A method for active exploration, sensing, and intelligent diagnosis of energy storage batteries

By using quality identification status profiling, electro-thermal mechanism residuals, and time-series joint diagnostic methods, the problem of existing energy storage battery monitoring systems being unable to accurately identify early faults has been solved. This enables efficient diagnosis of fault types and locations, reducing waste of sampling resources and monitoring interference.

CN123066577APending Publication Date: 2026-11-10JIANGSU SIJI TECH SERVICE CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611231757.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-14
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing energy storage battery monitoring systems cannot adjust the sensing objects according to the current operating conditions, model uncertainties, and fault risks. This results in early fault characteristics being misaligned, missing, and masked by channel noise. Furthermore, high-frequency sampling increases communication bandwidth, acquisition energy consumption, and computational load, making it difficult to accurately determine the health status, remaining service life, and fault location.

Method used

A method combining quality identification status profiling, electro-thermal mechanism residuals, time-series joint diagnosis, and fault hypothesis conditional action evaluation is adopted. By analyzing the expected diagnostic information gain before and after the execution of candidate actions, the action utility is formed based on the fault consequence risk, data quality, and resource consumption. Under safety and resource constraints, the perceived actions are issued, and the joint posterior and uncertainty are updated to output the diagnostic results.

Benefits of technology

It reduces the interference of data misalignment and channel noise on early fault evidence, optimizes sampling resource allocation, ensures monitoring continuity under abnormal conditions, and reduces premature stopping and invalid high-frequency sampling.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123066577A_ABST
    Figure CN123066577A_ABST
Patent Text Reader

Abstract

This invention discloses an active exploration and intelligent diagnosis method for energy storage batteries, belonging to the field of energy storage battery state monitoring and fault diagnosis. To solve the problem that fixed sampling and fixed thresholds are difficult to extract early weak fault features, this invention constructs a state profile with quality labels and electro-thermal mechanism residuals to form a joint posterior of health state value, remaining service life, fault type and fault location. Information gain and action utility are calculated according to the conditional observation distribution of competing fault assumptions. Sensing actions are executed under safety and resource constraints, and the method is updated or stopped cyclically based on conformal calibration evidence, thereby realizing the targeted allocation of limited sampling resources and joint diagnosis with confidence information.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of energy storage battery state monitoring and fault diagnosis, and in particular to a method for active exploration, sensing and intelligent diagnosis of energy storage batteries. Background Technology

[0002] Energy storage battery systems typically consist of a large number of cells or cell packs connected in series and parallel. Their operation is simultaneously affected by the charging and discharging phases, ambient temperature, connection status, aging degree, and sensor link quality. Existing monitoring systems mostly continuously collect cell voltage, temperature, and cluster current at a pre-set sampling frequency and use fixed thresholds or a single model to determine anomalies.

[0003] The above methods cannot adapt the sensing objects and sampling configurations to current operating conditions, model uncertainties, and the risk of competing failures. When internal short circuits, loose connections, abnormal capacity decay, thermal anomalies, or sensor failures are in their early stages, their weak characteristics are easily masked by multi-rate data misalignment, missing observations, and channel noise. Continuously sampling all channels at high frequencies increases communication bandwidth, acquisition energy consumption, and computational load. Fixed thresholds also struggle to provide consistent confidence evidence of health status, remaining service life, fault type, and fault location simultaneously, potentially leading to premature cessation or ineffective continuation of the exploration.

[0004] Therefore, there is a need for an active exploration, sensing, and intelligent diagnostic method for energy storage batteries that can address the shortcomings of existing technologies. Summary of the Invention

[0005] One objective of this invention is to propose an active exploration, sensing, and intelligent diagnosis method for energy storage batteries. Addressing the problem that existing technologies use fixed sampling frequencies, fixed monitoring parameters, and fixed thresholds, failing to adjust the sensing object based on operating conditions, model uncertainties, and anomaly risks, this invention proposes a technical solution combining quality identification status profiling, electro-thermal mechanism residuals, time-series joint diagnosis, and fault hypothesis-based action evaluation. Centering on the expected diagnostic information gain before and after candidate action execution, action utility is formed based on fault consequence risk, data quality, and resource consumption. Sensing actions are issued under safety and resource constraints, and the joint posterior and uncertainty are updated with new observations. Furthermore, the consistency of the prediction set, prediction interval, and mechanism residuals is calibrated to control whether exploration continues or stops. This invention achieves the technical effect of allocating limited sampling resources to cells, parameters, and operating conditions with current diagnostic discriminative power and outputting joint diagnostic results with confidence information.

[0006] This invention provides an active exploration, sensing, and intelligent diagnosis method for energy storage batteries, comprising: S1, acquiring multi-source sampling data of each target cell or cell group, performing time alignment and data quality assessment on the multi-source sampling data to form a multi-source state profile with quality labels; S2, inputting the multi-source state profile into an electro-thermal mechanism model to obtain the predicted state and the electro-thermal mechanism residual between the measured state and the predicted state; S3, inputting the multi-source state profile, the electro-thermal mechanism residual, cell connection relationships, and historical degradation sequences into a trained time-series joint diagnosis model to obtain the joint posterior distribution and uncertainty of health status value, remaining service life, fault type, and fault location; S4, based on the joint posterior distribution... Candidate sensing actions are constructed based on their uncertainties. The conditional observation distribution of each candidate sensing action is predicted for the competing fault hypothesis. The expected diagnostic information gain is calculated based on the conditional observation distribution. The action utility of each candidate sensing action is determined in combination with the fault consequence risk, data quality and resource consumption. S5. The target sensing action with the highest action utility ranking is selected from the candidate sensing actions that meet the safety constraints and resource constraints and sent to the acquisition module. The multi-source state profile, the electro-thermal mechanism residual, the joint posterior distribution and its uncertainty are updated using the new observation data. When the updated diagnostic evidence meets the stopping condition, the health status value, remaining service life, fault type and fault location and corresponding confidence information are output. Otherwise, the process returns to step S4.

[0007] Optionally, S1 includes: resampling a multi-rate sampling sequence including cell voltage, cell temperature and cluster current, using timestamps and charging / discharging phases as alignment references;

[0008] Set a missing marker for sampling positions that do not fall within the aligned time window;

[0009] The reliability of the sensor is calculated based on the variance of the sensor calibration residual, the effective sampling ratio, and the number of adjacent sampling jumps.

[0010] The aligned sampled values, the missing markers, and the sensor reliability are combined to form the multi-source state profile, wherein the sensor reliability serves as the input for determining data quality in step S4.

[0011] Optionally, S2 includes: using an equivalent circuit model to determine the predicted voltage and circuit model parameters based on the cell voltage and cluster current in the multi-source state profile;

[0012] A thermal model is used to determine the predicted temperature and thermal model parameters based on the cell temperature, cluster current and cell connection relationship in the multi-source state profile.

[0013] The difference between the measured cell voltage and the predicted voltage, and the difference between the measured cell temperature and the predicted temperature are calculated separately, and combined according to the aligned time window to form the electro-thermal mechanism residual. The electro-thermal mechanism residual is used for the joint diagnosis in step S3 and the candidate observation prediction in step S4.

[0014] Optionally, S3 includes: using a time-series coding network to extract the temporal features of the multi-source state profile, the electro-thermal mechanism residual, and the historical degradation sequence;

[0015] Based on the cell connection relationship, the time characteristics are propagated between connected cells or cell groups to form a spatiotemporal joint characteristic;

[0016] The spatiotemporal joint features are respectively input into the health status regression head, the remaining service life regression head, the fault type classification head, and the fault location classification head, and the output distributions are jointly normalized to obtain the joint posterior distribution.

[0017] The uncertainty is determined based on the differences in the outputs of multiple model instances to the same input;

[0018] Furthermore, the training of the time-series joint diagnostic model includes using historical degradation samples with timestamps, charging and discharging stages, cell voltage, cell temperature, cluster current, quality identification, cell connection relationships, fault type labels, fault location labels, health status labels, and remaining service life labels as training samples.

[0019] Based on the time-aligned measured cell voltage, measured cell temperature, and cluster current, the electro-thermal mechanism residuals corresponding to each training sample are generated by the electro-thermal mechanism model.

[0020] The parameters of the time-series joint diagnostic model are updated by weighting and summing the health status regression loss, remaining useful life regression loss, fault type classification loss, fault location classification loss, and mechanism residual consistency loss, and a model set for determining the uncertainty is formed by multiple model instances with different initial parameters.

[0021] Optionally, S4 includes: using a combination of the target cell or cell group, monitoring parameters, sampling frequency, sampling start phase, and continuous window as a candidate sensing action;

[0022] The physical fault state formed by combining the physical fault type (such as normal state, internal short circuit, loose connection, abnormal capacity decay or thermal abnormality) with the fault location in the target cell or cell group, and the sensor fault state formed by combining the sensor fault type with the monitoring parameters or sampling channel location are used as the competing fault assumptions.

[0023] The baseline model parameters are obtained by fitting the electro-thermal mechanism model with the multi-source state profile, and the baseline model parameters are corrected using the electro-thermal mechanism residuals. The baseline observation noise covariance is estimated based on the normal state calibration samples according to the monitoring parameters and sampling channels. The mapping between physical fault type, fault location and model parameter offset, as well as the mapping between sensor fault type, sampling channel location and observation deviation parameters, observation drift parameters and noise covariance correction parameters are established based on the fault calibration samples.

[0024] For each candidate sensing action, the corresponding target object and sampling configuration are input into the electro-thermal mechanism model. For the normal state, the corrected baseline model parameters and the baseline observation noise covariance of the corresponding sampling channel are used. For the physical fault state, the corresponding model parameter offset is superimposed on the corrected baseline model parameters. For the sensor fault state, the corresponding observation bias parameters, observation drift parameters and noise covariance correction parameters are applied to the model output and the corresponding baseline observation noise covariance to obtain the conditional observation distribution.

[0025] Furthermore, the expected diagnostic information gain is obtained by sampling the conditional observation distribution and calculating the expected value of the Kolb-Leibler divergence between the joint posterior distributions before and after adding the sampling results, or by calculating the difference between the joint posterior entropy before and after adding the sampling results.

[0026] The risk of failure consequences is determined based on the consequences of each competing failure hypothesis and the posterior probability.

[0027] For each candidate sensing action, the data integrity and sensor reliability corresponding to the target cell or cell group, monitoring parameters and sampling channels of the candidate sensing action are extracted from the quality identifier of the multi-source state profile to determine the data quality weight of the candidate sensing action.

[0028] The expected diagnostic information gain is multiplied by the fault consequence risk and the data quality weight, and the costs of communication bandwidth usage, data acquisition energy consumption, computational load and action switching are deducted to obtain the action utility.

[0029] Optionally, S5 includes: safety constraints generated from the voltage operating boundary, temperature operating boundary, and charge / discharge current operating boundary recorded by the energy storage battery management system, as well as the safety monitoring objects, monitoring parameters, minimum sampling frequency, and maximum sampling start delay corresponding to each operating boundary; and a fixed safety minimum sampling configuration is generated from the safety constraints, covering the corresponding safety monitoring objects and monitoring parameters, with a sampling frequency not less than the minimum sampling frequency and a sampling start delay not greater than the maximum sampling start delay; and resource constraints are generated from the currently available communication bandwidth, remaining acquisition energy, and available computing load.

[0030] Eliminate candidate sensing actions from all candidate sensing actions that exceed the corresponding available capacity due to excessive communication bandwidth usage, data acquisition energy consumption, or computing load requirements.

[0031] When the distance from the measured state or the predicted state to any working boundary is not greater than the corresponding safety monitoring trigger distance, candidate sensing actions that do not meet the safety constraints are excluded from the remaining candidate sensing actions.

[0032] When the set of candidate sensing actions after screening is not empty, the target sensing action is determined according to the order of the action utility.

[0033] When the candidate sensing action set is empty, safety monitoring is placed before non-safety monitoring, and non-safety monitoring tasks are paused to release communication bandwidth, energy acquisition, and computing load. The resource requirements of the fixed safety minimum sampling configuration are reviewed based on the released available resources. If the review passes, the fixed safety minimum sampling configuration is added to the candidate sensing action set and determined as the target sensing action. If the review fails, a resource shortage flag is generated and a safety alarm is sent to the energy storage battery management system. The active action selection in steps S4 and S5 is paused and the controlled degradation sampling configuration is executed. The controlled degradation sampling configuration retains the safety monitoring objects and monitoring parameters in ascending order of distance from the measured state or predicted state to the corresponding working boundary based on the available sampling channels and the released available resources. The sampling frequency that can be executed within the available resources is set for the retained monitoring parameters.

