A new energy vehicle power battery pack adaptive equalization control method

Abstract

This invention relates to an adaptive equalization control method for power battery packs in new energy vehicles, aiming to solve the problems of poor timeliness of existing battery equalization response and insufficient dynamic adaptability of path allocation. Its core scheme includes: real-time acquisition of multi-channel transient voltage and current signals of each cell and key node of the power battery; generation of a topology-sensitive feature vector set based on signal feature extraction; online inference of electrical functional adjacency relationships and centrality indicators between cells using a lightweight graph neural network model combined with massive aging test data; dynamic determination and limitation of the equalization influence domain; further calculation and issuance of precise parameterized equalization control commands; and driving the equalization circuit to complete millisecond-level closed-loop energy flow. The system achieves adaptive adjustment and feedback correction through real-time topology recalibration. This scheme can improve the response speed and local adaptability of equalization control, reduce the risk of thermal imbalance, and significantly enhance the operational safety and lifespan of the battery pack.

Description

Technical Field

[0001] This invention relates to the field of power battery system control and management technology, and in particular to an adaptive equalization control method for power battery packs in new energy vehicles. Background Technology

[0002] Currently, several mainstream technical solutions have emerged in the field of equalization control for new energy vehicle power battery management systems. These include static threshold judgment based on open-circuit voltage or SOC difference, and global equalization triggering mechanisms using fixed equalization current and fixed equalization duration. These methods largely rely on static parameters such as individual cell voltage and SOC as the basis for equalization decisions, working in conjunction with equalization circuits to achieve periodic energy transfer, thereby mitigating cell inconsistency degradation and performance decline under extreme conditions. In addition, some traditional solutions also achieve load data synchronization through CAN bus communication, or rely on hardware strategies such as passive resistor equalization and carrier phase-shift control to improve equalization efficiency. Recent technological development trends mainly focus on optimizing active equalization topologies, introducing adjustable equalization current parameters, and using algorithms to assist in determining equalization triggering conditions to cope with complex operating condition changes. However, the industry mainstream still relies on global static threshold judgment, lacking a deep understanding of the dynamic changes in the actual electrical topology of the power battery pack.

[0003] Representative technologies, such as equalization control methods based on the static SOC and voltage differences of individual cells, are suitable for battery management scenarios under standard operating conditions, such as ordinary passenger cars and short-distance commercial vehicles. These methods can periodically perform energy transfer by monitoring the equalization difference over a long period of time. However, under complex dynamic load conditions such as rapid acceleration, regenerative braking, short-term high-power discharge, or parallel shunt operation of multiple modules, the impedance of the connection path of each cell in the battery pack and the main discharge path change frequently. Traditional methods cannot identify the target cell that actually bears the main load in real time, which often leads to equalization trigger lag, misjudgment, or even conflicting equalization operations, thereby affecting the response speed of the power system and the long-term life of the battery.

[0004] The current technological shortcomings are mainly manifested in the following ways: During dynamic load changes, load identification suffers from a significant delay, especially under complex operating conditions. The actual load on the main load path and key nodes of the battery pack shifts rapidly, and existing load determination models based on open-circuit voltage estimation, static SOC, or general filtering algorithms are often limited by sampling intervals, signal delays, and model staticity, resulting in a severe lag in equalization parameter adjustment. Furthermore, traditional equalization strategies cannot perform domain-specific and node-specific parameterized adaptive adjustments for the actual main load path. The equalization operation is still mainly triggered globally, which cannot suppress redundant cell equalization in a timely manner, nor can it optimize energy flow, affecting overall equalization efficiency and battery health.

[0005] As the application scenarios of power batteries evolve towards advanced operating conditions such as extreme dynamic loads, complex multi-module topologies, and real-time energy splitting, existing technologies urgently need to improve their dynamic perception capabilities of power battery pack load states. There is a pressing need for a load-sensing equalization control mechanism capable of identifying the main load path in milliseconds, reconstructing electrical adjacency relationships in real time, and driving adaptive parameter adjustment based on local dynamic topology manifolds. This would further shorten the latency between load identification and equalization response, improve the system's dynamic adaptability, and ensure the consistency, safety, and long-term performance of the battery pack under varying operating conditions. Therefore, this invention proposes a load-sensing equalization parameter instantaneous reconstruction method based on dynamic mapping of the power battery pack topology. This aims to solve the problems of lag, low efficiency, and uncontrollable lifespan risks associated with traditional equalization control methods under dynamic load scenarios, providing a more efficient and accurate dynamic equalization solution for new energy vehicle power battery pack management systems. Summary of the Invention

[0006] This application provides an adaptive equalization control method for power battery packs of new energy vehicles, which aims to solve one of the problems or issues of the existing technology mentioned in the background.

[0007] This application provides an adaptive equalization control method for a new energy vehicle power battery pack, specifically including: S1: Obtain multi-channel voltage transient signals and current transient signals from the positive and negative terminals of each cell in the power battery pack and key nodes of the module-level busbar, and generate the original topology-coupled sampling dataset; S2: Perform sliding window difference enhancement processing on the original topological coupling sampling dataset to extract the topologically sensitive feature vector set; S3: Based on the prior knowledge base of massive real vehicle accelerated aging test data trained offline, a graph neural network reasoning model is constructed; wherein, the prior knowledge base of massive real vehicle accelerated aging test data stores several historical topology-sensitive feature vector sets. S4: Input the current topology-sensitive feature vector set into the graph neural network inference model to perform online mapping calculation and output the functional adjacency matrix; S5: Calculate the betweenness centrality and eigenvector centrality indices of each node in the electrical topology based on the functional adjacency matrix, and delineate the equilibrium influence domain; S6: Based on the power carrying weight of the cell assembly in the topology within the equalization influence domain, dynamically match and generate an instantaneous reconfiguration equalization control parameter set; S7: The instantaneous reconfiguration equalization control parameter group drives the equalization execution circuit to perform parameterized equalization operation on the target cell combination within the equalization influence domain, completing the closed-loop equalization action driven by the local topology manifold.

[0008] S8: Monitor the changes in voltage-current topology response characteristics after execution and feed them back to the graph neural network inference model for topology recalibration to generate an updated functional adjacency matrix for adaptive adjustment in the next sampling period.

[0009] The adaptive equalization control method for power battery packs in new energy vehicles provided in this application has the following beneficial effects: (1) The battery system equalization control method based on topology dynamic sensing provided in this application significantly improves the accuracy and real-time response of cell status identification under complex dynamic conditions compared with the traditional equalization strategy that relies on static series-parallel structure and fixed threshold judgment. Existing technologies generally use voltage difference or SOC estimation results as equalization trigger conditions, which are easily affected by open circuit voltage recovery delay, filtering inertia and current integral drift. In transient processes such as rapid acceleration and regenerative braking, misjudgment is frequently generated, resulting in ineffective or even reverse equalization of non-critical cells, exacerbating inconsistency within the group. This solution deploys a microsecond-level synchronous sampling circuit at the hardware layer to realize multi-channel high-frequency (≥10kHz) voltage and current coupling acquisition of cell-level and module-level nodes, and introduces sliding window differential enhancement technology to extract topology-sensitive features such as voltage jump slope, current phase shift and transient impedance envelope, effectively capturing subtle evolution of electrical behavior. Building upon this foundation, a lightweight graph neural network-driven topology dynamic mapping engine is constructed. This engine outputs a functional adjacency matrix reflecting the actual current path, thermal coupling strength, and disturbance propagation chain, overcoming the limitations of physical connection structures. This allows the system to accurately identify the core cell group that truly participates in the main energy exchange under specific operating conditions. For example, in the initial stage of high-current discharge, high-SOC cells exhibit a "pseudo-high voltage" due to internal resistance voltage drop hysteresis, making them prone to misjudgment as priority discharge targets using traditional methods. However, this solution analyzes the topological role change of these cells in the functional adjacency matrix—where they have detached from the main path and become weakly coupled side nodes—and automatically suppresses balancing operations on these cells, thus avoiding capacity waste and lifespan loss caused by malfunctions.

[0010] (2) Furthermore, by anchoring the balancing decision to the node centrality index in the dynamic topology, this invention achieves a paradigm shift from "globally unified rules" to "adaptive local influence domains," significantly improving the targeting and resource utilization efficiency of balancing execution. Unlike the widely used full-group scanning balancing or multi-objective optimization methods based on preset weights in existing technologies, this scheme dynamically identifies key cell combinations with high information transmission capacity and strong power carrying potential on the current main load path based on graph theory indices such as betweenness centrality and eigenvector centrality, and implements parameterized balancing control only on 2–4 core nodes with a sudden increase in centrality. This mechanism ensures that the balancing energy is concentrated on the local area with the greatest impact on overall performance, avoiding redundant intervention on edge nodes and significantly reducing system power consumption and heat accumulation. At the same time, the balancing current amplitude and pulse width duty cycle are adjusted in real time according to the power weight of the target node in the current topology, forming a dynamic response characteristic that strictly matches the operating conditions. The entire closed-loop process encompasses five stages: "topology perception → influence domain identification → parameter mapping → execution feedback → topology recalibration." The single reconstruction cycle is controlled within 3ms, far faster than the hundreds of milliseconds of response gaps in traditional BMS caused by CAN communication delays and model update lags, effectively overcoming the control blind spot problem in dynamic processes. Furthermore, this approach completely abandons externally dependent and computationally expensive techniques such as CAN bus frequency modulation, genetic algorithm optimization, TOPSIS multi-attribute decision-making, and cloud-based collaborative prediction, focusing instead on the measurable modeling and tracking of the battery pack's intrinsic electrical manifold, possessing high domain specificity and engineering deployability.

[0011] The aforementioned technical solutions collectively construct an intelligent balancing system centered on real-time electrical topology evolution. This system not only achieves a deep understanding and proactive intervention of the energy flow within the battery pack, but also achieves high-precision, low-latency, and low-power closed-loop control without the need for complex adjustments or external coordination. The system exhibits excellent interpretability—balancing decisions are always based on observable topology evolution, rather than black-box scoring or empirical thresholds. It also demonstrates superior scalability, suitable for power battery systems with different series-parallel configurations and significant differences in aging levels. It is particularly suitable for applications in new energy vehicles, energy storage power stations, and other scenarios with stringent requirements for safety, energy efficiency, and lifespan, providing a new technical path for next-generation high-performance battery management systems. Attached Figure Description

[0012] Figure 1 This is the main flowchart of an adaptive equalization control method for power battery packs in new energy vehicles. Figure 2 This is a sub-flowchart of an adaptive equalization control method for power battery packs in new energy vehicles; Figure 3This is another sub-flowchart of an adaptive equalization control method for power battery packs in new energy vehicles. Detailed Implementation

[0013] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0014] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0015] like Figure 1 As shown, this application provides an adaptive equalization control method for a new energy vehicle power battery pack, specifically including: S1: Obtain multi-channel voltage transient signals and current transient signals from the positive and negative terminals of each cell in the power battery pack and key nodes of the module-level busbar to generate the original topology-coupled sampling dataset.

