Battery anomaly identification and monitoring methods for battery operation and maintenance

Abstract

This invention discloses a battery anomaly identification and monitoring method for battery operation and maintenance, specifically relating to the field of battery testing technology. After a battery enters the site, it synchronously collects temperature difference array data, shell micro-deformation data, cavity gas indicators, and endplate near-end temperature data to form a skin observation vector. A slight current disturbance is applied, and the voltage response and temperature rise difference are recorded to generate a disturbance difference vector. Under a unified time reference, event alignment, same-window shadow trajectory debiasing, and quantile projection normalization are performed to obtain a stable and reliable equivalent short-window characterization vector. This vector is then weighted and matched with historical benchmark fingerprint quality to obtain a consistency score. A multi-physics disturbance coupling index and a historical trajectory deviation index are calculated. A short-window confidence coefficient is output through a discrimination model. The consistency score is scaled and combined with a consistency threshold coefficient to generate a risk label, which is then linked to isolation re-inspection, power limiting, temperature rise limiting, or release archiving. This solves the problem that anomalies within a short time window are easily masked by environmental common modes and artifacts.

Description

Technical Field

[0001] This invention relates to the field of battery testing technology, and more specifically, to a battery anomaly identification and monitoring method for battery maintenance. Background Technology

[0002] With the increasing adoption of battery swapping for heavy-duty vehicle refueling, batteries are frequently transferred between centralized charging stations and charging sites, requiring entry inspection and exit clearance to be completed within a very short time. Common practice within charging stations primarily relies on battery management system reporting, supplemented by rapid checks of appearance, temperature, and basic electrical parameters, to minimize vehicle waiting time and maintain station throughput. Because batteries are switching between different operational phases, environmental factors such as temperature and humidity, casing stress, and transportation vibrations can affect external measurements in a short period. This means that the signals observed within the station contain both the actual state and temporary changes brought about by operating conditions.

[0003] Current rapid inspection methods primarily rely on single measurements or threshold judgments, making it difficult to clearly separate early-stage internal problems from normal fluctuations within a limited timeframe. They also lack a combined approach that utilizes signals from the outer casing and differences before and after minor current disturbances. Because multiple external indicators cannot be consistently assessed simultaneously and compared with the battery's past operating trajectory, the system is prone to misjudging health conditions as either too good or too bad. This can lead to potentially high-risk batteries being put into service or healthy batteries being excessively restricted, ultimately impacting operational safety and asset turnover efficiency. Therefore, a novel anomaly identification and monitoring method is needed that is tailored to the battery swapping operation cycle, primarily using signals from the battery skin side, and combining minor disturbance comparisons with historical trajectory verification. This method should provide reliable release and isolation decisions within a short time window. Summary of the Invention

[0004] To overcome the aforementioned deficiencies in the prior art, this invention provides a battery anomaly identification and monitoring method for battery operation and maintenance. This method involves synchronously acquiring skin observation vectors and applying slight current disturbances to form a disturbance difference vector. After event alignment, debiasing of the same-window shadow trajectory, and quantile projection, an equivalent short-window representation vector is obtained. A consistency score is then obtained through quality-weighted matching. The short-window reliability coefficient is output by combining the multi-physics disturbance coupling index and the historical trajectory deviation index. The coefficient is then scaled and thresholded to generate risk labels and executable instructions, thereby solving the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a battery anomaly identification and monitoring method for battery operation and maintenance, comprising the following steps:

[0006] Step S1: After the battery enters the station, the battery outer surface temperature difference array, shell micro deformation, cavity gas indicator and end plate near-end temperature are synchronously collected at the work station according to the same time reference and encapsulated into a skin observation vector. During encapsulation, the fast variable component and the slow variable component are explicitly distinguished.

[0007] Step S2: Apply a preset slight current disturbance sequence within a safe range and keep the environment unchanged. Record the voltage response, temperature rise difference, and shell micro deformation change before and after the disturbance, and summarize to obtain the disturbance difference vector.

[0008] Step S3: The skin observation vector and disturbance difference vector are time-series aligned and debiased normalized with the station's environmental quantities and recent operational intensity information. For fast variable components, a fixed kernel is established based on the steady-state segment before the disturbance to generate shadow trajectories through window regression, and these are subtracted point by point to remove environmental common modes. The fallback segment is time-reversed at its center time and mirrored with the rise segment. The symmetric difference sequence is calculated to offset the slow drift and thermal hysteresis residuals. The controlled response residuals and symmetric differences are respectively subjected to rank-preserving quantile projection normalization and aggregated by region index to obtain the equivalent short-window representation vector in a unified dimensionless space. The rank-preserving quantile projection normalization is implemented based on a fixed quantile anchor grid. The quantile anchor grid is a set of several quantile control points solidified based on cross-station and cross-batch data during the offline period, and is read-only and not adjusted during the operational period. The projection is a monotonically order-preserving piecewise linear mapping used to map each channel sample to a unified dimensionless interval while maintaining the monotonically unchanged rank, thereby achieving comparability across workstations and batches.

[0009] Step S4: Use the equivalent short window representation vector to perform similarity matching with the battery history benchmark fingerprint stored in the cloud to obtain a consistency score. Calculate the multi-physical perturbation coupling index and the historical trajectory deviation index and input them together into the predefined consistency discrimination model to obtain the short window credibility coefficient. Scale the consistency score with the short window credibility coefficient and compare it with the consistency threshold to generate a risk label.

[0010] Step S5: Issue operation instructions based on the risk label.

[0011] Preferably, step S1 includes:

[0012] The completion of mechanical locking is used as a trigger to generate a unified physical examination time stamp, which is then associated with the environmental quantity within the station and recent operational intensity information.

[0013] Before data collection, perform health self-checks on various sensors for zero point, drift, and connectivity, and generate sensor health flags based on the check results. If any abnormality is found, the process is terminated.

[0014] Using the physical examination time marker as a hardware trigger, all sensor channels are controlled to collect data synchronously within the same sampling cycle.

[0015] Artifact suppression is performed on the acquired data, including: detecting and replacing isolated spikes, performing baseline wrapping on temperature sensing data to counteract condensation drift, smoothing electromagnetic jitter, and verifying data steady state;

[0016] The collected data is mapped to preset region indices according to the battery shape, forming a channel distribution with spatial meaning;

[0017] The processed multimodal data is encapsulated into the skin observation vector according to the regional index and time order, and metadata is attached to explicitly distinguish between fast variable components and slow variable components.

[0018] Preferably, the edge control unit is an industrial-grade embedded device with a built-in hardware time base for generating a unified physical examination time stamp and synchronously triggering multi-channel sensor acquisition and driving a preset current disturbance sequence; the edge control unit includes:

[0019] An integrated isolated acquisition and execution interface with a built-in real-time processing flow is used to complete event locking elastic registration, mirror overlay symmetry difference, window shadow trajectory subtraction and quantile projection normalization processing to synthesize the equivalent short window representation vector.

[0020] The built-in inference module is used to calculate the multi-physics disturbance coupling index and the historical trajectory deviation index locally, call the consistency discrimination model to generate risk labels and issue operation instructions;

[0021] It supports two-way communication with the cloud to obtain historical baseline fingerprints and transmit data back, while providing electrical isolation and time synchronization capabilities.

[0022] Preferably, the amplitude, single-pulse duration, number of pulses, and pulse interval of the slight current disturbance sequence are dynamically determined and applied based on temperature rise margin constraints, charge drift constraints, and terminal voltage disturbance constraints; the temperature rise margin constraints are based on the skin observation vector and on-site environmental quantities to estimate the current heat dissipation capacity to limit the upper limit of temperature rise; the charge drift constraints are based on the battery rated capacity to limit the upper limit of the total charge of the pulses; and the terminal voltage disturbance constraints are based on historical benchmark fingerprints to limit the transient range of terminal voltage.