[0034] After the fixed minimum sampling configuration satisfies both the security constraint and the resource constraint, it is added to the candidate sensing action set and the target sensing action selection in step S5 is restored.

[0035] Furthermore, using calibration samples independent of the time-series joint diagnostic model, fault type non-compliance score, fault location non-compliance score, health status value non-compliance score, and remaining service life non-compliance score are constructed respectively, and the calibration quantile value of each non-compliance score is determined according to the target coverage.

[0036] After each update of the joint posterior distribution and its uncertainty, the diagnostic outputs of the current wheel's fault type, fault location, health status value, and remaining service life are combined with the corresponding calibration quantile values ​​to form the current wheel's fault type conformal prediction set, fault location conformal prediction set, health status value conformal prediction interval, and remaining service life conformal prediction interval. The conformal prediction set and the conformal prediction interval are used as confidence information for the corresponding diagnostic output.

[0037] When the number of elements in the fault type conformal prediction set and the fault location conformal prediction set are not greater than the corresponding set threshold, the width of the health status value conformal prediction interval and the remaining service life conformal prediction interval are not greater than the corresponding interval width threshold, and the electro-thermal mechanism residual falls into the consistency interval determined by the calibration sample residual distribution, and the above conditions are repeatedly met within the continuous evidence window, the diagnostic evidence is determined to meet the stopping condition.

[0038] Furthermore, the energy storage battery management system sets the minimum number of available sampling channels required for safety monitoring according to cell voltage, cell temperature and cluster current, and uses the minimum number of available sampling channels as the channel requirements for the corresponding monitoring parameters in the fixed safety minimum sampling configuration.

[0039] When the communication link is unavailable, the number of available sampling channels for any monitoring parameter is less than the corresponding minimum number of available sampling channels, or the target sensing action fails to be issued, the action switching based on the action effectiveness is suspended.

[0040] When the number of available sampling channels for each monitoring parameter is not less than the corresponding minimum number of available sampling channels and the resource requirements of the fixed safety minimum sampling configuration are verified, the fixed safety minimum sampling configuration stored in the acquisition module is invoked for acquisition; otherwise, the controlled degradation sampling configuration is invoked and a safety alarm is sent to the energy storage battery management system.

[0041] After the communication link is available, the number of available sampling channels for each monitoring parameter is not less than the corresponding minimum number of available sampling channels, the acquisition module is in a ready state, and the test sensing action receives a confirmation response, the joint posterior distribution and its uncertainty are recalculated using the current observation data, the cumulative number of consecutive evidence windows is cleared, and then steps S4 and S5 are resumed.

[0042] The beneficial effects of this invention are:

[0043] 1. By performing time alignment, missing data labeling, and reliability evaluation on multi-rate sampling data, and by inputting the quality label and electro-thermal mechanism residuals into the joint diagnostic model, the interference of data misalignment, channel missing data, and sensor noise on early fault evidence can be reduced.

[0044] 2. By predicting the conditional observation distribution of candidate sensing actions based on normal, physical fault, and sensor fault assumptions, and by comprehensively considering information gain, fault consequence risk, action-related data quality, and resource cost to determine action utility, sampling resources can be directed to target objects and monitoring parameters that have discriminatory power against the current competing assumptions.

[0045] 3. By ensuring monitoring continuity under abnormal conditions through dual constraints of safety and resources, fixed minimum safety sampling, and controlled degradation branches, and by using the consistency of the current round shape-preserving prediction set, prediction interval, and mechanism residual to form stopping evidence, premature stopping and invalid continuous high-frequency sampling can be reduced. Attached Figure Description

[0046] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0047] Figure 1 This is a flowchart of the active exploration, sensing, and intelligent diagnosis method for energy storage batteries according to the present invention.

[0048] Figure 2 This is a flowchart of the S4 candidate perceived action construction and action utility evaluation of the present invention. Detailed Implementation

[0049] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0050] refer to Figures 1-2 A method for active exploration, sensing, and intelligent diagnosis of energy storage batteries includes: S1, acquiring multi-source sampling data of each target cell or cell group, performing time alignment and data quality assessment on the multi-source sampling data to form a multi-source state profile with quality labels; S2, inputting the multi-source state profile into an electro-thermal mechanism model to obtain the predicted state and the electro-thermal mechanism residual between the measured state and the predicted state; S3, inputting the multi-source state profile, the electro-thermal mechanism residual, cell connection relationships, and historical degradation sequences into a trained time-series joint diagnostic model to obtain the joint posterior distribution and its uncertainty of health status value, remaining service life, fault type, and fault location; S4, based on the joint posterior distribution and its uncertainty... The determination constructs candidate sensing actions, predicts the conditional observation distribution of each candidate sensing action for competing fault assumptions, calculates the expected diagnostic information gain based on the conditional observation distribution, and determines the action utility of each candidate sensing action in combination with fault consequence risk, data quality and resource consumption; S5, selects the target sensing action with the highest action utility ranking from the candidate sensing actions that meet safety constraints and resource constraints and sends it to the acquisition module, updates the multi-source state profile, the electro-thermal mechanism residual, the joint posterior distribution and its uncertainty using the new observation data, and outputs the health status value, remaining service life, fault type and fault location and corresponding confidence information when the updated diagnostic evidence meets the stopping condition, otherwise returns to step S4.

[0051] In this specific embodiment, S1 includes:

[0052] This specific implementation takes a battery cluster consisting of 240 lithium iron phosphate cells connected in series as the diagnostic object. The battery management system assigns a cell identifier to each cell and maintains its cell group, module position and series sequence. The acquisition module reads the cluster current at 10 Hz, the cell voltage at 1 Hz, the cell temperature at 0.2 Hz and the cluster-level ambient temperature at 0.2 Hz. The raw records are written to a circular sampling buffer indexed by the battery cluster identifier, cell identifier, sampling channel and timestamp. The recorded fields include timestamp, charging and discharging stage, cell voltage, cell temperature, cluster current, cluster-level ambient temperature, sampling channel identifier, calibration version and communication status. The cluster-level ambient temperature record uses the battery cluster identifier and is not bound to a single cell.

[0053] The time alignment module constructs a unified time scale with an alignment period of 1 second. , Indicates the first The alignment time, measured in seconds, is derived from the monotonic clock of the energy storage battery management system. This indicates the start time of the current diagnostic window, in seconds. This represents the alignment time index starting from 0 and is a dimensionless integer. The alignment period is represented by the unit s, which is set to 1 s in this specific embodiment. The cluster current is taken within 50 ms before and after each alignment moment, the closest sample in time is taken within 500 ms before and after each alignment moment, the cell voltage is taken within 500 ms before and after each alignment moment, and the cell temperature and cluster ambient temperature are linearly interpolated in time ratio between two adjacent effective temperature samples. Interpolation is performed only when the interval between two effective samples does not exceed 10 s and they are in the same charging and discharging stage.

[0054] To preserve high-frequency diagnostic information generated by active actions, the time alignment module also calculates the number of subsamples, effective proportion, arithmetic mean, maximum value, minimum value, range, sample standard deviation, and slope obtained by dividing the difference between the first and last effective samples by the actual time interval for all original subsamples belonging to the same parameter and channel within each 1-second time scale. These statistics, along with the start and end times of the original subsamples, are written into the high-frequency summary field. When there are fewer than two effective subsamples in the window, the standard deviation and slope are not calculated, and a summary insufficiency flag is set. When the frequency of active actions is not higher than 1 Hz, the alignment principal value is used as the average value, and the other high-frequency fields are set with an inapplicable flag. The original subsamples are still retained according to the action sequence number until the current round of diagnosis is completed.

[0055] The charging and discharging phases are determined by the median of the cluster current over 5 consecutive seconds. A median greater than 5 A is marked as the charging phase, and a median less than -5 A is marked as the discharging phase. The rest are marked as the resting phase. Samples taken 2 seconds before and after the phase switch are retained in the original phase and the new phase, respectively, and cross-phase interpolation is prohibited. For sampling positions that do not fall within the corresponding alignment time window, have abnormal communication status, have failed calibration version, or exceed the cell voltage range of 2.0 V to 3.8 V, cell temperature range of -20 ℃ to 70 ℃, cluster-level ambient temperature range of -30 ℃ to 60 ℃, and cluster current range of -300 A to 300 A, a missing marker is set, and the reason for the missing value is written into the quality field instead of being replaced by a zero value.

[0056] After each maintenance calibration, the system saves the calibration residuals of 120 consecutive zero-input or standard source readings for each sampling channel, and calculates the effective sampling ratio and the number of jumps between adjacent effective samples within the current 24-hour sliding window. Calculate sensor reliability. Indicates sampling channel The sensor reliability is a dimensionless number within the range [0,1]. A larger value indicates a more reliable channel. Indicates the sampling channel index. Indicates channel The sample variance of the calibration residuals is expressed in squares of the corresponding monitoring parameter units. This represents the upper limit of the residual variance of the same channel at the time of the most recent calibration acceptance and is related to... Same unit, This represents the proportion of valid samples within a 24-hour window to the planned sample size, and is a dimensionless number within the range [0,1]. This represents the number of times adjacent valid samples within the window exceed the channel transition threshold, and is a dimensionless integer. The reference number represents the number of jumps allowed within the same window and is a positive integer. Both references are read from the calibration version table by channel identifier. This indicates that the calculation result will be truncated to the given upper and lower bounds. This represents an exponential function with the natural constant as its base. Its input must be a dimensionless number and its output must be a positive dimensionless number.

[0057] The calibration version table stores channel identifiers, monitoring parameters, calibration time, upper limit of residual variance, jump threshold, baseline of allowed jump times, effective start and end time, and version number. The upper limit of residual variance is taken as the 95th percentile of the residual variance of the calibrated channel and must be strictly greater than 0. The jump threshold is taken as the 99th percentile of the absolute value of the difference between adjacent samples in the same charge and discharge phase. The version is updated after sensor replacement, sampling front-end gain adjustment, or quarterly calibration. One voltage channel records the channel identifier as V-C087, the upper limit of residual variance as 0.000004 V², the jump threshold as 0.08 V, and the baseline of allowed jump times as 12 times / 24 h. If a valid version cannot be found, or the upper limit of residual variance read is missing, non-numerical, or not greater than 0, the exponential ratio in the reliability formula is not executed, the corresponding reliability is set to 0, and a calibration pending flag and a calibration parameter abnormal flag are set.

[0058] For each target cell and monitoring parameters Calculate data integrity within the most recent 60-second diagnostic window. , Indicates battery cell Monitoring parameters The data completeness is a dimensionless number within [0,1]. Indicates the cell index. This index represents parameters such as cell voltage, cell temperature, cluster current, or cluster-level ambient temperature. The data integrity for cluster-level ambient temperature is shared among all cells within the same battery cluster. This represents the number of valid alignment positions within the window that are not marked with a missing marker, and is a dimensionless integer. This represents the number of planned alignment positions determined by the current sampling configuration and is a positive integer; when a parameter is completely missing within the window, the data integrity is set to 0, the temperature interpolation value retains the interpolation flag and is included in the valid positions, but is multiplied by an interpolation quality coefficient of 0.8 in subsequent model inputs;

[0059] When multiple valid records appear for the same cell, the same parameters, and the same alignment time, the profile construction module selects the main record in the order of valid calibration version, sensor reliability from largest to smallest, and time distance from smallest to largest, and retains the remaining records as supporting evidence. When the difference between the main record and the supporting evidence record exceeds the dual-channel consistency threshold in the calibration version table, a channel inconsistency flag is set at that position and the reliability of both channels is multiplied by 0.7. When the subsequent 10 consecutive alignment positions are all below the consistency threshold, the original reliability is restored, so that repeated sampling, conflict observation, and reliability update have a definite processing path.

[0060] The profile construction module generates multi-source state profile records according to the alignment time. The record keys are battery cluster identifier, cell identifier, and alignment time. The value fields include aligned cell voltage, cell temperature, cluster current, cluster-level ambient temperature, charging and discharging stage, missing parameter markers, interpolation markers, high-frequency summary field, data integrity, sensor reliability, calibration version, and profile version. 60 consecutive records form a diagnostic window. When the same parameter has two channels, the channel value with higher reliability is retained and the other channel is saved as supporting evidence. If the reliability is the same, the sample with the closer time distance is selected. The resulting multi-source state profile with quality label is written to the state profile cache and passed to step S2 along with the cell connection relationship. At the same time, the data integrity, sensor reliability, and high-frequency summary quality are used as data quality inputs for step S4.

