Energy storage battery management system cloud fault self-diagnosis and root cause positioning method and system

Abstract

The present application relates to the technical field of energy storage battery management system, in particular to a kind of energy storage battery management system cloud end fault self-diagnosis and root cause positioning method and system.The present application includes the following steps: S1, multi-source data acquisition and preprocessing;S2, hierarchical time series anomaly detection;S3, abnormal event reception and topological mapping;S4, abnormal propagation path backtracking;S5, causal reasoning and root cause candidate generation;S6, multi-source evidence fusion score;S7, knowledge base matching and auxiliary determination;S8, operation and maintenance work order automatic generation;S9, disposal result backfilling and knowledge base updating;S10, abnormal detection threshold adaptive updating.Compared with the prior art, through hierarchical time series anomaly detection, three types of anomalies, gradual change, sudden change and intermittency, are covered, and the fault detection coverage is expected to reach more than 95%, and the false alarm rate is expected to be controlled below 3%.Through dynamic topological correlation graph and causal reasoning engine, the fault root cause positioning time is expected to be compressed from the traditional manual troubleshooting hours to within 5 minutes.

Description

Technical Field

[0001] This invention relates to the field of energy storage battery management system technology, specifically to a cloud-based fault self-diagnosis and root cause localization method and system for energy storage battery management system. Background Technology

[0002] With the continuous growth of installed capacity of electrochemical energy storage systems, large-scale energy storage power stations typically contain dozens of battery clusters, hundreds of battery modules, and thousands of individual cells. During operation, the Battery Management System (BMS) generates massive amounts of multi-dimensional time-series data, including cell voltage, current, temperature, State of Charge (SOC), State of Health (SOH), equilibrium status, and communication status. When a fault occurs in the energy storage system, maintenance personnel need to quickly locate the fault and analyze its root cause from this massive amount of data in order to take effective measures. However, existing fault diagnosis technologies suffer from low diagnostic efficiency, narrow coverage, and lack of traceability of diagnostic results, making it difficult to meet the needs of efficient operation and maintenance of large-scale energy storage systems.

[0003] Existing energy storage BMS fault diagnosis technologies have the following defects and shortcomings: First, existing statistical threshold alarm methods use fixed thresholds or simple adaptive thresholds to determine anomalies in a single monitored quantity. Fixed thresholds cannot adapt to the normal fluctuation range under different operating conditions and aging stages, leading to frequent false alarms under some operating conditions and serious missed alarms under others. For slow-changing anomalies such as a gradual increase in battery internal resistance and gradual capacity degradation, as well as occasional faults such as intermittent poor contact, the detection capability of fixed threshold methods is very limited.

[0004] Second, existing anomaly detection methods typically operate at a single level, such as detecting timing anomalies only in the voltage sequence of individual cells or performing statistical analysis only on module temperature. This single-level detection cannot correlate fault propagation relationships across levels. For example, an abnormal increase in the internal resistance of a single cell might cause an elevated module temperature, which in turn could lead to abnormal cluster-level equalization. The lack of multi-level correlation analysis capabilities results in insufficient fault location accuracy, requiring maintenance personnel to still conduct extensive manual troubleshooting.

[0005] Third, existing fault diagnosis methods based on knowledge graphs or expert systems rely on manually predefined fault rule bases. Building these rule bases requires extensive experience from domain experts, and once the rules are fixed, they struggle to automatically adapt to new fault modes that arise in actual energy storage system operation. When new battery cells, new module structures, or new BMS hardware are introduced, the coverage of the existing rule base rapidly declines.

[0006] Fourth, while existing deep learning-based end-to-end fault diagnosis methods possess strong pattern recognition capabilities, their outputs are typically fault type labels, lacking an interpretable reasoning process from anomaly detection to fault location and root cause analysis. Operations personnel cannot understand the reasoning basis and evidence chain behind the diagnosis, making it difficult to verify and trust the diagnostic results, resulting in a low adoption rate in actual engineering projects.

[0007] Fifth, the output of the existing fault diagnosis system is disconnected from the operation and maintenance process. The diagnostic results are presented in the form of an alarm list, requiring maintenance personnel to manually interpret the meaning of the alarms, determine the priority of the faults, write maintenance work orders, and allocate resources for handling. This manual translation process is inefficient, error-prone, and makes it difficult to share and accumulate diagnostic experience across different sites.

[0008] Based on the above-mentioned defects and shortcomings, the technical problem to be solved by the present invention includes:

[0009] 1. How to achieve multi-level automatic detection of timing anomalies from individual cells to station-level systems, covering three types of fault modes: slow-change anomalies, sudden-change anomalies, and intermittent anomalies;

[0010] 2. How to dynamically construct a weighted topology graph based on the actual physical connection and logical jurisdiction of energy storage devices to support cross-level anomaly propagation path tracing and fault location;

[0011] 3. How to perform causal reasoning based on the temporal sequence and topological propagation path of anomalous events, and output a root cause candidate list with confidence scores and traceable evidence chains;

[0012] 4. How to achieve a fully automated closed loop from fault detection, fault location, root cause analysis to maintenance work order generation, and achieve continuous self-evolution of the fault knowledge base through feedback of handling results. Summary of the Invention

[0013] To overcome the shortcomings of existing technologies, this invention provides a cloud-based fault self-diagnosis and root cause localization method and system for energy storage battery management systems. It is applicable to intelligent operation and maintenance scenarios for commercial and industrial energy storage systems and grid-connected energy storage power stations that include multiple battery clusters and modules. The system adopts a cloud-edge collaborative architecture, with the edge device deployed in the local controller of the energy storage power station for real-time anomaly detection and emergency protection, and the cloud-based operation and maintenance platform server responsible for in-depth root cause analysis and knowledge base management.

[0014] To achieve the above objectives, a cloud-based fault self-diagnosis and root cause localization method for an energy storage battery management system is designed, comprising the following steps:

[0015] S1, Multi-source data acquisition and preprocessing: The edge device acquires multi-dimensional time-series data from the BMS controller at acquisition cycles corresponding to each level; the acquired raw data is timestamped to align data from different acquisition cycles onto a common time axis; missing data points are filled using linear interpolation of adjacent valid values; the sliding window mean is calculated for each dimension of data. and standard deviation Z-score standardization is performed to obtain the standardized multidimensional feature matrix. Where D is the number of feature dimensions and L is the length of the feature window;

[0016] S2, Hierarchical Temporal Anomaly Detection: Anomaly detection is performed on the standardized feature matrix output from step S1 at five levels, with each level containing three parallel detection channels.

[0017] S2A, Sliding Window Reconstruction of Error Channel: Reconstructs the feature window of this level. Input the autoencoder network and compute the reconstructed output. and reconstruction error Reconstruction error and adaptive threshold In comparison, if This is then determined to be a mutation anomaly;

[0018] S2B, Trend Deviation Channel: Performs linear regression on the sliding window data of key indicators at this level to calculate the slope. The absolute value of the slope is compared with the trend threshold. In comparison, if lasting longer than the minimum duration If so, it is determined to be a slow-change anomaly;

[0019] S2C, intermittent abnormal channel: within a short-period sliding window The number of jumps in an internal statistical unidimensional indicator that went out of the normal range and then recovered. If the frequency jumps Exceeding the intermittent threshold If so, it is determined to be an intermittent abnormality;

[0020] The union of the detection results from the three channels is taken, and each detected anomaly is recorded as a structured message. Report to the cloud;

[0021] S3, Abnormal Event Reception and Topology Mapping: The cloud receives the set of abnormal events reported by the edge devices. Each abnormal event The device IDs are mapped to the dynamic device topology graph. The corresponding node Mark the abnormal event attributes on the node; if the device number of the abnormal event does not exist in the graph, trigger an incremental update of the topology graph.