[0023] Preferably, the perturbation sequence is determined adaptively as follows:

[0024] Based on the skin observation vector, the environmental quantity in the station and the recent operating intensity, the candidate grid is discretized into parameters including amplitude, duration and interval under the three constraints of temperature rise margin, charge drift and terminal voltage disturbance. The visibility target band is set with the historical terminal voltage transient boundary as the hard limit.

[0025] The candidate cell with the smallest single-pulse energy injection (satisfying the three constraints of temperature rise margin, charge drift, terminal voltage disturbance, and historical transient hard limit) is selected and applied to the battery as a single calibration preamble pulse. Only the slight voltage drop and slight temperature rise difference during the rising phase are taken, and the smallest of the two individual visibility ratios is taken as the initial visibility score. At the same time, the smallest of the three margins of voltage, heat, and charge is taken as the safety margin score. The disturbance adaptive coefficient γ is set as the minimum of the initial visibility score and the safety margin score (γ=0 if the edge or any margin is zero).

[0026] The model prioritizes amplitude, followed by duration, and then uses interval as a fallback for reshaping. If the perturbation adaptive coefficient γ is low, the amplitude is increased in small steps, the duration is shortened, and the interval is increased. If the perturbation adaptive coefficient γ is high, the amplitude is fixed, the duration is shortened first, and the interval is increased. If the limit is exceeded, the model is locked back, and the output sequence and start and end markers are displayed.

[0027] Preferably, step S2 includes:

[0028] After the safety interlock verification is passed, the current disturbance sequence is applied synchronously and the high-frequency acquisition of voltage, temperature difference and deformation channels is started, with the disturbance start time mark as the unified trigger. The response segments are divided with the disturbance start and end marks as the boundary, and the event-locked elastic registration is used to correct the response hysteresis misalignment between different sensing channels to achieve timing alignment.

[0029] Robust increment values ​​relative to the disturbance start time marker are calculated for voltage response, temperature difference array, and shell micro-deformation, respectively. The voltage response increment, temperature rise difference increment, and shell micro-deformation increment are obtained by combining the difference between the high quantile and low quantile values ​​within the segment with window averaging, and encapsulated as a disturbance difference vector.

[0030] Preferably, step S3 includes:

[0031] Using the physical examination time marker and the disturbance start and end marker as a unified reference, the skin observation vector and the disturbance difference vector are aligned and their intersections are filtered at the object level.

[0032] Based on the slow variable components of the skin observation vector, the on-site environmental quantity and recent operational intensity information, a fixed kernel window regression relationship is established to generate the window shadow trajectory of the fast variable components. The controlled response residual sequence that suppresses environmental common mode interference is obtained by subtracting the fast variable components point by point.

[0033] After time reversal of the fall segment according to the center time of the segment, it is paired with the rise segment point by point, and the difference between the paired points is calculated to obtain the symmetric difference sequence.

[0034] The controlled response residual sequence and the symmetric difference sequence are subjected to rank-preserving quantile projection normalization to map each physical quantity to a unified dimensionless space, and the marked isolated peak points are subjected to gated replacement; spatial aggregation and minimization interpolation are performed according to the region index to assemble the equivalent short window representation vector.

[0035] Preferably, the battery historical benchmark fingerprint is a robust aggregation result of the equivalent short-window representation vectors of the same battery at multiple sites and multiple operating stages, after quality filtering, time reference unification, and regional index consistency. It includes the typical shape of the controlled response residual, the symmetry description of the symmetry difference, and the distribution profile of the slow variable baseline, and is bound to the frozen parameter vector to fix the quantile boundary and soften the function shape. The acquisition method is to periodically perform quality weight aggregation and outlier sample removal on the historical equivalent short-window representation vectors in the cloud, generate the corresponding region's statistical template with the regional index as the key value, solidify it into a searchable object, and call it in read-only mode during runtime.

[0036] Preferably, the consistency score is obtained in the following way:

[0037] Similarity matching recalls common regions using the region index as the key and generates region quality weights based on data quality labels. For each region, the morphological-symmetric joint score of its fast variable component and historical baseline, as well as the background distribution score of its slow variable component based on a first-order distribution shift metric, are calculated and multiplied to obtain the initial region similarity kernel. To suppress spatial isolation anomalies, the initial similarity kernel is smoothed by neighborhood coordination based on region adjacency. The smoothed region similarity kernel is then weighted and averaged using the region quality weights to obtain a consistency score.

[0038] Preferably, the multi-physics perturbation coupling index is obtained by weighted averaging of the phase lock-in degree and fast component ratio of the voltage, temperature rise, and deformation response residuals of each region; the historical trajectory deviation index is obtained by calculating the maximum value of the distribution offset measure of each region in different response segments and historical benchmarks and then performing a weighted average; the multi-physics perturbation coupling index and the historical trajectory deviation index are used as inputs and processed by a predefined consistency discrimination model with monotonicity constraints, and the output is a short-window confidence coefficient used to scale the consistency score.

[0039] Preferably, step S5 includes:

[0040] Based on the risk label, an operation instruction package is generated, which includes the action type, applicable scope and execution monitoring window. High risk triggers isolation and re-inspection and suspends release; medium risk triggers power limiting and temperature rise limiting strategies and sets re-test time markers; low risk triggers release and archiving.

[0041] Before issuing instructions, a security interlock check is performed. In cases of medium risk, a control token is issued based on the consistency score, short window confidence coefficient and historical trajectory deviation exponential boundary to determine the amplitude limit strength.

[0042] Feedback is collected within the monitoring window. Pre-outbound verification is performed on medium-risk strategies. If the strategy is not followed, the execution feedback is automatically downgraded to high-risk processing.

[0043] Based on the recent operational intensity and historical trajectory deviation index, the risk retest time stamp is calculated and linked to the station's scheduling system to issue routing restriction suggestions to ensure execution; maintenance work orders are generated and bound to the health check mark, and the asset ledger and historical benchmark fingerprint summary are updated.

[0044] The technical effects and advantages of this invention are as follows:

[0045] (1) This invention synchronously collects and encapsulates the skin observation vector under the work station identification and physical examination time mark, applies a slight current disturbance to obtain the disturbance difference vector, and then forms an equivalent short window characterization vector through time alignment and debias normalization, so as to realize the stable characterization of the battery's own state within a short dwell window, effectively solving the problems of environmental common mode, artifact interference and easy misjudgment of single threshold.

[0046] (2) This invention generates a consistency score by using the equivalent short window representation vector and the historical benchmark fingerprint for quality weighted matching, and introduces the multi-physical disturbance coupling index and the historical trajectory deviation index into the discrimination model to obtain the short window credibility coefficient. Combined with the consistency threshold coefficient, the risk label is output and the instruction is executed in conjunction with it, so as to achieve cross-site consistency and traceable handling, effectively solving the problems of insufficient short window credibility, unexecutable decision and decentralized asset management. Attached Figure Description

[0047] Figure 1 This is a flowchart of the battery anomaly identification and monitoring method of the present invention.

[0048] Figure 2 This is a flowchart of the process for obtaining the equivalent short-window representation vector of the present invention. Detailed Implementation

[0049] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0050] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0051] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0052] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0053] See Figure 1 The present invention provides a flowchart of a battery anomaly identification and monitoring method. Figure 1 The battery anomaly identification and monitoring method shown for battery operation and maintenance includes the following steps:

[0054] Step S1: After the battery enters the station, the battery outer surface temperature difference array, shell micro-deformation, cavity gas indicator and end plate near-end temperature are synchronously collected at the work station according to the same time reference and encapsulated into skin observation vector as initial evidence for subsequent processing.

[0055] Background: After the battery enters the station with the vehicle and completes parking positioning, the alignment and locking of the battery with the fixture at the testing station is completed within a short stable window. Within the stable window, the edge control unit issues a trigger command with the station as the main body, requiring the synchronous acquisition of skin-side signals such as the battery outer surface temperature difference array, shell micro-deformation, cavity gas indication, and end plate near-end temperature, without changing the on-site environment. These signals are used to construct external observation evidence at the same moment and serve as the basic input for subsequent slight current disturbances, short-window equivalent characterization, and historical matching, ensuring that subsequent steps are performed under the same time reference and the same spatial reference. The slow variable component will serve as the input base for the same window shadow trajectory, and the fast variable component will serve as the main output channel for controlled transient evidence.

