Lithium iron phosphate battery multi-type fault diagnosis method based on reinforcement learning

Abstract

The invention discloses a lithium iron phosphate battery multi-type fault diagnosis method based on reinforcement learning, and the method comprises the following steps: collecting multi-source data in the operation process of a lithium iron phosphate battery, and carrying out the preprocessing of the multi-source data; generating a candidate fault hypothesis according to the triggering observation result; constructing a corresponding short-time response prediction path, and generating a prediction response sequence; newly-added operation data are collected based on the triggering observation result, and a newly-added response sequence is extracted; retaining, degrading or eliminating the candidate fault hypotheses; inputting the candidate contraction result into a reinforcement learning observation action selection strategy, and updating the candidate contraction result; judging whether a convergence condition is met or not according to the candidate contraction result, and outputting a final fault diagnosis result; and outputting a fault type, a risk level and a corresponding diagnosis basis according to the final fault diagnosis result. According to the invention, reinforcement learning is adopted to realize multi-fault diagnosis of the lithium iron phosphate battery.

Description

Technical Field

[0001] This invention relates to the field of battery fault diagnosis technology, and in particular to a method for diagnosing multiple types of faults in lithium iron phosphate batteries based on reinforcement learning. Background Technology

[0002] Existing fault diagnosis methods for lithium iron phosphate batteries typically rely on operational data such as voltage, current, temperature, state of charge, and health status to identify abnormal battery conditions. Among existing technologies, one type of method uses threshold judgment, rule matching, or state parameter analysis to determine internal short circuits, thermal anomalies, capacity decay, and consistency anomalies; another type constructs data models to extract features and classify collected time-series data; and still others utilize the difference between predicted and actual values ​​to detect and warn of anomalies.

[0003] Most of the aforementioned existing technologies rely on current observation data for direct fault identification or simply compare a single prediction result with the actual result. When multiple types of faults coexist or their manifestations are similar, it is difficult to progressively distinguish between multiple candidate faults. Furthermore, existing methods typically depend on a fixed procedure for selecting subsequent observations, lacking a mechanism to dynamically adjust observation actions based on changes in candidate faults. This results in insufficient narrowing of candidate faults during the fault diagnosis process, ultimately affecting the specificity of fault type determination and the completeness of the diagnostic output.

[0004] Therefore, how to provide a method for diagnosing multiple types of faults in lithium iron phosphate batteries based on reinforcement learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a multi-type fault diagnosis method for lithium iron phosphate batteries based on reinforcement learning. This invention comprehensively utilizes multi-source operating data analysis, candidate fault hypothesis generation, short-time response prediction, consistency comparison, and reinforcement learning observation action selection methods. It details the implementation process of progressively narrowing the diagnosis and convergence determination for internal short-circuit faults, thermal anomaly faults, capacity decay faults, and consistency anomaly faults. It has the advantages of strong multi-type fault differentiation ability, high adaptability of the observation process, clear diagnostic basis, and high reliability of diagnostic results.

[0006] A method for diagnosing multiple types of faults in lithium iron phosphate batteries based on reinforcement learning according to an embodiment of the present invention includes the following steps: Multi-source data is collected during the operation of lithium iron phosphate batteries, the multi-source data is preprocessed, and abnormal triggering events are identified to obtain triggering observation results; Based on the trigger observation results, candidate fault hypotheses are generated, resulting in a candidate fault hypothesis set; For each candidate fault hypothesis in the candidate fault hypothesis set, a corresponding short-time response prediction path is constructed, and a prediction response sequence is generated to obtain a candidate prediction result set; New operational data is collected based on the trigger observation results, and new response sequences are extracted. These sequences are then compared with the predicted response sequences to obtain a candidate consistency result set. Based on the candidate consistency result set, candidate failure hypotheses are retained, downgraded, or eliminated, and the diagnostic priority of the remaining candidate failure hypotheses is updated to obtain the candidate shrinkage result. When the number of remaining candidate fault hypotheses in the candidate shrinkage results is greater than one, the candidate shrinkage results are input into the reinforcement learning observation action selection strategy to obtain the observation action selection results, and the candidate shrinkage results are updated according to the observation action selection results. Based on the candidate shrinkage results, determine whether the convergence condition is met, obtain the convergence determination result, and output the final fault diagnosis result based on the convergence determination result. Based on the final fault diagnosis results, the fault type, risk level, and corresponding diagnostic basis are output to obtain the fault diagnosis output results.

[0007] Optionally, the preprocessing includes noise reduction, outlier removal, time alignment, and normalization.

[0008] Optionally, the step of generating candidate fault hypotheses and obtaining a candidate fault hypothesis set based on the trigger observation results specifically includes: Extract the trigger discriminant corresponding to the current abnormal state from the trigger observation results to obtain the trigger discriminant set; Based on the set of trigger discriminant variables, calculate the support for the internal short-circuit fault hypothesis, the thermal anomaly fault hypothesis, the capacity decay fault hypothesis, and the consistency anomaly fault hypothesis respectively to obtain the set of hypothesis support. Normalize each support in the hypothesis support set to obtain a normalized support set; The candidate fault hypotheses are sorted according to the magnitude of each normalized support in the normalized support set to obtain the candidate hypothesis ranking result; Candidate fault hypotheses are selected based on the ranking of candidate hypotheses, and the candidate fault hypotheses are organized into a candidate fault hypothesis set.

[0009] Optionally, the step of constructing corresponding short-time response prediction paths and generating prediction response sequences for each candidate fault hypothesis in the candidate fault hypothesis set to obtain a candidate prediction result set specifically includes: For each candidate fault hypothesis in the candidate fault hypothesis set, the individual unit voltage, individual unit temperature, inter-unit voltage difference, and inter-unit temperature difference corresponding to the current sampling time of the trigger observation results are read as the initial response quantities. Based on the set of fault evolution coefficients, the predicted changes at each prediction time are generated, resulting in a set of predicted changes. Generate voltage change sequences corresponding to each candidate fault hypothesis based on the set of predicted changes; Generate temperature change sequences corresponding to each candidate fault hypothesis based on the set of predicted changes; Generate the individual unit deviation change sequence corresponding to each candidate fault hypothesis based on the voltage change sequence and temperature change sequence; Based on the candidate fault hypothesis types in the candidate fault hypothesis set, the paths of voltage change sequence, temperature change sequence and individual unit deviation change sequence are combined to obtain the short-time response prediction path; Based on each short-time response prediction path, a corresponding predicted response sequence is generated, and the results are summarized according to the candidate fault hypothesis type to obtain a candidate prediction result set.

[0010] Optionally, the set of fault evolution coefficients includes: Historical operating data of lithium iron phosphate batteries under different fault types were collected, and the historical operating data were classified according to the fault type to obtain a historical fault sample set. The voltage change, temperature change, and deviation change at multiple consecutive sampling times after the occurrence of the corresponding fault are extracted from various historical fault samples in the historical fault sample set to obtain the historical evolution feature set. For various historical fault samples in the historical evolution feature set, the average voltage change rate, average temperature change rate, and average deviation change rate under the corresponding fault type are calculated to obtain the fault evolution statistics. Based on the statistical results of fault evolution, the voltage evolution coefficient, temperature evolution coefficient, and deviation evolution coefficient corresponding to each fault type are determined to obtain the set of fault evolution coefficients.

[0011] Optionally, the step of collecting newly added operational data based on the triggered observation results, extracting newly added response sequences, and comparing them with the predicted response sequences to obtain a candidate consistency result set specifically includes: After triggering the current sampling time corresponding to the observation result, continue to collect new running data for multiple consecutive sampling times to obtain a new running data sequence; Extract the new voltage response sequence, new temperature response sequence, and new individual unit deviation response sequence corresponding to the candidate prediction result set from the new operational data sequence to obtain the new response sequence set; For any candidate fault hypothesis in the candidate prediction result set, the voltage consistency deviation, temperature consistency deviation, voltage difference consistency deviation, and temperature difference consistency deviation are calculated respectively to obtain the individual consistency deviation results. The overall consistency deviation of any candidate fault hypothesis at multiple consecutive new sampling times is calculated based on the individual consistency deviation results to obtain the overall consistency deviation result. The consistency score is calculated based on the overall consistency deviation results to obtain the consistency score result; The consistency scores corresponding to each candidate failure hypothesis in the candidate failure hypothesis set are summarized to obtain the candidate consistency result set.