[0016] S2: Perform sliding window differential enhancement processing on the original topology-coupled sampling dataset to extract a set of topology-sensitive feature vectors. The set of topology-sensitive feature vectors includes: voltage jump slope, current phase offset, and transient impedance response envelope.

[0017] S3: Based on a prior knowledge base of massive real-vehicle accelerated aging test data trained offline, a graph neural network inference model is constructed to characterize the main current path and strongly correlated thermal coupling regions between battery cells. The prior knowledge base of massive real-vehicle accelerated aging test data stores several historical sets of topology-sensitive feature vectors.

[0018] S4: Input the current topology-sensitive feature vector set into the graph neural network inference model to perform online mapping calculation, so as to output a functional adjacency matrix that reflects the electrical functional adjacency relationship of each cell at the current moment.

[0019] S5: Calculate the betweenness centrality and eigenvector centrality indices of each node in the electrical topology based on the functional adjacency matrix, so as to delineate the balanced influence domain that only includes cells with a sudden increase in centrality that are on the current load main path.

[0020] S6: Based on the power carrying weight of the cell assembly in the topology within the equalization influence domain, dynamically match and generate an instantaneous reconfiguration equalization control parameter set. The instantaneous reconfiguration equalization control parameter set includes: a specific equalization current amplitude and a pulse width duty cycle.

[0021] S7: The instantaneous reconfiguration equalization control parameter group drives the equalization execution circuit to perform parameterized equalization operation on the target cell combination within the equalization influence domain, so as to complete the millisecond-level closed-loop equalization action driven by the local topology manifold.

[0022] S8: Monitor the changes in voltage-current topology response characteristics after execution and feed them back to the graph neural network inference model for topology recalibration to generate an updated functional adjacency matrix for adaptive adjustment in the next sampling period.

[0023] Step S1: Obtain multi-channel voltage transient signals and current transient signals from the positive and negative terminal leads of each cell in the power battery pack and key nodes of the module-level busbar to generate the original topology-coupled sampling dataset. Specifically, this includes: S1.1: Based on the physical connection architecture of the power battery pack, obtain the location coordinates of the positive and negative leads of each cell and the distribution map of key nodes of the module-level bus. Use the hardware abstraction layer to perform electrical node logical addressing processing to generate a standardized node addressing list containing unique identifiers as the basic input for subsequent channel configuration.

[0024] The input conditions include the physical connection architecture data of the power battery pack, the coordinates of the positive and negative terminal leads of each cell, and the distribution information of key nodes in the module-level busbar, serving as the initial parameters for the hardware abstraction layer mapping process. Multi-dimensional coordinate analysis is performed on the above inputs, converting the three-dimensional spatial position vectors of each lead and key node into a unified engineering coordinate system data format, ensuring that the positioning accuracy meets millimeter-level requirements. Based on the converted engineering coordinate system data, the electrical topology construction function is called to synchronously load the physical connection relationships and spatial distribution into the data model of the hardware abstraction layer, forming a set of node prototypes with physical semantics. Logical addressing operations are applied to the node prototype set. In this operation, by recursively traversing the physical connection edges and combining node indexing rules, a unique logical address is assigned to each node, and this address is bidirectionally bound to the physical coordinates for storage. An address conflict detection mechanism scans the entire node set, performing conflict repair and completion processing for any duplicate or empty numbers to ensure the consistency and integrity of the node addressing list. The repaired logical address list and node physical coordinate binding information are encapsulated into a standardized node addressing list. This list contains unique identifiers and meets the accuracy requirements for node positioning in subsequent multi-channel sampling channel configurations. Through the above hardware abstraction layer mapping process, the physical connection architecture data is transformed into a standardized node addressing list that can be directly used to trigger timing configurations, achieving a high degree of consistency between node positioning and logical control.

[0025] S1.2: Configure the trigger timing parameters of the multi-channel synchronous sampling circuit according to the standardized node addressing list, and use a differential probe array to perform parallel acquisition processing of microsecond-level synchronous voltage transient signals and current transient signals at the positive and negative leads of each cell and key nodes of the module-level busbar to generate a set of original discrete sampling sequences with unified timestamps.

[0026] The input conditions include the standardized node addressing list generated in the previous step, which contains unique identifiers for the positive and negative leads of each cell and key nodes of the module-level bus. For these identifiers, a node index mapping table is first established in the control logic of the multi-channel synchronous sampling circuit to ensure that the hardware trigger signal can accurately locate the node under test. Based on the node index mapping table, the timing configuration module is called to set the microsecond-level trigger period through an internal high-precision clock source, and a phase-locked loop mechanism is used to strictly align the trigger times of each sampling channel to the same time base. For each node index, the physical channel position of the differential probe array is assigned in the sampling circuit, and the differential probe bias compensation module is called to adjust the probe zero-point drift and common-mode rejection ratio parameters to improve the accuracy of high-frequency transient signal acquisition. The differential probe array is simultaneously connected to the voltage measurement sub-channel and current measurement sub-channel of the corresponding node. When the trigger signal arrives, the analog-to-digital converter performs synchronous quantization to generate discrete voltage and current data with the node identifier as the primary key and a unified timestamp tag. The quantization results of each channel are arranged in ascending order by timestamp and bound to the corresponding node identifier to form discrete sampling data records for subsequent anti-aliasing filtering and time base correction processing. Through the above processing method, the node addressing list of the previous step is transformed into a trigger timing configuration that can directly drive the differential probe array, achieving the expected technical effect of node positioning accuracy and sampling synchronization.

[0027] S1.3: Perform anti-aliasing filtering and time base alignment correction on the original discrete sampling sequence set, and use oversampling reconstruction to eliminate inter-channel phase deviation and suppress high-frequency noise interference to generate a calibrated transient electrical waveform dataset with high signal-to-noise ratio.

[0028] For the original discrete sampling sequence set with a unified timestamp, anti-aliasing filtering is performed on the voltage and current transient signals output by the multi-channel differential probe array using a bandpass filter bank to suppress high-frequency components exceeding the Nyquist frequency during sampling and prevent other-frequency components from mixing into the target signal. After anti-aliasing processing, each channel's sampling sequence is time-base aligned using a time base correction module to compensate for microsecond-level trigger delay deviations caused by differences in flange or probe installation positions. Oversampling reconstruction is applied to the corrected multi-channel signals. With the sampling frequency increased to four times the original design frequency, the transition region between adjacent sampling points is numerically estimated using an interpolation function. Combining the phase response characteristics of each channel, the phase deviation between channels is calculated and eliminated to ensure strict synchronization of different channel signals on the time axis. An exponential smoothing window function is used to suppress high-frequency noise in the reconstructed sequence to eliminate transient glitches introduced by switching devices and the electromagnetic environment. Through the above chain processing method, the original discrete sampling sequence set of the previous step is transformed into a calibrated transient electrical waveform dataset with high signal-to-noise ratio, low phase difference, and high time synchronization, so as to achieve the benchmark data accuracy required for real-time feature extraction by the equalization strategy generation module.

[0029] For example, in a power battery module containing 96 cells, the original sampling frequency of each positive and negative terminal lead and key node of the module-level bus is 10kHz. The anti-aliasing filter bank is set with a bandwidth of 0.1Hz to 4.8kHz, and the damping factor is inserted within the bandwidth to ensure critical damping characteristics. Time base alignment correction adopts a global synchronization criterion with a maximum deviation of no more than 2μs, and an offset compensation vector is used to time-shift the data of each channel. In the oversampling reconstruction process, cubic spline interpolation is selected as the interpolation function, the oversampling factor is set to 4 times, and the phase difference calculation formula is applied to the reconstructed sequence: in, Let i be the phase difference between the i-th channel and the j-th channel. and These represent the sampled values ​​of channel i and channel j at the k-th sampling time, respectively. The phase difference is used for phase alignment compensation to ensure that all channels maintain consistent delay across the simulated real electrical topology. The high-frequency noise suppression process uses an exponential smoothing window with a window length of 256 points and a smoothing coefficient α of 0.15. After processing, the signal-to-noise ratio of the calibration dataset is significantly improved compared to the original sequence, and the phase difference fluctuation is controlled within 0.05°, providing a stable and accurate input basis for the subsequent S1.4 sliding time window slicing.

[0030] S1.4: Based on the calibrated transient electrical waveform dataset, perform a sliding time window slicing operation to cut the continuous waveform into data frames of fixed length according to a preset sampling frequency of not less than 10kHz, so as to generate a time-seriesd original topology-coupled sampling data subset containing complete dynamic operating condition segments.

[0031] S1.5: Perform structured encapsulation processing on the time-seriesd original topology-coupled sampling data subset, and bind the voltage transient signal component and the current transient signal component by key-value association according to the node identifier to generate the original topology-coupled sampling dataset that can be used in the feature extraction stage.

[0032] Step S2: Perform sliding window differential enhancement processing on the original topology coupling sampling dataset to extract a set of topology-sensitive feature vectors. The set of topology-sensitive feature vectors includes: voltage jump slope, current phase offset, and transient impedance response envelope. Specifically, it includes: S2.1: Perform a multi-scale sliding window truncation operation on the original topologically coupled sampling dataset to generate a set of local time series segments containing transient fluctuation information, ensuring that subsequent processing can capture microsecond-level voltage and current change details.

[0033] For each node signal stream in the original topology-coupled sampling dataset, precise time-axis indexing is performed according to a predetermined sampling frequency to ensure that the voltage transient signal component and the current transient signal component are fully aligned in the time domain so that phase consistency can be maintained during subsequent windowing processing.

[0034] The calibrated transient electrical waveform data is segmented according to a preset multi-scale sliding window group. The multi-scale refers to the window length covering three time spans: short, medium, and long, so that transient fluctuation characteristics can be captured at microsecond, millisecond, and longer scales. The division of different scale windows is determined based on the sampling frequency and the rate of change of the target features.

[0035] When performing window sliding, setting the window overlap ratio can improve the sensitivity of capturing transient changes. For example, an overlap ratio higher than 50% can ensure that the change is covered by the window in multiple sampling periods where the fluctuation occurs, avoiding loss caused by fluctuations crossing the window.

[0036] Boundary protection is performed on the data segments within each window. Smooth attenuation weight functions are introduced at the left and right ends of the window to suppress the interference of boundary effects on subsequent differential and spectral analysis. The shape of the weight function is selected according to the frequency components of the target signal, such as using Hanning or flat-top functions.

[0037] The set of local time series segments generated by multi-scale sliding window extraction is indexed and bound according to node identifiers to ensure that local segments of different nodes can be quickly matched by index within the same time interval, thereby enabling multi-node parallel processing during subsequent feature extraction.

