making it difficult to capture subtle differences in the early stages of degradation.

[0128] Compared to analysis in the time or frequency domain alone, the wavelet packet decomposition (WPD) method employed in this embodiment provides localized multi-scale decomposition in the time-frequency plane. This preserves temporal location information and distinguishes frequency components at different resolutions, which helps to decompose voltage relaxation sequences containing multiple time-scale features in the signal.

[0129] More specifically, the obtained discrete sequence of terminal voltage relaxation By applying the bandpass filter operator BP{·} and subtracting the final steady-state voltage V[M-1], the correction sequence is obtained:

[0130] ΔV[n]=BP{V[n]}-V[M-1], equation (7)

[0131] Where ΔV[n] is the corrected voltage attenuation value, representing the voltage relaxation amount after removing the steady-state baseline at the nth sampling time; V[n] is the voltage value measured at the nth sampling time; BP{·} is the bandpass filter operator with a preset cutoff frequency; M is the total number of discrete sampling times in the relaxation record; and V[M-1] is the steady-state voltage, representing the voltage baseline value corresponding to the last sampling point of the sequence.

[0132] Here, the digital bandpass filter BP{·} can typically be a second-order Butterworth or Chebyshev II, with the low cutoff frequency set to 0.05Hz (to filter out DC drift) and the high cutoff frequency set to 5Hz (to suppress high-frequency measurement noise). The filter design employs bidirectional zero-phase filtering to avoid phase distortion.

[0133] Furthermore, using the last sample V[M-1] of the sequence as the steady-state voltage baseline, the DC component is accurately subtracted through ΔV[n], ensuring that ΔV[n] contains only the relaxation decay component, and subsequent decomposition is unaffected by drift. Thus, removing DC drift avoids energy statistical shift and suppresses high-frequency abrupt changes, enhancing decomposition stability.

[0134] Set ΔV[n] as the initial approximation coefficient:

[0135] a (0) [n]=ΔV[n], n=0,...,M-1, Formula (8)

[0136] For a (0) [n] Perform U-layer wavelet packet decomposition and recursively obtain the approximation coefficients of the v-th subband at the (r+1)-th layer. and detail coefficient

[0137]

[0138] Where r is the decomposition layer index, r = 0, ..., U-1, U is the total number of decomposition layers, and v = 1, ..., 2 r is the subband index of the r-th layer; h[p] and g[p] are the low-pass filter coefficients and high-pass filter coefficients of the Db4 mother wavelet, respectively, and p is the filter coefficient index;

[0139] For the set of all approximation coefficients of the U-th layer and detail coefficient set Calculate the time-domain energy separately:

[0140]

[0141] In the formula, N v Let be the subband coefficient length of the v-th subband coefficient in the U-th layer.

[0142] Here, by statistically analyzing the energy of each sub-band coefficient, a set of multi-scale energy values ​​is obtained, which directly reflects the relative intensity and temporal evolution characteristics of different physical processes throughout the relaxation process, providing a quantitative and highly differentiated input.

[0143] More specifically, DaubechiesDb4 is chosen because its fourth-order support length is reasonable, and it possesses sufficient orthogonality and time-frequency localization characteristics, making it suitable for oscillating attenuated signals. The filter coefficients h[p] and g[p] are derived from the Daubechies definition and can be pre-calculated and hard-coded.

[0144] Furthermore, the number of decomposition layers is selected based on the total signal duration and the target frequency resolution, and can be 4 or 6 layers. Higher layers result in finer frequency band division and the ability to distinguish more degradation mechanisms, but the computational load and feature dimensions increase exponentially. Each layer simultaneously performs filtering and downsampling on the sub-band coefficients (approximation and detail), resulting in 2... U Each sub-band corresponds to a narrow frequency band.