[0061] In this specific embodiment, S2 includes:

[0062] The mechanism calculation module reads the 60-second multi-source state profile written in step S1 and obtains the rated capacity, initial state of charge, open-circuit voltage curve, circuit parameter version, and thermal parameter version according to the cell identifier. In this specific implementation, a first-order Thevenin equivalent circuit model and a single-node lumped thermal model are used. The model parameters are calibrated by the most recent capacity verification and pulse charge-discharge tests at different temperatures and under different states of charge, and are stored in the versioned parameter table using the state of charge range, temperature range, and charge-discharge stage as the joint lookup key.

[0063] The equivalent circuit model is first based on Update state of charge. and They represent battery cells respectively. In the and the The charged states at each alignment time are dimensionless numbers within the range [0,1]. Indicates the cell index. This represents the alignment time index, which is a dimensionless integer. This represents the coulombic efficiency and is a dimensionless number. It is read from the circuit parameter table according to the charging and discharging stages and is limited to being greater than 0 and not greater than 1. Indicates the first The arithmetic mean of the high-frequency summary of cluster currents within a certain alignment time, expressed in amperes (A). Charge is represented by positive values, and discharge by negative values; therefore, charging increases the state of charge (SOC) while discharging decreases it. This indicates the time interval between adjacent alignments, in seconds. Indicates battery cell The current available capacity, in Ah, is obtained from the most recent capacity verification result and must be strictly greater than 0. The constant 3600 is used to convert A·s to Ah. This means truncating the state of charge to 0 to 1; before calculation, the coulomb efficiency and current available capacity are checked. If any value is missing, non-numerical, or does not meet the above range, the formula is not entered, the previous effective state of charge is maintained, and the mechanism parameter is set to unavailable and the predicted degradation flag is set; when the cluster current is missing or the reliability is lower than 0.3, the previous state of charge is also maintained and the predicted degradation flag is set.

[0064] Polarization voltage according to Recursion, where , and They represent battery cells respectively. Polarization overpotential relative to the open-circuit voltage at adjacent time points, in units of V, with positive charging direction and negative discharging direction. This represents the polarization attenuation coefficient and is a dimensionless number within (0,1]. This represents polarization resistance, measured in Ω. The polarization capacitance is expressed in F. Polarization resistance and polarization capacitance are obtained from the circuit parameter table according to the state of charge, temperature, and charge / discharge stage, and both must be strictly greater than 0. Before calculating the polarization attenuation coefficient, two parameters are verified. If either is missing, non-numerical, or not greater than 0, the polarization attenuation and polarization voltage formulas are not executed. The most recent valid positive value parameter from the same cell and parameter version is used preferentially. If the parameter does not exist, the polarization prediction is set to invalid, and parameter anomaly and prediction degradation flags are set. The polarization attenuation coefficient is calculated in real-time using the listed exponential formula after obtaining the two valid parameters and the adjacent time interval, without being read from the parameter table. and Using the definition of the previous formula, The definition of the natural exponential function in step S1 is used, and its current input is a dimensionless number. At the beginning of each diagnostic window, if the previous effective polarization overpotential of the same cell and the same circuit parameter version exists in the mechanism state cache, it is used as the starting value of the window. If the state does not exist but the reliability of the terminal voltage at the first alignment moment is not less than 0.7, the difference between the open circuit voltage obtained by subtracting the voltage from the initial charge state and the ohmic voltage drop calculated by the cluster current at the first moment is used as the starting polarization overpotential. If neither source is available, the starting polarization overpotential position is set to 0 and a zero-value degradation initialization flag is set. The initialization source, parameter version, and time are recorded and saved with the circuit model state. When the lookup key falls between adjacent intervals, bilinear interpolation is performed on the resistance and capacitance. If it exceeds the calibration range, the nearest boundary parameter is used and a parameter extrapolation flag is set.

[0065] Before calculating the predicted voltage, first verify that the ohmic resistance read from the circuit parameter table is finite and non-negative. If the table value is missing, non-numerical, or negative, prioritize using the most recent valid non-negative value from the same cell and parameter version. If no valid value is found, do not proceed to the predicted voltage formula; instead, set the predicted voltage and corresponding voltage residual to invalid and set parameter anomaly and prediction degradation flags. Only when a valid non-negative table value is obtained is a fitting boundary of 0.5 to 2 times that table value established. After verification, proceed as follows: The predicted voltage is obtained. Indicates battery cell In the The predicted voltage at each alignment moment, in V. This represents the open-circuit voltage lookup function, whose input is the state of charge and whose output unit is V. The open-circuit voltage curve is obtained from a static calibration test. This represents the internal resistance in ohms, with the unit being Ω. , and Following the aforementioned definition of positive charging, the ohmic voltage rise and polarization overpotential increase the terminal voltage during charging and decrease the terminal voltage during discharging. Bounded least squares fitting is performed on the most recent 30 s window using voltage samples with a reliability of not less than 0.7 and no missing markers in the image. The ohmic internal resistance and polarization resistance are allowed to be adjusted within a range of 0.5 to 2 times the above effective current values. The fitting result is written into the circuit model state record along with the parameter version only when it is a finite and non-negative number.

[0066] thermal model Predict temperature, Indicates battery cell The predicted temperature at the next alignment time, in °C. Indicates battery cell The temperature used for recursion and parameter lookup at the current moment, in °C. Indicates the first The ambient temperature of the battery cluster at each alignment time, in °C, is read from the cluster-level ambient temperature field of the multi-source state profile in step S1, according to the battery cluster identifier and alignment time. This represents the equivalent thermal resistance from the battery cell to the environment, expressed in K / W, and must be strictly greater than 0. This indicates the equivalent heat capacity of the battery cell, expressed in J / K, and must be strictly greater than 0. This refers to the battery cells determined by their interconnection relationships. Hot-coupled neighbor index set, This represents the index of the neighboring battery cell in the set. Indicates battery cell with neighbors The thermal coupling coefficient between them, in W / K, is fitted by the module constant temperature chamber test and must be a non-negative number. Indicates neighbors In the The effective temperature at each alignment moment, in °C. , and Following the aforementioned definition, the ohmic internal resistance must be non-negative. Before calculating the thermal model, verify the equivalent thermal resistance, equivalent heat capacity, ohmic internal resistance, and all thermal coupling coefficients involved in the summation. If any value is missing, non-numerical, or does not meet the above range, prioritize using the most recent valid parameter of the same parameter version. If no valid parameter exists, invalidate the thermal prediction, set parameter anomaly and prediction degradation flags, and do not generate corresponding temperature residuals. The summation term is the term flowing into the cell. Thermal coupling power, in W, is positive when the neighbor temperature is higher than the current cell temperature. If there is no effective neighbor temperature, the neighbor is deleted from the set. When the set is empty, the summation item is set to 0. When the ambient temperature is missing or its reliability is lower than 0.3, the median effective temperature of the same module is used as the environmental proxy and the proxy source is marked.

[0067] When the current measured temperature of the target cell is missing or the reliability is below 0.3, the mechanism calculation module sequentially selects the previous effective measured temperature no more than 10 seconds from the current time, the previous effective model predicted temperature, and the median temperature of the cell with a reliability of not less than 0.7 in the same module as the recursive temperature. It sets the historical measured replacement, model predicted replacement, or module median replacement flags respectively, and uses the replacement temperature for looking up in the circuit parameter table and thermal parameter table. When using historical measured replacement, the temperature input variance is amplified by 1.2 times; when using model predicted replacement, it is amplified by 1.5 times; and when using module median replacement, it is amplified by 2 times. When all three sources are unavailable, the thermal prediction is set to invalid, the temperature missing and mechanism degradation flags are retained, and no temperature residual is generated. The missing flag is input in step S3, and the temperature sympathetic action of the same module is added to the candidate set in step S4.

[0068] Mechanism residual module calculates respectively and , Indicates battery cell In the Voltage mechanism residuals at each alignment moment, in V. This indicates the corresponding measured cell voltage, in volts (V). Using the definition of predicted voltage, Indicates battery cell In the The temperature mechanism residuals at each alignment time, in °C. This indicates the corresponding measured cell temperature, in °C. The predicted temperature definition is used; for locations with missing markers, the numerical residuals are not calculated but the residual missing markers are retained; for locations with parameter extrapolation or prediction degradation markers, the residual quality field is written at the same time.

[0069] One record in the circuit parameter table includes the state of charge range of 0.4 to 0.6, temperature range of 20 ℃ to 30 ℃, discharge stage, internal resistance of 0.0008 Ω, polarization resistance of 0.0015 Ω, polarization capacitance of 12000 F, coulombic efficiency of 0.995, and version ECM-2026Q2. Another record in the thermal parameter table includes the module position as the 8th module, equivalent thermal resistance of 1.8 K / W, equivalent thermal capacity of 980 J / K, and adjacent thermal coupling coefficient of 0.12. W / K and version TH-2026Q2; the parameter table is updated after capacity verification, quarterly pulse test or module structure change, and the old version is retained to support the recalculation of historical degradation sequences. When no record is hit, online generation of uncalibrated parameters is prohibited; when both the circuit parameter table and thermal parameter table are hit and there is no most recent valid parameter version for the same cell, the mechanism calculation module sets the predicted voltage, predicted temperature and model parameter fields in the predicted state of the cell to null values. The state of charge field preferentially adopts the most recent valid state of charge of the same cell in the mechanism state cache. If there is no cache value but the cluster current is reliable and the coulomb efficiency and current available capacity pass the valid domain verification, the coulomb integral result of the current cluster is used. When both sources are unavailable, the state of charge word is... The segment is set to null and a missing state of charge mask is written; at the same time, the unavailable mechanism parameter and the predicted degradation flag are written, and the corresponding voltage residual and temperature residual are set to missing instead of being replaced by zero value. Step S3 receives the sample with the unavailable parameter mask, the missing state of charge mask and the missing residual mask; the mechanism calculation module separately generates a parameter recalibration maintenance request containing cell identifier, missing parameter type, most recent valid version and request time and sends it to the energy storage battery management system. This maintenance request is not a candidate sensing action and does not participate in the action utility ranking; step S4 prohibits increasing the exploration intensity based on the current mechanism result of the cell, only retains the fixed safety minimum sampling configuration, and adds the same module sympathetic sampling actions that conform to the candidate action field structure to the candidate set;

[0070] The predicted state record saves the predicted voltage, predicted temperature, state of charge, circuit model parameters, thermal model parameters, parameter version, and quality flag according to the cell identifier and alignment time. The electro-thermal mechanism residual record saves the voltage residual, temperature residual, missing marker, and residual quality field with the same key. The records for 60 consecutive seconds form a residual tensor that corresponds one-to-one with the multi-source state profile. The mechanism calculation module writes the predicted state and electro-thermal mechanism residual into the mechanism state cache for joint diagnosis in step S3, and provides the corrected model state, parameter version, and residual statistics to step S4 to generate the conditional observation distribution under each competing fault assumption.

[0071] In this specific embodiment, S3 includes:

[0072] The joint diagnostic module takes the 60-second multi-source state profile tensor from step S1, the electro-thermal mechanism residual tensor from step S2, the cell connection relationship, and the most recent 90-day historical degradation sequence of each cell as input. It is based on the cell construction time step characteristics, including normalized voltage, temperature, cluster current, cluster-level ambient temperature, charge-discharge stage unique thermal encoding, voltage residual, temperature residual, missing data marker, data completeness, sensor reliability, state of charge, profile time interval, and the number of subsamples, effective proportion, mean, range, standard deviation, and slope in the high-frequency summary of each monitoring parameter. Voltage, temperature, current, ambient temperature, residual, and high-frequency summary with physical units are normalized using the median and interquartile range of normal samples in the training set and truncated to -5 to 5. Missing or inapplicable summary values ​​are filled with 0, and the true zero value is distinguished by the summary insufficient flag and the inapplicable flag.