[0012] Optionally, the step of retaining, downgrading, or eliminating candidate failure hypotheses based on the candidate consistency result set, and updating the diagnostic priority of the remaining candidate failure hypotheses to obtain the candidate shrinkage result specifically includes: Read the consistency scores corresponding to each candidate fault hypothesis in the candidate fault hypothesis set from the candidate consistency result set to obtain the candidate score set; For each consistency score in the candidate score set, combined with the consistency score of the corresponding candidate fault hypothesis in the previous diagnostic cycle, the score change of each candidate fault hypothesis in multiple consecutive diagnostic cycles is calculated to obtain the score change result. Based on the candidate score set and score change results, each candidate fault hypothesis in the candidate fault hypothesis set is subject to retention, downgrade, or elimination processing to obtain the candidate processing results. Candidate fault hypotheses that are to be retained are given a priority for diagnosis, candidate fault hypotheses that are to be downgraded are given a lower priority for diagnosis, and candidate fault hypotheses that are to be eliminated are removed from the set of candidate fault hypotheses, thus obtaining the priority adjustment results; Based on the diagnostic priorities in the priority adjustment results, the candidate fault hypotheses after retention processing and downgrade processing are reordered to obtain the candidate ranking update results. Based on the candidate ranking update results and the corresponding deletion results of the elimination process, a set of remaining candidate fault hypotheses is formed, and the candidate shrinkage results are obtained.

[0013] Optionally, when the number of remaining candidate fault hypotheses in the candidate shrinkage results is greater than one, inputting the candidate shrinkage results into the reinforcement learning observation action selection strategy to obtain the observation action selection results, and updating the candidate shrinkage results based on the observation action selection results specifically includes: When the number of remaining candidate fault hypotheses in the candidate shrinkage results is greater than one, the current observation decision state is constructed based on the candidate shrinkage results, and the observation decision state result is obtained. The observation decision state results are input into the reinforcement learning observation action selection strategy. The action value corresponding to transient voltage drop observation action, voltage recovery observation action, temperature-voltage coupling observation action and single-unit bias expansion observation action is calculated to obtain the action value set. Based on the action value set, the observation action with the highest action value is selected as the next round of observation action, and the observation action selection result is obtained; Based on the results of the action selection, the corresponding new running data is re-collected to obtain the new running data sequence corresponding to the action; Extract the corresponding new response sequence based on the new running data sequence corresponding to the action, update the candidate consistency result set, and obtain the updated candidate consistency result set; Based on the updated candidate consistency result set, the retention, downgrade, or elimination processes for each candidate fault hypothesis in the candidate fault hypothesis set are re-executed, and the diagnostic priority of the remaining candidate fault hypotheses is updated to obtain the updated candidate shrinkage result.

[0014] Optionally, the step of determining whether the convergence condition is met based on the candidate shrinkage results, obtaining the convergence determination result, and outputting the final fault diagnosis result based on the convergence determination result specifically includes: Read the set of remaining candidate fault hypotheses, the diagnostic priority corresponding to each remaining candidate fault hypothesis, the consistency score corresponding to each remaining candidate fault hypothesis, and the fault type corresponding to each remaining candidate fault hypothesis from the candidate shrinkage results to obtain the convergence determination input results; Based on the convergence determination input result, the number of candidate fault hypotheses in the remaining candidate fault hypothesis set is counted to obtain the candidate number determination result; Based on the convergence determination input results, high-risk candidate fault hypotheses are identified, and the number of consecutive cycles in which each high-risk candidate fault hypothese maintains the highest diagnostic priority is counted to obtain the high-risk priority determination results. Based on the convergence determination input results, we can determine whether all other candidate fault hypotheses in the candidate fault hypothesis set have been eliminated, and obtain the elimination status determination results. Based on the results of the candidate quantity determination, the high-risk priority determination, and the elimination status determination, the convergence condition is determined, and the convergence determination result is obtained. When the convergence determination result is that the convergence condition is met, the fault type corresponding to the candidate fault hypothesis with the highest diagnosis priority in the remaining candidate fault hypothesis set is read and output as the final fault diagnosis result.

[0015] Optionally, the step of outputting the fault type, risk level, and corresponding diagnostic basis based on the final fault diagnosis result to obtain the fault diagnosis output result specifically includes: Read the candidate fault hypothesis type corresponding to the final fault diagnosis result, the diagnosis priority of the corresponding candidate fault hypothesis in the candidate shrinkage result, the consistency score of the corresponding candidate fault hypothesis in the candidate consistency result set, and the convergence condition satisfaction status in the convergence judgment result to obtain the output judgment information. The fault type is determined based on the output judgment information, and the fault type output result is obtained. The risk level evaluation value corresponding to the final fault diagnosis result is calculated based on the output judgment information to obtain the risk evaluation result. The risk level is determined based on the risk assessment results, and the risk level output is obtained. Based on the final fault diagnosis result, candidate prediction result set, candidate consistency result set and convergence judgment result, extract the corresponding diagnostic criteria and obtain the diagnostic criteria output result. The fault diagnosis output results are obtained by summarizing the fault type output results, risk level output results, and diagnostic basis output results.

[0016] The beneficial effects of this invention are: This invention generates multiple candidate fault hypotheses and constructs corresponding short-time response prediction paths for each. Then, it compares the newly added operational data with the corresponding predicted response sequences for consistency, transforming the fault diagnosis process from a single-step determination to a progressively narrowing, continuous determination process. This approach allows for round-by-round screening of candidate faults when multiple fault behaviors are similar, thereby improving the ability to distinguish between different types of faults.

[0017] Compared to existing diagnostic methods that employ fixed observation procedures, this invention introduces a reinforcement learning-based observation action selection strategy when candidate fault hypotheses have not converged. Based on the candidate shrinkage results, the next round of observation actions is selected, and corresponding new operational data is re-collected for each selected action. This approach ensures that subsequent observations correspond to the current candidate fault state, making the acquisition of new response sequences during the diagnostic process more targeted, thereby improving the efficiency of updating the candidate consistency result set and candidate shrinkage results.

[0018] Furthermore, this invention performs retention, downgrading, or elimination processing on candidate fault hypotheses and generates the final fault diagnosis result by combining it with convergence conditions. It then outputs the fault type, risk level, and corresponding diagnostic basis, ensuring that the fault diagnosis output not only includes the fault category but also the basis for determining that category. This approach establishes a clear correspondence between the results of each stage in the diagnosis process, improving the completeness of the fault diagnosis output.

[0019] Furthermore, this invention introduces voltage change sequences, temperature change sequences, and individual unit deviation change sequences into the short-time response prediction path, and employs different path combinations for different candidate fault hypotheses, establishing a correspondence between the candidate prediction result set and the candidate fault hypothesis type. In this way, the consistency comparison between the newly added response sequence and the predicted response sequence can be based on responses matching the fault type, thereby enhancing the ability of the candidate consistency result set to identify different fault types. Attached Figure Description

[0020] 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: Figure 1 This is a flowchart of a multi-type fault diagnosis method for lithium iron phosphate batteries based on reinforcement learning proposed in this invention; Figure 2 This is a schematic diagram of the candidate fault prediction process in a multi-type fault diagnosis method for lithium iron phosphate batteries based on reinforcement learning proposed in this invention. Figure 3 This is a schematic diagram of the convergence determination process in a multi-type fault diagnosis method for lithium iron phosphate batteries based on reinforcement learning proposed in this invention. Detailed Implementation

[0021] 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.