[0145] Furthermore, the approximation coefficients and detail coefficients of the U-th layer decomposition output total 2. U Each sub-band is statistically analyzed independently, and the sum of squares is performed on all coefficients within each sub-band. This results in a large energy difference, which can affect model convergence. As a result, the energy value directly corresponds to the signal strength of different frequency bands, making it easy to compare and sort. High-frequency sub-bands naturally have low noise energy, while low-frequency sub-bands carry the main attenuation information. Energy statistics inherently suppress noise.

[0146] Approximating energies {Q} a,v} and detail energy {Q q,v Arranged in ascending order by sub-band index, these form multi-scale energy feature vectors.

[0147]

[0148] In the formula, e is the multi-scale voltage relaxation energy eigenvector, used to characterize the relaxation decay process at different time scales. U This represents the total number of subbands after decomposition.

[0149] Here, by arranging the multi-scale energy values ​​into a fixed-dimensional vector, they can be directly used as input to deep fusion networks or time-series regression models, which helps to simplify the data preprocessing process.

[0150] Compared to traditional exponential fitting, which can only distinguish two stages of attenuation, the embodiments of this application employ wavelet packet decomposition, which can perform up to 2... L The segment time-frequency decomposition refines the energy of each frequency band. By performing fine time-frequency decomposition and energy quantization on the terminal voltage relaxation signal, it accurately distinguishes various complex coupling mechanisms such as double layer, electrolyte diffusion, and solid-phase diffusion, thereby improving the distinguishability and robustness of multi-physical degradation processes.

[0151] Figure 5 A schematic diagram of the structural connections of an example battery SOH prediction model according to an embodiment of this application is shown.

[0152] It should be noted that although impedance morphology embedding, relaxation energy multi-scale decomposition, and dynamic equilibrium misalignment index together constitute a comprehensive characterization of the multi-mechanism degradation within the battery, they are essentially high-dimensional features or a set of temporal features, and cannot be directly used to guide a single numerical output on SOH. Therefore, a battery SOH prediction model is used to analyze these multimodal, high-dimensional comprehensive features to obtain the final SOH index output.

[0153] like Figure 5 As shown, the battery SOH prediction model 500 includes a static feature encoding module 510, a modulation coefficient generation module 520, a dynamic feature time series extraction module 530, a time series pooling module 540, and a regression prediction module 550.

[0154] The static feature encoding module 510 is used to concatenate the impedance shape embedding vector E with the multi-scale voltage relaxation energy feature vector e to form the static joint feature S. raw =[E;e]; Subsequently, the static feature encoding module uses a multi-layer fully connected network and a self-attention mechanism to process S raw Perform nonlinear mapping and reweighting to output the context vector S.

[0155] Here, the static feature encoding module is responsible for fusing two essentially independent, high-dimensional static information categories—"electrochemical morphology" and "relaxation energy"—into a contextual representation that can discriminate battery degradation. The impedance morphology embedding vector E captures the geometric morphological features of the EIS curve, such as the size of the semicircle, the low-frequency dip, and the Warburg line, while the multi-scale voltage relaxation energy feature vector e separates the energy distribution of the double layer, liquid phase, and solid-phase diffusion processes in the time-frequency domain. Since directly concatenating these two elements can easily lead to excessively high dimensionality and difficulty in balancing their weights, a nonlinear mapping and attention mechanism are used to automatically learn the cooperative and mutually exclusive relationships between different components, compressing and reweighting them into a fixed-dimensional context vector S. This effectively compresses the high-dimensional static features and automatically assigns higher weights to key channels through the self-attention module, improving sensitivity to subtle differences.

[0156] Modulation coefficient generation module 520 is used to map the context vector S into the parameter mapping network, and through at least two fully connected layers, generate a set of feature modulation coefficients {γ} corresponding to each dilated convolutional layer of the dynamic feature temporal extraction module. f ,β f |f=1,...,G};where G represents the total number of dilated convolutional layers, γ f and β f Let γ represent the channel scaling factor vector and channel offset factor vector of the f-th layer, respectively. f ,β f Used to modulate the channel activation of the f-th dilated convolutional layer.