[0073] The temporal coding network employs a two-layer gated recurrent unit (GRU) architecture, with each layer having a hidden dimension of 128. The first layer receives 60 consecutive time-step features aligned to each other and outputs a time-step hidden vector. The second layer encodes the current diagnostic window together with 90 degradation feature vectors obtained by summarizing the historical degradation sequence daily. The historical degradation features include daily average capacity estimation, internal resistance estimation, maximum temperature, temperature difference, voltage dispersion, effective cycle count, and residual quantiles. The time interval is used as the attenuation gate input. The same field order and normalized version are maintained during training and inference. The time window is less than 60 seconds. When s, zeros are padded at the front and a missing marker is set; a fixed daily bucket is established for the historical degradation sequence according to the 90 natural days before the current diagnosis date. Each daily bucket only uses the records that are qualified by the quality mark on the current day to summarize the above degradation features, and at the same time generates the daily bucket effective mask and the time interval from the nearest effective day; when the operation is less than 90 days, invalid daily buckets are padded at the front of the sequence. When the historical record is interrupted, invalid daily buckets are inserted on the corresponding date. The normalized features of the padded or inserted daily buckets are filled with 0 and must be distinguished by the daily bucket effective mask and the true zero value; when there are fewer than 30 effective daily buckets, a historical cold start flag is set. The second layer only uses the effective daily buckets and their time intervals to participate in the decay encoding, applies a cold start mask to the output of the historical branch, and amplifies the model integration uncertainty to 1.5 times the original value. During the training phase, the historical days are randomly masked with the same daily bucket mask rule so that the cold start input path has corresponding training samples;

[0074] The graph propagation layer establishes a cell connection graph based on the series sequence, module affiliation, and shared temperature sampling location. Nodes represent target cells or cell groups. Edges are established between adjacent cells in series and between cells sharing a thermal path within the same module. Each edge records the connection type and the correlation coefficient calibrated using historical normal data. Layer map features by spread, and They represent battery cells respectively. In the Layer and first The feature vectors of the layer are dimensionless vectors. Indicates the target cell index. This represents the graph propagation layer index, which is a dimensionless integer. Indicates the relationship with battery cells A set of connected neighbor indices. Indicates the neighbor index that participated in feature propagation. This indicates a neighbor index used only for calculating the sum of neighborhood edge weights. and These represent the corresponding neighbor edge weights, which are dimensionless numbers within the range [0,1], obtained from the connection type and the correlation coefficient of normal data. The fractional expression represents dividing each edge weight by the cell weight. The sum of all neighboring rights, and Indicates the first The training weight matrix of the layer's self-nodes and neighboring nodes has dimensionless elements. This represents the element-wise ReLU activation function;

[0075] The graph propagation process performs two layers and normalizes the edge weights of each neighbor item according to the formula. When the neighbor set is empty or the sum of edge weights is 0, the fractional summation is skipped and only the self-node item is retained, forming a spatiotemporal joint feature containing time degradation information, mechanism residual information, high-frequency summary, and connectivity. This feature is input into the health status regression head, remaining useful life regression head, fault type classification head, and conditional location classification head, respectively. The two regression heads each output the mean and logarithmic variance. The health status value is the ratio of the current available capacity to the rated capacity and is limited to 0 to 1. The remaining useful life is in units of equivalent complete cycles and is limited to non-negative values. The fault type classification head outputs the softmax probability for normal, internal short circuit, loose connection, abnormal capacity decay, thermal anomaly, and sensor fault. The conditional location classification head outputs the probability for 240 cell locations and unknown cell locations under a physical fault mask. Under the sensor fault mask, the probability of channel location labels composed of monitoring parameter identifiers and sampling channel identifiers and the probability of unknown channel locations are output. In addition to the identifiers, objects and versions of the main channel of cell voltage, lateral channel, module temperature channel, cluster current channel and cluster-level ambient temperature channel, the conditional location label table also sets a unique fault-free location label. When the fault type is normal, the conditional location classification head outputs a conditional probability of 1 for the fault-free location label and 0 for all physical fault locations and sensor channel locations. When the fault type is physical fault or sensor fault, it outputs 0 for the fault-free location label and normalizes only within the corresponding cell location or channel location label space. S3 writes the fault type, conditional location label, corresponding posterior probability and label table version into the joint diagnostic cache for step S4 to directly generate sensor fault hypotheses and query the mapping table.

[0076] The joint state discretizer divides the health state value range of 0 to 1 into 50 left-closed and right-open intervals with an equal width of 0.02, and makes the last interval contain 1. The remaining lifetime of 0 to 3000 equivalent complete cycles is divided into 60 left-closed and right-open intervals with 50 cycles each, and a tail interval from 3000 to positive infinity is set. The Gaussian output of the two regression heads is integrated on each interval to obtain the interval probability. The interval boundaries, tail merging rules and version GRID-2026Q2 are stored in the model version package. The fault type, conditional position containing the fault-free position label, health state interval and remaining lifetime interval Cartesian product constitute a finite joint state support set. The normal type is only compatible with the fault-free position label, and any abnormal fault type is incompatible with the fault-free position label. The same version support set is used before and after adding simulated observations or real observations in the same diagnostic session. When the model update causes the probability to fall outside the boundary, it is merged into the corresponding first interval or tail interval.

[0077] The outputs of the four diagnostic heads are passed through Form a joint posterior distribution. Indicates the given input Joint diagnostic variables The posterior probability is a dimensionless number within [0,1]. This represents the model input, consisting of the current multi-source state profile, mechanistic residuals, connectivity graph, and historical degradation sequence. This represents the combined value of the fault type, fault location, health status range, and remaining service life range. The output probability represents the fault type classification header. This indicates the probability output by the fault location classification header. This represents the Gaussian probability of the regression head of the health status within a given interval. This represents the Gaussian probability of the remaining useful life regression head within a given interval. This represents the normalization constant obtained by summing all joint values ​​and is a positive dimensionless number; after applying a zero-probability mask to the location combination that is incompatible with the fault type, it is re-normalized, where the location factor of the normal type is uniquely given by the unit probability of the fault-free location label and the normal probability is not copied at other locations; the same version of the location label table is used before and after the simulation observation is added.

[0078] This specific implementation uses five model instances with the same network structure but different initial parameters and training sample bootstrapping. In each round, the completeness of the output fields, finiteness of values, non-negativity of variance, and normalization of classification probabilities of each instance are checked. Instances that pass all checks constitute a valid set of model instances. A new aggregation result is generated only when the number of elements in this set is not less than 3, and output is performed for any scalar. according to Determine the uncertainty, and sum only iterate through valid model instances. If there are fewer than 3 valid instances, do not enter the formula and execute the model unavailable branch described later. Indicates output Total uncertainty, in units of output The square of a unit The output index represents the health status value, remaining lifespan, or probability of a certain category. This represents the number of valid model instances in the current round and is a positive integer not less than 3. Represents the local renumbering index of a valid model instance. Indicates the first A valid instance for output The given mean, in units of Consistent Represents the arithmetic mean of the means of all valid instances and is Same unit, Indicates the first The conditional variance of the output of each valid instance, in units of... The square of the unit; for health status values ​​and remaining useful life, the conditional variance is obtained by taking the natural exponent of the logarithmic variance output by the corresponding regression head; for failure type probability or conditional location probability, let equal to the The softmax probabilities output by each valid model instance and according to... Determine the Bernoulli conditional variance of the categorical indicator variable, where the categorical probability and corresponding uncertainty are both dimensionless quantities;

[0079] The model training samples include timestamps, charging / discharging stages, cell voltage, cell temperature, cluster current, cluster-level ambient temperature, sampling channel identifiers, original subsamples required to form high-frequency summaries of each monitoring parameter, quality identifiers, cell connection relationships, fault type labels, fault location labels, health status labels, and remaining service life labels. The training samples strictly adhere to the same version of the time alignment, missing data determination, sensor reliability, and high-frequency summarization process as the online step S1, generating training inputs according to the same field order of the online time step features. Historical records lacking original subsamples or cluster-level ambient temperatures are only retained if the corresponding missing mask has been covered by the random masking rules during the training phase; otherwise, they are excluded from the training set. Before training, the residuals for each sample are generated using the same version of the electro-thermal mechanism model as the online step S2. The total loss is calculated as follows: calculate, This represents the dimensionless total training loss. and Let represent the normalized Gaussian negative log-likelihood loss for health status and remaining useful life, respectively. and The cross-entropy loss represents the fault type and the fault location, respectively. This represents the consistency loss between the predicted category and the mechanism residual pattern. The residual pattern of the consistency loss is composed of the mean absolute value, standard deviation, 95th percentile absolute value, and linear slope of the effective voltage residual and temperature residual within the current window for each cell. Each statistic is normalized using the interquartile range of the training set. Positions with missing values, parameter extrapolation, or prediction degradation flags are excluded by a mask. For each fault type, the mean of the sample pattern confirmed by maintenance records and retests in the training set and with valid residuals is used as the prototype of that type. The squared Euclidean distance from the current sample pattern to each type prototype is calculated. The distance is negative and softmax is applied to obtain the residual pattern category probability. Then, the cross-entropy is calculated with the sample's actual fault type. The average is first applied among valid cells and then among batch samples to obtain the dimensionless consistency loss. Samples without any valid residual positions are not included in the denominator of this loss but still participate in the other four losses. The fault type prototype and normalized version are saved with the model version package. All five losses are dimensionless quantities. , , , and The values ​​represent the non-negative dimensionless weights determined by grid search on the independent validation set, which in this specific implementation are 1.0, 1.0, 1.2, 1.2 and 0.5 respectively;

[0080] Training data is divided into 70% training set, 15% validation set, and 15% test set according to battery clusters to avoid leakage of samples from the same cluster. Fault labeling is confirmed by maintenance records and retest results. Health status and remaining service life labeling are obtained by capacity verification and subsequent degradation trajectory. Each model instance is trained using the Adam optimizer. The training stops when the total loss of the validation set does not decrease for 10 consecutive rounds and the optimal parameters are retained. Input fields, normalized statistics, connectivity graph construction rules, model parameters, and mechanistic parameter versions together constitute the model version package. When online input field or parameter versions are incompatible, active diagnosis stops and the most recent effective diagnosis result is used, and a model incompatibility flag is set.

[0081] The training batch is stratified and sampled according to six categories: normal, internal short circuit, loose connection, abnormal capacity decay, thermal anomaly, and sensor failure. When the number of samples in any failure category is less than 20% of the number of samples in the normal category, resampling with time jitter and measurement noise disturbance is performed on that failure category. If the voltage, temperature, or current after disturbance exceeds the range of the acquisition in step S1, the enhanced sample is discarded. When a training batch is missing a certain fault location, position enhancement is performed only after the sampling channel configuration, module structure, series boundary conditions, and thermal adjacency of the source and target locations are consistent and pass the exchangeability verification. The original sampling fields at the cell level, multi-source state profile features, electro-thermal mechanism residuals, quality identifiers, historical degradation features, node identifiers, connection diagrams, and fault location labels are rearranged synchronously according to the same one-to-one position mapping. Cluster-level shared quantities such as cluster current, cluster-level ambient temperature, and charging / discharging stages remain unchanged. If any exchangeability condition is not met, position translation is canceled, and only same-location resampling without changing the evidence location is used. The above rules are only used for the training set and do not apply to the validation and test sets.

[0082] The validation module sets a display acceptance threshold of 0.5 for both fault type and fault location. This threshold controls the display of a single label without pruning the joint posterior distribution. When the maximum class probability is lower than the display acceptance threshold, the displayed label is marked as pending, and all classification probabilities are retained. When the regression variance of health status or remaining useful life exceeds the 95th percentile threshold of the independent validation set variance, a high uncertainty flag is set for the corresponding regression result. In each round, the aforementioned validity check is performed on 5 model instances, and the mean, conditional variance, classification probability, and joint posterior are recalculated using only valid instances. When the number of valid model instances is not less than 3, a new joint posterior version is allowed to be written; when it is less than 3, the current round of joint normalization is stopped, the previous valid joint posterior version is used, and the model becomes unavailable.

[0083] The model registry stores the input field list, normalization version, mechanism parameter version, connection graph rule version, network parameter hash, validation set loss, classification acceptance threshold, and regression variance threshold. When the online input fields are inconsistent with the model registry, the model instance is rejected. When there are fewer than 3 valid model instances, the joint diagnostic module outputs a model unusable flag, does not generate a new joint posterior version, and freezes and uses the most recent valid joint posterior version. If there is no most recent valid joint posterior version, the safe initial prior version that is included with the model version package and verified offline is read, and the initial prior is prohibited from being used as new diagnostic evidence. In step S4, the previous round of active exploration action plan is frozen in this state and the exploration frequency cannot be increased. If there is no previous round of action plan, only the fixed safe minimum sampling configuration is executed. In step S5, the current round of stopped evidence status is set to insufficient evidence and the continuous evidence count is cleared to zero. When the reloaded model passes the field, numerical range, and output normalization checks for 3 consecutive diagnostic windows, the model is restored to the model set and the restoration time is written to the diagnostic audit record.

[0084] The joint diagnostic module writes the output of each model instance, the normalized joint posterior distribution, health status value, remaining service life, fault type, fault location, and corresponding uncertainty into the joint diagnostic cache. The cache key includes the diagnostic round, battery cluster identifier, and profile version. Joint hypotheses with a posterior probability lower than 0.01 are only hidden in the display layer and are not deleted from the internal competing fault hypothesis set. When the probability of the unknown location is the highest, the cell group level location is retained. The joint posterior distribution and its uncertainty serve as direct inputs for constructing candidate sensing actions, predicting conditional observation distributions, and calculating information gain in step S4, thereby forming initial diagnostic evidence for an active exploration sensing and intelligent diagnostic method for energy storage batteries.