[0022] refer to Figures 1-3 A multi-type fault diagnosis method for lithium iron phosphate batteries based on reinforcement learning includes the following steps: Multi-source data is collected during the operation of lithium iron phosphate batteries, the multi-source data is preprocessed, and abnormal triggering events are identified to obtain triggering observation results; Based on the trigger observation results, candidate fault hypotheses are generated, resulting in a candidate fault hypothesis set; For each candidate fault hypothesis in the candidate fault hypothesis set, a corresponding short-time response prediction path is constructed, and a prediction response sequence is generated to obtain a candidate prediction result set; New operational data is collected based on the trigger observation results, and new response sequences are extracted. These sequences are then compared with the predicted response sequences to obtain a candidate consistency result set. Based on the candidate consistency result set, candidate failure hypotheses are retained, downgraded, or eliminated, and the diagnostic priority of the remaining candidate failure hypotheses is updated to obtain the candidate shrinkage result. When the number of remaining candidate fault hypotheses in the candidate shrinkage results is greater than one, the candidate shrinkage results are input into the reinforcement learning observation action selection strategy to obtain the observation action selection results, and the candidate shrinkage results are updated according to the observation action selection results. Based on the candidate shrinkage results, determine whether the convergence condition is met, obtain the convergence determination result, and output the final fault diagnosis result based on the convergence determination result. Based on the final fault diagnosis results, the fault type, risk level, and corresponding diagnostic basis are output to obtain the fault diagnosis output results.

[0023] In this embodiment, multi-source data is collected during the operation of the lithium iron phosphate battery, the multi-source data is preprocessed, and abnormal triggering events are identified to obtain the triggering observation results, specifically including: Multi-source data of lithium iron phosphate batteries during operation were collected at each sampling time. The multi-source data included the individual cell voltage, total current, temperature of each individual cell or corresponding module, state of charge, health status, voltage difference between cells, and temperature difference between cells at each sampling time, thus obtaining the original data sequence. The original data sequence is denoised to obtain a denoised data sequence. Specifically, for any original sample value in the original data sequence, a sliding window averaging method is used for denoising. Specifically, the average of multiple consecutive sample values ​​within the sliding window at the current sampling time is calculated, and the average value is used as the denoised sample value corresponding to the current sampling time. Outlier removal is performed on the denoised data sequence to obtain the outlier-removed data sequence. For any denoised sample value in the denoised data sequence, the difference between the denoised sample value and the mean of the corresponding sample variable within the statistical window is first calculated. Then, the difference is divided by the standard deviation of the corresponding sample variable within the statistical window to obtain the standardized deviation of the denoised sample value. When the absolute value of the standardization deviation is greater than three, the corresponding denoised sampled value is judged as an outlier and removed. The outlier removal data sequence is time-aligned to obtain a time-aligned data sequence. The sampling time of the total current is used as the reference time. The sampling data corresponding to the individual unit voltage, temperature, state of charge and health are resampled or interpolated to make each monitoring variable correspond to each other at the same sampling time. The time-aligned data sequence is normalized to obtain a normalized data sequence. Specifically, for any time-aligned sampled value in the time-aligned data sequence, extreme value normalization is used. The sampled value is first subtracted from the minimum value of the corresponding sampled variable in the current diagnostic period, and then the result is divided by the difference between the maximum and minimum values ​​of the sampled variable in the current diagnostic period to obtain the normalized sampled value. The current change rate, inter-cell deviation expansion, charge / discharge state switching signal, load change signal, and rest start signal are calculated based on the normalized data sequence. The current change rate is obtained by subtracting the total current at the previous sampling time from the total current at the current sampling time, and then dividing by the time interval between adjacent sampling times. The inter-cell deviation expansion is obtained by subtracting the difference between the maximum and minimum values ​​of the individual cell voltages at the previous sampling time from the difference between the maximum and minimum values ​​of the individual cell voltages at the current sampling time. Abnormal triggering events are identified based on current change rate, charge / discharge state switching signal, load change signal, rest start signal, and inter-cell deviation expansion, resulting in trigger observation results. An abnormal triggering event is determined to have occurred at the corresponding sampling time when one of the following conditions is met: The absolute value of the current change rate is greater than the current change rate judgment value; the charging / discharging status indicator at the current sampling time is different from that at the previous sampling time; the absolute value of the difference between the load power at the current sampling time and the load power at the previous sampling time is greater than the load change judgment value; the total current at the current sampling time is zero and the total current at the previous sampling time is not zero; or the inter-unit deviation expansion is greater than the inter-unit deviation expansion judgment value.

[0024] In this embodiment, candidate fault hypotheses are generated based on the trigger observation results, and the candidate fault hypothesis set specifically includes: The trigger discrimination quantity corresponding to the current abnormal state is extracted from the trigger observation results to obtain the trigger discrimination quantity set. The trigger discrimination quantity set includes the current change rate, the inter-cell deviation expansion amount, the charge and discharge state switching amount, the load change amount, and the static start indicator. Based on the set of trigger discriminant variables, calculate the support for the internal short-circuit fault hypothesis, the thermal anomaly fault hypothesis, the capacity decay fault hypothesis, and the consistency anomaly fault hypothesis respectively to obtain the set of hypothesis support. The support for the internal short-circuit fault hypothesis is obtained by multiplying the absolute value of the current change rate, the inter-unit deviation expansion, and the load change by their respective contribution coefficients and then summing them. The support for the thermal anomaly fault hypothesis is obtained by summing the values ​​of the static start mark, load change, and inter-unit deviation expansion, each multiplied by its corresponding contribution coefficient. The support for the capacity decay fault hypothesis is obtained by summing the values ​​of the charge / discharge state switching amount, the inter-cell deviation expansion amount, and the resting start indicator, each multiplied by its corresponding contribution coefficient. The support for the consistency anomaly fault hypothesis is obtained by multiplying the inter-unit deviation expansion, the charge / discharge state switching amount, and the load change amount by their respective contribution coefficients and then summing them. Normalize each support in the hypothesis support set to obtain a normalized support set. The normalized support for any candidate fault hypothesis is obtained by dividing the support for the candidate fault hypothesis by the sum of the support for the internal short-circuit fault hypothesis, the thermal anomaly fault hypothesis, the capacity decay fault hypothesis, and the consistency anomaly fault hypothesis. The candidate fault hypotheses are sorted according to the magnitude of each normalized support in the normalized support set to obtain the candidate hypothesis ranking result; Based on the candidate hypothesis ranking results, the candidate fault hypothesis with the top two normalized support values ​​is selected, or the candidate fault hypothesis with a normalized support value greater than the average support value is selected. The average support value is obtained by summing the normalized support values ​​corresponding to the internal short-circuit fault hypothesis, thermal anomaly fault hypothesis, capacity decay fault hypothesis, and consistency anomaly fault hypothesis and then dividing by four. Each candidate fault hypothesis is organized into a candidate fault hypothesis set, wherein the candidate fault hypothesis set includes at least two of the following: internal short-circuit fault hypothesis, thermal anomaly fault hypothesis, capacity decay fault hypothesis, and consistency anomaly fault hypothesis.