[0157] Here, inspired by the concept of Feature-by-Feature Linear Modulation (FiLM), a modulation coefficient generation module bridges the static context S with the dynamic temporal network, allowing static features to "guide" each layer of dynamic feature extraction, i.e., through γ. f ,β f The channel activations of the f-th temporal convolution are linearly scaled and shifted element-wise to achieve cross-modal information transfer and modulation. In this way, the dynamic temporal network is no longer a "blindly calculated" equilibrium misalignment exponential sequence, but rather dynamically adjusts the attention of each convolutional layer to changes at different time scales based on the context of the current electrochemical morphology and diffusion characteristics.

[0158] Thus, the impedance morphology embedding vector and the multi-scale voltage relaxation energy feature together construct a high-dimensional contextual representation of the electrochemical interface and diffusion process in the static coding module. By organically integrating the influence of static information on the dynamic extraction of features at each layer, the model can fundamentally distinguish different degradation mechanisms.

[0159] The dynamic feature time series extraction module 530 is used to receive the dynamic equilibrium imbalance index sequence M. bal(t), and then extracts temporal features layer by layer through G-layer dilated convolutional units, calculating at each layer.

[0160]

[0161] And closely follow the feature modulation

[0162]

[0163] Among them, h (f) (t) is the activation output of the f-th layer, representing the feature vector after convolution and activation at sampling time t, h (0) (t)=M bal (t); This indicates that the expansion rate d is 2. f-1 One-dimensional convolution operation; ⊙ represents element-wise multiplication operation. The activation value after modulation of the f-th layer represents the output feature vector of this layer after context-driven channel-level modulation at sampling time t.

[0164] It should be noted that the dynamic equilibrium imbalance index M bal (t) is a high-frequency time-series signal, whose degradation information is distributed across short-term abrupt changes, periodic fluctuations, and long-term trends. A Temporal Convolutional Network (TCN) is chosen instead of a Recurrent Neural Network (RNN) to enable parallel computation and capture different time dependencies through exponential dilation. Combined with a modulation coefficient module, context-driven linear modulation of channel activations is applied after each convolutional layer, allowing the model to emphasize responses at different time scales in different cyclic spheres.

[0165] Therefore, based on the context-driven channel-level modulation mechanism, the response of the dynamically balanced imbalance exponential sequence in each layer of the temporal convolutional network is precisely amplified or suppressed, ensuring that the model is sensitive to key inconsistency fluctuations rather than being smoothed out in general.

[0166] The temporal pooling module 540 is used to process the modulated feature sequence output by the dynamic feature temporal extraction module. Perform global average pooling along the time dimension to generate a pooling vector P.

[0167] It should be understood that the output of TCN is still a sequence that evolves over time. In order to integrate it with the static context and feed it into the regression network, the temporal features need to be summarized into a fixed-length vector. Global average pooling (or attention pooling) can highlight the overall trend of the sequence while ensuring time invariance.

[0168] The regression prediction module 550 is used to concatenate the pooling vector P with the context vector S to form a fused feature.

[0169] [P;S], and outputs the SOH estimate of the battery pack through a multi-layer fully connected regression network.

[0170] The regression prediction module 550 takes the fused static context S and the temporal pooling vector P as inputs. Through deep nonlinear mapping, it maps the multi-source feature space onto the SOH estimation scalar. Through multi-layer nonlinear regression mapping, while preserving the richness of the multi-source information, it effectively suppresses overfitting and improves generalization ability through hierarchical feature fusion and regularization.

[0171] Through the embodiments of this application, deep encoding and fusion of multi-source heterogeneous features are introduced to realize multi-angle, multi-scale and dynamic adaptive analysis of battery health status, effectively enhancing the accuracy and reliability of SOH assessment results.

[0172] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of combined actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Secondly, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application. In the above embodiments, the descriptions of each embodiment have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0173] Figure 6 A structural block diagram of an example of an electric vehicle battery health assessment system based on multi-source data fusion according to an embodiment of this application is shown.