[0022] S4, Anomaly Propagation Path Backtracking: For all nodes in the topology graph marked with anomaly events, sort them from earliest to latest by the time of the anomaly occurrence; starting from the earliest anomaly node, search upstream and downstream along the directed edges of the topology graph for other anomaly nodes; for each pair of anomaly nodes found... Check if the following conditions are met: S4A, and There are directed reachable paths between them and the path hop count does not exceed the maximum propagation hop count. S4B, The anomaly occurred earlier than The time of occurrence of the anomaly; S4C, arrive The time difference does not exceed the maximum propagation delay Node pairs that meet the conditions form causal candidate links. ;

[0023] S5, Causal Reasoning and Root Cause Candidate Generation: Based on the causal candidate links obtained in step S4, construct a causal propagation subgraph; identify all nodes with an in-degree of zero as root cause candidates in the propagation subgraph; for each root cause candidate... The influence range is calculated by traversing downstream along the causal propagation subgraph. The initial confidence of the root cause candidate is determined by the number of direct anomalous evidences associated with it and the severity of the anomalousness, and it decays hop by hop along the propagation path according to the edge weight.

[0024] S6, Multi-source evidence fusion scoring: For each root cause candidate Collect all anomalous evidence associated with it; encode each piece of evidence into an evidence vector. The weighted composite confidence score is calculated based on three components: hierarchical weight, channel weight, and time series weight. ; Root cause candidate Sort in descending order; simultaneously generate an evidence chain report for the candidate, recording the source level, detection channel, anomaly time, anomaly magnitude, and topological propagation path of each piece of evidence;

[0025] S7, Knowledge Base Matching and Assisted Judgment: The root cause candidates output in step S6 are matched with existing fault patterns in the fault knowledge base based on similarity. If the similarity exceeds the matching threshold, the candidate is determined to be the root cause. If an existing pattern is found, the historical handling methods and confirmation rates of that existing pattern are added to the root cause report; if no existing pattern is found, it is marked as a new type of failure mode and the operation and maintenance personnel are prompted to pay close attention to it in the root cause report.

[0026] S8, Automatic generation of maintenance work orders: Based on the root cause report in steps S6-7, maintenance work orders are automatically generated and automatically pushed to the maintenance management system.

[0027] S9, Feedback of Handling Results and Update of Knowledge Base: After the maintenance personnel complete the fault handling and close the work order, the system collects the actual root cause confirmation results and handling methods in the work order, sends the confirmed fault cases into the fault knowledge base self-evolution module in a structured form, and runs the online incremental clustering algorithm for pattern matching and knowledge base update; if the diagnosed root cause does not match the actual root cause, the deviation information is recorded for subsequent optimization and adjustment of anomaly detection thresholds and causal inference parameters;

[0028] S10, Adaptive update of anomaly detection threshold: Based on the deviation information fed back in step S9, the thresholds and parameters of each channel in the hierarchical time-series anomaly detection module are adjusted periodically; if the false alarm rate of a certain channel in a certain level is higher than the target value, the threshold of that channel is appropriately relaxed; if the false alarm rate is higher than the target value, the threshold of that channel is appropriately tightened.

[0029] In step S1, the Z-score standardization method is as follows:

[0030] ,in:

[0031] For the first Dimensional features in The standardized value of time is dimensionless;

[0032] For the first Dimensional features in The original sampled value at time step, the unit of which depends on the feature type;

[0033] For the first Dimensional features in The mean value within the sliding window at the endpoint, in units of Same, window length is ;

[0034] For the first Dimensional features in The standard deviation within the sliding window with the endpoint as the endpoint, in units of same;

[0035] ϵ is a zero constant, and its value is [value missing]. , dimensionless.

[0036] In step S2, the reconstruction error is calculated as follows:

[0037] ,in:

[0038] For the first hierarchical The reconstruction error at any given moment is dimensionless.

[0039] For the first hierarchical The standardized feature vector at time step, This represents the number of feature dimensions at this level.

[0040] For autoencoder pairs The reconstructed output;

[0041] for The d-th component;

[0042] for The d-th component;

[0043] It is an L2 norm;

[0044] Adaptive threshold The calculation method is as follows:

[0045] ,in:

[0046] For the first hierarchical The reconstruction error threshold at each time step is adaptive and dimensionless.

[0047] The reconstructed error is the mean value within the sliding window, and is dimensionless.

[0048] The standard deviation of the reconstruction error within the sliding window is dimensionless;

[0049] is the threshold sensitivity coefficient, with a value ranging from 2.0 to 4.0, and is dimensionless.

[0050] In step S2, the slope The specific calculation method is as follows:

[0051] ,in:

[0052] For the first Hierarchical key indicators The slope of the linear regression within the sliding window at time step 1, expressed in feature units per sampling point;

[0053] The default number of sampling points within the trend detection sliding window is 120 for a single layer, 60 for a module layer, and 30 for a cluster layer.

[0054] For the first Level 1 Dimensional indicators The standardized value of time is dimensionless;

[0055] This is a relative time index within the window, with values ​​ranging from 0 to... .

[0056] In step S5, root cause candidates The confidence level is calculated as follows:

[0057] ,in:

[0058] For the first The initial confidence level of each root cause candidate is dimensionless and ranges from 0 to positive infinity;

[0059] For the first One root cause candidate node;

[0060] For root cause candidates The set of all anomalous evidence events linked by a causal propagation path;

[0061] For the first An unusual evidence event;

[0062] For the first The severity score of each piece of abnormal evidence ranges from 0 to 1 and is dimensionless.

[0063] For root cause candidates To the evidence event Directed paths on a topological graph;

[0064] For the nodes on the path To the node The directed edges;

[0065] For the edge The fault propagation attenuation coefficient has a value range of 0 to 1 and is dimensionless.

[0066] In step S5, the comprehensive confidence score is calculated as follows:

[0067] ,in:

[0068] For the first The overall confidence score of each root cause candidate is dimensionless.

[0069] Candidate for root cause The total number of related pieces of evidence;

[0070] For the first The hierarchical weights of the evidence are dimensionless.

[0071] For the first The detection channel weights of each piece of evidence are dimensionless.

[0072] For the first The temporal weight of each piece of evidence is dimensionless; the calculation formula is as follows: ,in The time difference between the evidentiary event and the root cause candidate, in seconds; This is the time decay constant, which defaults to 0. ;

[0073] For the first The severity score of each piece of evidence ranges from 0 to 1 and is dimensionless.

[0074] The method for similarity calculation in step S7 is as follows:

[0075] ,in:

[0076] For new failure cases Compared with existing failure modes The similarity between them, with a value range of 100. to Dimensionless;

[0077] The feature vector for a new failure case is obtained by encoding the root cause type, evidence pattern, and device-level features. The dimension of the feature vector;

[0078] The prototype feature vector of the existing failure mode;

[0079] It is the dot product of two vectors;

[0080] It is an L2 norm;

[0081] The method for updating the knowledge base schema in step S9 is as follows:

[0082] ,in:

[0083] This is the updated fault mode prototype vector;

[0084] The original failure mode prototype vector before the update;

[0085] This represents the current cumulative number of cases for this failure mode;

[0086] This represents the feature vector of the newly merged case.

[0087] In step S8, the work order includes the fault level, faulty device number, root cause description, evidence chain summary, suggested handling method, priority score, and knowledge base matching result; wherein the priority score Scoring based on the scope of the fault's impact Urgency level score and repairability score The weighted calculation across three dimensions is as follows:

[0088] ,in:

[0089] The overall priority score for maintenance work orders ranges from 0 to 100 and is dimensionless.

[0090] The scope of the fault's impact is scored, with a value ranging from 0 to 100, and is dimensionless.

[0091] An urgency level score is given, with a value ranging from 0 to 100, and is dimensionless;

[0092] The repairability score ranges from 0 to 100 and is dimensionless.

[0093] The weight for the scope of influence is 0.4 by default, and it is dimensionless.

[0094] The urgency level is weighted by a default value of 0.4, which is dimensionless.

[0095] This is the repairability weight, defaulting to 0.2, dimensionless;

[0096] and .

[0097] In step S10, the anomaly detection threshold update uses an exponential moving average filter; specifically as follows:

[0098] ,in:

[0099] For the first Level 1 The updated detection threshold for the channel is dimensionless.