[0085] In this specific embodiment, S4 includes:

[0086] The action evaluator reads the joint posterior distribution and its uncertainty from step S3. First, it groups the joint states by fault type and conditional location label. Then, it sums the posterior probabilities of all healthy state intervals and remaining service life intervals in each group, obtaining a grouped posterior indexed by fault type and location. Within each group, it retains the conditional weights of the healthy state interval and remaining service life interval relative to the group's probability. The grouped posteriors are sorted from largest to smallest, and if probabilities are the same, they are sorted lexicographically by the key composed of fault type and conditional location label. First, it retains the unique normal group composed of normal fault types and fault-free location labels, and the group with the highest probability among all abnormal fault types with a posterior probability greater than 0. Then, it supplements this group with the above sorting to a maximum of 7 explicit groups. All remaining groups not explicitly expanded are merged into a tail hypothesis, with the tail posterior probability being the sum of the probabilities of all merged groups. Its internal components, healthy state, and remaining service life sub-states retain their original posterior proportions. The resulting set of no more than 8 members constitutes the current competing fault hypothesis set. , This represents the explicit grouping and tail hypothesis set, which sums the information gain and fault consequence risk of all participants in this round of conditional observation prediction, information gain, and fault consequence risk. The sum of the weights of all members is 1. When generating simulated observations for explicit groups, a healthy state and remaining service life sub-state are first extracted according to the conditional weights within the group. For tail hypotheses, the fault type and location group are first extracted according to the weights of the components within the tail, and then the sub-states are extracted according to the conditional weights of that group, thus preserving all posterior probability quality without enumerating millions of joint states. Before and after the addition of simulated observations, the posterior and divergence are updated on the same complete joint state support set fixed in step S3. The grouping key, tail members, and random seeds are recorded with each diagnostic round. The candidate sensing action record consists of the target cell or cell group, monitoring parameters, sampling channel, sampling frequency, sampling start phase, duration window, and action version. The frequency set for this specific implementation method is 1 Hz, 5 Hz, and 10 Hz for cell voltage, 0.2 Hz, 1 Hz, and 2 Hz for cell temperature, and 10 Hz and 20 Hz for cluster current. The duration window is 30. s or 60 s, the starting phase is taken as the current alignment time or the first alignment time after the next charge / discharge phase switching;

[0087] The action construction module first retains cells with the same location as the competing fault, cells in the same module, and their series-connected adjacent cells. Then, based on the fault type, it retains distinguishable monitoring parameters: internal short circuits retain voltage, temperature, and cluster current; loose connections retain voltage and cluster current; abnormal capacity decay retains voltage and cluster current; thermal anomalies retain temperature, voltage, and cluster current; and sensor faults retain fault parameters and corroborating channels in the same module. The candidate frequency is not lower than the current frequency and does not exceed the upper limit listed in the acquisition module's capability table. Actions with the same target, parameters, channels, frequencies, phases, and windows are deduplicated using a combination key. When the target cell set is empty, the fixed minimum safe sampling configuration is retained as a candidate.

[0088] The uncertainty threshold table is established based on the 95th percentile of the output uncertainty of correctly diagnosed samples on the independent validation set and stored according to model version. When the uncertainty of the fault type exceeds the corresponding threshold, all monitoring parameters and circumstantial evidence channels that can distinguish the top three types of faults are retained. When the uncertainty of the fault location exceeds the threshold, the top three cells with the highest posterior probability, cells in the same module, and adjacent cells in series are retained. When the uncertainty of the health status value exceeds the threshold, the voltage and cluster current operation lasting for 60 seconds is retained. When the uncertainty of the remaining service life exceeds the threshold, the start of the next stable charge and discharge phase lasting for 60 seconds is retained. The voltage and cluster current of s are activated; when multiple outputs exceed the threshold simultaneously, the actions are merged and resources are allocated in the order of safety monitoring, fault consequence risk, fault location uncertainty, fault type uncertainty, health status uncertainty, and remaining service life uncertainty. When the threshold table is not hit, the maximum value of the threshold of the same model family is used and a threshold fallback flag is set. After merging and deduplication, the actions are sorted in the following order: resource priority, number of covered objects from most to least, estimated resource consumption from least to most, and lexicographical order of action combination key. A maximum of the first 12 normal active candidate actions are retained in each round. Fixed safety minimum sampling configuration is not included in this upper limit. The truncated actions can only be evaluated when the sorting conditions are met again in the next round.

[0089] The parameter fitter uses the bounded least squares rule from step S2 to obtain the baseline model parameters on the current multi-source state profile, and performs recursive correction using the mean of the mechanism residuals over the last 30 seconds. , This represents the corrected electro-thermal mechanism parameter vector, whose components have the same units as the corresponding resistance, capacitance, thermal resistance, or thermal capacity parameters. Represents the baseline model parameter vector obtained by fitting the current image and is related to... Same unit, This represents the residual correction gain matrix, where each element has units such that the product has the same units as the corresponding parameter components. It is obtained by recursive least squares fitting on normal calibration samples, and the correction amount is limited to no more than 20% of the reference parameter. This represents the mean vector of voltage and temperature residuals over the last 30 seconds, with components in units of V and ℃, respectively. If the effective position of the residual is less than 70% of the planned position in the window, the current correction is not performed. The previous effective correction parameters of the same cell and the same mechanism parameter version are used first. If there are no previous effective correction parameters, the corrected parameters are directly initialized as the reference model parameters for this round, and a correction cold start flag is set. The initialization source, parameter version, and diagnostic round are written into the condition observation parameter record.

[0090] The normal state calibration table stores the baseline observation noise covariance according to monitoring parameters and sampling channels; the physical fault mapping table stores model parameter offsets according to fault type, fault location, state of charge range, and temperature range; and the sensor fault mapping table stores observation bias, drift per second, and noise covariance correction according to fault type, monitoring parameters, and channel location. All three tables are established using maximum likelihood estimation from reserved normal or fault calibration samples. Record fields include unit, effective range, number of samples, version, and effective time. Updates are made quarterly upon calibration, sensor replacement, or when 20 new confirmed fault samples are added. The query key for the sensor fault mapping table directly uses the monitoring parameter identifier and sampling channel identifier from the conditional location label in step S3. When the physical fault location is an unknown cell location, the fault is identified by cells within the same battery cluster that are not excluded by the quality mask and are related to the fault location. All cells with compatible fault types are considered as candidate locations. The posterior probability of each cell location is re-normalized within the candidate set as the mixing weight. For each candidate location, the model parameter offset is queried, and conditional observation components are generated. The conditional observation distribution for unknown cell locations is a finite mixture distribution of these components. When the sensor fault location is an unknown channel location, all currently available channels with the same monitoring parameters and the same module are considered as candidate channels and are mixed with equal weight. For each candidate channel, the observation bias, drift per second, and noise covariance correction are read. If individual candidate records are missing, conservative components are formed using the largest absolute value of the bias, the largest absolute value of the drift, and the largest eigenvalue of the covariance correction among existing records. A voltage sensor bias fault record includes channel V-C087, bias 0.012 V, drift 0.00002 V / s, noise variance correction 0.000009 V², and version OBS-2026Q2.

[0091] The joint state parameter mapping table is configured according to health state intervals, remaining service life intervals, temperature intervals, and storage capacity ratios during charge / discharge phases, ohmic internal resistance ratios, polarization resistance ratios, polarization capacitance ratios, thermal parameter ratios, and degradation rates per cycle. Its calibration samples are derived from capacity verification, pulse tests, and continuous degradation tests. The version is bound to the joint state support set version in step S3. For each explicit competing fault group, before generating each conditional observation sample, a health state interval and remaining service life interval are extracted according to the conditional weights within that group. The tail assumption first extracts a fault type and location group according to the weights of its internal components before extracting sub-states. The capacity ratio is taken as the midpoint of the Kang state interval, the midpoint of the remaining service life finite interval, and the lower bound plus 25 equivalent complete cycles as the representative value of the tail interval. First, the capacity ratio and degradation factor are applied to the calibration reference parameter in step S2, then the physical fault parameter offset is superimposed, and finally the observation bias, drift and noise correction are applied to the sensor fault state. The conditional observation mixture distribution of the same group is generated from all sampled states of the same group. The tail assumption is that a mixture distribution that retains the tail quality is generated by two-level sampling. When the mapping key is not hit or the representative value exceeds the calibration range, the nearest boundary record is used and the parameter covariance is amplified by 1.5 times.

[0092] Targeting actions and the competing failure hypothesis First, subtract the mean of the normal calibration sample from each physical observation component and divide by the calibration standard deviation of the same parameter and channel, then proceed according to... Generates conditional observation distribution, Indicates the execution of candidate perception actions And the fault assumption is Dimensionless Standardized Observations The probability density of the Lebesgue measure relative to the normalized observation space is dimensionless. Indicates the candidate perceived action index. Indicates the index of race failure assumptions. This represents a dimensionless standardized voltage, temperature, or current observation vector, sorted in ascending order by channel combination key and then by action timestamp. This represents the dimensionless conditional mean vector under the same permutation. Represents the dimensionless conditional covariance matrix. This represents the multivariate normal probability density function determined by the listed mean vector and covariance matrix. Before standardization, the calibration mean must be a finite number, the calibration standard deviation a finite and strictly positive number, the one-point variance a finite and strictly positive number, the autocovariance kernel terms a finite number, and the cross-channel covariance coefficients a finite number ranging from -1 to 1. When the cross-channel coefficients are invalid, they are set to 0 and an independent backoff flag is set. When the calibration mean, calibration standard deviation, one-point variance, or autocovariance kernel is invalid, the one-point variance with the largest value among the currently valid calibration records for the same parameter and model channel is used first. If a record is found but no valid record is found, the fault hypothesis is marked as conditionally uncalculated and enters the fixed safety minimum sampling and circumstantial sampling branch. The normal state calibration table stores the single-point variance, the autocovariance kernel indexed by time lag, and the cross-channel covariance coefficients of synchronous channel pairs in the same module for each monitoring parameter, sampling channel, and supported sampling frequency. When assembling the conditional covariance matrix, the matrix elements of any two timestamps in the same channel are queried according to their time difference to obtain the autocovariance kernel. The synchronous or asynchronous elements of different channels are queried according to the channel pair coefficients and the corresponding autocovariance kernels. The difference kernel is multiplied, and uncalibrated cross-channel terms are set to 0 with an independent backoff flag. The correction amounts for physical fault parameter offset covariance and sensor noise covariance are extended to the corresponding matrix blocks according to the same channel-timestamp arrangement. The correction amount for time-related terms is only added to the corresponding diagonal elements. After assembly, the matrix is ​​symmetric and non-positive eigenvalues ​​less than 1% of the standard variance of the same parameter are raised to the strictly positive lower limit. Then, the matrix is ​​verified to have a finite number of elements and the smallest eigenvalue is strictly greater than 0. When the verification fails, only the diagonal of the effective positive single-point variance of each channel is retained. The covariance matrix is ​​set and a covariance degradation flag is set. If the diagonal backtracking still fails, the conditional observation is marked as uncomputable. This makes the matrix dimension equal to the length of the observation vector generated by the action. In the normal state, the corrected reference parameters and the reference noise covariance are used. In the physical fault state, the mapping parameter offset is superimposed on the corrected parameters and the model in step S2 is run. In the sensor fault state, the deviation and the drift accumulated over time are applied to the model output and the covariance correction is superimposed on the reference covariance. Then, the same calibration mean and standard deviation are used to complete the dimensionless transformation.