[0174] like Figure 6 As shown, the electric vehicle battery health assessment system 600 based on multi-source data fusion includes a multi-source data acquisition unit 610, an impedance morphology encoding unit 620, a voltage relaxation multi-scale extraction unit 630, a dynamic balance misalignment index calculation unit 640, and a battery pack SOH prediction unit 650.

[0175] The multi-source data acquisition unit 610 is used to synchronously trigger electrochemical impedance spectroscopy scanning, terminal voltage relaxation recording, and battery pack equalization operation based on the constant current stage cutoff event of the battery pack for electric vehicles, so as to obtain the corresponding Nyquist impedance curve, terminal voltage relaxation time series data, and equalization current time series data respectively; the terminal voltage relaxation time series data includes multiple sampling times and corresponding terminal voltage values, and the equalization current time series data includes multiple sampling times and corresponding equalization current values.

[0176] The impedance morphology encoding unit 620 is used to extract the impedance morphology embedding vector corresponding to the Nyquist impedance curve based on the autoencoder.

[0177] The voltage relaxation multi-scale extraction unit 630 is used to perform wavelet packet decomposition on the terminal voltage relaxation time series data to obtain the corresponding multi-scale voltage relaxation energy feature vector.

[0178] The dynamic imbalance offset index calculation unit 640 is used to calculate the dynamic imbalance offset index based on the terminal voltage relaxation timing data and the equalization current timing data.

[0179]

[0180] In the formula, M bal (t) is the dynamic equilibrium imbalance index; Δt represents the sampling time interval, σ V (t) represents the standard deviation of the terminal voltages of all individual cells at sampling time t. Let ΔI be the regularization constant. bal (t) represents the equalization current increment, indicating the difference in equalization current between two adjacent sampling times t-Δt and t.

[0181] The battery pack SOH prediction unit 650 is used to input the impedance morphology embedding vector, the multi-scale voltage relaxation energy characteristic vector and the dynamic equilibrium misalignment index into the battery SOH prediction model to output the SOH estimate of the battery pack.

[0182] In some embodiments, this application provides a non-volatile computer-readable storage medium storing one or more programs including execution instructions. These execution instructions can be read and executed by an electronic device (including but not limited to a computer, server, or network device) to perform the steps of any of the above-described electric vehicle battery health assessment methods based on multi-source data fusion.

[0183] In some embodiments, this application also provides a computer program product, the computer program product including a computer program stored on a non-volatile computer-readable storage medium, the computer program including program instructions, which, when executed by a computer, cause the computer to perform the steps of any of the above-described electric vehicle battery health assessment methods based on multi-source data fusion.

[0184] In some embodiments, this application also provides an electronic device comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform steps of a method for assessing the health of electric vehicle batteries based on multi-source data fusion.

[0185] Figure 7 This is a schematic diagram of the hardware structure of an electronic device for implementing a multi-source data fusion-based electric vehicle battery health assessment method, as provided in another embodiment of this application. Figure 7 As shown, the device includes:

[0186] One or more processors 710 and memory 720, Figure 7 Take the 710 processor as an example.

[0187] The device for performing a method for assessing the health of electric vehicle batteries based on multi-source data fusion may further include an input device 730 and an output device 740.

[0188] The processor 710, memory 720, input device 730, and output device 740 can be connected via a bus or other means. Figure 7 Taking the example of a connection between China and Israel via a bus.

[0189] The memory 720, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the electric vehicle battery health assessment method based on multi-source data fusion in the embodiments of this application. The processor 710 executes various server functions and data processing by running the non-volatile software programs, instructions, and modules stored in the memory 720, thereby realizing the electric vehicle battery health assessment method based on multi-source data fusion described in the above embodiments.