[0100] The target threshold is calculated based on the false alarm rate and the false negative rate. It is dimensionless. When the false alarm rate is greater than 3%, When the false negative rate is greater than 5%, ;

[0101] The detection threshold before the update is dimensionless;

[0102] The exponential moving average filter coefficients updated for the threshold range from 0.1 to 0.3 and are dimensionless.

[0103] This is the step size for adjusting the threshold; the default value is 0.05, and it is dimensionless.

[0104] This is the step size for adjusting the threshold; the default value is 0.03, and it is dimensionless.

[0105] To achieve the above objectives, a cloud-based fault self-diagnosis and root cause localization system for an energy storage battery management system is designed, comprising:

[0106] (1) Module 1, Multi-source data acquisition and preprocessing module: Collects multi-dimensional time series data from BMS controllers at all levels. After preprocessing by timestamp alignment, linear interpolation of missing values ​​and Z-score standardization, the multi-dimensional time series data is sent to the hierarchical time series anomaly detection module.

[0107] (2) Module 2, Layered Time-Sequence Anomaly Detection Module: Anomaly detection sub-models are constructed at five levels: single-unit layer, module layer, cluster layer, PCS layer, and station-level layer. Each level's detection sub-model contains three parallel detection channels:

[0108] The sliding window reconstructs the error channel, and uses an autoencoder to reconstruct the temporal feature window of this level. The reconstruction error exceeding the adaptive threshold is used as the basis for anomaly judgment, mainly detecting abrupt anomalies.

[0109] If the trend deviates from the channel, the slope of the sliding linear regression is calculated for the key indicators of this level. The slope that continuously exceeds the normal range is used as the basis for anomaly judgment, mainly detecting slow-changing anomalies.

[0110] Intermittent abnormal channel: calculate the number of abnormal jumps in the short period of the indicator sequence of this level, and use the jump frequency exceeding the threshold as the basis for anomaly judgment, mainly to detect intermittent faults;

[0111] The detection results from the three channels are combined. If any channel detects an anomaly, that level, that device, and that point in time is marked as an abnormal event.

[0112] (3) Module 3, Dynamic Equipment Topology Graph Module: Based on the physical connection relationship and logical jurisdiction relationship of the energy storage system, construct a five-layer directed weighted graph containing "single unit → module → cluster → PCS → station level". ;

[0113] (4) Module 4, Causal Reasoning Root Cause Localization Engine: Receives the set of abnormal events output by the hierarchical time-series anomaly detection module as evidence nodes, and maps these evidence nodes to the dynamic device topology association graph;

[0114] (5) Module 5, Multi-source Evidence Fusion Scoring Module: Unifies the encoding of abnormal signals from different levels and different detection channels into evidence vectors. Each root cause candidate The evidence set of this candidate consists of all the associated evidence vectors. The overall confidence score for each root cause candidate is calculated through weighted fusion. It outputs a root cause report with ranking scores and a complete chain of evidence, which records the source level, detection channel, anomaly time, anomaly magnitude, and topological path of each piece of evidence.

[0115] (6) Module Six, Fault Knowledge Base Self-Evolution Module: Maintain a structured fault pattern knowledge base. When a new confirmed fault case is added to the knowledge base, run an online incremental clustering algorithm to match the new case with the existing fault patterns based on similarity. If the matching degree exceeds the merging threshold, the algorithm will merge the fault patterns. If the match is below the merging threshold, a new failure mode entry is created.

[0116] (7) Module 7, Automatic Generation and Closed-Loop Management of Operation and Maintenance Work Orders: Based on the root cause report output by the causal reasoning root cause localization engine and the multi-source evidence fusion scoring module, operation and maintenance work orders are automatically generated and automatically pushed to the operation and maintenance management system; after the work order is closed, the actual root cause and handling results filled in by the operation and maintenance personnel are automatically fed back into the fault knowledge base self-evolution module to form a closed loop.

[0117] (8) Module 8, Cloud-Edge Collaborative Communication Module: Defines the data interaction protocol and division of labor between the edge and the cloud;

[0118] In the above modules, (1) the preprocessed data of Module 1 is sent to (2) Module 2 for hierarchical anomaly detection, (2) the abnormal events of Module 2 are reported to (4) Module 4 and mapped to the topology map of (3) Module 3; (4) the root cause candidates of Module 4 are sent to (5) Module 5 for evidence fusion scoring, (5) the root cause report output by Module 5 is sent to (7) Module 7 to generate operation and maintenance work orders, (7) the work order closed-loop results of Module 7 are fed back to (6) Module 6 to update the knowledge base, (6) the knowledge base of Module 6 provides historical fault mode references to (4) Module 4 to assist causal reasoning, and (8) Module 8 runs through the data interaction between the edge end and the cloud in the above modules.

[0119] Compared with the prior art, the present invention has the following advantages:

[0120] 1. By using hierarchical time-series anomaly detection to cover three types of anomalies—gradual, sudden, and intermittent—the fault detection coverage rate is expected to reach over 95%, and the false alarm rate is expected to be controlled below 3%.

[0121] 2. Through dynamic topological association graphs and causal reasoning engines, the time for root cause localization of faults can be reduced from hours in traditional manual troubleshooting to less than 5 minutes.

[0122] 3. By using a multi-source evidence fusion scoring mechanism to output a traceable chain of root cause evidence, the interpretability of diagnostic results and the confidence of maintenance personnel are significantly improved.

[0123] 4. Through the self-evolution mechanism of the fault knowledge base, the system's diagnostic capabilities continue to improve over time, and the fault mode library is expected to automatically expand by 10% to 20% of new modes every quarter.

[0124] 5. The automatic generation rate of operation and maintenance work orders is expected to reach over 80%, significantly improving operation and maintenance response efficiency. Attached Figure Description

[0125] Figure 1 This is a diagram of the overall architecture of the present invention.

[0126] Figure 2 This is a schematic diagram of the process of the present invention.

[0127] Figure 3 This is a schematic diagram of the layered timing anomaly detection in step S2 of the present invention.

[0128] Figure 4 This is a schematic diagram of the dynamic device topology association map in step S3 of the present invention.

[0129] Figure 5 This is a schematic diagram of the causal reasoning path backtracking in step S5 of the present invention.

[0130] Figure 6 This is a schematic diagram of the end-to-end diagnostic process in Embodiment 1 of the present invention. Detailed Implementation

[0131] The present invention will now be further described with reference to the accompanying drawings.

[0132] The software control flowchart of the cloud-based fault self-diagnosis and root cause localization method for the energy storage battery management system in this embodiment is as follows: Figure 2 As shown, the execution sequence of 10 steps and three independently running process layers are illustrated: the edge real-time processing process (steps S1 to S2), the cloud-based deep analysis process (steps S3 to S8), and the knowledge base maintenance process (steps S9 to S10). The three-channel parallel detection in step S2 is represented by parallel branches. The knowledge base matching results in step S7 are represented by conditional branches, matching existing patterns and novel fault modes.

[0133] Specifically, the method includes the following steps:

[0134] S1, Multi-source data acquisition and preprocessing: The edge device acquires multi-dimensional time-series data from the BMS controller at acquisition cycles corresponding to each level; the acquired raw data is timestamped to align data from different acquisition cycles onto a common time axis; missing data points are filled using linear interpolation of adjacent valid values; the sliding window mean is calculated for each dimension of data. and standard deviation Z-score standardization is performed to obtain the standardized multidimensional feature matrix. Where D is the number of feature dimensions and L is the length of the feature window;

[0135] S2, Hierarchical Temporal Anomaly Detection: Anomaly detection is performed on the standardized feature matrix output from step S1 at five levels, with each level containing three parallel detection channels.