[0093] Before the simulation starts, the action evaluator reads the available computing load of the edge diagnostic processor for the next 60 seconds. It estimates the worst-case computational load using the 95th percentile of the measured times of the 100 previous S2 and S3 forward updates. Normal active action evaluation is only initiated if the available computing load does not exceed 80% of the current round and a maximum of 12 candidate actions, each with a maximum of 8 competing hypotheses, and 128 forward updates per hypothesis can be completed within the 45-second computation deadline. Action evaluation is only initiated if the computational entry conditions are not met and the evaluation process reaches 45 seconds. If the deadline or available computational load falls below the reserved limit midway, no information gain or action utility ranking of incomplete or partially completed actions is used. All ordinary active candidate actions in this round are marked as computationally incomplete, and a fixed minimum safe sampling configuration is deterministically output. If the resource verification of this configuration fails, the process proceeds to controlled degradation sampling in step S5 and the reason for the computation deadline is recorded. After passing the computational admission, the action evaluator generates 128 sets of standardized observation samples for each conditional observation distribution using a fixed random seed. Each set of samples is destandardized into simulated original sub-samples with physical units and action timestamps. The time alignment, missing data determination, reliability update, and high-frequency summary process corresponding to the current profile version in step S1 are strictly called to form a simulated profile. Specifically, using the 24-hour effective proportion and number of jumps saved in the current real profile as the baseline, the simulated sub-samples of the current group are included in the temporary statistics according to the same missing data, range, effective proportion, and jump rules as in S1. The temporary sensor reliability is calculated only for this forward update and is not written back to the real calibration table. h Rolling statistics or real profile version, so that simulated and real new observations use the same field semantics and update rules; then call a forward update in steps S2 and S3, according to Calculate the expected diagnostic information gain. Indicates candidate actions The expected diagnostic information gain, in units of nat and non-negative. Represents the current set of race failure hypotheses The fault hypothesis index is used to iterate through all members. Member The current posterior probability is a dimensionless number within [0,1], and the sum of the weights of all members is 1. This represents the number of conditional observations sampled for each hypothesis, and is a positive integer; in this specific implementation, it is 128. Indicates the sampling sequence number. Indicates joining the The joint posterior distribution over a fixed joint state support set after group simulation observations. This represents the current joint posterior distribution on the same support set before being included in the simulated observations. This represents the Kolb-Leibler divergence of the former distribution relative to the latter distribution; terms with a posterior probability less than 1e-8 are bounded to 1e-8 and renormalized on the same support set.

[0094] Failure consequences and risks according to Sure, This represents the risk of the current failure and is a dimensionless number within the range [0,1]. The larger the value, the more severe the consequences. Traverse the current set of competition failure hypotheses All members, Members The current posterior probability is defined as follows: Indicating fault assumptions The normalized consequence loss is a dimensionless number within [0,1], read from the consequence loss table according to fault type, location, and current state of charge interval. The consequence loss table is calibrated by thermal runaway propagation test, connection fault heating test, capacity verification, and operation and maintenance records, recording candidate location set rules, location label table version, and loss version. The consequence loss of a record for a cell with an internal short circuit and located in a high state of charge is 0.95. The normalized consequence loss of the normal group is set to 0. For the tail hypothesis, the consequence loss table is queried for all fault types and location components within it according to their respective fault type, location, and current state of charge interval, and the largest effective normalized loss is obtained. The loss is treated as a conservative consequence loss for the tail assumption. If no matching record is found for any internal component, the loss of that component is first set to 1 and then the maximum value is taken. The action evaluator writes the aggregation method of "taking the maximum value of the internal component loss at the tail", the list of internal components, the version of the internal weight, the version of the consequence loss table, and the obtained tail loss into the risk calculation record. For unknown cell locations, the maximum value of the loss of the same fault type in the current state of charge interval is used for its candidate locations. For unknown channel locations, the maximum value of the loss of the same sensor fault type in its candidate channels is used. If the candidate set is empty or there is no matching loss record, the corresponding consequence loss is conservatively set to 1 and a loss mapping missing flag is set.

[0095] For each candidate action, the target object, monitoring parameters, and data integrity and sensor reliability corresponding to the channel are read from the status profile quality field, and then... Obtain data quality weights. Indicates action The data quality weights are dimensionless numbers within the range [0,1]. A larger value indicates better expected observation availability. This represents the minimum value of the integrity of the action-covered objects and data within the window, and is a dimensionless number within [0,1]. It represents the minimum reliability of the sampling channel used in the action and is a dimensionless number within [0,1]. When the action involves multiple channels, the minimum value is used to reduce the weight of any critical channel with low quality. When a channel is to be calibrated or the data integrity is 0, the quality weight is reset to 0.

[0096] The resource estimator calculates communication bandwidth usage, acquisition energy consumption, computational load, and switching cost based on the number of sampling points, bytes per point, acquisition front-end power, measured inference time, and current configuration switching count. Before calculating the normalized ratio, it verifies that the currently available communication bandwidth, remaining acquisition energy, and available computational load are all greater than 0. If any available quantity is not greater than 0, it directly sets a resource out-of-bounds flag for candidate actions requiring corresponding resources and skips the action utility formula. Only actions where all three available quantities are positive are processed. Calculate the utility of the action. Indicates candidate actions The action utility is a dimensionless number; the larger the value, the better. and These represent the action communication bandwidth requirement and the strictly positive currently available communication bandwidth, respectively, both in kbit / s. and These represent the energy consumption requirement for motion acquisition and the strictly positive remaining acquisition energy for this round, respectively, both in J. and These represent the motion calculation requirements and the strictly positive available computational load, respectively, both in GFLOPs. This represents the normalized switching cost relative to the current sampling configuration and is a dimensionless number within the range [0,1]. , , and This represents a non-negative, dimensionless cost weight, determined by the verification results of historical actions in relation to diagnostic benefits and resource consumption, and in this specific implementation, it is 0.25, 0.20, 0.20, and 0.10 respectively. , and Using the aforementioned definition; when the risk of the current failure consequence is 0, the information gain benefit item is 0, only the cost item is retained, and the action sequencing is still subject to the fixed safety minimum sampling configuration and the safety constraint of step S5 as a fallback;

[0097] The action evaluator reads the quality field formed in step S1 using timestamps and charging / discharging phases as alignment benchmarks, and explicitly retains missing markers that do not fall within the alignment time window; when a candidate action has a missing marker at its corresponding position, that position is included in the planned sample but not in the valid sample; when a candidate action spans the charging / discharging phase, it is split into actions within two phases; when the duration window after splitting is less than 30 seconds, the sub-action is deleted; when the sensor reliability is below 0.3, only actions from the same parameter symmetric channel are retained; when the data integrity is below 0.5, the channel is prohibited from being used as the sole observation source, thereby allowing the alignment and missing information described in the claims to directly enter the data quality calculation of the candidate action;

[0098] During the candidate action combination phase, if the target object belongs to the current maximum posterior fault location, at least one action covering that object is retained; if the sum of the posterior probabilities of internal short circuit and thermal anomaly is not less than 0.2, actions with a temperature frequency not less than 1 Hz are retained; if the posterior probability of loose connection is not less than 0.2, actions with voltage and cluster current phase are retained; if the posterior probability of abnormal capacity decay is not less than 0.2, actions spanning 60 s of stable charging or discharging phase are preferentially retained; if the posterior probability of sensor fault is not less than 0.2, both fault channel and circumstantial channel actions are retained simultaneously; if a competing fault hypothesis has no distinguishable actions, the fixed safety minimum sampling configuration is added to the action coverage list of that hypothesis and an insufficient coverage flag is recorded.

[0099] During the conditional observation generation phase, when the number of physical fault mapping records is less than 20, the parameter offset covariance is amplified to 1.5 times its original value; when the cumulative deviation of sensor fault drift exceeds 10% of the acquisition range, the mean is truncated to the range boundary and a truncation flag is added; when the covariance matrix has non-positive eigenvalues, the minimum eigenvalue is raised to 1% of the standard noise variance for the same parameter; when the simulated observation falls outside the acquisition range, it is truncated according to the saturation rules of the actual acquisition module and a saturation flag is set; if any fault hypothesis fails to hit the physical fault mapping table or sensor on its first attempt... When mapping the fault table, the assumption is retained instead of being replaced by the normal state. At the same time, it is marked as conditional observation that cannot be computed, so that the stopping condition of this round is not met and the continuous evidence count is cleared. Step S4 only outputs the fixed safety minimum sampling configuration and the circumstantial sampling action that can supplement the corresponding mapping. The continuous miss count is only used to generate a mapping maintenance alarm when the third round is reached. It is not used as a waiting period for the first two rounds of calculation. When the same query key hits a valid version record, the count is cleared. When the candidate set of unknown positions is empty or the candidate records are insufficient to form the aforementioned conservative components, it is processed as the same uncomputable branch.

[0100] During the information gain calculation stage, when the simulated observation contains missing locations, the joint input rule of zero-filling and missing labeling from step S3 is used; when the updated joint posterior normalization constant is less than 1e-12, the simulated sample is discarded; when the effective simulated samples are less than 80% of the planned number, the candidate action is marked as insufficient and its information gain is set to 0; when the relative difference in information gain between two independent random seed calculations exceeds 10%, 128 additional samples are added for re-estimation; if the difference still exceeds 10% after adding, the smaller value of the two estimates is used as the conservative information gain.

[0101] The action evaluator synchronously reads the safety constraint metadata required in step S5. When a candidate action does not cover a safety monitoring object near the working boundary, a safety coverage failure flag is set; when the sampling frequency is lower than the minimum safe frequency, a frequency failure flag is set; when the predicted start delay exceeds the maximum sampling start delay, a delay failure flag is set; when the number of available sampling channels for cell voltage, cell temperature, or cluster current is lower than the minimum number set by the energy storage battery management system, utility-based action switching is frozen; when any resource requirement of the fixed minimum safe sampling configuration exceeds the current available amount, a resource insufficiency flag is preset. These flags enter step S5 with the candidate action table and do not replace the final constraint screening in step S4.

[0102] The action evaluator also reads the continuous evidence window status returned from the previous round in step S5. When the conformal prediction set of fault type or fault location exceeds the set threshold, the location action that can distinguish the category within the set is retained. When the prediction range of health status or remaining service life exceeds the width threshold, the voltage and current actions with a duration window of 60 s are retained. When the mechanism residual does not fall into the consistency range, the action that can retest the corresponding residual parameter is retained. When the continuous evidence window has not repeatedly met the stop condition, the generation of stop command is prohibited and candidate actions continue to be output. When the stop condition has been met in the previous round, only the fixed safety minimum sampling configuration is retained and the active expansion of the candidate set is stopped.

[0103] The action evaluator receives constraint snapshots from the energy storage battery management system, where the energy storage battery management system sets the minimum number of available sampling channels required for safety monitoring according to cell voltage, cell temperature, and cluster current, and uses the minimum number of available sampling channels as the channel requirements for the corresponding monitoring parameters in the fixed safety minimum sampling configuration; the action evaluator also receives the diagnostic evidence status returned in step S5. When the number of conformal prediction set elements, the width of conformal prediction interval, and the consistency of mechanism residual repeatedly meet the corresponding conditions within the continuous evidence window, the diagnostic evidence determined by step S5 meets the stopping condition, and the generation of active exploration actions is stopped while the safety monitoring actions are retained;

[0104] The resource cost table stores the number of bytes per point, acquisition front-end power, model forward propagation time, computational load, and configuration switching latency according to sampling channels, sampling frequency, and duration window. The data is formed from the 95th percentile of the first 100 measurements before deployment and is retested after firmware updates or hardware replacements. When a resource record version is earlier than the current firmware version, the record is prohibited from use, and the corresponding ordinary active action is marked as an unestimated resource and excluded from the utility sorting and executable set. The fixed minimum safe sampling configuration is verified using a conservative upper bound of communication, energy, and computational requirements that is fixed by the acquisition module and bound to the current firmware hash. If this conservative upper bound record is also missing or version incompatible, proceed to step [step number missing]. S5 performs controlled degradation sampling and sends resource model failure alarms until retesting generates the current firmware version resource record; when any normalized ratio of action cost is greater than 1, a resource out-of-bounds flag is set; when action utility is the same, actions with more risk coverage objects, less resource consumption, and smaller action combination key lexicographical order are selected in sequence; the action evaluator writes candidate action records, competing fault hypothesis versions, conditional observation distribution parameters, information gain, fault consequence risk, data quality weight, various resource requirements, action utility, and generation time into the candidate action table, and the candidate action table, together with the fixed safety minimum sampling configuration, is passed to step S5 for safety and resource constraint screening.

[0105] In this specific embodiment, S5 includes:

[0106] The constraint controller reads the candidate action table from step S4 and reads the operating boundaries of cell voltage (2.5 V to 3.65 V), cell temperature (0 ℃ to 55 ℃), and cluster current (-240 A to 240 A) from the safety boundary table of the energy storage battery management system, as well as the corresponding safety monitoring objects, monitoring parameters, minimum sampling frequency, maximum sampling start delay, and safety monitoring trigger distance. The safety boundary table is generated by cell specifications, system rated power, thermal management capability, and field protection settings. The records include physical units, operating conditions, version, and effective time. It is updated synchronously after the protection settings are changed. The fixed minimum safety sampling configuration covers all safety monitoring objects and parameters in the table, and the sampling frequency is not lower than the corresponding minimum frequency and the start delay does not exceed the corresponding maximum delay.