[0025] In this embodiment, for each candidate fault hypothesis in the candidate fault hypothesis set, a corresponding short-time response prediction path is constructed, and a prediction response sequence is generated to obtain the candidate prediction result set, which specifically includes: For each candidate fault hypothesis in the candidate fault hypothesis set, the individual unit voltage, individual unit temperature, inter-unit voltage difference, and inter-unit temperature difference corresponding to the current sampling time of the trigger observation results are read as the initial response quantities. Based on the set of fault evolution coefficients, the predicted changes at each prediction time are generated, resulting in a set of predicted changes. The fault evolution coefficients include voltage evolution coefficients, temperature evolution coefficients, and deviation evolution coefficients. The voltage change sequence corresponding to each candidate fault hypothesis is generated based on the set of predicted changes. For any candidate fault hypothesis, the predicted voltage change at each prediction time is obtained by multiplying the individual voltage at the current sampling time by the voltage evolution coefficient corresponding to the candidate fault hypothesis, multiplying the current change rate by the voltage evolution coefficient corresponding to the candidate fault hypothesis, multiplying the inter-individual deviation expansion by the voltage evolution coefficient corresponding to the candidate fault hypothesis, and summing the above parts. The predicted voltage value corresponding to the prediction time is obtained by sequentially accumulating the individual unit voltage at the current sampling time with the predicted voltage changes from the first prediction time to the current prediction time, thereby obtaining the voltage change sequence corresponding to the candidate fault hypothesis. The temperature change sequence corresponding to each candidate fault hypothesis is generated based on the set of predicted changes. For any candidate fault hypothesis, the predicted temperature change at each prediction time is obtained by multiplying the individual temperature at the current sampling time by the temperature evolution coefficient corresponding to the candidate fault hypothesis, multiplying the absolute value of the current change rate by the temperature evolution coefficient corresponding to the candidate fault hypothesis, multiplying the inter-individual deviation expansion by the temperature evolution coefficient corresponding to the candidate fault hypothesis, and summing the above parts. The predicted temperature value corresponding to the prediction time is obtained by sequentially accumulating the individual temperature at the current sampling time with the predicted temperature changes from the first prediction time to the current prediction time, thereby obtaining the temperature change sequence of the corresponding candidate fault hypothesis. Based on the voltage change sequence and temperature change sequence, a single-unit deviation change sequence corresponding to each candidate fault hypothesis is generated. The voltage difference between predicted single units at each prediction time is obtained by subtracting the minimum value of the predicted voltage of each single unit at the prediction time from the maximum value of the predicted voltage of each single unit at the prediction time. The temperature difference between predicted cells at each prediction time is obtained by subtracting the minimum predicted temperature of each cell at each prediction time from the maximum predicted temperature of each cell at each prediction time. The change in prediction deviation at each prediction time is obtained by multiplying the voltage difference between predicted cells at the prediction time by the deviation evolution coefficient of the corresponding candidate fault hypothesis, and then multiplying the temperature difference between predicted cells at the prediction time by the deviation evolution coefficient of the corresponding candidate fault hypothesis, and summing the two, thus obtaining the cell deviation change sequence of the corresponding candidate fault hypothesis. Based on the candidate fault hypothesis types in the candidate fault hypothesis set, the voltage change sequence, temperature change sequence, and individual unit deviation change sequence are combined to obtain the short-time response prediction path corresponding to each candidate fault hypothesis. When the corresponding candidate fault hypothesis is an internal short-circuit fault hypothesis, the voltage change sequence and temperature change sequence are combined in the order of each prediction time to form the corresponding short-time response prediction path. When the corresponding candidate fault hypothesis is the thermal anomaly fault hypothesis, the temperature change sequence and the individual deviation change sequence are combined in the order of each prediction time to form the corresponding short-time response prediction path. When the corresponding candidate fault hypothesis is the capacity decay fault hypothesis, the voltage change sequence is combined into the corresponding short-time response prediction path according to the order of each prediction time. When the corresponding candidate fault hypothesis is the consistency anomaly fault hypothesis, the individual deviation change sequence is combined into the corresponding short-time response prediction path according to the order of each prediction time. The corresponding predicted response sequence is generated based on each short-time response prediction path, and then summarized according to the candidate fault hypothesis type to obtain the candidate prediction result set. The predicted response sequence corresponding to any candidate fault hypothesis is formed by arranging the predicted response values ​​at each prediction time in the short-time response prediction path corresponding to the candidate fault hypothesis in chronological order.

[0026] In this embodiment, the set of fault evolution coefficients includes: Historical operating data of lithium iron phosphate batteries under different fault types are collected. The historical operating data includes cell voltage, total current, cell temperature or module temperature, state of charge, state of health, voltage difference between cells and temperature difference between cells. The historical operating data is classified according to the fault type to obtain a historical fault sample set. The fault types include internal short circuit fault, thermal abnormality fault, capacity decay fault and consistency abnormality fault. The voltage change, temperature change, and deviation change are extracted from various historical fault samples in the historical fault sample set at multiple consecutive sampling times after the fault occurs, to obtain the historical evolution feature set. The voltage change is obtained by the difference of the individual voltage at adjacent sampling times, the temperature change is obtained by the difference of the individual temperature at adjacent sampling times, and the deviation change is obtained by one or both of the difference of the voltage difference between individuals and the difference of the temperature difference between individuals at adjacent sampling times. For various historical fault samples in the historical evolution feature set, the average voltage change rate, average temperature change rate, and average deviation change rate under the corresponding fault type are calculated to obtain the fault evolution statistics. The average voltage change rate is obtained by averaging the voltage changes of each historical fault sample under the corresponding fault type, the average temperature change rate is obtained by averaging the temperature changes of each historical fault sample under the corresponding fault type, and the average deviation change rate is obtained by averaging the deviation changes of each historical fault sample under the corresponding fault type. Based on the statistical results of fault evolution, the voltage evolution coefficient, temperature evolution coefficient, and deviation evolution coefficient corresponding to each fault type are determined to obtain the set of fault evolution coefficients. Among them, the voltage evolution coefficient, temperature evolution coefficient, and deviation evolution coefficient corresponding to internal short-circuit faults are determined based on the average voltage change rate, average temperature change rate, and average deviation change rate of internal short-circuit faults; the voltage evolution coefficient, temperature evolution coefficient, and deviation evolution coefficient corresponding to thermal anomaly faults are determined based on the average voltage change rate, average temperature change rate, and average deviation change rate of thermal anomaly faults; the voltage evolution coefficient, temperature evolution coefficient, and deviation evolution coefficient corresponding to capacity decay faults are determined based on the average voltage change rate, average temperature change rate, and average deviation change rate of capacity decay faults; and the voltage evolution coefficient, temperature evolution coefficient, and deviation evolution coefficient corresponding to consistency anomaly faults are determined based on the average voltage change rate, average temperature change rate, and average deviation change rate of consistency anomaly faults.

[0027] In this embodiment, newly added operational data is collected based on the triggered observation results, and newly added response sequences are extracted. These sequences are then compared with the predicted response sequences to obtain a candidate consistency result set, which specifically includes: After triggering the current sampling time corresponding to the observation result, continue to collect new operating data for multiple consecutive sampling times to obtain a new operating data sequence. The new operating data sequence includes the individual unit voltage, total current, individual unit temperature or module temperature, state of charge, health status, voltage difference between individual units, and temperature difference between individual units corresponding to the new sampling time. The newly added voltage response sequence, newly added temperature response sequence, and newly added individual deviation response sequence corresponding to the candidate prediction result set are extracted from the newly added running data sequence to obtain the set of newly added response sequences. Among them, the newly added voltage difference between individual units at any newly added sampling time is obtained by subtracting the minimum value of the newly added voltage of each individual unit at the newly added sampling time from the maximum value of the newly added voltage of each individual unit at the newly added sampling time. The newly added temperature difference between individual units at any newly added sampling time is obtained by subtracting the minimum value of the newly added temperature of each individual unit at the newly added sampling time from the maximum value of the newly added temperature of each individual unit at the newly added sampling time. For any candidate fault hypothesis in the candidate prediction result set, calculate the voltage consistency deviation, temperature consistency deviation, voltage difference consistency deviation, and temperature difference consistency deviation between the corresponding newly added response sequence and the predicted response sequence in the newly added response sequence set, and obtain the sub-item consistency deviation results. Specifically, the voltage consistency deviation at any new sampling time is obtained by the absolute value of the difference between the new voltage value and the predicted voltage value of the corresponding unit; the temperature consistency deviation at any new sampling time is obtained by the absolute value of the difference between the new temperature value and the predicted temperature value of the corresponding unit or module; the voltage difference consistency deviation at any new sampling time is obtained by the absolute value of the difference between the voltage difference between new units and the predicted voltage difference between units; and the temperature difference consistency deviation at any new sampling time is obtained by the absolute value of the difference between the temperature difference between new units and the predicted temperature difference between units. The overall consistency deviation of any candidate fault hypothesis at multiple consecutive new sampling times is calculated based on the individual consistency deviation results. The overall consistency deviation corresponding to any candidate fault hypothesis is obtained by the following method: first, averaging the voltage consistency deviation of each individual unit at each new sampling time, and then multiplying it by the voltage consistency deviation weight corresponding to the candidate fault hypothesis; then, averaging the temperature consistency deviation of each individual unit at each new sampling time, and then multiplying it by the temperature consistency deviation weight corresponding to the candidate fault hypothesis. The voltage difference consistency deviation at each new sampling time is multiplied by the voltage difference consistency deviation weight corresponding to the candidate fault hypothesis, the temperature difference consistency deviation at each new sampling time is multiplied by the temperature difference consistency deviation weight corresponding to the candidate fault hypothesis, and the above parts are summed over multiple consecutive new sampling times to obtain the result. The consistency score corresponding to any candidate fault hypothesis is calculated based on the overall consistency deviation result. The consistency score corresponding to any candidate fault hypothesis is obtained by dividing one by the sum of the overall consistency deviations corresponding to the candidate fault hypothesis. The smaller the overall consistency deviation, the larger the consistency score of the corresponding candidate fault hypothesis. The consistency scores corresponding to each candidate failure hypothesis in the candidate failure hypothesis set are summarized to obtain the candidate consistency result set.