[0056] Furthermore, step S1 includes the following sub-steps:

[0057] Step 101: Inbound Positioning and Time Base Establishment

[0058] After the vehicle stops and completes alignment, the fixture at the inspection station mechanically locks the battery. The edge control unit generates a station identifier and an inspection identifier for this inspection, and triggers an inspection time marker with a unified time base using the inspection identifier. At the same time, it reads the environmental data and the battery's recent operating intensity information from the station system and associates them with the inspection time marker for subsequent steps to call and align under the same time base.

[0059] Step 102: Sensor Placement and Health Self-Check

[0060] After the physical examination time mark is triggered, four types of sensors, namely temperature difference array, shell micro deformation, cavity gas indicator and end plate near-end temperature, enter the in-place self-test state, respectively complete zero point check, drift check and channel connectivity check, generate sensor health mark and bind it to workstation identifier and physical examination time mark. If any channel health mark is abnormal, manual review is prompted and data acquisition is stopped to avoid abnormal channels from introducing artifacts.

[0061] Step 103: Synchronous Acquisition and Hardware Triggering:

[0062] Under the premise that all sensor health indicators are normal, the edge control unit uses the physical examination time mark as the only trigger, requiring the four types of channels to collect synchronously within the same sampling cycle, and to record the start and end marks of the physical examination time mark at the beginning and end of sampling, respectively, to ensure that the observed values ​​of the temperature difference array, shell micro deformation, cavity gas indication and end plate near-end temperature belong to the same time period of the same sample.

[0063] Step 104: Artifact Suppression and Steady-State Confirmation:

[0064] After data collection, using the short windows adjacent to the physical examination time mark as references, the following steps are taken: First, isolated spikes are detected in each channel. If the spike duration is shorter than the window length and the neighborhood variance is higher than the steady-state threshold, the neighborhood median is used to replace the spike and a replacement marker is generated. Then, baseline wrapping is performed on the temperature difference array and the near-end plate temperature to extrapolate the linear trend of the steady-state segment to the physical examination time mark to offset the slow negative drift caused by condensation. Electromagnetic jitter is smoothed with a fixed length and the smoothing marker is retained. After processing, the steady-state segment variance is verified to enter the steady-state threshold with the physical examination time mark as the center. If it is not achieved, the physical examination is stopped and a retest is requested.

[0065] Step 105: Spatial Mapping and Region Indexing:

[0066] The temperature difference array and the micro-deformation of the casing are mapped to the preset area index according to the battery shape, the area position relationship of each channel is clarified, and the end plate near-end temperature and cavity gas indication are assigned and labeled according to the area index.

[0067] Step 106: Data encapsulation and vector generation:

[0068] After artifact suppression and spatial mapping are completed, the temperature difference array, shell micro-deformation, cavity gas indicator and endplate near-end temperature are encapsulated into skin observation vectors according to region index and time sequence. The station identifier, physical examination time mark, sensor health mark, station environmental quantity and recent operation intensity information are added as metadata to the skin observation vectors. Fast variable components and slow variable components are explicitly distinguished in the skin observation vectors. The fast variable components include the temperature difference array and shell micro-deformation, and the slow variable components include the steady-state baseline components of cavity gas indicator and endplate near-end temperature.

[0069] The edge control unit is an industrial-grade embedded device installed at the inspection station, consisting of a processor, erasable and rewritable memory, hardware time base, isolated acquisition and execution interface, communication unit, and power supply and protection structure. After the vehicle battery is aligned, it generates a station identifier and inspection time marker, synchronously triggers the temperature difference array, shell micro-deformation, cavity gas indication, and endplate near-end temperature acquisition according to a unified time base, and drives the battery to execute a slight current disturbance sequence within a safe range. The device has a built-in real-time processing flow to complete event-locked elastic registration, mirror superposition symmetry difference, same-window shadow trajectory subtraction, and quantile projection normalization, forming a skin observation vector and a disturbance difference vector, and synthesizing an equivalent short-window representation vector. It calculates the multi-physical disturbance coupling index and historical trajectory deviation index through local inference, calls a predefined consistency discrimination model to obtain the short-window confidence coefficient, compares it with the consistency threshold to generate a risk label, and then issues issuance, restriction, or isolation instructions. The device supports bidirectional communication with the cloud to read historical benchmark fingerprints and transmit inspection data back, providing time synchronization capability and electrical isolation to ensure measurement consistency and safety.

[0070] Step S2: Apply a preset slight current disturbance sequence within a safe range and keep the environment unchanged. Record the voltage response, temperature rise difference, and shell micro-deformation change before and after the disturbance, and calculate the disturbance difference vector.

[0071] Abstract: A preset sequence of slight current disturbances is applied to the battery through an isolated current injection loop, and all observation channels are synchronously triggered using a hardware time base. This disturbance aims to induce a controllable and measurable multi-physical response within a short period of time, distinguishing between the battery's own state and the common-mode influence of the environment, providing evidence of controlled-source changes for subsequent equivalent short-window characterization and historical matching. After completing elastic registration by marking the start / removal time of the disturbance, each channel is divided into three segments according to the time axis: rising segment, steady-state segment, and falling segment. Based on this, the voltage response increment, temperature rise differential increment, and casing micro-deformation increment are calculated and encapsulated as a disturbance difference vector. The purpose of these three segments is to quantify the response of the same fast variable component at different stages, facilitating subsequent mirror superposition and segment shift measurement. The slow variable component is used to record the background at that time and does not participate in mirror or incremental calculations.

[0072] Furthermore, step S2 includes the following sub-steps:

[0073] Step 201: Disturbance preparation and safety interlock verification

[0074] After the physical examination time mark is triggered, the edge control unit first confirms that the temperature difference array, shell micro deformation, cavity gas indicator and end plate near-end temperature channel are in a usable state based on the sensor health mark. Then, it confirms that the cooling, ventilation and electrical isolation meet the safety conditions based on the station environmental quantity and recent operation intensity information. After confirming that there are no abnormalities, it generates disturbance start time mark and disturbance cancellation time mark, and binds the two with the work station identifier and physical examination time mark as time anchor points for subsequent synchronous acquisition and segmented processing.

[0075] Step 202: Disturbance application and synchronous acquisition triggering:

[0076] While keeping the environmental quantities within the station constant, the edge control unit applies a slight current disturbance sequence through the isolated current injection loop according to the disturbance start time marker, and uses this marker as the sole trigger to synchronously start the high sampling acquisition of the voltage response channel, the temperature difference array channel, and the shell micro-deformation channel. At the same time, it records the acquisition time corresponding to the disturbance cancellation time marker, ensuring that the three types of observations belong to the same time reference and the same sampling cycle throughout the entire disturbance process.

[0077] Step 203: Fragment segmentation and event anchoring alignment

[0078] After data acquisition, the observation sequence is divided into three segments—rising segment, steady-state segment, and falling segment—based on the disturbance start time marker and disturbance withdrawal time marker. Elastic registration with event locking is performed on these three segments across the voltage response, temperature difference array, and shell micro-deformation channel. This corrects the misalignment caused by response lag in different channels to the same event time axis, outputting the aligned segment sequence to ensure that subsequent change calculations are based on comparable time-series positions.

[0079] Step 204: Extraction of basic changes

[0080] With the segment division and event anchoring alignment completed, the edge control unit calculates the basic statistics within the segment for the voltage response channel, temperature difference array channel and shell micro-deformation channel in the rising segment, steady state segment and falling segment respectively, based on the disturbance start time mark and disturbance cancellation time mark, and forms a basic change sequence.