[0107] Every 1 second, the resource monitor reads the currently available communication bandwidth from the communication scheduler, the remaining acquisition energy from the acquisition module, and the available computing load for the next 60 seconds from the edge diagnostic processor, and forms a record of resource availability. First, it eliminates candidate actions where any one of the following exceeds the corresponding available amount: communication bandwidth usage, acquisition energy consumption, or computing load demand. Then, it defines the smaller of the difference between subtracting the lower working boundary from each state value and subtracting the state value from the upper working boundary as a signed safety margin. This margin is non-negative within the working range and negative when crossing any boundary. For cell voltage and cell temperature, the smaller of the two safety margins is used when both the measured state and the predicted state in step S2 are available; the smaller margin is used when only one is available; and the smaller margin is used when both are missing. The safety margin is conservatively set to the negative value of the safety monitoring trigger distance; the cluster current is the exogenous input of step S2 rather than the predicted output, and its safety margin is calculated only using the reliable measured cluster current in the profile of step S1. The safety margin is conservatively set to the negative value of the safety monitoring trigger distance only when the measured cluster current is missing or the reliability is less than 0.3; when the signed safety margin is not greater than the safety monitoring trigger distance in the safety boundary table, only actions that cover the corresponding objects and parameters and meet the minimum sampling frequency and maximum sampling start delay are retained. Any negative safety margin will not be allowed to enter the safety monitoring branch.

[0108] When the candidate action set is not empty after filtering, the controller sorts the actions from largest to smallest according to their utility and selects the first action. If the utility is the same, the risk coverage, resource consumption and combination key order rules of step S4 are followed. The obtained target perception actions are written into action instructions with an incrementing sequence number, configuration check code and 1200 ms confirmation timeout. After the acquisition module verifies the channel capability, frequency, phase and duration window, it returns a confirmation response and starts acquisition in the specified phase. The controller only records the action as executed after receiving a confirmation response with the same sequence number and check code. If the response is repeated, acquisition will not be started again.

[0109] When the set of candidate actions after screening is empty, the controller pauses non-safety monitoring tasks such as image recording, non-critical environmental monitoring, and batch uploading of historical data to release resources, and uses the released available resources to verify and fix the minimum safe sampling configuration. If the verification passes, the configuration is added to the candidate action set and directly determined as the target sensing action. If the verification fails, a resource shortage flag is generated, a safety alarm is sent to the energy storage battery management system, the active action selection in steps S4 and S5 is paused, and the safety monitoring objects and parameters that have exceeded the boundary, are close to the boundary, or are missing in state are preferentially retained in the order of the aforementioned signed safety margin from small to large. Within the range allowed by the combined communication, energy, and computing availability, the frequency is allocated from the lowest frequency in the safety boundary table downwards according to the level supported by the acquisition module to form a controlled degradation sampling configuration.

[0110] The energy storage battery management system specifies the minimum number of available sampling channels for cell voltage, cell temperature, and cluster current. In this specific implementation, the minimum number of channels is all 240 voltage channels, at least 2 temperature channels per module, and 1 cluster current channel. This number is written into the channel requirement field of the fixed safety minimum sampling configuration. The communication status records the configuration transmission link from the controller to the acquisition module and the data transmission link from the acquisition module to the controller. If any link is unavailable, the number of available channels for any parameter is lower than the corresponding requirement, or the target sensing action transmission fails, the controller suspends the switching based on action utility and determines the sampling output according to a unique priority: if the fixed safety minimum sampling configuration that has been locally fixed by the acquisition module and verified by the check code is available, and its channel, local data link, acquisition energy, and computing resource requirements have all been verified, then the fixed configuration is executed, and no new transmission command is required even if the remote configuration transmission link is interrupted; only when any of the above conditions are not met will a controlled degradation sampling configuration be formed according to the currently available channels and resources, and a safety alarm be maintained.

[0111] After a new observation arrives, it is appended to the original sampling buffer of step S1 according to its action sequence number, sampling channel, and timestamp. Using the same profile version as the simulation update in step S4, time alignment, missing data determination, reliability update, and high-frequency summary every 1 second window are performed sequentially. Then, mechanism prediction and joint diagnosis are performed to form a new profile version, a new residual version, and a new joint posterior version. If the effective reception ratio of the same action plan sample is less than 0.8, the received sub-samples and high-frequency summaries are retained, but the action execution status is recorded as incomplete. The stopping condition must not be met in this round. In the next round, step S4 will still construct the action based on the updated high-frequency summary quality, sensor reliability, and resource status.

[0112] Calibration samples, completely separated from the training samples of the joint diagnostic model, are divided by battery clusters. Dimensionless non-compliance scores are constructed by subtracting the true class prediction probability from one for each fault type and fault location. For each regression task, an original valid calibration sample set is first formed, ensuring compatibility with the current model parameter hash, input normalization version, and non-compliance definition version. For each sample, the true label, model prediction mean, and model output standard deviation must all be finite numbers, and the model output standard deviation must be non-negative. Samples with missing true labels, missing prediction means, or quality exclusion markers are not included in this set. When there are at least 20 original valid samples, the 5th percentile is first calculated from the valid model output standard deviations of these samples, and then... Construct scale-normalized non-compliance scores, where , Indicates the first Dimensionless non-compliance score of each regression calibration sample Indicates the index of the original valid calibration sample. This indicates the true health status value or the true remaining lifespan. This indicates the predicted mean of the corresponding model and is consistent with... Same unit, This indicates that the model output standard deviation is equal to... Same unit, This indicates the larger of the pre-specified strict positive scaling lower bound and the 5th percentile of the original effective standard deviation mentioned above, and also... Within the same unit, the preset lower limit for the health status value task is 0.001, and the preset lower limit for the remaining lifespan task is one equivalent complete cycle. Therefore... Strictly greater than 0, Indicates the prediction scale after applying the lower bound and is consistent with... Same unit, This indicates that the larger of the two inputs is selected; when the original valid samples are less than 20, the new scale lower limit, non-compliance score and calibration quantile are not calculated, and the branch that is compatible with the previous version or allows the full range of output is directly executed; the calibration quantile to be calculated does not participate in the determination of the validity of the original samples;

[0113] Target coverage according to Determine the calibration quantile value. This represents the dimensionless calibration quantile value for the corresponding diagnostic task. This represents the target coverage rate and is a dimensionless number within (0,1), which is 0.05 in this specific implementation. This indicates the number of valid calibration samples for the task, and is a positive integer. to This represents the dimensionless non-compliance score for each calibration sample. This indicates the selection of quantile values ​​based on the rank of a given finite sample. This means selecting the smaller of the two inputs and capping the quantile level at 1. This indicates rounding up; the calibration version records the model parameter hash, input normalization version, joint state support set version, location label table version, and non-compliance definition version. Each task requires at least 20 valid calibration samples. If there are fewer than 20, the previous valid calibration version is used only if all five identifiers are consistent with the current diagnostic version. If any identifier is incompatible or there is no previous compatible version, the classification set includes the current complete compatible label space, the regression interval is set to its full range, and the stopping condition is set to not met, until a new calibration version that matches the current diagnostic version and has a sufficient sample size is generated.

[0114] After each update, the conformal prediction set of fault types is determined by satisfying... The composition of fault types, Indicates the fault type of the current wheel. The posterior probability is a dimensionless number within [0,1]. Indicates the fault type index. The calibration quantile value representing the degree of non-compliance of the fault type is a dimensionless number within [0,1]; the fault location conformal prediction set is constructed using the posterior probability of the location and the location quantile value according to the same rule, and the health status value and remaining service life are respectively calculated according to... Forming a regression-preserving prediction interval, Indicates diagnostic quantity The original shape-preserving prediction interval of the current wheel and the unit is the same as Consistent Indicates the health status value or remaining lifespan. This represents the current round's predicted mean of the diagnostic quantity and its correlation with... Same unit, This represents the dimensionless calibration quantile value for the corresponding regression task. This indicates the larger of the standard deviation of the current model output and the corresponding lower bound of the scale, and... Same unit; the interval width determination uses the original interval before boundary truncation, and when outputting, the health status interval is truncated to 0 to 1 and the lower limit of the remaining lifespan interval is truncated to 0. The four results are written into the confidence information record and bound to the current joint posterior version.

[0115] The stop controller requires that the number of elements in the conformal prediction set for fault type not exceed 2, the number of elements in the conformal prediction set for fault location not exceed 3, the width of the original conformal prediction interval before the health status value boundary truncation not exceed 0.05, and the width of the original conformal prediction interval before the remaining service life boundary truncation not exceed 100 equivalent complete cycles; residual evidence in the most recent 60 Within the S window, the following conditions must be met simultaneously: there must be at least 48 valid voltage residual positions, accounting for at least 0.8% of the current planned positions; there must be at least 9 valid temperature residual positions, accounting for at least 0.8% of the current planned positions. Additionally, at least 90% of the valid values ​​for each type of residual must fall within the 2.5% to 97.5% quantile consistency range of the corresponding residuals in the independent calibration samples. Residuals marked with missing values, environmental or target temperature proxy flags, parameter extrapolation flags, or prediction degradation flags are not included in the valid quantity but are included in the denominator of the planned positions. If any type of residual fails to reach the quantity or coverage threshold, the residual consistency status is recorded as insufficient evidence, the stop condition is set to invalid, the cumulative number of consecutive evidence windows is cleared, and the missing monitoring parameters and target objects are returned to step S4 to construct supplementary actions. When a new observation reduces the standard deviation of the current model, the regression interval is narrowed accordingly using the aforementioned scaling formula. If all conditions are met simultaneously in three consecutive diagnostic rounds and the action execution status of each round is complete, the diagnostic evidence is deemed to meet the stop condition. If any condition is not met, the current diagnostic cache is carried back to step S4.

[0116] Before issuing an action, the action execution state machine recalculates the requirements based on the candidate action resource records. If the bandwidth requirement exceeds the currently available bandwidth, the action is deleted; if the energy consumption for data collection exceeds the remaining energy, the action is deleted; if the computational requirement exceeds the available computational load, the action is deleted; if the action has a security coverage failure flag, the action is deleted; if the action has a frequency failure or delay failure flag, the action is deleted; if the remaining set is not empty, the target sensing action is selected based on the first priority of the action utility ranking; if the remaining set is empty, the fixed security minimum sampling configuration is reviewed.

[0117] Every 1 second, the security recovery state machine verifies the configuration of the distribution link, data transmission link, channel, acquisition module, and resource status. If any communication link remains unavailable, it continues to select the local fixed security minimum sampling configuration or controlled degradation sampling according to the unique priority mentioned in S5. When the number of available channels for any parameter is still insufficient, a security alarm is maintained. When the acquisition module is not ready, the sending of test sensing actions is prohibited. A separate test action confirmation failure flag is set for the test sensing action. If a confirmation response with the same sequence number and check code is not obtained within 1200 ms, the reason is recorded as missing confirmation response, inconsistent sequence number or check code, acquisition module not ready, or link unavailable. The failure is not directly written as insufficient resources. The insufficient resources flag is only set when the recalculated communication bandwidth, acquisition energy consumption, or calculation requirements actually exceed the corresponding available amount. After each status is restored, the flag is cleared item by item according to the corresponding reason, and the test action is sent again and the joint posterior and uncertainty are recalculated using the current observation. If the model unavailable flag appears during recalculation, the action switching based on action utility is not restored.

[0118] In each round, the confidence-stop state machine writes the fault type and fault location conformal set size, health status and remaining service life interval width, number of valid voltage residual locations, number of valid temperature residual locations, the proportion of each type of residual to the currently configured planned location, the consistency ratio of the two types of residuals, residual evidence status, and action completeness. When the size of any set or the interval width exceeds the threshold, the continuous evidence count is cleared to zero. When the number of valid voltage residual locations is less than 48, the number of valid temperature residual locations is less than 9, the proportion of any residual to the planned location is less than 0.8, the consistency ratio of any residual is less than 90%, or the residual evidence status is insufficient, the continuous evidence count is cleared to zero and a stop flag is not generated. When the action execution status is incomplete, the continuous evidence count is also cleared to zero. Only when both sets, two intervals, the number of two types of residuals, the coverage of two types of residuals, the consistency ratio of two types of residuals, and the action completeness are all qualified, is the continuous evidence count incremented by 1. When the continuous evidence count reaches 3, a stop flag is generated. If the stop flag is not generated, the current evidence version is returned to step S4.