[0038] By using the multi-scale sliding window extraction and index binding described above, the results of the previous step are transformed into a set of local time series segments containing transient fluctuation information, thus achieving the expected technical effect that subsequent slope calculations and phase analyses at each node can capture microsecond-level voltage and current abrupt changes.

[0039] S2.2: Based on the set of local time series segments, high-order finite difference processing is applied to calculate the slope, so as to extract the voltage jump slope parameter that characterizes the transient change rate of the cell terminal voltage, and eliminate the interference of static DC component on dynamic feature extraction.

[0040] Based on the voltage transient signal components contained in the local time series fragment set, the computational core of high-order finite difference processing is invoked to perform sequence index-level data reading and buffering processing on the discrete sampling point sequence within each fragment, ensuring that the calculation process can obtain complete amplitude information of adjacent points.

[0041] After the data has been read, a differential operator matrix is ​​constructed according to the preset differential order. The voltage amplitude differences of adjacent sampling points are combined into an ordered differential vector in sequence, and the sampling interval time Δt is introduced as a time reference to establish the denominator reference value for slope calculation.

[0042] The difference operator matrix is ​​multiplied by the ordered difference vector to obtain a series of consecutive difference results. Then, the results are scaled by the power of Δt to ensure the accurate conversion of the time base in high-order difference calculations and to calibrate the physical quantity of transient rate of change.

[0043] Cumulative weighting coefficients and distribution smoothing are applied to the differential result sequence to eliminate the bias effect caused by local static DC components and enhance the sensitivity to microsecond-level voltage fluctuations, so that the obtained parameters more strictly represent the actual transient rate of change of the cell terminal voltage.

[0044] The differential calculation results, after weight correction and smoothing, are encapsulated into voltage jump slope parameters, with added node identifiers and time labels, and stored in a feature parameter buffer for subsequent cross-correlation time-domain analysis. Through high-order finite-difference calculations and DC component suppression, the set of local time series segments obtained in the previous step is transformed into voltage jump slope parameters with physical meaning and significant noise suppression, achieving accurate quantification of the transient change rate at the cell end.

[0045] S2.3: Perform cross-correlation time-domain analysis processing using the voltage jump slope parameter and the current transient signal in the original topology coupled sampling dataset to calculate the current phase offset value reflecting the inductive effect and line impedance characteristics, and establish the time-series correlation between voltage disturbance and current response.

[0046] A cross-correlation analysis model is established between the voltage jump slope parameter calculated using high-order finite difference processing and the transient current signal in the original topology-coupled sampling dataset to achieve accurate time-domain registration of the two types of time-series signals. The voltage jump slope sequence and the transient current signal sequence are respectively mean-reduced to eliminate the influence of DC bias on the cross-correlation results. The cross-correlation function is calculated in the time domain using the fast convolution method to determine the time delay corresponding to the maximum correlation peak of the current signal relative to the voltage signal. A sampling rate correction factor is used to convert the time delay into a phase offset, which is then normalized by combining known power frequency or high-frequency components to obtain the instantaneous current phase offset value. This value, combined with the electrical equivalent model parameters, quantifies the time lag caused by inductive effect between voltage disturbances and current responses, and reflects the influence of line impedance on signal propagation.

[0047] in, The time delay of the cross-correlation peaks. The operating frequency of the signal. The calculated current phase offset angle is used to establish a time-series correlation matrix between voltage disturbance and current response, providing necessary input parameters for subsequent impedance spectrum fitting. Through cross-correlation time-domain analysis, the voltage jump slope parameter from the previous step is converted into a current phase offset value, enabling quantitative identification of inductive effects and line impedance characteristics.

[0048] S2.4: Based on the current phase offset value and the voltage jump slope parameter, perform complex domain impedance spectrum fitting operation to reconstruct transient impedance response envelope data that characterizes the dynamic changes of the internal connection path of the battery pack, and quantify the electrical path impedance characteristics under different operating conditions.

[0049] When performing complex-domain impedance spectrum fitting on the current phase shift value and voltage jump slope parameter obtained from the cross-correlation analysis in step S2.3, the two types of time-domain features are used as the basis for constructing the real and imaginary parts of the complex data, and a unified frequency sampling vector is defined. Feature pairing is performed, mapping the voltage jump slope parameter to the voltage amplitude change rate at each frequency point, and the current phase shift value to the phase delay angle at the corresponding frequency point. The Fast Fourier Transform module is called to convert the paired time-series features into a frequency-domain complex vector, resulting in a dual-channel data structure containing the amplitude spectrum and phase spectrum. Based on this frequency-domain complex vector, an impedance spectrum estimation formula is constructed according to the definition of complex impedance: in, It is a complex vector of voltage amplitude. The current amplitude is represented as a complex vector. A polar coordinate transformation is performed to map the complex impedance to an amplitude and phase representation. The phase curve in the high-frequency band is corrected based on the current phase offset value to eliminate phase distortion caused by inductive components. Window function smoothing and continuity-constrained interpolation methods are applied to curve fit the impedance values ​​at adjacent frequencies, forming a transient impedance response envelope that continuously represents the impedance from low to high frequencies. This envelope is indexed and encapsulated according to node identifiers and operating condition labels, resulting in a structured dataset that quantifies the impedance characteristics of electrical paths under different dynamic operating conditions. Through complex domain impedance spectrum fitting, the current phase offset and voltage jump slope parameters from the previous step are transformed into a transient impedance response envelope that directly represents the dynamic changes in the internal connection paths of the battery pack, achieving the expected technical effects of impedance characteristic quantification and traceable operating condition differences.

[0050] S2.5: Perform multi-dimensional feature fusion and normalization mapping on transient impedance response envelope data, current phase offset values ​​and voltage jump slope parameters to generate a set of topology-sensitive feature vectors containing complete electrical topology dynamic information, which serves as the standardized input object for the graph neural network inference model.

[0051] like Figure 2 As shown, step S3: Based on the prior knowledge base of massive real-vehicle accelerated aging test data trained offline, a graph neural network inference model is constructed to characterize the main current path and strongly correlated thermal coupling region between battery cells. The prior knowledge base of massive real-vehicle accelerated aging test data stores several historical topology-sensitive feature vector sets; these historical topology-sensitive feature vector sets include: historical voltage jump slope sequences, current phase offset sequences, and transient impedance response envelope data. Specifically, they include: S3.1: Obtain the prior knowledge base of massive real vehicle accelerated aging test data generated during the offline training phase, and perform multi-modal alignment and noise cleaning on the historical voltage jump slope sequence, current phase offset sequence and transient impedance response envelope data in the massive real vehicle accelerated aging test data prior knowledge base to generate a standardized graph neural network training sample set containing complete battery life cycle degradation characteristics.

[0052] In this embodiment, the massive real-vehicle accelerated aging test data prior knowledge base specifically refers to a multimodal time-series database that integrates the electrochemical characteristics of the entire battery life cycle with the actual operating conditions of the vehicle. This knowledge base was acquired and constructed in the following ways: First, real vehicles with significant operational characteristics, such as typical urban buses, taxis, and ride-hailing vehicles, were selected as test samples. The Battery Management System (BMS) was used to collect raw operational data from these vehicles at high frequency over a real-world operation period of 12 to 24 months, under different temperature environments, load conditions, and charge / discharge rates. Second, state decoupling based on a Hidden Markov Model was performed on the raw data to extract deep feature sequences reflecting the internal aging mechanism of the battery. These sequences include the voltage jump slope sequence reflecting changes in the lattice stress of the electrode material, the current phase shift sequence reflecting the electrolyte impedance growth trend, and the transient impedance response envelope data reflecting the dynamic process of the growth and fracture of the solid electrolyte interphase (SEI). Finally, the above multi-dimensional feature data was spatiotemporally aligned with corresponding aging labels such as battery cycle count and internal resistance growth factor, and stored in a structured manner at the cell level to form a standardized prior knowledge base that can be directly accessed by graph neural networks.

[0053] S3.2: Based on the standardized graph neural network training sample set, perform graph structure initialization and construction operation, define each cell node as a graph vertex and the potential electrical connection between cells as a graph edge, and perform node feature embedding operation to map the historical voltage jump slope sequence, current phase offset sequence and transient impedance response envelope data into high-dimensional node feature vectors to generate historical initial cell node feature tensors with physical semantic information.

[0054] Based on the standardized graph neural network training sample set, the initial modeling of the physical connection relationships of battery cells is performed using graph-oriented structure construction to achieve global association mapping of nodes and edges for subsequent inference calculations. The unique identifier of each battery cell unit in the training sample set is loaded into the node definition module, generating a corresponding graph vertex index set according to its physical location. The potential electrical connection information recorded in the training samples is parsed, mapping node pairs that satisfy the conditions for effective current conduction and thermal coupling to graph edges, and assigning initial unweighted connection attributes to each edge. The voltage jump slope parameter, current phase offset value, and transient impedance response envelope value contained in the topology-sensitive feature vector set are matched according to the node identifier and mapped to a high-dimensional feature space. Principal component orthogonalization and normalization are used to combine the three physical semantic features into a node feature vector, preserving the consistency of physical dimensions and numerical stability of each feature during dimensional expansion. An embedded vector generator is used to form an initial battery cell node feature tensor, where each row of the tensor matrix corresponds to a battery cell node vertex, and each column corresponds to a different normalized physical feature value. After the embedding mapping is completed, the node feature tensor is bound to the graph vertex index set to form the input ground state data structure that can be used for subsequent graph convolution aggregation and connection prediction.

[0055] By using graph structure initialization and embedding mapping, the results of the standardized training sample set from the previous step are transformed into an initial cell node feature tensor with physical semantic information, thereby enabling the model to represent electrical topology features in a high dimension during the input stage.

[0056] For example, in a training sample set containing 96 battery cells, the node definition module generates a graph vertex index set from 1 to 96 according to the physical arrangement number. Parsing the current path data in this sample set yields a total of 144 potential electrical connection edges. Edges satisfying the current conduction validity condition are assigned a connection flag of 1 during initialization. In the topology-sensitive feature vector set, the voltage jump slope parameter of each battery cell node ranges from -0.45 to 0.38 V / μs, the current phase offset ranges from -12 to 15 degrees, and the transient impedance response envelope ranges from 0.8 to 3.5 mΩ. When mapped to a high-dimensional space, each feature class is normalized to the interval [-1, 1]. Principal component orthogonalization is used to retain the orthogonal components of the three physical features, and a uniform dimension adjustment is performed. The resulting initial battery cell node feature tensor is a 96×3 matrix, where the matrix row index corresponds to the battery cell number, and the column index corresponds to the normalized voltage slope, phase offset, and impedance value. After being bound to the graph vertex index set, this matrix is ​​input into the subsequent graph convolutional network. During the validation set operation, it can significantly improve the accuracy of main path identification and ensure that the model can stably output the true functional adjacency relationship between each cell under complex load conditions.