[0190] The memory 720 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the electronic device. Furthermore, the memory 720 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 720 may optionally include memory remotely located relative to the processor 710, and these remote memories can be connected to the electronic device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0191] Input device 730 can receive input digital or character information and generate signals related to user settings and function control of the electronic device. Output device 740 may include display devices such as a display screen.

[0192] The one or more modules are stored in the memory 720, and when executed by the one or more processors 710, they execute the electric vehicle battery health assessment method based on multi-source data fusion in any of the above method embodiments.

[0193] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0194] The electronic devices in this application embodiments exist in various forms, including but not limited to:

[0195] (1) Mobile communication devices: These devices are characterized by their mobile communication capabilities and primarily aim to provide voice and data communication. These terminals include smartphones, multimedia phones, feature phones, and low-end phones.

[0196] (2) Ultra-mobile personal computer devices: These devices fall under the category of personal computers, possessing computing and processing capabilities, and generally also have mobile internet access features. These terminals include: PDAs, MIDs, and UMPCs, etc.

[0197] (3) Portable entertainment devices: These devices can display and play multimedia content. This category includes audio and video players, handheld game consoles, e-book readers, as well as smart toys and portable car navigation devices.

[0198] (4) Other airborne electronic devices with data interaction capabilities, such as vehicle-mounted systems installed on vehicles.

[0199] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0200] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0201] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for assessing the health of electric vehicle batteries based on multi-source data fusion, comprising: Based on the synchronous triggering of electrochemical impedance spectroscopy scanning, terminal voltage relaxation recording, and battery pack equalization operation during the constant current phase cutoff event of the battery pack for electric vehicles, the corresponding Nyquist impedance curves, terminal voltage relaxation time series data, and equalization current time series data are obtained respectively. The terminal voltage relaxation timing data includes multiple sampling times and corresponding terminal voltage values, and the equalization current timing data includes multiple sampling times and corresponding equalization current values. The impedance shape embedding vector corresponding to the Nyquist impedance curve is extracted based on the autoencoder. Wavelet packet decomposition is performed on the aforementioned terminal voltage relaxation time series data to obtain the corresponding multi-scale voltage relaxation energy feature vector; Based on the aforementioned terminal voltage relaxation timing data and the aforementioned equalization current timing data, the dynamic equalization offset index is calculated: In the formula, M bal (t) is the dynamic equilibrium imbalance index; Δt represents the sampling time interval, σ V (t) represents the standard deviation of the terminal voltages of all individual cells at sampling time t. Let ΔI be the regularization constant. bal (t) represents the equalization current increment, indicating the difference in equalization current between two adjacent sampling times t-Δt and t. The impedance morphology embedding vector, the multi-scale voltage relaxation energy characteristic vector, and the dynamic equilibrium misalignment index are input into the battery SOH prediction model to output the SOH estimate of the battery pack.

2. The method according to claim 1, wherein, The simultaneous triggering of electrochemical impedance spectroscopy scanning, terminal voltage relaxation recording, and battery pack equalization operations based on the constant current phase cutoff event for electric vehicle battery packs includes: The charging and discharging current is sampled in real time at the shunt resistor of the battery pack, and the sampled signal is processed by first-order low-pass filtering and moving average to obtain a smooth current curve. Will With preset threshold I thr When comparing, And the duration is not less than the dejitter time t deb When the constant current phase ends, it is determined that the current-constant phase has ended. If the constant current phase is determined to be cut off, a hardware trigger message is generated and broadcast via the vehicle communication bus to synchronously trigger electrochemical impedance spectroscopy scanning, terminal voltage relaxation recording, and battery pack equalization operation; the hardware trigger message includes a cycle number and a timestamp of the sampling time.