[0119] When the stopping condition is met, output the health status value, remaining service life, fault type, fault location, corresponding conformal prediction set or prediction interval, joint posterior probability, model uncertainty, profile version, mechanism parameter version, and diagnosis time, and send the results to the energy storage battery management system; after the configuration download link and data upload link are restored, the number of available channels for each monitoring parameter meets the minimum requirements, the acquisition module is ready, and the test sensing action receives a confirmation response with the same sequence number and check code, recalculate the joint posterior and uncertainty based on the current observation. Only when the recalculation is successful, the number of valid model instances is not less than 3, and no... Only when the model is unavailable flag appears and a new joint posterior version has been written to the joint diagnostic cache, is the cumulative count of continuous evidence windows cleared and steps S4 and S5 restored. If any restoration condition is not met, the unique priority mentioned above in S5 shall apply, i.e., if the local fixed minimum security sampling configuration channel, local data link, and resources are all satisfied, the configuration shall continue to be executed; otherwise, controlled degradation sampling shall be executed. In both branches, security alarms shall be maintained and the stop evidence status shall be set to invalid. The resource insufficiency flag shall be retained only when the recalculated resource requirements exceed the available amount, and the degradation process and security alarm records shall be retained for traceability.

[0120] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0121] This invention links multi-source data quality processing, electro-thermal mechanism prediction, time-series joint diagnosis, and active sensing action evaluation into a closed-loop processing chain, so that new observations can be directly used to update the state profile, mechanism residuals, joint posterior, and uncertainty, thereby enabling the sensing action to change with the current diagnostic evidence.

[0122] This invention sets up a fault hypothesis-conditional action evaluator and evidence controller, enabling physical faults and sensor faults to adopt corresponding conditional observation generation mechanisms, and controlling the cessation of exploration through conformal calibration results; when communication, sampling channels or resource conditions are abnormal, a safety-priority fixed configuration or controlled degradation sampling is adopted to maintain the continuous updating of diagnostic evidence.

Claims

1. A method for active exploration, sensing, and intelligent diagnosis of energy storage batteries, characterized in that, include: S1. Acquire multi-source sampling data for each target cell or cell group, perform time alignment and data quality assessment on the multi-source sampling data, and form a multi-source status profile with quality labeling. S2. Input the multi-source state profile into the electro-thermal mechanism model to obtain the predicted state and the electro-thermal mechanism residual between the measured state and the predicted state. S3. Input the multi-source state profile, electro-thermal mechanism residual, cell connection relationship, and historical degradation sequence into the trained time-series joint diagnostic model to obtain the joint posterior distribution and uncertainty of the health status value, remaining service life, fault type, and fault location. S4. Construct candidate sensing actions based on the joint posterior distribution and uncertainty. Predict the conditional observation distribution of each candidate sensing action for competing fault hypotheses. Calculate the expected diagnostic information gain based on the conditional observation distribution and determine the action utility of each candidate sensing action by considering fault consequence risk, data quality, and resource consumption. S5. Select the target sensing action with the highest action utility ranking from the candidate sensing actions that satisfy safety and resource constraints and send it to the acquisition module. Update the multi-source state profile, electro-thermal mechanism residual, joint posterior distribution, and uncertainty using the newly added observation data. Output the health status value, remaining service life, fault type, fault location, and corresponding confidence information when the updated diagnostic evidence meets the stopping condition; otherwise, return to step S4.

2. The active exploration, sensing, and intelligent diagnosis method for energy storage batteries according to claim 1, characterized in that, S1 includes: The multi-rate sampling sequence, including cell voltage, cell temperature, and cluster current, is resampled based on timestamps and charging / discharging stages as alignment references. Set a missing marker for sampling positions that do not fall within the aligned time window; The sensor reliability is calculated based on the variance of the sensor calibration residual, the effective sampling ratio, and the number of adjacent sampling jumps; the aligned sampled values, the missing marker, and the sensor reliability are combined to form the multi-source state profile, wherein the sensor reliability serves as the input for determining data quality in step S4.

3. The active exploration, sensing, and intelligent diagnosis method for energy storage batteries according to claim 1, characterized in that, S2 includes: using an equivalent circuit model to determine the predicted voltage and circuit model parameters based on the cell voltage and cluster current in the multi-source state profile; using a thermal model to determine the predicted temperature and thermal model parameters based on the cell temperature, cluster current, and cell connection relationship in the multi-source state profile; calculating the difference between the measured cell voltage and the predicted voltage, and the difference between the measured cell temperature and the predicted temperature, respectively, and combining them according to the aligned time window to form the electro-thermal mechanism residual, which is used for the joint diagnosis in step S3 and the candidate observation prediction in step S4.

4. The active exploration, sensing, and intelligent diagnosis method for energy storage batteries according to claim 1, characterized in that, S3 includes: extracting the temporal features of the multi-source state profile, the electro-thermal mechanism residual, and the historical degradation sequence using a time-series coding network; propagating the temporal features between connected cells or cell groups based on the cell connection relationship to form a spatiotemporal joint feature; inputting the spatiotemporal joint feature into the health status regression head, the remaining service life regression head, the fault type classification head, and the fault location classification head respectively, jointly normalizing each output distribution and obtaining the joint posterior distribution; and determining the uncertainty based on the output differences of multiple model instances for the same input.

5. The active exploration, sensing, and intelligent diagnosis method for energy storage batteries according to claim 1, characterized in that, S4 includes: using a combination of the target cell or cell group, monitoring parameters, sampling frequency, sampling start phase, and continuous window as a candidate sensing action; taking the physical fault state formed by combining the physical fault type (normal state, internal short circuit, loose connection, abnormal capacity decay, or thermal anomaly) with the fault location in the target cell or cell group, and the sensor fault state formed by combining the sensor fault type with the monitoring parameters or sampling channel location as the competing fault hypothesis; fitting the electro-thermal mechanism model to obtain the baseline model parameters based on the multi-source state profile, and correcting the baseline model parameters using the electro-thermal mechanism residual; estimating the baseline observation noise covariance based on the normal state calibration sample according to the monitoring parameters and sampling channel; and based on the fault... The calibration samples establish mappings between physical fault types, fault locations, and model parameter offsets, as well as mappings between sensor fault types, sampling channel locations, and observation bias parameters, observation drift parameters, and noise covariance correction parameters. For each candidate sensing action, the corresponding target object and sampling configuration are input into the electro-thermal mechanism model. For the normal state, the corrected baseline model parameters and the baseline observation noise covariance of the corresponding sampling channel are used. For the physical fault state, the corresponding model parameter offset is superimposed on the corrected baseline model parameters. For the sensor fault state, the corresponding observation bias parameters, observation drift parameters, and noise covariance correction parameters are applied to the model output and the corresponding baseline observation noise covariance to obtain the conditional observation distribution.

6. The active exploration, sensing, and intelligent diagnosis method for energy storage batteries according to claim 5, characterized in that, The expected diagnostic information gain is obtained by sampling the conditional observation distribution and calculating the expected value of the Kolb-Leibler divergence between the joint posterior distribution before and after adding the sampling results, or by calculating the difference in the joint posterior entropy before and after adding the sampling results; the failure consequence risk is determined based on the consequence loss and posterior probability of each competing failure hypothesis; for each candidate sensing action, the data integrity and sensor reliability corresponding to the target cell or cell group, monitoring parameters and sampling channels of the candidate sensing action are extracted from the quality identifier of the multi-source state profile to determine the data quality weight of the candidate sensing action; The expected diagnostic information gain is multiplied by the fault consequence risk and the data quality weight, and the costs of communication bandwidth usage, data acquisition energy consumption, computational load and action switching are deducted to obtain the action utility.

7. The active exploration, sensing, and intelligent diagnosis method for energy storage batteries according to claim 6, characterized in that, The safety constraints in step S5 are generated from the voltage operating boundary, temperature operating boundary, and charge / discharge current operating boundary recorded by the energy storage battery management system, as well as the safety monitoring objects, monitoring parameters, minimum sampling frequency, and maximum sampling start delay corresponding to each operating boundary. A fixed safety minimum sampling configuration is generated from these safety constraints, covering the corresponding safety monitoring objects and parameters, with a sampling frequency not less than the minimum sampling frequency and a sampling start delay not greater than the maximum sampling start delay. Resource constraints are generated from the currently available communication bandwidth, remaining acquisition energy, and available computing load. Candidate sensing actions that exceed the corresponding available amount in terms of communication bandwidth usage, acquisition energy consumption, or computing load requirements are excluded from all candidate sensing actions. When the distance from the measured state or predicted state to any operating boundary is not greater than the corresponding safety monitoring trigger distance, candidate sensing actions that do not meet the safety constraints are excluded from the remaining candidate sensing actions. When the set of filtered candidate sensing actions is not empty, the target sensing action is determined by sorting the actions according to their utility. When the set is empty, safety monitoring is placed before non-safety monitoring, and non-safety monitoring tasks are paused to release communication bandwidth, energy acquisition, and computing load. The resource requirements of the fixed safety minimum sampling configuration are reviewed based on the released available resources. If the review passes, the fixed safety minimum sampling configuration is added to the candidate sensing action set and determined as the target sensing action. If the review fails, a resource insufficiency flag is generated and a safety alarm is sent to the energy storage battery management system. The active action selection in steps S4 and S5 is paused, and the controlled degradation sampling configuration is executed. The controlled degradation sampling configuration retains safety monitoring objects and monitoring parameters in ascending order of distance from the measured state or predicted state to the corresponding working boundary based on the available sampling channels and the released available resources. The sampling frequency that can be executed within the available resources is set for the retained monitoring parameters. After the fixed safety minimum sampling configuration satisfies both the safety constraint and the resource constraint, it is added to the candidate sensing action set, and the target sensing action selection in step S5 is resumed.

8. The active exploration, sensing, and intelligent diagnosis method for energy storage batteries according to claim 7, characterized in that, Using calibration samples independent of the time-series joint diagnostic model, fault type non-compliance score, fault location non-compliance score, health status value non-compliance score, and remaining service life non-compliance score are constructed respectively, and the calibration quantile value of each non-compliance score is determined according to the target coverage. After each update of the joint posterior distribution and its uncertainty, the diagnostic outputs of the current wheel's fault type, fault location, health status value, and remaining service life are combined with the corresponding calibration quantile values ​​to form the current wheel's fault type conformal prediction set, fault location conformal prediction set, health status value conformal prediction interval, and remaining service life conformal prediction interval. The conformal prediction set and the conformal prediction interval are used as confidence information for the corresponding diagnostic outputs. When the number of elements in the fault type conformal prediction set and the fault location conformal prediction set are not greater than the corresponding set threshold, the width of the health status value conformal prediction interval and the remaining service life conformal prediction interval are not greater than the corresponding interval width threshold, and the electro-thermal mechanism residual falls into the consistency interval determined by the calibration sample residual distribution, and the above conditions are repeatedly met within the continuous evidence window, the diagnostic evidence is determined to meet the stopping condition.

9. The active exploration, sensing, and intelligent diagnosis method for energy storage batteries according to claim 8, characterized in that, The energy storage battery management system sets the minimum number of available sampling channels required for safety monitoring according to cell voltage, cell temperature, and cluster current, and uses the minimum number of available sampling channels as the channel requirements for the corresponding monitoring parameters in the fixed safety minimum sampling configuration. When the communication link is unavailable, the number of available sampling channels for any monitoring parameter is less than the corresponding minimum number of available sampling channels, or the target sensing action fails to be issued, the action switching based on the action effectiveness is suspended. When the number of available sampling channels for each monitoring parameter is not less than the corresponding minimum number of available sampling channels and the resource requirements of the fixed safety minimum sampling configuration are verified, the fixed safety minimum sampling configuration stored in the acquisition module is invoked for acquisition; otherwise, the controlled degradation sampling configuration is invoked and a safety alarm is sent to the energy storage battery management system. After the communication link is available, the number of available sampling channels for each monitoring parameter is not less than the corresponding minimum number of available sampling channels, the acquisition module is in a ready state, and the test sensing action receives a confirmation response, the joint posterior distribution and its uncertainty are recalculated using the current observation data, the cumulative number of consecutive evidence windows is cleared, and steps S4 and S5 are resumed.

10. The active exploration, sensing, and intelligent diagnosis method for energy storage batteries according to claim 4, characterized in that, The training of the time-series joint diagnostic model includes: using historical degradation samples with timestamps, charging / discharging stages, cell voltage, cell temperature, cluster current, quality identification, cell connection relationships, fault type labels, fault location labels, health status labels, and remaining service life labels as training samples; generating electro-thermal mechanism residuals for each training sample based on time-aligned measured cell voltage, measured cell temperature, and cluster current using the electro-thermal mechanism model; updating the parameters of the time-series joint diagnostic model with a weighted sum of health status regression loss, remaining service life regression loss, fault type classification loss, fault location classification loss, and mechanism residual consistency loss; and forming a model set for determining the uncertainty using multiple model instances with different initial parameters.