[0136] S2A, Sliding Window Reconstruction of Error Channel: Reconstructs the feature window of this level. Input the autoencoder network and compute the reconstructed output. and reconstruction error Reconstruction error and adaptive threshold In comparison, if This is then determined to be a mutation anomaly;

[0137] S2B, Trend Deviation Channel: Performs linear regression on the sliding window data of key indicators at this level to calculate the slope. The absolute value of the slope is compared with the trend threshold. In comparison, if lasting longer than the minimum duration If so, it is determined to be a slow-change anomaly;

[0138] S2C, intermittent abnormal channel: within a short-period sliding window The number of jumps in an internal statistical unidimensional indicator that went out of the normal range and then recovered. If the frequency jumps Exceeding the intermittent threshold If so, it is determined to be an intermittent abnormality;

[0139] The union of the detection results from the three channels is taken, and each detected anomaly is recorded as a structured message. Report to the cloud;

[0140] S3, Abnormal Event Reception and Topology Mapping: The cloud receives the set of abnormal events reported by the edge devices. Each abnormal event The device IDs are mapped to the dynamic device topology graph. The corresponding node Mark the abnormal event attributes on the node; if the device number of the abnormal event does not exist in the graph, trigger an incremental update of the topology graph.

[0141] S4, Anomaly Propagation Path Backtracking: For all nodes in the topology graph marked with anomaly events, sort them from earliest to latest by the time of the anomaly occurrence; starting from the earliest anomaly node, search for other anomaly nodes upstream (parent node direction) and downstream (child node direction) along the directed edges of the topology graph; for each pair of anomaly nodes found... Check if the following conditions are met: S4A, and There are directed reachable paths between them and the path hop count does not exceed the maximum propagation hop count. S4B, The anomaly occurred earlier than The time of occurrence of the anomaly; S4C, arrive The time difference does not exceed the maximum propagation delay The default time is 300 seconds; node pairs that meet the conditions form causal candidate links. ;

[0142] S5, Causal Reasoning and Root Cause Candidate Generation: Based on the causal candidate links obtained in step S4, construct a causal propagation subgraph; identify all nodes with an in-degree of zero (i.e., no upstream abnormal cause) as root cause candidates in the propagation subgraph; for each root cause candidate... The influence range is calculated by traversing downstream along the causal propagation subgraph, i.e. the number of affected devices and the hierarchical span. The initial confidence of the root cause candidate is determined by the number of direct anomalous evidences associated with it and the severity of the anomalousness, and it decays hop by hop along the propagation path according to the edge weight.

[0143] S6, Multi-source evidence fusion scoring: For each root cause candidate Collect all anomalous evidence associated with it; encode each piece of evidence into an evidence vector. The weighted composite confidence score is calculated based on three components: hierarchical weight, channel weight, and time series weight. ; Root cause candidate Sort in descending order; simultaneously generate an evidence chain report for the candidate, recording the source level, detection channel, anomaly time, anomaly magnitude, and topological propagation path of each piece of evidence;

[0144] S7, Knowledge Base Matching and Assisted Judgment: The root cause candidates output in step S6 are matched with existing fault patterns in the fault knowledge base based on similarity. If the similarity exceeds the matching threshold, the candidate is determined to be the root cause. The default value is 0.7. If an existing pattern is matched, the historical handling method and confirmation rate of the existing pattern will be added to the root cause report. If no existing pattern is matched, it will be marked as a new fault mode and the operation and maintenance personnel will be prompted to pay special attention to it in the root cause report.

[0145] S8, Automatic generation of maintenance work orders: Based on the root cause report in steps S6-7, maintenance work orders are automatically generated and automatically pushed to the maintenance management system.

[0146] S9, Feedback of Handling Results and Update of Knowledge Base: After the maintenance personnel complete the fault handling and close the work order, the system collects the actual root cause confirmation results and handling methods in the work order, sends the confirmed fault cases into the fault knowledge base self-evolution module in a structured form, and runs the online incremental clustering algorithm for pattern matching and knowledge base update; if the diagnosed root cause does not match the actual root cause, the deviation information is recorded for subsequent optimization and adjustment of anomaly detection thresholds and causal inference parameters;

[0147] S10, Adaptive update of anomaly detection threshold: Based on the deviation information fed back in step S9, the thresholds and parameters of each channel in the hierarchical time-series anomaly detection module are adjusted periodically; if the false alarm rate of a certain channel in a certain level is higher than the target value, the threshold of that channel is appropriately relaxed; if the false alarm rate is higher than the target value, the threshold of that channel is appropriately tightened.

[0148] In step S1, the Z-score standardization method is as follows:

[0149] ,in:

[0150] For the first Dimensional features in The standardized value of time is dimensionless;

[0151] For the first Dimensional features in The original sampled value at time t, the unit depends on the feature type; for example, voltage is... The current is The temperature is ;

[0152] For the first Dimensional features in The mean value within the sliding window at the endpoint, in units of Same, window length is The default is 300 sampling points;

[0153] For the first Dimensional features in The standard deviation within the sliding window with the endpoint as the endpoint, in units of same;

[0154] ϵ is a zero constant, and its value is [value missing]. , dimensionless.

[0155] In step S2, the reconstruction error is calculated as follows:

[0156] ,in:

[0157] For the first hierarchical The reconstruction error at any given time is dimensionless, and the input has been standardized.

[0158] For the first hierarchical The standardized feature vector at time step, This represents the number of feature dimensions at this level.

[0159] For autoencoder pairs The reconstructed output;

[0160] for The d-th component;

[0161] for The d-th component;

[0162] The L2 norm is the Euclidean distance between vectors.

[0163] Adaptive threshold The calculation method is as follows:

[0164] ,in:

[0165] For the first hierarchical The reconstruction error threshold at each time step is adaptive and dimensionless.

[0166] The reconstructed error is the mean value within the sliding window, and is dimensionless.

[0167] The standard deviation of the reconstruction error within the sliding window is dimensionless;

[0168] is the threshold sensitivity coefficient, with a value ranging from 2.0 to 4.0, and is dimensionless. The smaller the value, the higher the sensitivity, but the higher the false alarm rate. The default value is 3.0.

[0169] In step S2, the slope The specific calculation method is as follows:

[0170] ,in:

[0171] For the first Hierarchical key indicators The slope of the linear regression within the sliding window at time step 1, expressed in feature units per sampling point;

[0172] The default number of sampling points within the trend detection sliding window is 120 for a single layer, 60 for a module layer, and 30 for a cluster layer.

[0173] For the first Level 1 Dimensional indicators The standardized value of time is dimensionless;

[0174] This is a relative time index within the window, with values ​​ranging from 0 to... .

[0175] In step S5, root cause candidates The confidence level is calculated as follows:

[0176] ,in:

[0177] For the first The initial confidence level of each root cause candidate is dimensionless and ranges from 0 to positive infinity;

[0178] For the first One root cause candidate node;

[0179] For root cause candidates The set of all anomalous evidence events linked by a causal propagation path;

[0180] For the first An unusual evidence event;

[0181] For the first The severity score of each piece of anomalous evidence, ranging from 0 to 1, is dimensionless and is obtained by normalizing the anomalous amplitude.

[0182] For root cause candidates To the evidence event Directed paths on a topological graph;

[0183] For the nodes on the path To the node The directed edges;

[0184] For the edge The fault propagation attenuation coefficient, ranging from 0 to 1, is dimensionless. Edges within the same level default to 0.9, while edges across levels default to 0.7.

[0185] In step S5, the comprehensive confidence score is calculated as follows:

[0186] ,in:

[0187] For the first The overall confidence score of each root cause candidate is dimensionless.

[0188] Candidate for root cause The total number of related pieces of evidence;

[0189] For the first The weight of each piece of evidence is dimensionless; the weight of a single layer is 1.0 by default, the weight of a module layer is 0.9, the weight of a cluster layer is 0.8, the weight of a PCS layer is 0.7, and the weight of a station-level layer is 0.6. The lower the level, the closer it is to the physical source of the fault, and the higher the weight.

[0190] For the first The detection channel weights for each piece of evidence are dimensionless; the reconstruction error channel defaults to 1.0, the trend deviation channel to 0.8, and the intermittent channel to 0.7.