[0081] Step 205: Extraction of Changes and Assembly of Regions

[0082] On the basic change sequence, robust increment values ​​of voltage response increment, temperature rise differential increment, and shell micro deformation increment are calculated respectively. The robust increment value is obtained by combining the difference between the high quantile and low quantile within the segment with window averaging to suppress isolated spikes. The robust increment value is obtained by combining the difference between the high quantile and low quantile within the segment with window averaging, and is denoted as voltage response increment, temperature rise differential increment, and shell micro deformation increment respectively. The segment type, region index, and reference relationship of disturbance start time mark and disturbance removal time mark corresponding to each increment are recorded.

[0083] Step 206: Vector Encapsulation and Availability Registration

[0084] The field order of the disturbance differential vector is fixed as [region index, voltage response increment, temperature rise differential increment, shell micro-deformation increment, fragment type, replacement tag set], and is registered together with the workstation identifier, physical examination time tag, disturbance start time tag, disturbance cancellation time tag, station environmental quantity and recent operating intensity information.

[0085] The explanation states that the slight current disturbance sequence refers to a combination of restricted current pulses or steps applied by the edge control unit within the physical examination time stamp window. Its amplitude, single pulse duration, number of pulses and pulse interval simultaneously satisfy the temperature rise margin constraint, charge drift constraint and terminal voltage disturbance constraint, and ensure that the sampling bandwidth can completely cover the frequency components of the sequence without triggering any protection or alarm.

[0086] Furthermore, taking the three physical constraints directly related to safety and measurability—temperature rise, charge, and terminal voltage—as boundaries, the sequence parameters are jointly limited by the current heat dissipation capacity given by the on-site environmental quantity and the skin observation vector, the charge perturbation given by the rated capacity and the pulse charge quantity, and the voltage perturbation margin given by the historical benchmark fingerprint. This ensures that the parameters are small enough to not change the environmental and safety status, and large enough to obtain a distinguishable controlled response within a short window, which facilitates the subsequent calculation of the perturbation difference vector and the equivalent short window characterization vector.

[0087] The temperature rise margin constraint is based on the skin observation vector and the environmental quantity in the station obtained in step S1. The current heat dissipation capacity and the upper limit of short-term temperature rise are estimated, and the current amplitude and single pulse duration are selected to ensure that the predicted temperature rise does not exceed the safe temperature rise margin. If it is not satisfied, the amplitude is reduced or the duration is shortened proportionally until it is satisfied.

[0088] Among them, the charge drift constraint estimates the charge change based on the battery's rated capacity and the charge of a single pulse, and limits the total charge of each pulse and the sequence to not exceed the small drift limit; if it exceeds the limit, the number of pulses is reduced or the duration of a single pulse is shortened.

[0089] Among them, the terminal voltage disturbance constraint compares the predicted value of the terminal voltage transient with the steady-state fluctuation range of the historical benchmark fingerprint to limit the terminal voltage change to not exceed the safety disturbance margin; if it is critical, the pulse interval is appropriately lengthened or the current amplitude is reduced.

[0090] Among them, the recovery and interval setting estimates the controlled response fall-off time based on recent operating intensity information, ensures that the interval between adjacent pulses is not shorter than the fall-off time, and limits the total energy of the sequence to within the safe energy limit, ensuring that the rising segment, steady-state segment and fall-off segment can be clearly separated and observed.

[0091] Among them, spectrum and sampling matching ensure that the main frequency corresponding to the pulse rising edge and duration falls within the effective range of the current sampling bandwidth, and retains an observation margin of at least one recovery period to ensure the distinguishability of the subsequent controlled response residual and the disturbance difference vector.

[0092] The amplitude table, time table, and upper limit of the number are obtained by taking the most stringent feasible combination of the above three constraints, which is the quantitative configuration of this slight current disturbance sequence. This configuration is executed by the edge control unit at the disturbance start time mark and archived together with the physical examination time mark, workstation identifier, station environmental quantity and recent operating intensity information for direct call in step S3 alignment and debiasing normalization.

[0093] In one possible embodiment, the sequence of slight current disturbances is determined through the following adaptive process:

[0094] Constraint-driven candidate lattice construction: Based on the skin observation vector and on-site environmental quantities, the current heat dissipation capacity is derived. Combined with recent operational intensity information and health status assessment, the acceptable ranges of amplitude, single-pulse duration, and pulse interval are determined under temperature rise margin constraints, charge drift constraints, and terminal voltage perturbation constraints, respectively. The intersection of the three constraints (temperature rise margin constraints, charge drift constraints, and terminal voltage perturbation constraints) is discretized into feasible candidate lattices of amplitude-duration-interval. The terminal voltage transient boundary given by the historical benchmark fingerprint is used as a hard limit. Visibility is defined as the voltage drop required for target detection that is measurable and the temperature rise difference is within the threshold. The target band (visibility target band is a preset, quantified numerical range that defines the minimum measurable response and the maximum allowable safe response that must be achieved after applying a perturbation for subsequent analysis to be effective. Its lower limit is determined by the noise level and recognition sensitivity of the measurement system, while its upper limit is determined by the normal fluctuation range and safety margin of the historical benchmark fingerprint); thus, subsequent selection within this candidate range can ensure that both safety and measurability are simultaneously achieved; intention: use three constraints to first converge the parameter space, and then set a unified target band to provide a bounded and operable solution domain for subsequent single calibration and one-time selection;

[0095] Single calibration and perturbation adaptive coefficient γ generation: Select the candidate cell with the lowest perturbation energy to the battery as the calibration leader pulse. After application, record the two direct quantities of voltage drop and temperature rise difference of the battery in the rising phase. Calculate the response visibility ratio relative to the visibility target band. If the terminal voltage transient touches the historical reference fingerprint boundary, it is immediately marked as a hard constraint warning.

[0096] The explanation is that the candidate cell with the lowest applicable perturbation energy has the least impact on the battery, making it the safest and most conservative choice among all feasible options, perfectly aligning with the initial intention of a trial-and-error, low-risk calibration preamble. The response visibility ratio is a dimensionless scalar used to quantify the actual effect of the calibration preamble relative to the desired target position. The calculation involves two steps:

[0097] Calculate the individual ratios for voltage drop and temperature rise difference separately: For voltage drop, the individual ratio r_voltage = (actual measured voltage drop value - target band lower limit) / (target band midline value - target band lower limit); map the measured values ​​to a scale based on the lower half of the target band; calculate the individual ratio r_temperature for temperature rise difference = (actual measured temperature rise difference - target band lower limit) / (target band midline value - target band lower limit);

[0098] The two individual ratios are combined to generate an initial visibility score. The combination strategy needs to be defined in advance as taking the minimum of the two: R = min(r_voltage, r_temperature). This means that the initial visibility score is determined by the physical quantity with the weakest response. An initial visibility score less than 0 indicates that the response has not reached a measurable level, equal to 1 indicates that the response has just reached the ideal target midline, and greater than 1 indicates that the response is too strong.

[0099] The voltage margin, thermal margin, and charge margin are calculated based on the transient boundary of the terminal voltage, the upper limit of the temperature rise margin, and the upper limit of the single-pulse charge, respectively (all are inverse quantizations of the monotonic mapping of "used / available", ranging from zero to one); the minimum value of the three margins is taken to obtain the safety margin score, and then the disturbance adaptive coefficient is obtained by γ=min(initial visibility score, safety margin score); if the transient voltage has reached the historical reference fingerprint boundary or any margin is zero, the disturbance adaptive coefficient γ is directly set to zero; the disturbance adaptive coefficient γ represents the excitation amplitude margin that can be increased without exceeding the boundary; the advantage is that a single adjustment amount for the current operating condition can be obtained without multiple rounds of iteration, avoiding the introduction of additional thermal burden by frequent trials;

[0100] One-time selection and shaping based on the perturbation adaptive coefficient γ: Read the perturbation adaptive coefficient γ, candidate grid, and terminal voltage transient hard constraint, and perform one-time selection and shaping in a fixed priority order of amplitude first, duration second, and interval as a fallback; when the perturbation adaptive coefficient γ is too low, indicating insufficient visibility, adjust it step by step along the amplitude axis without touching the terminal voltage transient boundary, and reduce the single pulse duration to the feasible lower limit and take the pulse interval above the candidate grid to suppress temperature rise accumulation; when the perturbation adaptive coefficient γ is too high, indicating sufficient visibility, keep the amplitude unchanged, prioritize shortening the duration to reduce energy injection, and increase the pulse interval to further dilute the thermal burden; if any adjustment will touch the temperature rise margin constraint or charge drift constraint, immediately fall back to the previous feasible point and lock the parameters, output a slight current perturbation sequence and a unified perturbation start and end mark for subsequent steps to call; Intended explanation: drive unidirectional one-time shaping with a single coefficient, obtain a visible and safe sequence under the protection of three constraints and hard boundaries, and avoid forming closed loop repetition and additional heat accumulation.