[0028] In this embodiment, based on the candidate consistency result set, candidate fault hypotheses are retained, downgraded, or eliminated, and the diagnostic priority of the remaining candidate fault hypotheses is updated to obtain the candidate shrinkage result, specifically including: Read the consistency scores corresponding to each candidate fault hypothesis in the candidate fault hypothesis set from the candidate consistency result set to obtain the candidate score set; For each consistency score in the candidate score set, combined with the consistency score of the corresponding candidate fault hypothesis in the previous diagnostic cycle, the score change of each candidate fault hypothesis in multiple consecutive diagnostic cycles is calculated to obtain the score change result. The score change of any candidate fault hypothesis in the current diagnostic cycle is obtained by subtracting the consistency score of the candidate fault hypothesis in the previous diagnostic cycle from the consistency score of the candidate fault hypothesis in the current diagnostic cycle. Based on the candidate score set and score change results, each candidate fault hypothesis in the candidate fault hypothesis set is subject to retention, downgrade or elimination processing to obtain the candidate processing results. Among them, when the consistency score of any candidate fault hypothesis is in the top two of the current candidate fault hypothesis set in multiple consecutive diagnostic cycles, the candidate fault hypothesis is subject to retention processing. When the consistency score of any candidate fault hypothesis is in the last two positions in the current candidate fault hypothesis set and the score change is less than zero, the candidate fault hypothesis is downgraded. When any candidate fault hypothesis has the lowest consistency score in the current candidate fault hypothesis set for multiple consecutive diagnostic cycles, and the overall consistency deviation between the corresponding new response sequence and the predicted response sequence continues to increase, the candidate fault hypothesis is eliminated. Candidate fault hypotheses that are to be retained are given a priority for diagnosis, candidate fault hypotheses that are to be downgraded are given a lower priority for diagnosis, and candidate fault hypotheses that are to be eliminated are removed from the set of candidate fault hypotheses. The priority adjustment results are obtained by multiplying the consistency score of the candidate fault hypothesis by the consistency score weight, multiplying the change in the score of the candidate fault hypothesis by the change in the score weight, multiplying the normalized support corresponding to the candidate fault hypothesis by the normalized support weight, and summing the above parts. Based on the diagnostic priorities in the priority adjustment results, the candidate fault hypotheses after retention processing and downgrade processing are reordered to obtain the candidate ranking update results. Based on the candidate ranking update results and the corresponding deletion results of the elimination process, a set of remaining candidate fault hypotheses is formed, and the candidate shrinkage results are obtained.

[0029] In this embodiment, when the number of remaining candidate fault hypotheses in the candidate shrinkage results is greater than one, the candidate shrinkage results are input into the reinforcement learning observation action selection strategy to obtain the observation action selection results, and the candidate shrinkage results are updated based on the observation action selection results. Specifically, this includes: When the number of remaining candidate fault hypotheses in the candidate shrinkage results is greater than one, the set of remaining candidate fault hypotheses, the diagnostic priority corresponding to each remaining candidate fault hypothesis, the consistency score corresponding to each remaining candidate fault hypothesis, and the normalized support corresponding to each remaining candidate fault hypothesis are read from the candidate shrinkage results. All of the above items are used together to constitute the current observation decision state, and the observation decision state result is obtained. The results of the observation decision state are input into the reinforcement learning observation action selection strategy. The reinforcement learning observation action selection strategy calculates the action value corresponding to the transient voltage drop observation action, voltage recovery observation action, temperature-voltage coupling observation action, and single-unit bias expansion observation action, respectively, to obtain the action value set. The action value corresponding to any observation action is obtained by multiplying the discrimination of the observation action to the remaining candidate fault hypothesis by the discrimination weight, multiplying the recognition contribution of the observation action to the high-risk candidate fault hypothesis by the recognition contribution weight, multiplying the observation cost corresponding to the observation action by the observation cost weight, and combining the sum of the first two with the last one to obtain the corresponding action value. The observation action with the highest action value is selected as the next observation action based on the action value set, and the observation action selection result is obtained. The observation action selection result is the observation action with the highest action value in the action value set. Based on the observation action selection result, the corresponding new operation data is re-acquired to obtain the new operation data sequence corresponding to the action. Among them, when the observation action selection result is transient voltage drop observation action, the individual voltage and total current data of multiple consecutive sampling times after the current change time are collected. When the observation action selection result is voltage recovery observation action, the individual unit voltage data are collected at multiple consecutive sampling times after the load change ends; When the observation action is selected as temperature-voltage coupled observation action, the unit temperature, module temperature and unit voltage data are collected at multiple consecutive sampling times; When the observation action selection result is the single-unit deviation expansion observation action, the single-unit voltage difference and single-unit temperature difference data are collected at multiple consecutive sampling times; Extract the corresponding new response sequence based on the new running data sequence corresponding to the action, compare the consistency of the extracted new response sequence with the corresponding predicted response sequence in the candidate prediction result set, update the candidate consistency result set, and obtain the updated candidate consistency result set. Based on the updated candidate consistency result set, the retention, downgrade, or elimination processes for each candidate fault hypothesis in the candidate fault hypothesis set are re-executed, and the diagnostic priority of the remaining candidate fault hypotheses is updated to obtain the updated candidate shrinkage result.