[0191] For the first The temporal weight of each piece of evidence is dimensionless; the calculation formula is as follows: ,in The time difference between the evidentiary event and the root cause candidate, in seconds; This is the time decay constant, which defaults to 0. ;

[0192] For the first The severity score of each piece of evidence ranges from 0 to 1 and is dimensionless.

[0193] The method for similarity calculation in step S7 is as follows:

[0194] ,in:

[0195] For new failure cases Compared with existing failure modes The similarity between them, with a value range of 100. to Dimensionless;

[0196] The feature vector for a new failure case is obtained by encoding the root cause type, evidence pattern, and device-level features. The dimension of the feature vector is 64 by default;

[0197] The prototype feature vector of the existing failure mode;

[0198] It is the dot product of two vectors;

[0199] It is an L2 norm;

[0200] The method for updating the knowledge base schema in step S9 is as follows:

[0201] ,in:

[0202] This is the updated fault mode prototype vector;

[0203] The original failure mode prototype vector before the update;

[0204] This represents the current cumulative number of cases for this failure mode;

[0205] This represents the feature vector of the newly merged case.

[0206] In step S8, the work order includes the fault level, faulty device number, root cause description, evidence chain summary, suggested handling method, priority score, and knowledge base matching result; wherein the priority score Scoring based on the scope of the fault's impact Urgency level score and repairability score The weighted calculation across three dimensions is as follows:

[0207] ,in:

[0208] The overall priority score for maintenance work orders ranges from 0 to 100, is dimensionless, and the higher the score, the higher the priority.

[0209] The scope of the fault impact is scored, with a value ranging from 0 to 100. It is dimensionless and is obtained by mapping the number of affected devices and the number of levels across.

[0210] An urgency score is assigned, ranging from 0 to 100, and is dimensionless. It is derived by mapping the severity of the anomaly to its duration.

[0211] The repairability score ranges from 0 to 100. It is dimensionless and is mapped from the historical repair success rate of similar faults in the knowledge base. If there is no historical data, the default value is 50.

[0212] The weight for the scope of influence is 0.4 by default, and it is dimensionless.

[0213] The urgency level is weighted by a default value of 0.4, which is dimensionless.

[0214] This is the repairability weight, defaulting to 0.2, dimensionless;

[0215] and .

[0216] In step S10, the anomaly detection threshold update uses an exponential moving average filter to avoid abrupt changes, and the updated parameters are sent to the edge to take effect; specifically as follows:

[0217] ,in:

[0218] For the first Level 1 The updated detection threshold for the channel is dimensionless.

[0219] The target threshold is calculated based on the false alarm rate and the false negative rate. It is dimensionless. When the false alarm rate is greater than 3%, When the false negative rate is greater than 5%, ;

[0220] The detection threshold before the update is dimensionless;

[0221] The exponential moving average filter coefficients updated for the threshold, with values ​​ranging from 0.1 to 0.3, and a default value of 0.2 for dimensionless values;

[0222] This is the step size for adjusting the threshold; the default value is 0.05, and it is dimensionless.

[0223] This is the step size for adjusting the threshold; the default value is 0.03, and it is dimensionless.

[0224] Steps S1 to S2 above constitute the real-time edge processing flow, running according to the collection cycle of each level. Steps S3 to S8 constitute the cloud-based deep analysis flow, running in response to abnormal events. Steps S9 to S10 constitute the knowledge base maintenance flow, running in response to work order closure events. The three flows run independently and do not block each other.

[0225] in Figure 3 This diagram illustrates a hierarchical time-series anomaly detection system. From bottom to top, it shows five levels: Individual Layer → Module Layer → Cluster Layer → PCS Layer → Station-level Layer. Within each level, three detection channels are displayed side-by-side: Reconstruction Error Channel, Trend Deviation Channel, and Intermittent Anomaly Channel. Each channel is labeled with its primary anomaly type. The feature window length for each level is numerically labeled.

[0226] Figure 4 This is a schematic diagram of the dynamic device topology graph. The diagram shows a five-layer directed weighted graph: the bottom layer consists of individual nodes, followed by module nodes, cluster nodes, PCS nodes, and station-level nodes. Directed edges point from child nodes to parent nodes, and fault propagation attenuation coefficients are labeled on the edges: 0.9 within the same layer and 0.7 across layers. Several abnormal event nodes and causal propagation paths are highlighted in the diagram.

[0227] Figure 5 This diagram illustrates the backtracking of the causal reasoning path. The diagram shows multiple anomalous event nodes arranged chronologically on a local subgraph of the topology, propagating downstream along directed edges from the earliest node (root cause candidate). The diagram labels the occurrence time, anomaly type, and severity score of each anomalous event, as well as the confidence decay process along the causal link. The final output is a list of root cause candidates and their overall confidence scores, ranked accordingly.

[0228] The overall architecture diagram of the system implementing the cloud-based fault self-diagnosis and root cause localization method for the above-mentioned energy storage battery management system is shown below. Figure 1 As shown in the diagram, the area is divided into two regions: the edge and the cloud. The edge includes a multi-source data acquisition and preprocessing module and a hierarchical time-series anomaly detection module, connected to the cloud via a cloud-edge collaborative communication module. The cloud includes a dynamic device topology association graph module, a causal reasoning root cause localization engine, a multi-source evidence fusion scoring module, a fault knowledge base self-evolution module, and an automatic generation and closed-loop management module for maintenance work orders. Data flow is indicated by arrowed lines: BMS data edge acquisition preprocessing hierarchical detection anomaly event reporting cloud topology mapping causal reasoning evidence fusion root cause reporting work order generation disposal backfeeding knowledge base update. Among them:

[0229] (1) Module 1, Multi-source Data Acquisition and Preprocessing Module: Acquires multi-dimensional time-series data from BMS controllers at all levels, including individual unit voltages. Module current Monomer temperature Module temperature Monomer SOC estimates Estimated SOH value of monomer Equalization switch status Communication frame count and alarm signs at all levels Multidimensional time-series data, after being preprocessed by timestamp matching, linear interpolation of missing values, and Z-score standardization, is sent to the hierarchical time-series anomaly detection module. In specific use, the acquisition period is configurable, with a default of 1 second for single-unit level, 5 seconds for module level, and 10 seconds for cluster level.

[0230] (2) Module 2, Layered Time-Sequence Anomaly Detection Module: Anomaly detection sub-models are constructed at five levels: single-unit layer, module layer, cluster layer, PCS layer, and station-level layer. Each level's detection sub-model contains three parallel detection channels:

[0231] The sliding window reconstructs the error channel, and uses an autoencoder to reconstruct the temporal feature window of this level. The reconstruction error exceeding the adaptive threshold is used as the basis for anomaly judgment, mainly detecting abrupt anomalies.

[0232] If the trend deviates from the channel, the slope of the sliding linear regression is calculated for the key indicators of this level. The slope that continuously exceeds the normal range is used as the basis for anomaly judgment, mainly detecting slow-changing anomalies.

[0233] Intermittent abnormal channel: calculate the number of abnormal jumps in the short period of the indicator sequence of this level, and use the jump frequency exceeding the threshold as the basis for anomaly judgment, mainly to detect intermittent faults;

[0234] The detection results from the three channels are combined into a union. An anomaly detected in any channel marks that level, device, and time point as an abnormal event. In practice, the feature window length differs for each level: 60 sampling points by default for the single-unit layer, 30 for the module layer, 20 for the cluster layer, 10 for the PCS layer, and 5 for the station-level layer. The layered detection module is deployed at the edge, with a detection latency not exceeding 30 seconds.

[0235] (3) Module 3, Dynamic Equipment Topology Graph Module: Based on the physical connection relationship and logical jurisdiction relationship of the energy storage system, construct a five-layer directed weighted graph containing "single unit → module → cluster → PCS → station level". ;like Figure 4 As shown, each node in the diagram This represents a device entity, with node attributes including device type, device number, hierarchy, and current operating status. Each directed edge... Represents a dependency or connection relationship, edge weight This represents the fault propagation attenuation coefficient of the connection. The topology graph module is automatically built based on the device configuration file during system initialization, and nodes and edges are automatically added or removed through an incremental update algorithm when the device goes online or offline, requiring no manual maintenance. The topology graph module is deployed in the cloud.