[0101] The division of the rising segment, steady-state segment, and falling segment is as follows: using the disturbance start time mark and disturbance removal time mark as boundaries, and according to the joint criterion of voltage response derivative and variance: the rising segment is when the derivative turns from zero to positive and exceeds the quantile threshold until the derivative falls back to near zero and the variance drops to the steady-state threshold; the steady-state segment is when the derivative is near zero and the variance is stable; and the falling segment is when the derivative is negative after removal and exceeds the quantile threshold until the derivative and variance simultaneously return to the steady-state threshold.

[0102] The elastic registration implementation method of event locking is as follows: using the voltage response as the reference channel, the derivative peak and slope change point are extracted as event anchor points in the rising and falling segments respectively; the cross-correlation delay between the temperature difference array and the shell micro-deformation channel and the reference channel is calculated in the same time window and limited to a small range to obtain the channel delay estimate; a monotonic time mapping is applied to the time axis of each channel to align the segment endpoints with the event anchor points while keeping the relative order within the segment unchanged, and the aligned segment sequence is output.

[0103] The processing method for the peer shadow trajectory and the environmental common mode term is as follows: taking the steady-state period before the disturbance as the training window, using the in-station environmental quantity and recent operational intensity information as inputs and each observation channel as outputs to establish the peer regression relationship, and generating the predicted trajectory as the peer shadow trajectory along the same time axis throughout the entire disturbance process; the component obtained by subtracting the peer shadow trajectory from the observation sequence is the controlled response residual sequence after the environmental common mode term is removed.

[0104] The explanation states that the on-site environmental parameters refer to the external environment and workstation conditions related to the observation during the on-site period, including ambient temperature, humidity, ventilation and cooling conditions, vibration acceleration and electromagnetic background; the recent operating intensity information refers to the battery's load, power level, braking and climbing frequency, temperature trajectory and rate trajectory during a recent period of driving and charging / discharging.

[0105] Step S3: The skin observation vector and disturbance difference vector are time-series aligned and debiased and normalized with the on-site environmental quantities and recent operational intensity information to synthesize an equivalent short-window representation vector, which is used to stably represent the battery's own state.

[0106] Background Summary: Step S3 uses the physical examination time marker as the time reference and the workstation identifier as the spatial reference, calling the skin observation vector, disturbance difference vector, on-site environmental quantity, and recent operational intensity information to perform time-series alignment and de-biasing normalization processing. The aim is to construct an equivalent short-window representation vector that can reflect the battery's own state and can be directly compared with historical benchmark fingerprints, providing input with the same dimensions, time axis, and spatial index for the consistency matching and index calculation in Step S4. A fixed-kernel window regression is established using slow variable components, on-site environmental quantity, and recent operational intensity information to generate fast variables. The shadow trajectory of the component under the same window; the two are subtracted point by point to obtain the controlled response residual sequence, realizing the quantitative removal of background common mode; the time reversal of the fall segment of the fast variable component and the mirror image of the rise segment are superimposed to calculate the symmetric difference sequence, which cancels the slow drift and thermal hysteresis residual (the slow variable component does not participate in the mirror operation, only its baseline is retained for comparison); the controlled response residual sequence and the symmetric difference sequence are respectively subjected to rank-preserving quantile projection normalization, which is mapped to a unified dimensionless interval; the slow variable baseline component is not normalized, and is output as the zero reference / background distribution along with the equivalent short window representation vector.

[0107] For further details, please refer to [link / reference]. Figure 2 The flowchart for obtaining the equivalent short-window representation vector is shown below. Step S3 includes the following sub-steps:

[0108] Step 301: Object-level time reference unification and regional intersection filtering

[0109] Using the physical examination time marker, disturbance start time marker, and disturbance withdrawal time marker as a unified time reference, the skin observation vector and the disturbance difference vector are aligned at the object level to verify that the two vectors have consistent regional indices, consistent time references, and consistent field order. Inconsistent items are filtered by intersection according to regional indices, records without common areas are removed and removal marks are generated, and the dockable object pairs (skin observation vector entries, disturbance difference vector entries) are output. Event-level registration is no longer performed on the original channels.

[0110] Step 302: Fixed-kernel window regression, shadow trajectory generation, and controlled residual extraction

[0111] Using the slow variable components of the skin observation vector, the on-site environmental quantity, and recent operational intensity information as inputs, and the fast variable components of the skin observation vector as outputs, a fixed-kernel window regression relationship is established based on the steady-state data before the disturbance. During the entire disturbance process, the window shadow trajectory of the fast variable is generated along the same time axis. The controlled response residual sequence is obtained by subtracting the corresponding shadow trajectory from the fast variable observation value in the object pair. The environmental common mode term is defined as the point-by-point estimate of the shadow trajectory.

[0112] The explanation of the fixed-kernel peer regression relationship refers to establishing a deterministic, non-parametric mapping from input to output within the same time window (same window) of the same physical examination, using only data collected during the steady-state period before the disturbance: The input consists of slow variable baseline components, on-site environmental quantities, and recent operational intensity information; the output is the fast variable channel (the original channel corresponding to the controlled response residual). Weighted values ​​are calculated for each moment on a unified time axis using a kernel function of fixed shape and bandwidth to obtain the predicted value for that moment (the peer shadow trajectory). The fixed kernel indicates that the kernel function type and bandwidth are determined and frozen during the offline phase based on historical samples and cross-validation, and no further parameter adjustments are made during runtime. The regression relationship represents the contribution of any time point within the same window, determined by the kernel weights based on its distance from the steady-state sample in time and input space. The output is a predicted sequence with the same dimension, region index, and sampling rhythm as the equivalent short-window representation vector. This approach ensures that the shadow trajectory only reflects the common mode changes caused by the current environment and intensity, without mixing in information after disturbance, thus providing a reproducible and auditable baseline for common mode elimination and dimensionless processing of subsequent shadow reduction observations;

[0113] Step 303: Ascending-Falling Mirror Pairing and Symmetric Difference Construction

[0114] The falling segment is time-reversed with its center time as the axis and then paired with the rising segment point by point. The difference between the paired points is calculated to obtain a symmetric difference sequence, which is used to offset the slow drift and thermal hysteresis residuals and retain the controlled fast component (the symmetric difference sequence is only generated in the fast variable channel, and the slow variable component does not participate in the mirror operation); the corresponding steady-state baseline component is retained as a zero reference for subsequent trade-offs.

[0115] Step 304: Rank-preserving quantile projection normalization and spike gating replacement

[0116] Rank-preserving quantile projection normalization is performed on the controlled response residual sequence and the symmetric difference sequence respectively, mapping each channel to a unified dimensionless interval. At the same time, the quantile projection parameters are output and bound to the corresponding region index. For isolated spikes caused by acceleration bursts, the median of the neighborhood is replaced according to the replacement marker trigger gate, and the replacement marker serial number is added in this step to ensure subsequent traceability.