[0057] S3.3: Using the initial cell node feature tensor, perform multi-level neighbor node information aggregation and nonlinear transformation processing, and dynamically allocate the state contribution weights of different neighbor nodes to the central node through an attention mechanism to generate a historical intermediate layer hidden state feature map that can characterize the actual current main path strength and thermal coupling transmission probability between cells.

[0058] Based on the initial cell node feature tensor, the input base layer data is determined in the lightweight graph convolutional network inference framework. Through a multi-level neighbor node information aggregation mechanism, the node weight initialization matrix is ​​called in the first layer convolution operation to perform linear weighted accumulation of the features of adjacent nodes, and a nonlinear activation function is applied to the accumulation result to enhance the representation ability of complex electrical correlation patterns.

[0059] In the second convolutional operation, a multi-level neighborhood expansion strategy is used to introduce the features of neighboring nodes on two-hop or higher paths into the aggregation calculation. The weight ratio of different path lengths is adjusted by the normalized coefficient matrix to ensure that the influence of far-end highly correlated nodes on the state of the central node is effectively preserved.

[0060] An attention weight calculation unit is introduced during the aggregation process. It uses the similarity vector between the features of neighboring nodes and the features of the center node as input to the attention score, and calculates the attention coefficient using the following formula: in, For trainable projection matrices, and These are the feature vectors of the center node and its neighboring nodes, respectively. concat is the feature concatenation operation, and softmax is the normalization function.

[0061] Normalization is applied to the attention coefficients of each group of neighbor nodes to suppress the impact of large weight differences on the network convergence stability. The normalized attention coefficients are then multiplied element-wise with the corresponding neighbor node feature vectors and summed to generate an aggregated feature vector set containing the contributions of each neighbor to the state of the central node.

[0062] After the convolution and attention aggregation operations are completed, the outputs of each layer are concatenated along the channel dimension, and batch normalization is called to equalize the feature distribution, forming an intermediate hidden state feature map that can characterize the strength of the actual main current path and the probability of thermal coupling between cells.

[0063] By using the above-mentioned multi-layer convolution and attention aggregation processing method, the initial cell node feature tensor of the previous step is transformed into a high-dimensional hidden state feature map with weighted neighbor information, thereby achieving accurate representation of the load main path and thermal coupling region in the electrical topology.

[0064] For example, for a power battery pack composed of 64 cells, the initial cell node feature tensor dimension is set to 64×128, where the 128-dimensional features include a normalized combination of voltage jump slope, current phase offset, and transient impedance response parameters. In the first convolutional layer, the neighbor aggregation weight matrix dimension is set to 128×64, the activation function is LeakyReLU, and the negative slope coefficient is 0.2. In the second convolutional layer, two-hop neighborhood expansion is used, and the path weight decay factor is set to 0.85. In the attention calculation, the W matrix dimension is 256×1, and the attention coefficients, after softmax normalization, fall within the range of 0.05–0.12. The hidden state feature map obtained after aggregation calculation has a dimension of 64×256. After batch normalization, the kurtosis and skewness of the hidden state feature map are significantly reduced. Its characterization ability is significantly improved in complex load change tests, with a significant increase in the stability of the main current path identification results and a substantial improvement in the accuracy of thermal coupling region identification. The output results, as input to the edge weight prediction decoding in S3.4, can ensure the validity and physical rationality of the original adjacency weight coefficient matrix.

[0065] S3.4: Based on the hidden state feature map of the intermediate layer, perform edge weight prediction and decoding operation, use bilinear mapping function to calculate the similarity score between the feature vectors of any two cell nodes, and convert the similarity score into a historical original adjacency weight coefficient matrix that quantifies the possibility of an effective current main path between cells, so as to complete the initial mapping from the hidden layer feature space to the explicit topological connection space.

[0066] Based on the intermediate layer hidden state feature map, the edge weight prediction and decoding module is invoked. The hidden state feature vector of each cell node is used as input to the bilinear mapping function for pairwise combination calculation to obtain the similarity score between any two nodes. For the bilinear mapping function, the hidden state feature vectors of the source and target nodes are extracted respectively, and dimension-consistent processing is performed in matrix form. Weight coefficients optimized during offline training are loaded into the mapping kernel matrix, and an unnormalized similarity quantization value is generated through bidirectional multiplication. A smoothing normalization process is applied to this similarity quantization value, mapping its range to between 0 and 1 to ensure comparability and stability between different node pairs. The normalized similarity score is used to determine the effectiveness of the functional path. When the score exceeds a preset association threshold, it is determined that the node pair has a potential main path or a strong thermal coupling relationship under the current operating condition, and its value is filled into the corresponding position in the original adjacency weight coefficient matrix. For node pairs below the threshold, a value of zero is set to avoid interference from false edges. The similarity score is calculated using the following formula: Where S is the similarity score. The weight matrix is ​​a bilinear mapping. Let be the hidden state feature vector of the target node at the previous time step. This is the transpose of the hidden state vector of the source node at the previous time step. Through bilinear mapping transformation and threshold determination, the hidden state feature space of the previous step is mapped to the original adjacency weight coefficient matrix in the explicit topological connection space, realizing a quantifiable expression of the electrical connection strength between node pairs.

[0067] For example, in a power battery module containing 96 cells, the dimension of the bilinear mapping weight matrix W is set to 64×64, the length of the hidden state feature vectors of the source node and the target node is 64, the mapping kernel is derived from the parameters optimized under high-speed discharge and deep recharge conditions during the offline training stage, and the association threshold is set to 0.72. For any cell node i and cell node j, the hidden state vectors at time t-1 are (0.15,0.23,0.11,…,0.09) and (0.17,0.21,0.14,…,0.08) respectively. After matrix multiplication, the unnormalized similarity quantization value is 23.7, and the normalized score is 0.81, which is greater than the threshold. Therefore, the value of this pair of nodes is assigned to 0.81 in the original adjacency weight coefficient matrix. This process generates a sparse original adjacency weight coefficient matrix across the entire module, with values ​​concentrated between 0 and 0.85. High values ​​correspond to main current paths and high thermal coupling pairs, while low values ​​correspond to weak coupling or non-main path node pairs. This matrix provides the original input for subsequent sparsification pruning and symmetry constraint correction in S3.5, achieving a significant improvement in the accuracy of topology mapping.

[0068] S3.5: Perform sparse pruning and symmetry constraint correction on the original adjacency weight coefficient matrix to eliminate false weak coupling connections below the dynamic confidence threshold and force the undirected graph physical properties to be satisfied. Finally, output a graph neural network inference model to characterize the actual current main path and thermally coupled strong correlation region between cells, which serves as the core computing engine for online real-time inference of the adjacency matrix.

[0069] The input object is the original adjacency weight coefficient matrix obtained by edge weight prediction decoding. This matrix corresponds to the quantized similarity score between cell nodes and may contain high-confidence valid electrical connections and spurious weak coupling connections introduced by noise or transient fluctuations.

[0070] The original adjacency weight coefficient matrix is ​​compared element-wise with the dynamic confidence threshold set. Edge weights below the threshold are extracted and marked with deletion operation flags in the sparse label matrix.

[0071] The sparse label matrix is ​​applied to the original adjacency weight coefficient matrix, and a matrix mask operation is performed to set the corresponding low-confidence edge weights to zero in order to eliminate the path influence of weakly coupled connections.

[0072] Symmetry constraint correction is performed on the sparsified matrix, and a symmetric verification operator is constructed to ensure the physical properties of the undirected graph. The difference between the values ​​of w(i,j) and w(j,i) at any position of the matrix is ​​compared. If there is a discrepancy, the average value is taken to replace the two.

[0073] Energy normalization is performed on the symmetry-corrected matrix. A normalization factor is constructed using the sum of node degree values. The weights of each edge are converted into dimensionless ratios through fractional operations to ensure that the topology mappings of battery packs of different sizes are comparable.

[0074] The normalized and symmetric sparse adjacency matrix is ​​encapsulated into the online computational core structure of the graph neural network inference model and embedded into the subsequent functional adjacency matrix generation process to achieve accurate characterization of the actual main current path and thermally coupled strongly correlated region between battery cells.

[0075] By using sparsification, masking to zero, symmetry correction, and normalization, the original adjacency weight coefficient matrix from the previous step is transformed into node connection data that satisfies the physical constraints of the undirected graph and has high confidence, thereby significantly improving the accuracy of false connection removal and topology mapping during model inference.

[0076] like Figure 3 As shown, step S4: Input the current topology-sensitive feature vector set into the graph neural network inference model to perform online mapping calculation, so as to output a functional adjacency matrix reflecting the electrical functional adjacency relationship of each cell at the current moment. Specifically, this includes: S4.1: Perform graph structure initialization construction operation on the current topology sensitive feature vector set to generate the current initial cell node feature tensor containing the current voltage jump slope, current phase offset and transient impedance response envelope attributes, as the underlying input data base of the graph neural network inference model.

[0077] The graph structure initialization construction operation for the current topology-sensitive feature vector set includes: defining each cell node as a graph vertex and defining the potential electrical connections between cells as graph edges, and performing a node feature embedding operation to map the current voltage jump slope sequence, current phase offset sequence and transient impedance response envelope data into high-dimensional node feature vectors.

[0078] S4.2: Based on the initial cell node feature tensor and calling the prior knowledge base parameters of massive real vehicle accelerated aging test data trained offline, perform graph convolution aggregation operation to generate the current intermediate layer hidden state feature map characterizing the potential electrical connection strength and thermal coupling transmission probability between cells.

[0079] The initial cell node feature tensor, which includes voltage jump slope, current phase offset, and transient impedance response envelope properties, is processed by graph convolution kernel to aggregate neighbor node information. The size of the aggregation window is set to be consistent with the maximum number of hops in the electrical topology to ensure complete capture of local load path states.

[0080] The neighbor feature vectors obtained from the aggregation process are combined with the central node feature vectors by a weighted combination. The weight coefficients are dynamically generated by the prior knowledge base parameters of the massive real vehicle accelerated aging test data obtained in the offline training phase, so as to reflect the probability of current main path participation and thermal coupling transmission strength under different operating conditions.

[0081] A normalized linear transformation is applied to the weighted combination result to map the dimensionless combination features into a unit-consistent hidden layer feature representation. This transformation eliminates the influence of the gain difference between different sensing channels on the network's computational output.

[0082] By using a nonlinear activation function to perform state nonlinear enhancement processing on the normalized hidden layer feature representation, an intermediate layer hidden state feature map that can highlight the potential high-strength electrical connection between cells is generated. This feature map also retains the heat transfer probability information accumulated in the multi-hop topology path.