[0030] In this embodiment, the convergence determination result is obtained by determining whether the convergence condition is met based on the candidate shrinkage results. Based on the convergence determination result, the final fault diagnosis result is output, specifically including: Read the set of remaining candidate fault hypotheses, the diagnostic priority corresponding to each remaining candidate fault hypothesis, the consistency score corresponding to each remaining candidate fault hypothesis, and the fault type corresponding to each remaining candidate fault hypothesis from the candidate shrinkage results to obtain the convergence determination input results; Based on the convergence determination input result, the number of candidate fault hypotheses in the remaining candidate fault hypothesis set is counted to obtain the candidate number determination result. The candidate number determination result is obtained in the following way: each candidate fault hypothesis in the candidate fault hypothesis set is judged one by one to see if it is retained in the remaining candidate fault hypothesis set. If a candidate fault hypothesis is retained in the remaining candidate fault hypothesis set, it is recorded as a valid candidate fault hypothesis. If a candidate fault hypothesis has been eliminated, it will not be counted in the number of valid candidate fault hypotheses; and the total number of valid candidate fault hypotheses will be used as the result of the candidate number determination. Based on the convergence judgment input results, high-risk candidate fault hypotheses are identified, and the number of consecutive cycles in which each high-risk candidate fault hypotheses maintain the highest diagnostic priority is counted to obtain the high-risk priority judgment result. The high-risk priority judgment result is obtained in the following way: for any high-risk candidate fault hypotheses, it is determined whether the high-risk candidate fault hypotheses have the highest diagnostic priority in the corresponding diagnostic cycle for each diagnostic cycle. If a high-risk candidate fault hypothesis has the highest diagnostic priority within a certain diagnostic cycle, then the diagnostic cycle is included in the duration cycle count. If a high-risk candidate fault hypothesis does not have the highest diagnostic priority within a certain diagnostic cycle, the diagnostic cycle will not be included in the duration cycle count. The number of duration cycles obtained from the statistics will be used as the result of the high-risk priority determination of the corresponding high-risk candidate fault hypothesis; Based on the convergence determination input result, we can determine whether all other candidate fault hypotheses in the candidate fault hypothesis set have been eliminated, and obtain the elimination status determination result. The elimination status determination result is obtained in the following way: first, identify the candidate fault hypothesis with the highest diagnostic priority, and then determine whether the other candidate fault hypotheses have been eliminated one by one. When all other candidate fault hypotheses except the one with the highest diagnostic priority have been eliminated, the elimination status determination result is that the elimination condition is met. If there are still other candidate fault hypotheses that have not been eliminated besides the candidate fault hypothesis with the highest diagnostic priority, the elimination status determination result is that the elimination condition is not met. Based on the candidate number determination result, the high-risk priority determination result, and the elimination status determination result, the convergence condition determination is performed to obtain the convergence determination result. Among them, the convergence determination result is determined to meet the convergence condition when one of the following conditions is met: the candidate number determination result shows that the number of candidate fault hypotheses in the remaining candidate fault hypothesis set is one. The results of the high-risk priority determination indicate that a certain high-risk candidate fault hypothesis maintains the highest diagnostic priority for multiple consecutive diagnostic cycles, and the number of consecutive cycles reaches the cycle determination value for the high-risk candidate fault hypothesis to maintain the highest diagnostic priority continuously. Alternatively, the elimination status determination result indicates that all other candidate fault hypotheses, except for the candidate fault hypothesis with the highest diagnostic priority, have been eliminated. When the convergence determination result is that the convergence condition is met, the fault type corresponding to the candidate fault hypothesis with the highest diagnosis priority in the remaining candidate fault hypothesis set is read and output as the final fault diagnosis result. If the convergence determination result is that the convergence condition is not met, the new running data collection corresponding to the observation action selection result and the update of the candidate shrinkage result will continue to be executed.

[0031] In this embodiment, the fault diagnosis output results, including the fault type, risk level, and corresponding diagnostic criteria, are output based on the final fault diagnosis results. Read the candidate fault hypothesis type corresponding to the final fault diagnosis result, the diagnosis priority of the corresponding candidate fault hypothesis in the candidate shrinkage result, the consistency score of the corresponding candidate fault hypothesis in the candidate consistency result set, and the convergence condition satisfaction status in the convergence judgment result to obtain the output judgment information. The fault type is determined based on the output judgment information, and the fault type output result is obtained. When the candidate fault hypothesis corresponding to the final fault diagnosis result is the internal short circuit fault hypothesis, the internal short circuit fault is output. When the candidate fault hypothesis corresponding to the final fault diagnosis result is the thermal anomaly fault hypothesis, output thermal anomaly fault; When the candidate fault hypothesis corresponding to the final fault diagnosis result is the capacity decay fault hypothesis, output the capacity decay fault. When the candidate fault hypothesis corresponding to the final fault diagnosis result is a consistent anomaly fault hypothesis, output the consistent anomaly fault. The risk level evaluation value corresponding to the final fault diagnosis result is calculated based on the output judgment information to obtain the risk evaluation result. The risk level evaluation value is obtained by multiplying the diagnosis priority of the candidate fault hypothesis corresponding to the final fault diagnosis result by the diagnosis priority weight, multiplying the consistency score of the candidate fault hypothesis corresponding to the final fault diagnosis result by the consistency score weight, multiplying the convergence strength corresponding to the final fault diagnosis result by the convergence strength weight, and summing the above parts. The risk level is determined based on the risk assessment results, and the risk level output result is obtained. When the risk level assessment value is greater than or equal to the high risk judgment value, the high risk level is output. When the risk level assessment value is less than the high risk judgment value but greater than or equal to the medium risk judgment value, the medium risk level is output. When the risk level assessment value is less than the medium risk judgment value, the low risk level is output. Based on the final fault diagnosis result, candidate prediction result set, candidate consistency result set and convergence judgment result, extract the corresponding diagnostic criteria and obtain the diagnostic criteria output result. The diagnostic criteria include at least one of the following: the short-time response prediction path type corresponding to the final fault diagnosis result, the consistency score corresponding to the final fault diagnosis result, the diagnostic priority corresponding to the final fault diagnosis result, and the convergence condition type that makes the convergence judgment result satisfy the convergence condition. The fault diagnosis output results are obtained by summarizing the fault type output results, risk level output results, and diagnostic basis output results.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to an online fault diagnosis scenario for a lithium iron phosphate energy storage battery system. This system consists of multiple battery modules. The battery management system continuously collects data on individual cell voltage, total current, module temperature, SOC, SOH, inter-cell voltage difference, and inter-cell temperature difference. During long-term operation, internal short-circuit faults, thermal anomaly faults, capacity decay faults, and consistency anomaly faults may simultaneously manifest as slight voltage fluctuations, slow temperature increases, or increased individual cell deviations in their early stages. Using a method based on direct judgment based on a single observation can easily lead to problems such as confusion of fault types, repeated changes in diagnostic conclusions, and incomplete output data.

[0033] In this scenario, the system first performs noise reduction, outlier removal, time alignment, and normalization on the collected multi-source data. It then identifies abnormal triggering events based on current change rate, charging / discharging state switching, load changes, inactivity start-up, and inter-unit deviation expansion. When an abnormal triggering event occurs, the system does not directly output the fault type. Instead, it generates multiple candidate fault hypotheses around the current abnormal state, forming a candidate fault hypothesis set. Next, the system constructs short-time response prediction paths for internal short-circuit fault hypotheses, thermal anomaly fault hypotheses, capacity decay fault hypotheses, and consistency anomaly fault hypotheses, generating corresponding predicted response sequences. As new operational data continuously enters, the system extracts the new response sequences and compares them with the predicted response sequences to obtain a candidate consistency result set. Finally, it performs retention, downgrade, or elimination processing on the candidate fault hypotheses, forming a candidate shrinkage result. If the candidate fault hypothesis has not yet converged, the candidate shrinkage result is input into the reinforcement learning observation action selection strategy. The next round of observation action is selected from transient voltage drop observation action, voltage recovery observation action, temperature-voltage coupling observation action, and single-unit deviation expansion observation action. Based on the action, new running data is collected again, and the candidate consistency result set and candidate shrinkage result are updated until the convergence condition is met and the final fault diagnosis result is output.

[0034] Within one verification cycle, a total of 240 sets of abnormal triggering event samples were obtained, including 54 sets of internal short-circuit faults, 48 ​​sets of thermal anomaly faults, 71 sets of capacity decay faults, and 67 sets of consistency anomaly faults. After diagnosis using the present invention, 228 sets of samples obtained final fault diagnosis results consistent with the actual state, with an overall diagnostic accuracy of 95.0%. Among them, the diagnostic accuracy of internal short-circuit faults was 92.6%, the diagnostic accuracy of thermal anomaly faults was 95.8%, the diagnostic accuracy of capacity decay faults was 95.8%, and the diagnostic accuracy of consistency anomaly faults was 95.5%. For the 182 sets of samples that required further observation before diagnosis could be completed, the present invention achieved candidate fault narrowing and output after an average of 2.7 diagnostic cycles, with an average output time of 1.57s; the average output time of the comparison method was 1.84s. At the same time, the average number of switching times of the diagnostic results of the present invention was 0.6 times / set, while that of the comparison method was 1.9 times / set, indicating that the present invention can reduce the fluctuation of conclusions during the diagnostic process.