[0236] (4) Module 4, Causal Reasoning Root Cause Localization Engine: Receives the set of abnormal events output by the hierarchical temporal anomaly detection module as evidence nodes, maps these evidence nodes to the dynamic device topology association graph; traces back the anomaly propagation path along the directed edges of the topology graph, and performs root cause candidate screening based on the following three causal reasoning rules: a. Temporal precedence rule: On the anomaly propagation path, the occurrence time of the root cause event must be earlier than that of the consequence event; b. Topological adjacency rule: There must be a directed reachable path between the root cause event and the consequence event on the topology graph, and the path length must not exceed the preset maximum propagation hop count. The default is 3 hops; c. Confidence decay rule: The confidence of root cause candidates decays hop by hop along the propagation path according to the edge weight. The longer the path and the smaller the edge weight, the lower the confidence. Root cause candidates are output in descending order of confidence. The engine is deployed in the cloud.

[0237] (5) Module 5, Multi-source Evidence Fusion Scoring Module: Unifies the encoding of abnormal signals from different levels and different detection channels into evidence vectors. Each root cause candidate The evidence set of this candidate consists of all the associated evidence vectors. The overall confidence score for each root cause candidate is calculated through weighted fusion. The weighting consists of three parts: the hierarchical weighting, the channel weighting, and the temporal weighting. The output is a root cause report with a ranking score and a complete evidence chain. The evidence chain records the source level, detection channel, anomaly time, anomaly magnitude, and topological path of each piece of evidence. The multi-source evidence fusion scoring module is deployed in the cloud.

[0238] (6) Module Six, Fault Knowledge Base Self-Evolution Module: Maintains a structured fault mode knowledge base. Each knowledge record includes a fault mode identifier, root cause type, typical evidence pattern, recommended handling method, and historical frequency of occurrence. When a new confirmed fault case, i.e., after the root cause has been confirmed by the maintenance personnel, is added to the knowledge base, an online incremental clustering algorithm is run to match the new case with existing fault modes based on similarity. If the matching degree exceeds the merging threshold... The default value is 0.8, which merges new cases into existing patterns and updates the pattern prototype; if the matching degree is lower than the merging threshold, a new fault pattern entry is created; the knowledge base update cycle is configurable, and a full clustering optimization is performed once a month by default. The fault knowledge base self-evolution module is deployed in the cloud.

[0239] (7) Module 7, Automatic Generation and Closed-Loop Management of Maintenance Work Orders: Based on the root cause report output by the causal reasoning root cause localization engine and the multi-source evidence fusion scoring module, maintenance work orders are automatically generated and automatically pushed to the maintenance management system. After the work order is closed, the actual root cause and handling results filled in by the maintenance personnel are automatically fed back into the fault knowledge base self-evolution module to form a closed loop. The work order content includes fault level, fault equipment number, root cause description, evidence chain summary, suggested handling method, and priority score. The priority score is calculated by weighting the fault impact scope, urgency, and repairability. The automatic generation and closed-loop management module of maintenance work orders is deployed in the cloud.

[0240] (8) Module 8, Cloud-Edge Collaborative Communication Module: Defines the data interaction protocol and division of labor between the edge and the cloud; the edge is responsible for real-time data acquisition, preprocessing, and hierarchical anomaly detection, and reports detected anomalies to the cloud in a structured message format; the cloud is responsible for topology map management, causal reasoning, evidence fusion, root cause report generation, work order management, and knowledge base maintenance. When communication between the edge and the cloud is interrupted, the edge can independently run anomaly detection and cache anomalies, and report them in batches after communication is restored; the cloud performs delayed root cause analysis on the cached events. Emergency protection actions, such as over-temperature cutoff and over-current protection, are always executed locally by the edge and do not depend on the cloud.

[0241] In the above modules, (1) the preprocessed data of Module 1 is sent to (2) Module 2 for hierarchical anomaly detection, (2) the abnormal events of Module 2 are reported to (4) Module 4 and mapped to the topology map of (3) Module 3; (4) the root cause candidates of Module 4 are sent to (5) Module 5 for evidence fusion scoring, (5) the root cause report output by Module 5 is sent to (7) Module 7 to generate operation and maintenance work orders, (7) the work order closed-loop results of Module 7 are fed back to (6) Module 6 to update the knowledge base, (6) the knowledge base of Module 6 provides historical fault mode references to (4) Module 4 to assist causal reasoning, and (8) Module 8 runs through the data interaction between the edge end and the cloud in the above modules.

[0242] Example 1:

[0243] This embodiment describes the end-to-end diagnosis of single-cell over-temperature faults in a commercial and industrial energy storage station. Specifically, a commercial and industrial energy storage station is configured with 4 battery clusters, each cluster containing 16 battery modules, and each module containing 16 individual cells, totaling 1024 cells. The system operates in parallel through 4 PCS (Power Control System) units, with the edge deployed on an on-site industrial control computer and the cloud deployed on a remote operation and maintenance platform. The system has been running stably for 6 months, and the fault knowledge base has accumulated 47 fault modes. Among them, cell number 11 in module 7 of cluster 2, numbered C2-M7-S11, experienced an abnormal increase in the contact resistance of the electrode tab welding, resulting in a gradual increase in internal resistance. The fault evolution process lasted approximately 3 hours, with the internal resistance rising from 2.5mΩ to 4.8mΩ. The temperature of the affected cell gradually increased from 28°C to 36°C, while the temperature of adjacent cells remained between 27°C and 29°C. The equalization circuit of module 7 continuously triggered equalization due to the low voltage of this cell.

[0244] Based on the aforementioned problems, the edge module 1 of this embodiment collects voltage, temperature data and equalization status of 1024 individual units at a 1-second cycle, and sends them to module 2 after Z-score standardization.

[0245] Module 2 operates a three-channel detection system at the single-unit layer: the reconstruction error channel first detects that the voltage sequence reconstruction error of C2-M7-S11 exceeds the adaptive threshold 45 minutes after the fault occurs. Marked as mutation abnormal The trend deviation channel detected the temperature slope of the individual cell 80 minutes after the fault occurred. Exceeding the trend threshold for 40 consecutive minutes Marked as a slow-varying anomaly Module 2 detected an abnormal equalization status in module 7 at the module level. The trigger frequency of the equalization switch increased from the normal 2 times per hour to 12 times per hour, and intermittent channels were marked as abnormal. .

[0246] Abnormal events , , Report to the cloud via Module 8.

[0247] Module 3 in the cloud maps the three abnormal events to the corresponding nodes in the topology graph: , Mapped to node "C2-M7-S11" Mapped to node "C2-M7".

[0248] Module 4 execution exception propagation path backtracking: (Voltage anomaly, occurring earliest) Located in the single-cell layer, (Temperature anomaly, which occurred later) Located in the same monomer, (Balance anomaly, occurring last) Located in the upper-level module. Timing relationships satisfy... The topology path is "C2-M7-S11→C2-M7", 1 hop, not exceeding [number missing]. The causal link is established.

[0249] Module 4 identifies root cause candidates Node "C2-M7-S11" has an in-degree of zero and no upstream anomalies. Initial confidence calculation:

[0250]

[0251] Module 5 Calculation Overall Score:

[0252] Root cause report output: The root cause is a single-cell abnormality of C2-M7-S11, with a confidence score of 1.76 and a chain of evidence containing 3 pieces of evidence.

[0253] Module 7 matches the knowledge base pattern F-012 "Abnormally high internal resistance of a single cell", with a similarity of 0.83 and a history of 9 occurrences. Recommended action: Replace the faulty cell, automatically generate a maintenance work order, and prioritize the score. .