[0083] Through continuous aggregation and nonlinear enhancement iteration, the hidden state feature map of the intermediate layer is gradually formed in multiple convolutional layers into a multi-scale feature expression that reflects the potential electrical connection strength and thermal coupling transmission probability between battery cells.

[0084] Through the above processing method, the results of the previous step are transformed into an intermediate hidden state feature map that can be directly used for subsequent electrical path confidence calculation, thereby realizing the real-time disclosure of the functional adjacency relationship of the battery cell under dynamic load conditions.

[0085] For example, in a power battery pack consisting of 96 cells, the initial cell node feature tensor has a dimension of 96×3, where the three dimensions correspond to the voltage jump slope, current phase offset, and transient impedance response envelope, respectively. The aggregation window size is set to 3 hops, and the prior weight coefficients are obtained through training with real vehicle aging data. Under rapid acceleration conditions, the main path participation probability is 0.82, and the thermal coupling transmission strength is 0.65. Normalization transformation is performed on the aggregated features, and the mapping interval is set to [-1,1] to eliminate dimensional differences. ReLU is used as a nonlinear activation function, and values ​​below 0 in the normalization result are set to zero, thereby highlighting high-strength connection features. After three layers of graph convolution operation iterations, the feature amplitude of each node in the hidden state feature map of the intermediate layer on the main path is significantly improved compared to the original input, and the thermal coupling transmission probability remains stable above 0.6 under multiple hop counts in the topology. This output result effectively avoids the risk of static models misjudging non-main path nodes in the subsequent confidence matrix generation and improves the accuracy of equalization parameter adjustment under dynamic load conditions.

[0086] S4.3: Utilize the hidden state feature map of the intermediate layer and the mapping mechanism of the nonlinear activation function to perform the electrical path confidence calculation operation to generate the current original adjacency weight coefficient matrix that quantifies the possibility of an effective current main path between any two cell nodes.

[0087] In this embodiment, the mapping mechanism of the nonlinear activation function refers to the process of converting the linear weighted sum of input data into a nonlinear output by introducing a nonlinear differentiable mathematical function at the output of a neuron, specifically using the hyperbolic tangent function, the sigmoid function, or the ReLU function.

[0088] Based on the hidden state feature map of the intermediate layer, a candidate connection set of electrical paths between cell nodes is constructed, clarifying the contribution value of each node to the potential conduction path of other nodes. For each pair of nodes in the candidate connection set, a mapping mechanism using a nonlinear activation function is applied to transform the linear aggregation result in the hidden state feature vector into a nonlinear confidence score, thereby enhancing the detection sensitivity of weak nonlinear coupling relationships. A uniform normalization process is applied to the output value in the mapping mechanism to eliminate the distortion effect of differences in feature amplitudes between different nodes on the scoring range. Based on the normalized confidence score, a probability threshold determination function is applied, marking node pairs below the threshold as invalid, and retaining node pairs above the threshold as original adjacency relationships. Weight coefficients are calculated for the retained node pairs, directly quantizing the confidence score into original adjacency weight coefficients, and filling the matrix according to the matrix format to generate the original adjacency weight coefficient matrix.

[0089] By using nonlinear activation mapping and weight calculation, the intermediate hidden state feature map result from the previous step is transformed into an original adjacency weight coefficient matrix that quantifies the possibility of an effective current main path between any two cell nodes, thus realizing an explicit topological connection probability characterization based on the hidden state feature vector.

[0090] For example, under the dynamic load condition of a new energy vehicle battery pack, the hidden state feature map of the intermediate layer contains hidden state vectors of 32 cell nodes, each with a dimension of 64. A hyperbolic tangent nonlinear activation function is used to map the differences in hidden state vectors between node pairs, with the mapped output values ​​limited to the range of -1 to 1. Normalization is applied to each output value using the following formula: in The difference between the hidden state vectors of the two nodes. For the output of the activation function, and These represent the minimum and maximum values ​​of the fully connected score set, respectively. A confidence threshold of 0.6 is set; when the normalized score is greater than 0.6, the node pair is retained, and its weight coefficient is recorded as the normalized score value. The matrix is ​​filled into a 32×32 matrix according to the node index to obtain the original adjacency weight coefficient matrix. In this example scenario, the non-zero elements in the matrix are mainly concentrated at the cell locations on the rapidly accelerating load path, verifying that this step can significantly improve the accuracy of effective current main path identification under dynamic operating conditions and provide accurate topological input for subsequent equalization influence domain delineation.

[0091] S4.4: Perform sparsity threshold filtering and symmetry correction on the current original adjacency weight coefficient matrix to eliminate false weak coupling connections caused by measurement noise and eliminate directional deviations, thereby generating a standardized electrical function adjacency matrix that conforms to the undirected characteristics of physical circuits.

[0092] S4.5: Perform topology integrity verification based on the standardized electrical function adjacency matrix to output the final functional adjacency matrix that reflects the actual electrical function adjacency relationship of each cell under complex dynamic conditions at the current moment, and complete the online mapping calculation from feature space to topology space.

[0093] Based on the standardized electrical function adjacency matrix input, the topology integrity verification module is invoked to perform a global traversal analysis of the node connectivity in the matrix, generating a topology connectivity index sequence reflecting network connectivity. Redundancy loop detection is performed on the topology connectivity index sequence, using depth-first search and edge label inversion mechanisms to identify and mark the set of closed paths that may cause electrical loop overcurrent. For the marked closed path set, edge weights are trimmed according to preset physical safety rules, eliminating edges with weight values ​​below the safety threshold to remove false loops. Node degree statistics are performed on the trimmed adjacency matrix, and the global average node degree is calculated using the following formula: in, For each node's degree value, The total number of nodes in the matrix is ​​given. The average node degree is compared with the physical connection reference value. The set of nodes with deviations exceeding the limit is output, and dynamic edge weight repair is performed to restore their physical rationality. Symmetry verification is applied to the repaired adjacency matrix. The undirected graph characteristics are verified through matrix transpose and difference operations with the original matrix. Elements with non-zero difference values ​​undergo bidirectional edge weight filling correction. The corrected adjacency matrix is ​​input into the path redundancy detection module. Path coverage verification processing based on minimum spanning tree is used to confirm that all critical load paths exist in the matrix, eliminating redundant parallel false paths. After completing the above verification, the final matrix is ​​output as the true electrical function adjacency matrix at the current time, realizing online mapping and transformation from feature space to topology space.

[0094] By employing global connectivity traversal, redundant loop removal, node degree balancing correction, symmetry verification, and critical path coverage verification, the standardized electrical function adjacency matrix from the previous step is transformed into real adjacency relationship data that meets physical feasibility and dynamic operating condition adaptability, thereby achieving the expected technical effects of accuracy and usability in topology space mapping.

[0095] For example, in a power battery pack containing 96 cells and 12 module bus nodes, the initial total number of nodes in the matrix is ​​configured to be 108. The global average node degree of the statistical results of each node degree value is 3.25, compared with the average node degree of 3.30 in the reference connection state. The deviation is within the limit. Some nodes, such as ID=35, are detected to have closed loop paths. The edge weights are pruned to 0 and recorded as safety correction flags. The matrix is ​​transposed and differentially verified. There are 12 non-zero element positions in the differential. After filling the bidirectional edge weights, the differential matrix is ​​all zero, which satisfies the characteristics of an undirected graph. In the path redundancy detection, the generated minimum spanning tree covers all main load paths, and 8 parallel false paths are eliminated. The final output real electrical function adjacency matrix, under rapid acceleration conditions, shows a significantly improved consistency between the path distribution calculated by online mapping and the measured main current path, effectively ensuring the accuracy of the subsequent equalization influence domain delineation.

[0096] Step S5: Calculate the betweenness centrality and eigenvector centrality indices of each node in the electrical topology based on the functional adjacency matrix, to delineate the equilibrium influence domain containing only cells with a sudden increase in centrality located on the current load main path. Specifically, this includes: S5.1: Perform shortest path topology traversal processing on the functional adjacency matrix to obtain the minimum electrical hop distance set between any two cell nodes, and count the number of optimal current conduction paths traversing each target cell node based on the minimum electrical hop distance set, thereby generating the original betweenness centrality data sequence characterizing the frequency of a node as a current hub.

[0097] Based on the standardized electrical connection data of the functional adjacency matrix, all cell nodes and their electrical functional connections are extracted from the node set and edge weight set as the basic input for traversal. Weighted shortest path calculation is performed on each pair of cell nodes in the node pair set, using the weight coefficient of the corresponding edge in the functional adjacency matrix as the path length metric, generating a set of minimum electrical hop distances between any two cell nodes. The weighted path calculation results are stored in a structured matrix, where matrix elements represent the minimum hop values ​​between nodes and serve as the index base for subsequent path statistics. Path traversal counting is performed on the minimum electrical hop distance set to count the number of optimal current conduction paths passing through the target node. Using the betweenness centrality fundamental formula, the path statistics data are transformed into the original betweenness centrality data sequence: in, This represents the total number of optimal current paths from node s to node t. This represents the number of nodes v traversed in these paths, where n is the total number of cell nodes. Through a chain-like process from traversal statistics to formula solution, the functional adjacency matrix is ​​mapped to the original betweenness centrality data sequence, thereby achieving a quantitative output of the hub status of nodes in the electrical topology.

[0098] For example, in a power battery pack consisting of 96 cells, the edge weight coefficients represented by the functional adjacency matrix range from 0.05 to 0.95. The minimum electrical hop distance between each pair of nodes is calculated using the weighted Dijkstra algorithm, where high-weight paths correspond to low-hop values, and low-weight paths correspond to high-hop values. The generated distance set is stored in a 96×96 matrix. For target node number 48, the number of shortest paths between all pairs of nodes passing through this node is counted. Path links are retrieved from the distance matrix, and the number of paths passing through node 48 is recorded as 1520, while the number of paths not passing through node 48 is recorded as 2440. Substituting this statistical data into the betweenness centrality formula, with n=96, the original betweenness centrality value of node 48 is calculated as follows: The calculated value is approximately 0.17. This output result is used in the subsequent S5.2 normalization weighting process to eliminate scale effects and dynamically correct the path weights in combination with the load current amplitude. During the verification phase, the equilibrium influence domain of node 48 under high load conditions was successfully identified, the equilibrium response speed was significantly improved, and the system stability remained within the optimization range.

[0099] S5.2: Perform normalized weighting operation based on the original betweenness centrality data sequence to eliminate the influence of battery pack size differences on numerical dimensions, and dynamically correct the path weights in combination with real-time load current amplitude, thereby generating a standard betweenness centrality index vector that reflects the actual shunting ratio of each cell node in the main load path under the current operating conditions.