[0035] Further statistical analysis revealed that among the 228 correctly diagnosed samples, 214 samples simultaneously output short-term response prediction path type, consistency score, diagnostic priority, and convergence condition type, accounting for 93.86% of the correctly diagnosed samples. This indicates that the present invention can not only differentiate and diagnose multiple types of faults but also provide diagnostic criteria corresponding to the final fault diagnosis result. Therefore, in this implementation scenario, the present invention can solve the problem of difficulty in distinguishing between early-stage multiple fault types when their manifestations are similar, and while ensuring diagnostic accuracy, it improves the relevance of new observations, shortens diagnostic output time, reduces fluctuations in diagnostic results, and enhances the completeness of fault diagnosis output results.

[0036] Table 1. Implementation verification results of the present invention in a lithium iron phosphate energy storage battery system.

[0037] As can be seen from the table above, this invention demonstrates good application results across multiple key indicators for diagnosing various types of faults in lithium iron phosphate batteries. Regarding diagnostic accuracy, the invention achieves diagnostic accuracy rates of 92.6%, 95.8%, 95.8%, and 95.5% for four types of faults: internal short-circuit faults, thermal anomaly faults, capacity decay faults, and consistency anomaly faults, respectively, with an overall accuracy of 95.0%. Notably, the identification accuracy rates for capacity decay faults and thermal anomaly faults are both close to 96%, indicating that this invention, through candidate fault hypothesis generation, short-time response prediction path construction, and consistency comparison of newly added response sequences, can effectively distinguish between fault types with similar performance, making it particularly suitable for complex operating scenarios where multiple types of faults coexist.

[0038] In terms of diagnostic efficiency, the average output time of this invention is 1.57s, significantly lower than the 1.84s of the comparative method. Further examining the fault types, the average output time is 1.22s for internal short-circuit faults, 1.31s for thermal anomalies, 1.68s for capacity decay faults, and 1.55s for consistency anomalies, all lower than the corresponding times of 1.68s, 1.74s, 1.92s, and 2.01s of the comparative method. This indicates that although this invention employs a process of progressively narrowing down and converging candidate faults, with the support of a reinforcement learning observation action selection strategy, it can select more discriminative observation actions based on the current diagnostic state, thereby reducing invalid observation processes and improving the overall output efficiency of fault diagnosis.

[0039] Regarding diagnostic stability, the average number of switching times for the diagnostic results of this invention is 0.6 times / group, while the comparative method reaches 1.9 times / group, showing a significant difference. Looking at the individual components, the average number of switching times for the diagnostic results of this invention in internal short-circuit faults, thermal anomaly faults, capacity decay faults, and consistency anomaly faults are 0.5 times / group, 0.6 times / group, 0.7 times / group, and 0.6 times / group, respectively, all significantly lower than the corresponding 1.8 times / group, 2.1 times / group, 1.9 times / group, and 1.8 times / group for the comparative method. This result indicates that this invention does not rely on a single observation to directly output conclusions, but rather uses a candidate consistency result set to retain, downgrade, and eliminate candidate fault hypotheses, ensuring that the final fault diagnosis result is based on the shrinking of candidates over a continuous diagnostic cycle. Therefore, it can reduce repeated changes in intermediate diagnostic conclusions.

[0040] From the perspective of the percentage of complete diagnostic data output, this invention achieves an overall accuracy of 93.86%, with internal short-circuit faults accounting for 94.0%, thermal anomaly faults for 93.5%, capacity decay faults for 92.7%, and consistency anomaly faults for 95.3%. This result indicates that the fault diagnosis results output by this invention are not limited to the fault type itself, but also simultaneously provide corresponding diagnostic evidence, demonstrating good result completeness. In summary, the above data shows that this invention achieves good results in terms of accuracy, output time, result stability, and diagnostic evidence completeness, effectively solving the problems of difficulty in distinguishing between multiple fault types exhibiting similar symptoms, large fluctuations in the diagnostic process, and incomplete output evidence in existing technologies.

[0041] The above are merely preferred embodiments 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.

Claims

1. A multi-type fault diagnosis method for lithium iron phosphate batteries based on reinforcement learning, characterized in that, Includes the following steps: Multi-source data is collected during the operation of lithium iron phosphate batteries, the multi-source data is preprocessed, and abnormal triggering events are identified to obtain triggering observation results; Based on the trigger observation results, candidate fault hypotheses are generated, resulting in a candidate fault hypothesis set; For each candidate fault hypothesis in the candidate fault hypothesis set, a corresponding short-time response prediction path is constructed, and a prediction response sequence is generated to obtain a candidate prediction result set; New operational data is collected based on the trigger observation results, and new response sequences are extracted. These sequences are then compared with the predicted response sequences to obtain a candidate consistency result set. Based on the candidate consistency result set, candidate failure hypotheses are retained, downgraded, or eliminated, and the diagnostic priority of the remaining candidate failure hypotheses is updated to obtain the candidate shrinkage result. When the number of remaining candidate fault hypotheses in the candidate shrinkage results is greater than one, the candidate shrinkage results are input into the reinforcement learning observation action selection strategy to obtain the observation action selection results, and the candidate shrinkage results are updated according to the observation action selection results. Based on the candidate shrinkage results, determine whether the convergence condition is met, obtain the convergence determination result, and output the final fault diagnosis result based on the convergence determination result. Based on the final fault diagnosis results, the fault type, risk level, and corresponding diagnostic basis are output to obtain the fault diagnosis output results.

2. The method for diagnosing multiple types of faults in lithium iron phosphate batteries based on reinforcement learning according to claim 1, characterized in that, The preprocessing includes noise reduction, outlier removal, time alignment, and normalization.

3. The method for diagnosing multiple types of faults in lithium iron phosphate batteries based on reinforcement learning according to claim 1, characterized in that, The process of generating candidate fault hypotheses based on the trigger observation results, and obtaining a candidate fault hypothesis set, specifically includes: Extract the trigger discriminant corresponding to the current abnormal state from the trigger observation results to obtain the trigger discriminant set; Based on the set of trigger discriminant variables, calculate the support for the internal short-circuit fault hypothesis, the thermal anomaly fault hypothesis, the capacity decay fault hypothesis, and the consistency anomaly fault hypothesis respectively to obtain the set of hypothesis support. Normalize each support in the hypothesis support set to obtain a normalized support set; The candidate fault hypotheses are sorted according to the magnitude of each normalized support in the normalized support set to obtain the candidate hypothesis ranking result; Candidate fault hypotheses are selected based on the ranking of candidate hypotheses, and the candidate fault hypotheses are organized into a candidate fault hypothesis set.

4. The method for diagnosing multiple types of faults in lithium iron phosphate batteries based on reinforcement learning according to claim 1, characterized in that, The process of constructing corresponding short-time response prediction paths for each candidate fault hypothesis in the candidate fault hypothesis set and generating prediction response sequences to obtain a candidate prediction result set specifically includes: For each candidate fault hypothesis in the candidate fault hypothesis set, the individual unit voltage, individual unit temperature, inter-unit voltage difference, and inter-unit temperature difference corresponding to the current sampling time of the trigger observation results are read as the initial response quantities. Based on the set of fault evolution coefficients, the predicted changes at each prediction time are generated, resulting in a set of predicted changes. Generate voltage change sequences corresponding to each candidate fault hypothesis based on the set of predicted changes; Generate temperature change sequences corresponding to each candidate fault hypothesis based on the set of predicted changes; Generate the individual unit deviation change sequence corresponding to each candidate fault hypothesis based on the voltage change sequence and temperature change sequence; Based on the candidate fault hypothesis types in the candidate fault hypothesis set, the paths of voltage change sequence, temperature change sequence and individual unit deviation change sequence are combined to obtain the short-time response prediction path; Based on each short-time response prediction path, a corresponding predicted response sequence is generated, and the results are summarized according to the candidate fault hypothesis type to obtain a candidate prediction result set.