[0254] Final output results: Fault detection time: First detected by the reconstruction error channel 45 minutes after the fault occurred. Root cause location time: Root cause location completed 2.3 minutes after the abnormal event was reported to the cloud. Root cause location result: Abnormally high internal resistance of C2-M7-S11 cell (confidence score 1.76, ranked 1st). Evidence chain: 3 pieces of evidence (abnormal sudden change in cell voltage → abnormal gradual change in cell temperature → intermittent abnormality in module balancing). Maintenance work order: Automatically generated, priority 55 points, recommended action is to replace the faulty cell.

[0255] A schematic diagram of the end-to-end diagnostic process in Example 1 is shown below. Figure 6 As shown in the figure, the fault evolution timeline is displayed from left to right: the internal resistance of individual cell C2-M7-S11 increases → the voltage sequence reconstruction error exceeds the threshold. →Temperature trend slope exceeds range →Module balance anomaly → Reporting abnormal events to the cloud → Topology mapping → Causal reasoning to locate the root cause → Evidence fusion and scoring → Knowledge base matching → Generation of operation and maintenance work orders. Key values ​​for each step, such as reconstruction error value, slope value, confidence score, and priority score, are displayed in an annotated manner.

[0256] In Example 1, the system completes the entire automated diagnosis chain from slow-varying anomaly detection to cross-level root cause localization without human intervention. The localization time is approximately 2 minutes, far less than the several hours required for traditional manual investigation. The chain of evidence is clear and traceable, and maintenance personnel can directly verify the diagnostic reasoning process.

Claims

1. A cloud-based fault self-diagnosis and root cause localization method for an energy storage battery management system, characterized in that: Includes the following steps: S1, Multi-source data acquisition and preprocessing: The edge device acquires multi-dimensional time-series data from the BMS controller at acquisition cycles corresponding to each level; the acquired raw data is timestamped to align data from different acquisition cycles onto a common time axis; missing data points are filled using linear interpolation of adjacent valid values; the sliding window mean is calculated for each dimension of data. and standard deviation Z-score standardization is performed to obtain the standardized multidimensional feature matrix. Where D is the number of feature dimensions and L is the length of the feature window; S2, Hierarchical Temporal Anomaly Detection: Anomaly detection is performed on the standardized feature matrix output from step S1 at five levels, with each level containing three parallel detection channels. S2A, Sliding Window Reconstruction of Error Channel: Reconstructs the feature window of this level. Input the autoencoder network and compute the reconstructed output. and reconstruction error Reconstruction error and adaptive threshold In comparison, if This is then determined to be a mutation anomaly; S2B, Trend Deviation Channel: Performs linear regression on the sliding window data of key indicators at this level to calculate the slope. The absolute value of the slope is compared with the trend threshold. In comparison, if lasting longer than the minimum duration If so, it is determined to be a slow-change anomaly; S2C, intermittent abnormal channel: within a short-period sliding window The number of jumps in an internal statistical unidimensional indicator that went out of the normal range and then recovered. If the frequency jumps Exceeding the intermittent threshold If so, it is determined to be an intermittent abnormality; The union of the detection results from the three channels is taken, and each detected anomaly is recorded as a structured message. Report to the cloud; S3, Abnormal Event Reception and Topology Mapping: The cloud receives the set of abnormal events reported by the edge devices. Each abnormal event The device IDs are mapped to the dynamic device topology graph. The corresponding node Mark the abnormal event attributes on the node; if the device number of the abnormal event does not exist in the graph, trigger an incremental update of the topology graph. S4, Anomaly Propagation Path Backtracking: For all nodes in the topology graph marked with anomaly events, sort them from earliest to latest by the time of the anomaly occurrence; starting from the earliest anomaly node, search upstream and downstream along the directed edges of the topology graph for other anomaly nodes; for each pair of anomaly nodes found... Check if the following conditions are met: S4A, and There are directed reachable paths between them and the path hop count does not exceed the maximum propagation hop count. S4B, The anomaly occurred earlier than The time of the abnormal occurrence; S4C arrive The time difference does not exceed the maximum propagation delay Node pairs that meet the conditions form causal candidate links. ; S5, Causal Reasoning and Root Cause Candidate Generation: Based on the causal candidate links obtained in step S4, construct a causal propagation subgraph; identify all nodes with an in-degree of zero as root cause candidates in the propagation subgraph; For each root cause candidate The influence range is calculated by traversing downstream along the causal propagation subgraph. The initial confidence of the root cause candidate is determined by the number of direct anomalous evidences associated with it and the severity of the anomalousness, and it decays hop by hop along the propagation path according to the edge weight. S6, Multi-source evidence fusion scoring: For each root cause candidate Collect all anomalous evidence associated with it; encode each piece of evidence into an evidence vector. The weighted composite confidence score is calculated based on three parts: hierarchical weight, channel weight, and time series weight. ; Root cause candidate Sort in descending order; simultaneously generate an evidence chain report for the candidate, recording the source level, detection channel, anomaly time, anomaly magnitude, and topological propagation path of each piece of evidence; S7, Knowledge Base Matching and Auxiliary Judgment: The root cause candidates output in step S6 are matched with existing fault patterns in the fault knowledge base based on similarity. If the similarity exceeds the matching threshold... If an existing pattern is found, the historical handling method and confirmation rate of that existing pattern are added to the root cause report; if no existing pattern is found, it is marked as a new type of failure mode and the operation and maintenance personnel are prompted to pay close attention to it in the root cause report. S8, Automatic generation of maintenance work orders: Based on the root cause report in steps S6-7, maintenance work orders are automatically generated and automatically pushed to the maintenance management system. S9, Feedback of Handling Results and Update of Knowledge Base: After the maintenance personnel complete the fault handling and close the work order, the system collects the actual root cause confirmation results and handling methods in the work order, sends the confirmed fault cases into the fault knowledge base self-evolution module in a structured form, and runs the online incremental clustering algorithm for pattern matching and knowledge base update; if the diagnosed root cause does not match the actual root cause, the deviation information is recorded for subsequent optimization and adjustment of anomaly detection thresholds and causal inference parameters; S10, Adaptive update of anomaly detection threshold: Based on the deviation information fed back in step S9, the thresholds and parameters of each channel in the hierarchical time-series anomaly detection module are periodically adjusted; If the false alarm rate of a certain channel at a certain level is higher than the target value, the threshold of that channel should be appropriately relaxed. If the false negative rate is higher than the target value, the threshold of this channel should be tightened appropriately.

2. The cloud-based fault self-diagnosis and root cause localization method for an energy storage battery management system according to claim 1, characterized in that: In step S1, the Z-score standardization method is as follows: ,in: For the first Dimensional features in The standardized value of time is dimensionless; For the first Dimensional features in The original sampled value at time step, the unit of which depends on the feature type; For the first Dimensional features in The mean value within the sliding window at the endpoint, in units of Same, window length is ; For the first Dimensional features in The standard deviation within the sliding window with the endpoint as the endpoint, in units of same; ϵ is a zero constant, and its value is [value missing]. , dimensionless.

3. The cloud-based fault self-diagnosis and root cause localization method for an energy storage battery management system according to claim 1, characterized in that: In step S2, the reconstruction error is calculated as follows: ,in: For the first hierarchical The reconstruction error at any given moment is dimensionless. For the first hierarchical The standardized feature vector at time step, This represents the number of feature dimensions at this level. For autoencoder pairs The reconstructed output; for The d-th component; for The d-th component; It is an L2 norm; Adaptive threshold The calculation method is as follows: ,in: For the first hierarchical The reconstruction error threshold at each time step is adaptive and dimensionless. The reconstructed error is the mean value within the sliding window, and is dimensionless. The standard deviation of the reconstruction error within the sliding window is dimensionless; is the threshold sensitivity coefficient, with a value ranging from 2.0 to 4.0, and is dimensionless.

4. The cloud-based fault self-diagnosis and root cause localization method for an energy storage battery management system according to claim 1, characterized in that: In step S2, the slope The specific calculation method is as follows: ,in: For the first Hierarchical key indicators The slope of the linear regression within the sliding window at time step 1, expressed in feature units per sampling point; The default number of sampling points within the trend detection sliding window is 120 for a single layer, 60 for a module layer, and 30 for a cluster layer. For the first Level 1 Dimensional indicators The standardized value of time is dimensionless; This is a relative time index within the window, with values ​​ranging from 0 to... .