[0117] Step 305: Region Aggregation, Minimum Imputation, and Slow Variable Alignment

[0118] Using the regional index of step S1 as the master table, the dimensionless results of the fast variables are aggregated by region. For the missing regions, the estimated values ​​of the window shadow trajectories in the region are interpolated with the neighborhood quantile projection parameters, and interpolation marks are generated. At the same time, the baseline components of the slow variables are aligned according to the same regional index to ensure that the fast and slow variables are consistent in spatial dimension.

[0119] Step 306: Assembly and cache registration of equivalent short-window representation vectors

[0120] Assemble the equivalent short-window representation vector according to a unified field order. The fields include [region index, dimensionless component of controlled response residual, dimensionless component of symmetric difference, dimensionless component of slow variable baseline, quantile projection parameter reference, replacement tag serial number, interpolation tag, and removal tag], and are accompanied by workstation identifier, physical examination time tag, disturbance start time tag, and disturbance withdrawal time tag. After completing the availability registration, write it to the edge cache for direct reading in step S4 for the calculation of consistency score, multi-physical disturbance coupling index, and historical trajectory deviation index.

[0121] Step S4: Use the equivalent short window representation vector to perform similarity matching with the battery history benchmark fingerprint stored in the cloud to obtain a consistency score. Calculate the multi-physical perturbation coupling index and the historical trajectory deviation index and input them together into the predefined consistency discrimination model to obtain the short window credibility coefficient. Scale the consistency score with the short window credibility coefficient and compare it with the consistency threshold to generate a risk label.

[0122] Background: This project utilizes an equivalent short-window representation vector as the current evidence source and a battery historical benchmark fingerprint as the reference baseline. It combines workstation identifiers and examination time stamps to complete a continuous computation chain from representation to matching to credibility to judgment at the edge control unit. The technical motivation lies in the fact that short-stay detection inevitably suffers from environmental common-mode residues and limited sample windows. Directly using point-to-point errors for threshold judgment can easily lead to inflated consistency or false alarms. Therefore, this step first generates a consistency score through quality weighting and neighborhood coordination. Then, it constructs a multi-physical perturbation coupling index and a historical trajectory deviation index from two orthogonal dimensions: whether multi-channel synchronization under control is normal and whether the trajectory is abnormal relative to itself. A predefined consistency discrimination model outputs a short-window credibility coefficient. The consistency score is then adaptively scaled and compared with the consistency threshold coefficient to finally provide a risk label. The battery historical benchmark fingerprint is a summary of cross-site and cross-batch operational health checks of the same battery accumulated in the cloud over a long period of time. It includes controlled response residual morphological elements, symmetry difference elements, and slow variable baseline distribution elements that correspond one-to-one with the regional index. It is obtained by quality filtering, time alignment, and robust aggregation of the equivalent short window representation vectors of each time, and then freezing them into the reference objects referenced by the frozen parameter vector. The core objective of this step is to transform the uncertainty of short window data into interpretable and reliable corrections while maintaining the consistency threshold coefficient of the unique adjustable factor, and to transform the judgment results into the basis for executable operational instructions.

[0123] Furthermore, the consistency score is obtained through the following methods:

[0124] Step 411: Using the region index as the retrieval key, recall the common region set from the equivalent short window representation vector and the battery history benchmark fingerprint; regions with removal marks are directly removed and not included in the calculation; for the remaining regions, read the coverage ratio of replacement marks and interpolation marks, and obtain the region quality weight according to the predetermined mapping. The mapping rule is that the larger the replacement coverage and the larger the interpolation coverage, the smaller the weight. The weight is limited to between zero and one and is not zero to preserve evidence of spatial continuity; after completion, output the common region set and region quality weight.

[0125] Step 412: Within the public area, read the dimensionless components of the controlled response residual and the rising and falling segments of the dimensionless components of the symmetry difference; calculate the morphological similarity score and the symmetry similarity score respectively, and combine them with fixed weights to form a morphological-symmetry joint score, which is output for the next step of synthesis with background distribution evidence;

[0126] The morphological similarity score can be obtained in any of the following ways:

[0127] One is the trend direction consistency rate, which is the proportion of the two sequences whose first-order difference signs are consistent;

[0128] The second is the amplitude ratio closeness, which is the amplitude ratio deviation between two sequence segments converted to a fraction of zero to one according to a fixed mapping;

[0129] The explanation is that the symmetry similarity score is calculated by comparing the falling segment with the rising segment point by point after mirroring it at the center time, calculating the normalized absolute deviation and taking their complementary score.

[0130] The explanation is as follows: the fixed weights are taken from the frozen parameter vector. During the offline training period, the labeled pairs of the historical equivalent short window representation vector and the battery historical benchmark fingerprint are used as samples. The target of minimizing the misjudgment risk is minimized through cross-validation and then fixed after site consistency verification. During the runtime, it is read-only and not adjusted.

[0131] Step 413: Within the same region, read the dimensionless components of the slow variable baseline, calculate its first-order distribution shift measure on the same time axis as the historical baseline, and normalize it to obtain the background distribution score (value range 0-1); multiply the morphological-symmetric joint score with the background distribution score to obtain the initial similarity kernel of the region; to suppress isolated anomalies, perform a weighted average of the initial similarity kernels of adjacent regions based on the neighborhood set of the region index with a fixed smoothing coefficient to obtain the region similarity kernel after neighborhood coordination;

[0132] The explanation is as follows: The first-order distribution shift metric is used to quantify the difference between the current short window and the historical baseline at the distribution level. Specifically, under a unified time reference, the dimensionless component samples of the slow variable baseline of the current short window and the historical baseline are taken respectively. An equally spaced quantile grid (a number of fixed quantile points within a closed interval from zero to one) is constructed. The mean of the absolute values ​​of the differences between the current quantile function and the historical quantile function at the corresponding quantile points is calculated, and then normalized using a scale reference of the historical distribution (such as a fixed quantile span) to obtain a dimensionless, non-negative distribution shift metric. The larger the metric, the more significant the current distribution shift relative to the historical distribution.

[0133] Step 414: Calculate a weighted average of the similarity kernels and regional quality weights for all regions to obtain a consistency score; at the same time, record the consistency input elements used for calculation, including the common region set, regional quality weights, neighborhood set identifiers, and fixed smoothing coefficient identifiers.

[0134] Furthermore, the multi-physics perturbation coupling index is obtained as follows:

[0135] Using the region index as a unit, the rising and falling segments of the dimensionless components of the three controlled response residuals in the equivalent short window characterization vector—voltage response, temperature rise difference, and shell micro-deformation—are read, and the voltage response is used as the reference channel.

[0136] Calculate the phase lock-in degree: In the rising and falling segments, respectively, calculate the normalized time delay consistency degree corresponding to the cross-correlation peak position of the three channels to the reference channel and perform an equal-weighted average with the peak position order consistency rate to obtain the phase lock-in degree of the segment. Then, perform an equal-weighted average of the segment results to obtain the regional phase lock-in degree.

[0137] Calculate the fast component ratio: sum the absolute integrals of the dimensionless components of the three-channel symmetrical difference in the rising and falling segments, and then sum the absolute integrals of the dimensionless components of the controlled response residuals in the same segment. Take the regional average of the summation values ​​to obtain the regional fast component ratio.

[0138] The regional coupling score is obtained by multiplying the regional phase lock-in degree by the regional fast component ratio. The regional coupling scores of all common areas are weighted and averaged according to the regional quality weight to output the multi-physical disturbance coupling index, which is then bound and archived with the workstation identifier and physical examination time mark.

[0139] Furthermore, the historical trajectory deviation index is obtained as follows:

[0140] Using regional indexes as units, the equivalent short-window representation vector and the battery historical baseline fingerprint are distributed and shifted in the rising segment, steady-state segment, and falling segment, respectively. Specifically, a fixed quantile grid is constructed for the dimensionless components of the slow variable baseline and the dimensionless components of the controlled response residual under a unified time reference. The uniform average of the absolute values ​​of the differences between the current quantile function and the historical quantile function at each quantile point is calculated. The segment distribution offset metric is obtained by dimensionless transformation using a frozen scale reference. The segment correction coefficient is obtained by looking up the table according to the recent operating intensity information and then multiplied to form the segment deviation. The maximum value of the three types of segment deviation in the same region is taken as the regional deviation score. The regional deviation scores of all common regions are weighted according to the regional quality weight to calculate the weighted average, output the historical trajectory deviation index, and bind it to the workstation identifier and physical examination time mark for archiving.