[0100] Based on the original betweenness centrality data sequence, a normalized weighting operation is performed, such as using Min-Max normalization or Z-Score normalization. The real-time load current amplitude is obtained by directly reading the bus current measurement value of the battery management system (BMS) in the current sampling period, and multiplied by the normalized betweenness centrality value as a weighting coefficient, thereby dynamically adjusting the weight ratio of each path.

[0101] S5.3: Construct a recursive eigenvalue iterative equation using the functional adjacency matrix to simulate the multi-level propagation effect of voltage disturbance in the cell network, and use the cumulative influence value of neighbor nodes generated in the previous iteration as the input basis for this round of calculation, thereby generating an eigenvector centrality score list characterizing the coupling strength between each cell node and its high-weight neighbor nodes.

[0102] Based on the calculation results of the standard betweenness centrality index vector and the topological information of the functional adjacency matrix, the initial state of the recursive eigenvalue iterative equation is constructed, and the initial influence value of each cell node is set as a unified reference constant to ensure the numerical stability of the iterative process.

[0103] The cumulative influence value of neighboring nodes is a core state variable characterizing the comprehensive influence of a battery cell node in the global network topology, and its generation mechanism is as follows: Normalization is performed on the edge weight coefficients in the functional adjacency matrix to convert the connection strength between different nodes into dimensionless proportional values ​​as propagation coefficients in the iterative equation, ensuring that the coupling strength calculation is not affected by absolute dimensions. In the first round of calculation, the iterative equation is called to perform a neighbor node weight accumulation and summation operation on the influence value of each cell node. The current influence of each neighbor node is multiplied by its corresponding propagation coefficient and summed to form the first round accumulated influence vector of that node. The output of the first round accumulated influence vector is used as the input basis for the second round of iteration. In the second round of iteration, the same propagation coefficient matrix is ​​applied to perform neighbor weight summation, realizing the simulated propagation of voltage disturbances between two levels of neighbors and further accumulating the node's influence value. The iterative process is repeated until the change in the node influence vector between two adjacent iterations is lower than a preset convergence threshold. The eigenvector centrality score is calculated using the following recursive iterative formula: Where x is the node influence vector and A is the normalized functional adjacency matrix. The converged node influence vector is normalized and mapped to obtain the feature vector centrality score list for each cell node. By recursively iterating to simulate the multi-level propagation of voltage disturbances, the functional adjacency matrix of the previous step is transformed into a centrality score representing the coupling strength between a node and its high-weight neighbor nodes, thus realizing the second core indicator required for the delineation of the balanced influence domain.

[0104] S5.4: Perform multivariate temporal difference comparison processing on the standard betweenness centrality index vector and the feature vector centrality score list to extract the instantaneous change rate of the centrality index of each cell node in adjacent sampling periods, and compare the instantaneous change rate with the preset dynamic mutation threshold to generate a set of candidate key cell nodes that identify a significant leap in centrality.

[0105] Based on the standard betweenness centrality index vector and eigenvector centrality score list, a multivariate time-series analysis data frame with unified sampling period alignment is established to ensure strict synchronization of each index on the time axis. First-order time-series difference operations are performed on each cell node index sequence in this data frame to calculate the change in centrality values ​​within adjacent sampling periods. The changes are then normalized to eliminate comparison biases between different centrality dimensions. An absolute value-based rate-of-change judgment model is used to compare the normalized changes element-wise with a preset dynamic mutation threshold, and the node positions meeting the mutation conditions are marked in Boolean judgment matrix form. The Boolean judgment matrix is ​​mapped back to the cell node index space to generate a list of instantaneous mutation event markers, and a preliminary set of candidate key cell nodes is constructed by combining the node physical identifiers. Through a multivariate cross-validation mechanism, isolated events caused by mutations in a single index are removed from the candidate node set, retaining nodes that simultaneously mutate in both betweenness centrality and eigenvector centrality, forming a set of candidate key cell nodes that satisfy bidirectional topological importance. By using the aforementioned time-series differential and threshold determination processing method, the centrality quantitative results of the previous step are transformed into a set of candidate key cell nodes that can directly indicate the transition of the dynamic load main path, thus achieving the expected technical effect of accurately capturing significant changes in the state of the topological hub within the millisecond range.

[0106] For example, in a new energy vehicle system equipped with 96 battery cells, the sampling period is configured to 1ms, and the dynamic mutation threshold is set to 0.15. The collected betweenness centrality sequence of a certain battery cell node is 0.48, 0.63, and 0.81, and the eigenvector centrality sequence is 0.35, 0.52, and 0.69, respectively. Performing a first-order difference operation, the betweenness centrality change sequence is obtained as 0.15 and 0.18, and the eigenvector centrality change sequence is 0.17 and 0.17, respectively. After performing amplitude standardization on the above changes, the threshold determination formula is used. Where I is the Boolean value of the mutation determination result, ΔC is the rate of change of centrality, and T is the preset dynamic mutation threshold. Substituting the data into the formula, the rate of change of both types of indicators is greater than 0.15, and the Boolean determination result is true. Therefore, this cell node is marked as a transient mutation node and enters the candidate key cell node set because it simultaneously meets the mutation conditions of both indicators. Within the 5ms window of switching from low-speed operation to rapid acceleration operation, the mutation of the centrality indicator of this node matches the actual load main path change tested, proving that the execution effect can significantly improve the accuracy and response speed of the equilibrium influence domain delineation, and in the subsequent equilibrium start-up parameter reconstruction process, this node is preferentially included in the target combination.

[0107] S5.5: Perform a logical intersection filtering operation based on the candidate key cell node set to remove redundant nodes in the candidate key cell node set that only have a single centrality surge feature but lack bidirectional topological importance, and spatially aggregate the remaining cell nodes that simultaneously satisfy the high betweenness hub characteristic and the high feature vector correlation characteristic, thereby generating the final defined equilibrium influence domain boundary range.

[0108] In this embodiment, the redundant nodes with a single centrality surge feature but lacking bidirectional topological importance refer to candidate nodes that only show a transient change in one of the metrics, betweenness centrality or eigenvector centrality, but perform poorly in the other. These nodes are usually caused by transient noise or local perturbations and are not true structural critical points.

[0109] The cell node that satisfies the characteristics of high betweenness centrality and high eigenvector correlation refers to a node that simultaneously meets two conditions: it is located at the intersection of multiple main load current paths, which means it has high betweenness centrality, and it has strong connections with other high-influence nodes in its sub-network, which means it has high eigenvector centrality.

[0110] Step S6: Based on the power carrying weight of the cell assembly in the topology within the equalization influence domain, dynamically match and generate an instantaneous reconfiguration equalization control parameter set. The instantaneous reconfiguration equalization control parameter set includes: a specific equalization current amplitude and pulse width duty cycle. Specifically, it includes: S6.1: Perform node weight parsing processing on the functional adjacency matrix of the target cell combination within the balanced influence domain to extract a set of power carrying weight coefficients that characterize the power diversion ratio of each cell in the current load main path.

[0111] Based on the input conditions of the functional adjacency matrix of the target cell combination within the balanced influence domain, the matrix parsing module is called to read the edge weight data of each node in the matrix and establish an edge weight mapping table between the node and its neighboring nodes.

[0112] Perform path location calculations on the mapping table to filter out the set of nodes that belong only to the current load main path, and extract the edge weight values ​​of the nodes connected to the main path neighboring nodes to eliminate the interference of weakly coupled nodes that are not on the main path.

[0113] The edge weights are multiplied and weighted by the betweenness centrality index of the node in the main path to generate a preliminary list of node weights that combines physical path carrying information and traffic hub degree.

[0114] The initial node weight list is processed by normalized vector transformation, so that all weight values ​​are mapped to the [0,1] interval, and the stability of the weight vector is improved by noise suppression filter.

[0115] The power shunt ratio calculation formula is called to convert the normalized weight values ​​into a set of power carrying capacity weight coefficients: in, This is the node power carrying capacity weighting coefficient. The denominator is the sum of the normalized node weights, and the denominator is the sum of the normalized weights of all nodes within the current equilibrium influence domain.

[0116] Through the above chain-like analysis, the equilibrium influence domain data from the previous step is transformed into a set of power carrying weight coefficients that can be quantified and directly mapped to the actual power splitting relationship, thus enabling accurate input for subsequent equilibrium current amplitude calculation and pulse width duty cycle mapping.

[0117] S6.2: Perform a multivariate linear mapping operation based on the power carrying weight coefficient set to convert the dimensionless weight ratios into a series of benchmark equalization current demand values ​​that reflect actual heat loss constraints.

[0118] Based on the input conditions of the power carrying weight coefficient set, the dimensionless weight ratio is used as the initial variable for the mapping operation. It is clarified that the weight ratio originates from the quantitative data of the power diversion ratio of the target cell combination in the current load main path within the equilibrium influence domain. The weight ratio is then converted into a series of benchmark equilibrium current demand values ​​that can reflect the actual heat loss constraints through a multivariate linear mapping method.

[0119] Input the power carrying weight coefficient set into the linear mapping matrix, and the elements in each column of the matrix correspond to the cell heat loss correction factor and the thermal current limiting factor.

[0120] A thermal correction is applied to each weight ratio, and the power carrying weight is mapped to the safe current amplitude by multiplying it by a reduction factor based on the current cell temperature.

[0121] The multivariable linear transformation function is invoked to adjust the relative amplitude of the reference values ​​of each cell according to the topology load current shunting characteristics, ensuring that the mapped current demand value sequence does not exceed the system safety limit under the overall power balance constraint.

[0122] The baseline equalization current requirement for a single battery cell is calculated using the following mapping formula: in, As the benchmark equalization current demand value, This is the heat loss correction factor. For power carrying capacity weight ratio, This represents the current total system load power.

[0123] Normalization is performed on each current demand value in the sequence to eliminate scaling errors caused by differences in absolute power values ​​under different operating conditions, resulting in a standardized reference equalization current demand value vector.

[0124] Through the above processing method, the power bearing weight coefficient set of the previous step is transformed into a series of benchmark equalization current demand values ​​that can be directly used for pulse energy calculation, thereby realizing the preset equalization current amplitude based on dynamic load status and heat loss constraints.

[0125] S6.3: Using the aforementioned benchmark equalization current demand value sequence in conjunction with the real-time terminal voltage fluctuation range of the battery pack, pulse energy equivalent conversion is performed to generate an initial pulse width duty cycle candidate set that meets the instantaneous power balance requirements.

[0126] S6.4: Apply switching frequency harmonic suppression constraints and dead time correction processing to the initial pulse width duty cycle candidate set to output a timing-optimized modified pulse width duty cycle control vector.

[0127] The input object is the initial pulse width duty cycle candidate set, which is generated by pulse energy equivalent conversion based on the reference equalization current demand value sequence and real-time terminal voltage fluctuation range in the previous sub-step S6.3. It includes the duty cycle parameter vector of each target cell combination that meets the power balance constraint under the current load condition.