3. The method according to claim 1, wherein, The encoder structure of the autoencoder includes a cascaded input preprocessing layer, a multi-scale dilated convolutional layer, a channel attention pooling layer, and a fully connected dimensionality reduction layer. The input preprocessing layer is used to perform the following operations: The obtained discrete impedance spectrum point set The Nyquist curve is obtained through interpolation and arc length parameterization, with the arc length set as follows: The Nyquist curve is divided into a resampling point sequence P according to equal arc length intervals L / N. j =(x j ,y j ), where L is the total arc length of the Nyquist curve, N is the total number of resampling points, j = 1,...,N; and ω i This represents the frequency of the i-th sinusoidal excitation signal. and The measured excitation frequency ω i The real and imaginary parts of the electrochemical impedance of the battery pack are given, where K represents the total number of discrete frequency points collected from the Nyquist curve, τ is the curve parameter, and s(t) represents the geometric arc length from the curve start point to the parameter value t; P j Let x represent the Nyquist plane coordinates of the j-th resampled point. j Corresponding to the real part, y j Corresponding imaginary part; After stacking the resampling points row by row, the input matrix H is obtained. (0) Each row of the matrix is ​​(x j ,y j ); The multi-scale dilated convolutional layer is used to perform the following operations: Input matrix H (0) Input three layers of one-dimensional dilated convolutions, each layer is... Iteratively obtained Among them, H (l) This represents an intermediate feature of the l-th layer. This indicates that the expansion rate d is 2. l One-dimensional convolution operation, b (l) The offset is for the l-th layer, and C is the number of channels; The channel attention pooling layer is used to perform the following operations: For H (3) Each row vector Calculate attention weights In the formula, α j This represents the attention weight corresponding to the j-th resampling point. Represents the attention mapping matrix. This is the attention weight vector; Calculate the global morphological feature vector: The fully connected dimensionality reduction layer is used to perform the following operations: The global morphological feature vector z is mapped through a two-level fully connected layer to form an output impedance morphological embedding vector. In the formula, Represents the first-level dimensionality reduction mapping matrix, Represents the first layer bias vector, Represents the second-level dimensionality reduction mapping matrix, The second-layer bias vector is represented by D, where D is the dimension of the impedance shape embedding vector, and E represents the impedance shape embedding vector.

4. The method according to claim 1, wherein, The process of performing wavelet packet decomposition on the terminal voltage relaxation time series data to obtain the corresponding multi-scale voltage relaxation energy feature vector includes: The obtained terminal voltage relaxation discrete sequence By applying the bandpass filter operator BP{·} and subtracting the final steady-state voltage V[M-1], the correction sequence is obtained: ΔV[n]=BP{V[n]}-V[M-1], Where ΔV[n] is the corrected voltage attenuation value, representing the voltage relaxation amount after removing the steady-state baseline at the nth sampling time; V[n] is the voltage value measured at the nth sampling time; BP{·} is the bandpass filter operator with a preset cutoff frequency; M is the total number of discrete sampling times in the relaxation record; V[M-1] is the steady-state voltage, representing the voltage baseline value corresponding to the last sampling point of the sequence; Set ΔV[n] as the initial approximation coefficient: a (0) [n]=ΔV[n],n=0,...,M-1, For a (0) [n] Perform U-layer wavelet packet decomposition and recursively obtain the approximation coefficients of the v-th subband at the (r+1)-th layer. and detail coefficient Where r is the decomposition layer index, r = 0, ..., U-1, U is the total number of decomposition layers, and v = 1, ..., 2 r is the subband index of the r-th layer; h[p] and g[p] are the low-pass filter coefficients and high-pass filter coefficients of the Db4 mother wavelet, respectively, and p is the filter coefficient index; For the set of all approximation coefficients of the U-th layer and detail coefficient set Calculate the time-domain energy separately: In the formula, N v Let be the subband coefficient length of the v-th subband coefficient in the U-th layer; Approximating energies {Q} a,v } and detail energy {Q q,v Arranged in ascending order by sub-band index, these form multi-scale energy feature vectors. In the formula, e is the multi-scale voltage relaxation energy feature vector, used to characterize the relaxation decay process at different time scales. U This represents the total number of subbands after decomposition.