[0141] Furthermore, the predefined consistency discrimination model uses the multi-physics perturbation coupling index and the historical trajectory deviation index as dual inputs, and outputs a short-window confidence coefficient. During the training period, the predefined consistency discrimination model is fitted with the labels of faulty and normal samples, and the parameters are frozen into a frozen parameter vector, which is not adjusted during the runtime. The predefined consistency discrimination model has a monotonicity constraint: when the multi-physics perturbation coupling index increases, the short-window confidence coefficient increases; when the historical trajectory deviation index increases, the short-window confidence coefficient decreases. The short-window confidence coefficient scales the consistency score, suppressing accidental high scores with low confidence and amplifying true consistency with high confidence, thereby improving the stability of the judgment.

[0142] Furthermore, the consistency score is scaled using the short-window confidence coefficient to obtain the scaled score. The scaled score is then compared with the consistency threshold. If the scaled score is significantly higher than the consistency threshold, it indicates that the battery is consistent with the historical benchmark fingerprint within the interpretable range and the controlled response is reliable. If it is significantly lower than the consistency threshold, it indicates insufficient consistency and low reliability.

[0143] Furthermore, when the consistency score is high but the short-window confidence coefficient is low, it is judged as medium risk and a power limiting and temperature rise limiting strategy is required in step S5, with the retest time marked. When the consistency score is low and the short-window confidence coefficient is high, the system still judges it as medium risk and requires a second check to prevent false alarms caused by local deviations in historical benchmark fingerprints. This constraint rule enables the judgment to be self-explanatory regarding conflicts between the two types of indices.

[0144] Step S5: Issue operation instructions based on risk labels. High-risk areas are transferred to isolation for re-inspection and release is suspended. Medium-risk areas are arranged for short-term use according to power and temperature rise limits and the re-test time is marked. Low-risk areas are allowed to release and the equivalent short window representation vector, consistency score and risk label are archived.

[0145] Background Summary: In step S4, after outputting risk labels and archiving consistency scores, scaled scores, multi-physical disturbance coupling indices, historical trajectory deviation indices, and short-window confidence coefficients, the edge control unit acts as the execution entity. Within the time window when the vehicle is still parked at the inspection station and the station is effectively locked, it reads the aforementioned judgment quantities from the edge storage based on the station identifier and the inspection time stamp, and reads the battery's historical benchmark fingerprint summary and asset ledger record from the cloud. The task of step S5 is to translate the risk labels into directly identifiable operation instruction packages without introducing new adjustment factors. This ensures that the three types of actions—isolation re-inspection, power limiting and temperature rise limiting strategies, and release archiving—have clear execution targets, execution times, and feedback collection methods. The execution feedback is bound to the inspection time stamp in the form of execution feedback tags, providing a unified reference for subsequent in-station scheduling and re-inspection arrangements.

[0146] Furthermore, step S5 includes the following sub-steps:

[0147] Step 501 uses the risk label as a trigger condition to generate an operation instruction package according to the workstation identifier and the physical examination time mark. The instruction package includes the action type, the scope of application of the action, the start and end time of execution, the execution monitoring window, and the execution feedback mark placeholder. When the risk label is high risk, the action type is isolation re-inspection and suspension of release. The scope of application covers all outbound paths of the battery. Execution starts at the current workstation. The execution monitoring window is set to a short period before isolation transfer to confirm that the isolation transfer conditions are met. When the risk label is medium risk, the action type is power limiting and temperature rise limiting strategy and a retest time mark is set. The scope of application is limited to the operation phase from the subsequent departure of this vehicle to the arrival of the retest time mark. The execution monitoring window is set to a short period of confirmation before departure and the first stage of operation to confirm that the power and temperature rise limits are followed. When the risk label is low risk, the action type is release and archiving. The scope of application is the current departure process. The execution monitoring window is set to a very short period of confirmation before departure to verify the consistency of transmission and archiving.

[0148] Step 502: After the operation instruction package is issued, the workstation status and electrical status are checked for safety interlocks. The verification includes the clamp locking status, connector temperature rise status, and whether the environmental parameters within the station are maintained near the set values. In case of high risk, the transfer unit is authorized to perform isolation re-inspection only if the interlock is passed and the isolation transfer path is unobstructed, and the outbound permission of the current workstation is frozen. In case of medium risk, after the interlock is passed, a power limit token and a temperature rise limit token are issued to the vehicle-side controller. The tokens reference the quantile boundaries of three values: consistency score, short window reliability coefficient, and historical trajectory deviation index, to determine the limiting strength. In case of low risk, after the interlock is passed, the vehicle is authorized to leave and enters the pre-exit confirmation window.

[0149] Step 503: After the authorization takes effect, feedback is collected according to the execution monitoring window defined in the operation instruction package; the feedback collection for the isolation re-inspection action is the confirmation of the workstation status before transfer and the status confirmation during the transfer process, generating an execution feedback mark and binding it with the physical examination time mark; the feedback collection for the power limiting and temperature rise limiting strategies is the voltage response in the confirmation window before leaving the station, the extremely short window sampling of the temperature difference array and the micro deformation of the shell, and the comparison and verification of the vehicle-end power request and the actual power. If it is found that the limit is not followed, the execution feedback is marked as failed and automatically downgraded to high-risk processing; the feedback collection for release and archiving is the data transmission consistency verification. If it is found that the archiving fails, the retransmission process is triggered and the retransmission completion time is recorded in the execution feedback mark;

[0150] Step 504: When the risk label is medium risk, calculate the retest time stamp based on recent operating intensity information and historical trajectory deviation index without introducing new factors, and associate the retest time stamp with the workstation schedule management system and the station vehicle dispatching system; In order to ensure that the retest time stamp is actually executed, the edge control unit sends a routing restriction suggestion to the dispatching system, so that the battery is preferentially allocated to the regular load and the regular temperature and humidity environment before the retest time stamp arrives, and registers the acceptance and execution status of the routing restriction suggestion in the edge storage;

[0151] Step 505: When the risk label is high risk or medium risk and the execution feedback mark is not passed, a maintenance work order is generated and a reference to the consistency score, scaled score, multi-physical disturbance coupling index, historical trajectory deviation index, short window confidence coefficient and execution feedback mark is attached; for all risk labels, the edge control unit writes the results of this physical examination into the asset ledger record and updates the most recent equivalent short window representation vector summary at the historical benchmark fingerprint corresponding to the battery for comparison in the subsequent step S4.