[0128] The initial candidate set of pulse width duty cycles is processed by switching frequency harmonic suppression constraint calculation. Based on the switching frequency of the target equalization execution circuit, the harmonic amplitude index of each candidate duty cycle at that frequency and its harmonics is calculated. The spectral distribution of the duty cycle waveform is analyzed using Fourier expansion, and candidate parameters whose values ​​exceed the preset harmonic amplitude tolerance are eliminated. Dead time correction is performed on the remaining candidate duty cycle parameters. Based on the turn-on delay and turn-off delay parameters of the power switching devices in the equalization execution circuit, the dead time is calculated, and the duty cycle value is corrected using the following formula: in The original duty cycle. To correct the duty cycle, For the switching cycle, This refers to the dead time. The dead time is a fixed delay duration set to prevent shoot-through of the power switching devices in the upper and lower bridge arms. The duty cycle parameter vector, corrected for the dead time, is sequentially mapped to a unified time base and merged with the preceding harmonic suppression constraints to form a timing optimization matrix. A binding relationship is established between each duty cycle parameter and the target cell combination power carrying capacity weight coefficient. Boundary truncation is performed on the timing optimization matrix to ensure that all duty cycle values ​​are within the hardware-allowed conduction ratio range. Through the above chained processing method, the initial pulse width duty cycle candidate set from the previous step is transformed into a corrected pulse width duty cycle control vector that satisfies harmonic suppression and dead time safety constraints, thereby improving the stability of the switching waveform and enhancing the accuracy of energy transfer in the equalization execution circuit under high-frequency dynamic conditions.

[0129] S6.5: Based on the modified pulse width duty cycle control vector and the reference equalization current demand value sequence, perform parameter encapsulation and recombination to generate an instantaneous reconstructed equalization control parameter set containing a specific equalization current amplitude and pulse width duty cycle.

[0130] Step S7: Utilize the instantaneous reconfiguration equalization control parameter set to drive the equalization execution circuit to perform parameterized equalization operations on the target cell assembly within the equalization influence domain, thereby completing a millisecond-level closed-loop equalization action driven by the local topology manifold. Specifically, this includes: S7.1: Obtain the equalization current amplitude and pulse width duty cycle data in the instantaneous reconfiguration equalization control parameter group, and use the pulse width modulation signal generation method to perform time-series mapping processing on the equalization current amplitude and pulse width duty cycle data to generate an original equalization drive pulse sequence containing specific on-time and off-time.

[0131] S7.2: Receive the original equalization drive pulse sequence, and perform channel routing allocation processing on the original equalization drive pulse sequence based on the electrical adjacency relationship of the target cell combination within the equalization influence domain, so as to generate a multi-channel independent equalization drive instruction set corresponding one-to-one with the physical connection port of each target cell combination.

[0132] The input to the received original equalization drive pulse sequence is the multi-cycle pulse sequence data generated in step S7.1, which includes parameters for on-time and off-time. This sequence maintains consistency with the equalization current amplitude and pulse width duty cycle in the instantaneous reconfiguration equalization control parameter group. For this input, the functional adjacency matrix of the target cell combination within the equalization influence domain is first invoked. A matrix index matching mechanism maps the pulse signals to specific physical connection ports, ensuring topological consistency in signal allocation. Subsequently, a port mapping table is used to perform channel number conversion on each pulse sequence, transforming logical channel identifiers into hardware channel numbers to achieve interface compatibility with the equalization execution circuit. Next, timing conflict detection processing is performed on the mapped pulse sequences. Potential concurrent conduction risks are screened based on the node coupling relationships in the adjacency matrix, and mutual exclusion rules are used to mark channel combinations requiring phase misalignment to prevent energy short circuits. Further, a routing allocation algorithm is invoked. Under the premise of satisfying the mutual exclusion rules, each pulse sequence is sorted according to its power carrying weight and sequentially bound to the corresponding hardware channel, forming multiple independent drive pulse signal groups. Finally, a port-specific electrical control parameter package is added to each drive pulse signal group, including trigger edge delay, duty cycle fine-tuning correction value, and amplitude limit threshold, generating a multi-channel independent equalization drive instruction set that corresponds one-to-one with the physical connection port of each target cell. Through the above routing allocation and parameter binding processing method, the result of the previous step is transformed into port-level control instructions that can directly drive the equalization execution circuit, achieving the equalization signal allocation technology effect of millisecond-level matching topology path.

[0133] For example, for a set of equalization influence domains containing four target cell combinations, the functional adjacency matrix shows that nodes 1 and 2 are the main discharge paths, and nodes 3 and 4 are the secondary discharge paths. The input original equalization drive pulse sequence contains four signals with different duty cycles: 0.45, 0.38, 0.52, and 0.40. The port mapping table maps logical channels CH1, CH2, CH3, and CH4 to physical ports P1, P2, P3, and P4, respectively. The routing allocation algorithm sorts the signals according to power carrying weights of 0.32, 0.28, 0.25, and 0.15, and marks P1 and P2 as needing to be phase-staggered to prevent concurrent conduction. In the final output instruction set, P1 is bound to a duty cycle of 0.45 and delayed for 2μs, P2 is bound to a duty cycle of 0.38 and delayed for 5μs, P3 is bound to a duty cycle of 0.52 with no delay, and P4 is bound to a duty cycle of 0.40 with amplitude limited to 1.5A. Collision detection uses the following formula to calculate the energy short-circuit risk coefficient: Where I is the instantaneous current amplitude of the channel, V is the voltage drop at the channel terminals, and P... maxThis represents the maximum allowable instantaneous power for this combination. Actual measurements showed that channel P1 had I=1.8A and V=3.6V at a duty cycle of 0.45, with calculated R=0.432, which is below the preset risk threshold of 0.5, allowing execution. After this step, the equalization execution circuit can accurately drive each target cell combination within 3ms, significantly improving topology path matching and the safety of the equalization action.

[0134] S7.3: Input the multi-channel independent equalization drive instruction set, and use the dead time insertion mechanism of the power switch gate drive circuit to perform safety logic verification and waveform shaping on the multi-channel independent equalization drive instruction set to generate a standardized power switch control signal with anti-shoot-through protection characteristics.

[0135] An input buffer loading operation is performed on the multi-channel independent equalization drive instruction set to ensure that the instruction data of each channel has a unified time base and synchronous triggering conditions before entering the gate drive circuit.

[0136] The logic verification module is invoked to check the turn-on and turn-off timing of each channel according to the preset switching mutual exclusion rules, and illegal timing combinations that may cause the power switching devices in the same bridge arm to turn on simultaneously are eliminated.

[0137] The instruction set that has undergone logical verification is sent to the dead-time insertion calculation unit. Based on the modified pulse width duty cycle control vector generated in the previous step S6.4, the minimum dead-time delay value to be inserted in each conduction cycle is calculated, as follows: in, For the current switching cycle, This is the dead zone ratio coefficient. The minimum dead time required for safe operation of the device. This indicates the actual dead time delay that needs to be inserted during the current conduction cycle.

[0138] The dead time delay value is combined with the original turn-on and turn-off edge information according to the time axis to generate a turn-on / turn-off waveform timing table after inserting the dead time, and this timing table is sent to the waveform shaping module.

[0139] In the waveform shaping module, amplitude limiting is performed on the turn-on and turn-off edges of each channel to suppress overshoot voltage. At the same time, edge slope control is applied to limit the dv / dt of the rising / falling edges to meet EMI constraints.

[0140] After dead-time insertion and waveform shaping, the multiple signals are encapsulated into a standardized power switch control signal set according to the physical channel order. Each of these signals has shoot-through protection and optimized edge characteristics.

[0141] By using dead-time insertion and waveform shaping, the multi-channel independent drive instruction set from the previous step is transformed into a standardized power switch control signal with anti-shoot-through protection, thereby achieving safe drive conditions for the balanced execution circuit.

[0142] S7.4: Apply the standardized power switch control signal to the bidirectional DC-DC converter switching device in the equalization execution circuit, and trigger the high-frequency switching action of the power switch device according to the rising and falling edges of the standardized power switch control signal, so as to establish a controlled transient energy transfer path between the target cell combinations within the equalization influence domain.

[0143] S7.5: Monitor the real-time current flow direction and voltage drop characteristics in the controlled transient energy transfer path, and perform closed-loop feedback correction processing on the real-time current flow direction and voltage drop characteristics in combination with the expected power carrying weight in the instantaneous reconfiguration equalization control parameter group, so as to complete the millisecond-level parameterized equalization operation driven by the local topology manifold and output the final equalization execution status flag.

[0144] The system acquires real-time current flow direction and voltage drop characteristic signals of each target cell combination in the controlled transient energy transfer path established by the equalization execution circuit, and calls high-precision current sensor and voltage sampling module to obtain time-synchronized multi-channel data stream.

[0145] Directional analysis processing is performed on the real-time current flow data, decomposing the current vector into components along the main path of the topology and components along the branch path, forming a vector structured dataset for power load matching.

[0146] Multi-scale piecewise linear fitting is applied to the voltage drop characteristic signal to extract the voltage drop rate and amplitude during transient conduction and turn-off, thereby achieving a quantitative characterization of the instantaneous load state.

[0147] By combining the expected power load weight coefficient set in the instantaneous reconfiguration equalization control parameter set, closed-loop deviation calculation is performed, and the power matching error value is obtained through the following formula: Where ΔP is the power matching error value, I is the measured current component, ΔV is the measured voltage drop, W is the expected power carrying capacity weighting coefficient, and P ref This is the reference power.

[0148] The power matching error value is corrected by the proportional-integral controller to obtain the optimized equalization current amplitude and pulse width duty cycle correction amount, and the equalization execution circuit drive signal is updated in real time to form a closed-loop regulation chain.

[0149] The updated execution action is monitored in real time to verify that the current flow direction and voltage drop characteristics after the correction amount is applied meet the expected power carrying capacity matching range. A balanced execution status flag is generated and used as the status input for the subsequent topology recalibration step.

[0150] By using a closed-loop feedback correction method, the execution signal from the previous step is transformed into real-time adjustment parameters that match the actual power load state, thereby achieving millisecond-level local topology manifold parameterized equalization control.

[0151] Step S8: Monitor the changes in voltage-current topology response characteristics after execution and feed them back to the graph neural network inference model for topology recalibration to generate an updated functional adjacency matrix for adaptive adjustment in the next sampling period. Specifically, this includes: S8.1: Perform secondary synchronous acquisition of multi-channel voltage transient signals and current transient signals at the positive and negative terminals of each cell and key nodes of the module-level bus after the equalization execution circuit has completed its operation, so as to generate a post-equalization topology coupling sampling dataset containing equalization disturbance response information.