5. The method according to claim 1, wherein, After calculating the dynamic equalization offset index based on the terminal voltage relaxation timing data and the equalization current timing data, the method further includes: The dynamic imbalance index is compared with a preset imbalance threshold, and when the dynamic imbalance index exceeds the imbalance threshold, an alarm operation for battery cell inconsistency fault is triggered.

6. The method according to claim 1, wherein, The battery SOH prediction model includes a static feature encoding module, a modulation coefficient generation module, a dynamic feature time series extraction module, a time series pooling module, and a regression prediction module. The static feature encoding module is used to concatenate the impedance morphology embedding vector E with the multi-scale voltage relaxation energy feature vector e to form a static joint feature S. raw =[E;e]; Subsequently, the static feature encoding module uses a multi-layer fully connected network and a self-attention mechanism to process S... raw Perform nonlinear mapping and reweighting to output the context vector S; The modulation coefficient generation module is used to map the context vector S into a parameter network and generate a set of feature modulation coefficients {γ} corresponding to each dilated convolutional layer of the dynamic feature temporal extraction module through at least two fully connected transformations. f ,β f |f=1,...,G};where G represents the total number of dilated convolutional layers, γ f and β f Let γ represent the channel scaling factor vector and channel offset factor vector of the f-th layer, respectively. f ,β f Used to modulate the channel activation of the f-th dilated convolutional layer; The dynamic feature time series extraction module is used to receive the dynamic equilibrium imbalance index sequence M. bal (t), and then extracts temporal features layer by layer through G-layer dilated convolutional units, calculating at each layer. And closely follow the feature modulation Among them, h (f) (t) is the activation output of the f-th layer, representing the feature vector after convolution and activation at sampling time t, h (0) (t)=M bal (t); This indicates that the expansion rate d is 2. f-1 One-dimensional convolution operation; ⊙ represents element-wise multiplication operation. The activation after modulation of the f-th layer represents the output feature vector of this layer after context-driven channel-level modulation at sampling time t. The temporal pooling module is used to process the modulated feature sequence output by the dynamic feature temporal extraction module. Perform global average pooling along the time dimension to generate a pooled vector P; The regression prediction module is used to concatenate the pooling vector P with the context vector S to form a fused feature. [P;S], and outputs the SOH estimate of the battery pack through a multi-layer fully connected regression network.

7. A battery health assessment system for electric vehicles based on multi-source data fusion, comprising: The multi-source data acquisition unit is used to simultaneously trigger electrochemical impedance spectroscopy scanning, terminal voltage relaxation recording, and battery pack equalization operation based on the constant current stage cutoff event of the battery pack for electric vehicles, so as to obtain the corresponding Nyquist impedance curve, terminal voltage relaxation time series data, and equalization current time series data, respectively. The terminal voltage relaxation timing data includes multiple sampling times and corresponding terminal voltage values, and the equalization current timing data includes multiple sampling times and corresponding equalization current values. Impedance morphology encoding unit, used to extract the impedance morphology embedding vector corresponding to the Nyquist impedance curve based on an autoencoder. A voltage relaxation multi-scale extraction unit is used to perform wavelet packet decomposition on the terminal voltage relaxation time series data to obtain the corresponding multi-scale voltage relaxation energy feature vector. The dynamic imbalance offset index calculation unit is used to calculate the dynamic imbalance offset index based on the terminal voltage relaxation timing data and the equalization current timing data. In the formula, M bal (t) is the dynamic equilibrium imbalance index; Δt represents the sampling time interval, σ V (t) represents the standard deviation of the terminal voltages of all individual cells at sampling time t. Let ΔI be the regularization constant. bal (t) represents the equalization current increment, indicating the difference in equalization current between two adjacent sampling times t-Δt and t. The battery pack SOH prediction unit is used to input the impedance morphology embedding vector, the multi-scale voltage relaxation energy characteristic vector, and the dynamic equilibrium misalignment index into the battery SOH prediction model to output the SOH estimate of the battery pack.

Images