5. The method for diagnosing multiple types of faults in lithium iron phosphate batteries based on reinforcement learning according to claim 4, characterized in that, The set of fault evolution coefficients includes: Historical operating data of lithium iron phosphate batteries under different fault types were collected, and the historical operating data were classified according to the fault type to obtain a historical fault sample set. The voltage change, temperature change, and deviation change at multiple consecutive sampling times after the occurrence of the corresponding fault are extracted from various historical fault samples in the historical fault sample set to obtain the historical evolution feature set. For various historical fault samples in the historical evolution feature set, the average voltage change rate, average temperature change rate, and average deviation change rate under the corresponding fault type are calculated to obtain the fault evolution statistics. Based on the statistical results of fault evolution, the voltage evolution coefficient, temperature evolution coefficient, and deviation evolution coefficient corresponding to each fault type are determined to obtain the set of fault evolution coefficients.

6. The method for diagnosing multiple types of faults in lithium iron phosphate batteries based on reinforcement learning according to claim 1, characterized in that, The process of collecting newly added operational data based on triggered observation results, extracting newly added response sequences, and comparing their consistency with predicted response sequences to obtain a candidate consistency result set specifically includes: After triggering the current sampling time corresponding to the observation result, continue to collect new running data for multiple consecutive sampling times to obtain a new running data sequence; Extract the new voltage response sequence, new temperature response sequence, and new individual unit deviation response sequence corresponding to the candidate prediction result set from the new operational data sequence to obtain the new response sequence set; For any candidate fault hypothesis in the candidate prediction result set, the voltage consistency deviation, temperature consistency deviation, voltage difference consistency deviation, and temperature difference consistency deviation are calculated respectively to obtain the individual consistency deviation results. The overall consistency deviation of any candidate fault hypothesis at multiple consecutive new sampling times is calculated based on the individual consistency deviation results to obtain the overall consistency deviation result. The consistency score is calculated based on the overall consistency deviation results to obtain the consistency score result; The consistency scores corresponding to each candidate failure hypothesis in the candidate failure hypothesis set are summarized to obtain the candidate consistency result set.

7. The method for multi-type fault diagnosis of lithium iron phosphate batteries based on reinforcement learning according to claim 1, characterized in that, The process of retaining, downgrading, or eliminating candidate failure hypotheses based on the candidate consistency result set, and updating the diagnostic priority of the remaining candidate failure hypotheses to obtain the candidate shrinkage result specifically includes: Read the consistency scores corresponding to each candidate fault hypothesis in the candidate fault hypothesis set from the candidate consistency result set to obtain the candidate score set; For each consistency score in the candidate score set, combined with the consistency score of the corresponding candidate fault hypothesis in the previous diagnostic cycle, the score change of each candidate fault hypothesis in multiple consecutive diagnostic cycles is calculated to obtain the score change result. Based on the candidate score set and score change results, each candidate fault hypothesis in the candidate fault hypothesis set is subject to retention, downgrade, or elimination processing to obtain the candidate processing results. Candidate fault hypotheses that are to be retained are given a priority for diagnosis, candidate fault hypotheses that are to be downgraded are given a lower priority for diagnosis, and candidate fault hypotheses that are to be eliminated are removed from the set of candidate fault hypotheses, thus obtaining the priority adjustment results; Based on the diagnostic priorities in the priority adjustment results, the candidate fault hypotheses after retention processing and downgrade processing are reordered to obtain the candidate ranking update results. Based on the candidate ranking update results and the corresponding deletion results of the elimination process, a set of remaining candidate fault hypotheses is formed, and the candidate shrinkage results are obtained.

8. The method for diagnosing multiple types of faults in lithium iron phosphate batteries based on reinforcement learning according to claim 1, characterized in that, When the number of remaining candidate fault hypotheses in the candidate shrinkage results is greater than one, the candidate shrinkage results are input into the reinforcement learning observation action selection strategy to obtain the observation action selection results, and the candidate shrinkage results are updated based on the observation action selection results. Specifically, this includes: When the number of remaining candidate fault hypotheses in the candidate shrinkage results is greater than one, the current observation decision state is constructed based on the candidate shrinkage results, and the observation decision state result is obtained. The observation decision state results are input into the reinforcement learning observation action selection strategy. The action value corresponding to transient voltage drop observation action, voltage recovery observation action, temperature-voltage coupling observation action and single-unit bias expansion observation action is calculated to obtain the action value set. Based on the action value set, the observation action with the highest action value is selected as the next round of observation action, and the observation action selection result is obtained; Based on the results of the action selection, the corresponding new running data is re-collected to obtain the new running data sequence corresponding to the action; Extract the corresponding new response sequence based on the new running data sequence corresponding to the action, update the candidate consistency result set, and obtain the updated candidate consistency result set; Based on the updated candidate consistency result set, the retention, downgrade, or elimination processes for each candidate fault hypothesis in the candidate fault hypothesis set are re-executed, and the diagnostic priority of the remaining candidate fault hypotheses is updated to obtain the updated candidate shrinkage result.

9. A multi-type fault diagnosis method for lithium iron phosphate batteries based on reinforcement learning according to claim 1, characterized in that, The process of determining whether the convergence condition is met based on the candidate shrinkage results, obtaining the convergence determination result, and outputting the final fault diagnosis result based on the convergence determination result specifically includes: Read the set of remaining candidate fault hypotheses, the diagnostic priority corresponding to each remaining candidate fault hypothesis, the consistency score corresponding to each remaining candidate fault hypothesis, and the fault type corresponding to each remaining candidate fault hypothesis from the candidate shrinkage results to obtain the convergence determination input results; Based on the convergence determination input result, the number of candidate fault hypotheses in the remaining candidate fault hypothesis set is counted to obtain the candidate number determination result; Based on the convergence determination input results, high-risk candidate fault hypotheses are identified, and the number of consecutive cycles in which each high-risk candidate fault hypothese maintains the highest diagnostic priority is counted to obtain the high-risk priority determination results. Based on the convergence determination input results, we can determine whether all other candidate fault hypotheses in the candidate fault hypothesis set have been eliminated, and obtain the elimination status determination results. Based on the results of the candidate quantity determination, the high-risk priority determination, and the elimination status determination, the convergence condition is determined, and the convergence determination result is obtained. When the convergence determination result is that the convergence condition is met, the fault type corresponding to the candidate fault hypothesis with the highest diagnosis priority in the remaining candidate fault hypothesis set is read and output as the final fault diagnosis result.

10. A method for diagnosing multiple types of faults in lithium iron phosphate batteries based on reinforcement learning according to claim 1, characterized in that, The specific steps for outputting the fault type, risk level, and corresponding diagnostic criteria based on the final fault diagnosis result include: Read the candidate fault hypothesis type corresponding to the final fault diagnosis result, the diagnosis priority of the corresponding candidate fault hypothesis in the candidate shrinkage result, the consistency score of the corresponding candidate fault hypothesis in the candidate consistency result set, and the convergence condition satisfaction status in the convergence judgment result to obtain the output judgment information. The fault type is determined based on the output judgment information, and the fault type output result is obtained. The risk level evaluation value corresponding to the final fault diagnosis result is calculated based on the output judgment information to obtain the risk evaluation result. The risk level is determined based on the risk assessment results, and the risk level output is obtained. Based on the final fault diagnosis result, candidate prediction result set, candidate consistency result set and convergence judgment result, extract the corresponding diagnostic criteria and obtain the diagnostic criteria output result. The fault diagnosis output results are obtained by summarizing the fault type output results, risk level output results, and diagnostic basis output results.

Images