5. The cloud-based fault self-diagnosis and root cause localization method for an energy storage battery management system according to claim 1, characterized in that: In step S5, root cause candidates The confidence level is calculated as follows: ,in: For the first The initial confidence level of each root cause candidate is dimensionless and ranges from 0 to positive infinity; For the first One root cause candidate node; For root cause candidates The set of all anomalous evidence events linked by a causal propagation path; For the first An unusual evidence event; For the first The severity score of each piece of abnormal evidence ranges from 0 to 1 and is dimensionless. For root cause candidates To the evidence event Directed paths on a topological graph; For the nodes on the path To the node The directed edges; For the edge The fault propagation attenuation coefficient has a value range of 0 to 1 and is dimensionless.

6. The cloud-based fault self-diagnosis and root cause localization method for an energy storage battery management system according to claim 1, characterized in that: In step S5, the comprehensive confidence score is calculated as follows: ,in: For the first The overall confidence score of each root cause candidate is dimensionless. Candidate for root cause The total number of related pieces of evidence; For the first The hierarchical weights of the evidence are dimensionless. For the first The detection channel weights of each piece of evidence are dimensionless. For the first The temporal weight of each piece of evidence is dimensionless; the calculation formula is as follows: ,in The time difference between the evidentiary event and the root cause candidate, in seconds; This is the time decay constant, which defaults to 0. ; For the first The severity score of each piece of evidence ranges from 0 to 1 and is dimensionless.

7. The cloud-based fault self-diagnosis and root cause localization method for an energy storage battery management system according to claim 1, characterized in that: The method for similarity calculation in step S7 is as follows: ,in: For new failure cases Compared with existing failure modes The similarity between them, with a value range of 100. to Dimensionless; The feature vector for a new failure case is obtained by encoding the root cause type, evidence pattern, and device-level features. The dimension of the feature vector; The prototype feature vector of the existing failure mode; It is the dot product of two vectors; It is an L2 norm; The method for updating the knowledge base schema in step S9 is as follows: ,in: This is the updated fault mode prototype vector; The original failure mode prototype vector before the update; This represents the current cumulative number of cases for this failure mode; This represents the feature vector of the newly merged case.

8. The cloud-based fault self-diagnosis and root cause localization method for an energy storage battery management system according to claim 1, characterized in that: In step S8, the work order includes the fault level, faulty device number, root cause description, evidence chain summary, suggested handling method, priority score, and knowledge base matching result; wherein the priority score Scoring based on the scope of the fault's impact Urgency level score and repairability score The weighted calculation across three dimensions is as follows: ,in: The overall priority score for maintenance work orders ranges from 0 to 100 and is dimensionless. The scope of the fault's impact is scored, with a value ranging from 0 to 100, and is dimensionless. An urgency level score is given, with a value ranging from 0 to 100, and is dimensionless; The repairability score ranges from 0 to 100 and is dimensionless. The weight for the scope of influence is 0.4 by default, and it is dimensionless. The urgency level is weighted by a default value of 0.4, which is dimensionless. This is the repairability weight, defaulting to 0.2, dimensionless; and .

9. The cloud-based fault self-diagnosis and root cause localization method for an energy storage battery management system according to claim 1, characterized in that: In step S10, the anomaly detection threshold update uses an exponential moving average filter; specifically as follows: ,in: For the first Level 1 The updated detection threshold for the channel is dimensionless. The target threshold is calculated based on the false alarm rate and the false negative rate. It is dimensionless. When the false alarm rate is greater than 3%, When the false negative rate is greater than 5%, ; The detection threshold before the update is dimensionless; The exponential moving average filter coefficients updated for the threshold range from 0.1 to 0.3 and are dimensionless. This is the step size for adjusting the threshold; the default value is 0.05, and it is dimensionless. This is the step size for adjusting the threshold; the default value is 0.03, and it is dimensionless.

10. A cloud-based fault self-diagnosis and root cause localization system for an energy storage battery management system, characterized in that: include: (1) Module 1, Multi-source data acquisition and preprocessing module: Collects multi-dimensional time series data from BMS controllers at all levels. After preprocessing by timestamp alignment, linear interpolation of missing values ​​and Z-score standardization, the multi-dimensional time series data is sent to the hierarchical time series anomaly detection module. (2) Module 2, Layered Time-Sequence Anomaly Detection Module: Anomaly detection sub-models are constructed at five levels: single-unit layer, module layer, cluster layer, PCS layer, and station-level layer. Each level's detection sub-model contains three parallel detection channels: The sliding window reconstructs the error channel, and uses an autoencoder to reconstruct the temporal feature window of this level. The reconstruction error exceeding the adaptive threshold is used as the basis for anomaly judgment, mainly detecting abrupt anomalies. If the trend deviates from the channel, the slope of the sliding linear regression is calculated for the key indicators of this level. The slope that continuously exceeds the normal range is used as the basis for anomaly judgment, mainly detecting slow-changing anomalies. Intermittent abnormal channel: calculate the number of abnormal jumps in the short period of the indicator sequence of this level, and use the jump frequency exceeding the threshold as the basis for anomaly judgment, mainly to detect intermittent faults; The detection results from the three channels are combined. If any channel detects an anomaly, that level, that device, and that point in time is marked as an abnormal event. (3) Module 3, Dynamic Equipment Topology Association Module: Based on the physical connection relationship and logical jurisdiction relationship of the energy storage system, construct a five-layer directed weighted graph containing "individual unit → module → cluster → PCS → station level". ; (4) Module 4, Causal Reasoning Root Cause Localization Engine: Receives the set of abnormal events output by the hierarchical time-series anomaly detection module as evidence nodes, and maps these evidence nodes to the dynamic device topology association graph; (5) Module 5, Multi-source Evidence Fusion Scoring Module: Unifies the encoding of abnormal signals from different levels and different detection channels into evidence vectors. Each root cause candidate The evidence set of this candidate consists of all the associated evidence vectors. The overall confidence score for each root cause candidate is calculated through weighted fusion. It outputs a root cause report with ranking scores and a complete chain of evidence, which records the source level, detection channel, anomaly time, anomaly magnitude, and topological path of each piece of evidence. (6) Module Six, Fault Knowledge Base Self-Evolution Module: Maintain a structured fault pattern knowledge base. When a new confirmed fault case is added to the knowledge base, run an online incremental clustering algorithm to match the new case with the existing fault patterns based on similarity. If the matching degree exceeds the merging threshold, the algorithm will merge the fault patterns. If so, the new case will be merged into the existing pattern and the pattern prototype will be updated; If the matching degree is lower than the merging threshold, a new fault mode entry is created; (7) Module 7, Automatic Generation and Closed-Loop Management of Operation and Maintenance Work Orders: Based on the root cause report output by the causal reasoning root cause localization engine and the multi-source evidence fusion scoring module, operation and maintenance work orders are automatically generated and automatically pushed to the operation and maintenance management system; after the work order is closed, the actual root cause and handling results filled in by the operation and maintenance personnel are automatically fed back into the fault knowledge base self-evolution module to form a closed loop. (8) Module 8, Cloud-Edge Collaborative Communication Module: Defines the data interaction protocol and division of labor between the edge and the cloud; In the above modules, (1) the preprocessed data of Module 1 is sent to (2) Module 2 for hierarchical anomaly detection, (2) the abnormal events of Module 2 are reported to (4) Module 4 and mapped to the topology map of (3) Module 3; (4) the root cause candidates of Module 4 are sent to (5) Module 5 for evidence fusion scoring, (5) the root cause report output by Module 5 is sent to (7) Module 7 to generate operation and maintenance work orders, (7) the work order closed-loop results of Module 7 are fed back to (6) Module 6 to update the knowledge base, (6) the knowledge base of Module 6 provides historical fault mode references to (4) Module 4 to assist causal reasoning, and (8) Module 8 runs through the data interaction between the edge end and the cloud in the above modules.

Images