[0152] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A battery anomaly identification and monitoring method for battery operation and maintenance, characterized in that, Includes the following steps: Step S1: After the battery enters the station, the temperature difference array on the outer surface of the battery, the micro deformation of the shell, the gas indicator in the cavity and the near-end temperature of the end plate are collected synchronously and encapsulated into a skin observation vector. During encapsulation, the fast variable component and the slow variable component are explicitly distinguished. Step S2: Apply a preset slight current disturbance sequence within a safe range and keep the environment unchanged. Record the voltage response, temperature rise difference, and shell micro deformation change before and after the disturbance, and summarize to obtain the disturbance difference vector. Step S3: The skin observation vector and disturbance difference vector are time-series aligned and debiased normalized with the station's environmental quantities and recent operational intensity information. For fast variable components, a fixed kernel is established based on the steady-state segment before the disturbance to generate shadow trajectories through window regression, and these are subtracted point by point to remove environmental common modes. The fall segment is time-reversed at its center time and superimposed on the rise segment to calculate the symmetric difference sequence to offset the slow drift and thermal hysteresis residuals. The controlled response residuals and symmetric differences are respectively subjected to rank-preserving quantile projection normalization and aggregated by region index to obtain the equivalent short-window representation vector. Step S4: Use the equivalent short window representation vector to perform similarity matching with the battery history benchmark fingerprint stored in the cloud to obtain a consistency score. Calculate the multi-physical perturbation coupling index and the historical trajectory deviation index and input them together into the predefined consistency discrimination model to obtain the short window credibility coefficient. Scale the consistency score with the short window credibility coefficient and compare it with the consistency threshold to generate a risk label. The battery historical benchmark fingerprint is a robust aggregation result of the equivalent short window representation vectors of the same battery at multiple sites and multiple operating stages, after quality filtering, time reference unification and regional index consistency. This includes the typical form of controlled response residuals, the symmetry description of symmetric differences, and the distribution profile of slow variable baselines, which are bound to frozen parameter vectors to fix quantile boundaries and soften function shapes. The acquisition method involves periodically aggregating quality weights and removing outlier samples from historical equivalent short-window representation vectors in the cloud, generating statistical templates for corresponding regions using regional indexes as keys, solidifying them into searchable objects, and allowing read-only access during runtime. The consistency score is obtained as follows: similarity matching recalls common regions using the region index as the key and generates region quality weights based on data quality labels; for each region, the morphological-symmetric joint score of its fast variable component and historical benchmark, and the background distribution score of its slow variable component based on the first-order distribution shift metric are calculated, and the two are multiplied to obtain the initial region similarity kernel. To suppress spatially isolated anomalies, the initial similarity kernel is smoothed by neighborhood coordination based on regional adjacency relationships; the smoothed regional similarity kernel is then weighted and averaged using regional quality weights to obtain a consistency score. The multi-physics perturbation coupling index is obtained by weighted averaging the phase lock-in degree and fast component ratio of the voltage, temperature rise, and deformation response residuals of each region; the historical trajectory deviation index is obtained by calculating the maximum value of the distribution offset measure of each region in different response segments and historical benchmarks and then weighted averaging them; the multi-physics perturbation coupling index and the historical trajectory deviation index are used as inputs and processed by a predefined consistency discrimination model with monotonicity constraints, and the output is a short-window confidence coefficient used to scale the consistency score; Step S5: Issue operation instructions based on risk labels. For high-risk cases, transfer to isolation for re-inspection and suspend release. For medium-risk cases, arrange short-term use according to power and temperature rise limits and mark the re-test time. For low-risk cases, allow release and archive.

2. The battery anomaly identification and monitoring method for battery operation and maintenance according to claim 1, characterized in that, Step S1 includes: The completion of mechanical locking is used as a trigger to generate a unified physical examination time stamp, which is then associated with the environmental quantity within the station and recent operational intensity information. Before data collection, a health self-check is performed on the sensor for zero point, drift, and connectivity. A sensor health flag is generated based on the check results. If any abnormality is found, the process is terminated. Using the physical examination time marker as a hardware trigger, all sensor channels are controlled to collect data synchronously within the same sampling cycle. Artifact suppression is performed on the acquired data, including: detecting and replacing isolated spikes, performing baseline wrapping on temperature sensing data to counteract condensation drift, smoothing electromagnetic jitter, and verifying data steady state; The collected data is mapped to preset region indices according to the battery shape, forming a channel distribution with spatial meaning; The processed multimodal data is encapsulated into the skin observation vector according to the regional index and time order, and metadata is attached.

3. The battery anomaly identification and monitoring method for battery operation and maintenance according to claim 2, characterized in that, It includes an edge control unit, which is an industrial-grade embedded device with a built-in hardware time base. This unit is used to generate a unified physical examination time stamp and synchronously trigger multi-channel sensor acquisition and drive a preset current disturbance sequence. The edge control unit includes: An integrated isolated acquisition and execution interface with a built-in real-time processing flow is used to complete event locking elastic registration, mirror overlay symmetry difference, window shadow trajectory subtraction and quantile projection normalization processing to synthesize the equivalent short window representation vector. The built-in inference module is used to calculate the multi-physics disturbance coupling index and the historical trajectory deviation index locally, call the consistency discrimination model to generate risk labels and issue operation instructions; It supports two-way communication with the cloud to obtain historical baseline fingerprints and transmit data back, while providing electrical isolation and time synchronization capabilities.

4. The battery anomaly identification and monitoring method for battery operation and maintenance according to claim 1, characterized in that, The amplitude, single-pulse duration, number of pulses, and pulse interval of the slight current disturbance sequence are dynamically determined and applied based on temperature rise margin constraints, charge drift constraints, and terminal voltage disturbance constraints; the temperature rise margin constraints are based on the skin observation vector and the environmental parameters within the station to estimate the current heat dissipation capacity and limit the upper limit of temperature rise. The charge drift constraint limits the upper limit of the total pulse charge based on the battery's rated capacity. The terminal voltage perturbation constraint limits the transient range of the terminal voltage based on historical benchmark fingerprints.

5. The battery anomaly identification and monitoring method for battery operation and maintenance according to claim 4, characterized in that, The perturbation sequence is determined adaptively as follows: Based on the skin observation vector, the environmental quantity in the station and the recent operating intensity, the candidate grid is discretized into parameters including amplitude, duration and interval under the three constraints of temperature rise margin, charge drift and terminal voltage disturbance. The visibility target band is set with the historical terminal voltage transient boundary as the hard limit. The candidate cell with the smallest single-pulse energy injection is selected and applied to the battery as a single calibration preamble pulse. Only the slight voltage drop and slight temperature rise difference in the rising phase are taken. The minimum of the two individual visibility ratios is taken as the initial visibility score. At the same time, the minimum of the three margins of voltage, heat and charge is taken as the safety margin score. The disturbance adaptive coefficient γ is set as the minimum of the initial visibility score and the safety margin score. The model prioritizes amplitude, followed by duration, and then uses interval as a fallback for reshaping. If the perturbation adaptive coefficient γ is low, the amplitude is increased in small steps, the duration is shortened, and the interval is increased. If the perturbation adaptive coefficient γ is high, the amplitude is fixed, the duration is shortened first, and the interval is increased. If the limit is exceeded, the model is locked back, and the output sequence and start and end markers are displayed.

6. The battery anomaly identification and monitoring method for battery operation and maintenance according to claim 5, characterized in that, Step S2 includes: After the safety interlock verification is passed, the current disturbance sequence is applied synchronously and the high-frequency acquisition of voltage, temperature difference and deformation channels is started, with the disturbance start time mark as the unified trigger. The response segments are divided with the disturbance start and end marks as the boundary, and the event-locked elastic registration is used to correct the response hysteresis misalignment between different sensing channels to achieve timing alignment. Robust increment values ​​relative to the disturbance start time marker are calculated for voltage response, temperature difference array, and shell micro-deformation, respectively. The voltage response increment, temperature rise difference increment, and shell micro-deformation increment are obtained by combining the difference between the high quantile and low quantile values ​​within the segment with window averaging, and encapsulated as a disturbance difference vector.

7. The battery anomaly identification and monitoring method for battery operation and maintenance according to claim 3, characterized in that, Step S3 includes: Using the physical examination time marker and the disturbance start and end marker as a unified reference, the skin observation vector and the disturbance difference vector are aligned and their intersections are filtered at the object level. Based on the slow variable components of the skin observation vector, the on-site environmental quantity and recent operational intensity information, a fixed kernel window regression relationship is established to generate the window shadow trajectory of the fast variable components. The controlled response residual sequence that suppresses environmental common mode interference is obtained by subtracting the fast variable components point by point. After time reversal of the fall segment according to the center time of the segment, it is paired with the rise segment point by point, and the difference between the paired points is calculated to obtain the symmetric difference sequence. The controlled response residual sequence and the symmetric difference sequence are subjected to rank-preserving quantile projection normalization to map each physical quantity to a unified dimensionless space, and the marked isolated peak points are subjected to gated replacement; spatial aggregation and minimization interpolation are performed according to the region index to assemble the equivalent short window representation vector.

Images