[0152] S8.2: Perform sliding window differential enhancement processing based on the post-equalized topology coupling sampling dataset to extract a set of post-equalized topology sensitive feature vectors that reflect the recovery rate of voltage jump slope, current phase shift convergence, and transient impedance response envelope changes after equalization current injection.

[0153] Based on the multi-channel voltage transient signals and current transient signals in the post-equalization topology coupling sampling dataset, a sliding time window truncation mechanism is invoked to form a set of local time series segments containing waveform changes before and after the equalization disturbance within a millisecond-level sampling period.

[0154] The transient voltage components in each local time series segment set are input into the high-order finite difference operation module to calculate the voltage jump slope recovery rate after equalization current injection, and the influence of operating condition background changes is eliminated through dynamic baseline correction.

[0155] By comparing the waveform characteristics and corresponding transient current components during the voltage jump slope recovery process, the convergence of current phase shift is calculated using cross-correlation time-domain analysis, and the ability of equalization disturbance to eliminate phase difference caused by inductive effect and line impedance is evaluated.

[0156] The voltage jump slope recovery rate and the current phase shift convergence are synchronously input into the complex domain impedance spectrum fitting module to generate the updated curve of the connection path impedance after equalization current injection, and the change in the transient impedance response envelope compared with the impedance curve before execution is calculated.

[0157] The voltage jump slope recovery rate, current phase shift convergence, and transient impedance response envelope change are fused and normalized using multidimensional features to form a structured set of post-equalization topology-sensitive feature vectors, which are used to support the model input for the topology recalibration stage.

[0158] Through the above chain processing method, the secondary synchronous acquisition data of the previous step is transformed into a standardized post-equilibrium topology sensitive feature vector that can quantify the impact of equilibrium disturbance on the trend of electrical topology state recovery, thus realizing a high-precision feature basis for topology recalibration.

[0159] S8.3: Construct a topology evolution residual tensor using the post-equilibrium topology sensitive feature vector set and the original functional adjacency matrix output from the preceding steps to quantify the dynamic deviation between the current measured electrical connection state and the model predicted state.

[0160] S8.4: Perform gradient backpropagation fine-tuning on the edge weight parameters of the graph neural network inference model based on the topology evolution residual tensor to correct the mapping logic of the actual main current path between cells and the strongly correlated thermal coupling region within the model and generate a topology recalibration intermediate model.

[0161] The graph neural network inference model parameter access interface is called on the topology evolution residual tensor, and the initial values ​​of the edge weight coefficients of the current model are loaded as the baseline input for fine-tuning operations.

[0162] The topology evolution residual tensor is split into edge weight error submatrices according to the graph edge association relationship. The edge-by-edge error mapping function is used to transform the submatrices into loss gradient reference values ​​to ensure that the gradient direction is consistent with the actual topology deviation.

[0163] The loss gradient reference value is normalized and scaled, and the gradient magnitude is adjusted using a decay factor to prevent overshoot in weight updates, resulting in a dynamically corrected gradient matrix.

[0164] The dynamically corrected gradient matrix is ​​used as the backpropagation input, and a chain gradient is passed through the weight nodes of the multi-layer convolution and attention mechanism of the graph neural network to calculate the gain correction value of the edge weight coefficients of each layer.

[0165] The weight coefficients of each side are additively adjusted using a weight update rule, the mathematical expression of which is: Where w is the current edge weight coefficient, η is the learning rate constant, and Δw is the calculated gain correction matrix.

[0166] The updated edge weight coefficient matrix is ​​written into the model weight storage area, and structural consistency verification is performed to eliminate the asymmetric connections caused by the update, forming a topology recalibration intermediate model.

[0167] Through the gradient backpropagation and weight update processing methods described above, the residual tensor result of the previous step is transformed into a modified edge weight coefficient matrix, thereby achieving accurate recalibration of the lightweight graph neural network in the mapping logic of the actual current main path of the battery cell and the thermally coupled strongly correlated region.

[0168] S8.5: Apply the aforementioned topology recalibration intermediate model to the topology-sensitive feature vector set of the next sampling period for online mapping calculation, so as to output an updated functional adjacency matrix that eliminates historical prediction lag errors and accurately reflects the latest load path distribution.

[0169] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0170] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0171] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An adaptive equalization control method for a power battery pack in a new energy vehicle, characterized in that, Specifically, it includes: S1: Obtain multi-channel voltage transient signals and current transient signals from the positive and negative terminals of each cell in the power battery pack and key nodes of the module-level busbar, and generate the original topology-coupled sampling dataset; S2: Perform sliding window difference enhancement processing on the original topological coupling sampling dataset to extract the topologically sensitive feature vector set; S3: Based on the prior knowledge base of massive real vehicle accelerated aging test data trained offline, a graph neural network reasoning model is constructed; wherein, the prior knowledge base of massive real vehicle accelerated aging test data stores several historical topology-sensitive feature vector sets. S4: Input the current topology-sensitive feature vector set into the graph neural network inference model to perform online mapping calculation and output the functional adjacency matrix; S5: Calculate the betweenness centrality and eigenvector centrality indices of each node in the electrical topology based on the functional adjacency matrix, and delineate the equilibrium influence domain; S6: Based on the power carrying weight of the cell assembly in the topology within the equalization influence domain, dynamically match and generate an instantaneous reconfiguration equalization control parameter set; S7: The instantaneous reconfiguration equalization control parameter group drives the equalization execution circuit to perform parameterized equalization operation on the target cell combination within the equalization influence domain, completing the closed-loop equalization action driven by the local topology manifold.

2. The adaptive equalization control method for a new energy vehicle power battery pack according to claim 1, characterized in that, The topology-sensitive feature vector set includes: voltage jump slope, current phase offset, and transient impedance response envelope.

3. The adaptive equalization control method for a new energy vehicle power battery pack according to claim 2, characterized in that, Step S3 specifically includes: A prior knowledge base of massive real vehicle accelerated aging test data generated during the offline training phase is obtained. Multi-modal alignment and noise cleaning are performed on the historical voltage jump slope sequence, current phase offset sequence and transient impedance response envelope data in the massive real vehicle accelerated aging test data prior knowledge base to generate a standardized graph neural network training sample set. Based on the standardized graph neural network training sample set, a graph structure initialization and construction operation is performed to generate a historical initial cell node feature tensor. Using the initial cell node feature tensor and calling the graph convolutional network architecture, multi-level neighbor node information aggregation and nonlinear transformation processing are performed to generate historical intermediate layer hidden state feature maps. Based on the hidden state feature map of the intermediate layer, the edge weight prediction and decoding operation is performed. The similarity score between the feature vectors of any two cell nodes is calculated using a bilinear mapping function, and the similarity score is converted into a historical original adjacency weight coefficient matrix. The original adjacency weight coefficient matrix is ​​subjected to sparse pruning and symmetry constraint correction to output the graph neural network inference model.

4. The adaptive equalization control method for a new energy vehicle power battery pack according to claim 3, characterized in that, Step S4 specifically includes: A graph structure initialization construction operation is performed on the current topology-sensitive feature vector set to generate a current initial cell node feature tensor containing the current voltage jump slope, current phase offset, and transient impedance response envelope attributes. Based on the initial cell node feature tensor and calling the prior knowledge base parameters of massive offline real vehicle accelerated aging test data, graph convolution aggregation operation is performed to generate the current intermediate layer hidden state feature map. Using the hidden state feature map of the intermediate layer and through the mapping mechanism of the nonlinear activation function, the electrical path confidence calculation operation is performed to generate the current original adjacency weight coefficient matrix; The current original adjacency weight coefficient matrix is ​​subjected to sparsification threshold filtering and symmetry correction to generate a standardized electrical function adjacency matrix. Perform topology integrity verification based on the standardized electrical function adjacency matrix and output the functional adjacency matrix.

5. The adaptive equalization control method for a new energy vehicle power battery pack according to claim 4, characterized in that, The nonlinear activation function is a hyperbolic tangent function, a sigmoid function, or a ReLU function.

6. The adaptive equalization control method for a new energy vehicle power battery pack according to claim 4, characterized in that, Based on the standardized graph neural network training sample set, a graph structure initialization and construction operation is performed, including: Each cell node is defined as a graph vertex and the potential electrical connections between cells are defined as graph edges. A node feature embedding operation is performed to map the historical voltage jump slope sequence, current phase offset sequence and transient impedance response envelope data into a high-dimensional node feature vector. The graph structure initialization construction operation for the current topology-sensitive feature vector set includes: defining each cell node as a graph vertex and defining the potential electrical connections between cells as graph edges, and performing a node feature embedding operation to map the current voltage jump slope sequence, current phase offset sequence and transient impedance response envelope data into high-dimensional node feature vectors.

7. The adaptive equalization control method for a new energy vehicle power battery pack according to claim 1, characterized in that, Step S5 specifically includes: The functional adjacency matrix is ​​subjected to shortest path topology traversal processing to obtain the minimum electrical hop distance set between any two cell nodes. Based on the minimum electrical hop distance set, the number of optimal current conduction paths traversing each target cell node is counted to generate the original betweenness centrality data sequence. Normalized weighting is performed based on the original betweenness centrality data sequence, and the path weights are dynamically corrected in combination with the real-time load current amplitude to generate a standard betweenness centrality index vector. The recursive eigenvalue iterative equation is constructed using the functional adjacency matrix, and the cumulative influence value of neighbor nodes generated in the previous iteration is used as the input basis for the current calculation to generate an eigenvector centrality score list. Multivariate temporal difference comparison processing is performed on the standard betweenness centrality index vector and the feature vector centrality score list to extract the instantaneous change rate of the centrality index of each cell node in adjacent sampling periods, and the instantaneous change rate is compared with the preset dynamic mutation threshold to generate a set of candidate key cell nodes. A logical intersection filtering operation is performed on the candidate key cell node set to generate the final defined boundary range of the equilibrium influence domain.

8. The adaptive equalization control method for a new energy vehicle power battery pack according to claim 7, characterized in that, Perform a logical intersection filtering operation based on the set of candidate key battery cell nodes, including: Redundant nodes in the candidate key cell node set that only have a single centrality surge feature but lack bidirectional topological importance are removed, and the remaining cell nodes that simultaneously satisfy the high betweenness hub property and the high feature vector correlation property are spatially aggregated.

9. The adaptive equalization control method for a new energy vehicle power battery pack according to claim 1, characterized in that, S8: Monitor the changes in voltage-current topology response characteristics after execution and feed them back to the graph neural network inference model for topology recalibration to generate an updated functional adjacency matrix for adaptive adjustment in the next sampling period.

10. The adaptive equalization control method for a new energy vehicle power battery pack according to claim 1, characterized in that, The instantaneous reconfiguration equalization control parameter set includes: a specific equalization current amplitude and a pulse width duty cycle.

Images