Energy storage power station graded early warning and fault diagnosis method and system

Abstract

This invention discloses a hierarchical early warning and fault diagnosis method and system for energy storage power stations. In this method, the edge-side real-time monitoring module uses a dynamic multi-threshold table to judge the operating data of the energy storage power station, generates alarms, and pushes the raw time-series data to edge nodes. Edge nodes use mechanistic models and lightweight time-series models to extract fault features and generate edge early warnings. When trigger conditions are met, they initiate deep inference requests to the cloud and upload data. The cloud-based intelligent agent converts the raw data into a statistical indicator table, completes deep diagnosis based on the ReAct framework and fault tree analysis, generates or updates diagnostic skills according to path benefits, synchronously updates the knowledge base and knowledge graph, generates structured diagnostic reports, and matches disposal strategies. The cloud distributes the reports and strategies to edge nodes for alarm escalation; the cloud periodically optimizes the edge model through knowledge distillation to achieve continuous evolution of diagnostic capabilities. This invention solves the problems of lack of self-evolutionary diagnostic capabilities and closed-loop optimization of cloud-edge models in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and safety early warning technology for energy storage power stations, specifically to a hierarchical early warning and fault diagnosis method and system for energy storage power stations. Background Technology

[0002] With the large-scale and clustered operation of energy storage power stations, the operating conditions of key equipment such as battery clusters, converters, and temperature control equipment in energy storage systems are complex, and the volume of time-series telemetry data is enormous. Fault causes are characterized by multiple couplings, strong concealment, and rapid chain propagation. Currently, the industry generally adopts a hierarchical early warning and fault diagnosis architecture that combines end-side threshold alarms, simple edge model analysis, and cloud-based post-event review. Relying on fixed threshold rules, traditional mechanism models, or shallow time-series models, it conducts operational status monitoring and anomaly identification. This architecture has been widely applied in the daily operation and maintenance, status inspection, and initial fault screening scenarios of grid-side, user-side, and wind and solar integrated energy storage power stations. It can achieve basic alarms and simple fault location for routine over-limit operating conditions, becoming an important technical means for the routine safety management of energy storage power stations.

[0003] However, existing technologies still have significant shortcomings: First, fixed static thresholds cannot adapt to the dynamic operating characteristics of energy storage power stations, easily leading to missed alarms and false alarms, and insufficient accuracy of hierarchical early warning; Second, the edge side only uses mechanistic models or time-series models alone, lacking a dual-model fusion judgment mechanism, making it impossible to quantify the confidence of edge early warnings and effectively identify difficult hidden faults; Third, the cloud relies solely on manual rules or simple model reasoning, without combining LLM structured parsing, fault tree logic deduction, and the ability to reuse historical diagnostic experience, resulting in an incomplete deep fault diagnosis reasoning chain; Fourth, there is a lack of diagnostic path benefit evaluation, diagnostic skill iteration, and knowledge graph dynamic update mechanisms, making it impossible to achieve self-evolution of diagnostic capabilities; Fifth, there is a lack of reporting, policy distribution, and model closed-loop optimization mechanisms between the cloud and the edge, making it difficult for the edge-side mechanistic and time-series models to continuously iterate and adapt with maintenance data, limiting long-term diagnostic stability and generalization capabilities.

[0004] Therefore, there is an urgent need for a graded early warning and fault diagnosis method for energy storage power stations to solve the problems of lack of diagnostic self-evolution and closed-loop optimization of cloud-edge models in existing technologies. Summary of the Invention

[0005] To address these issues, this invention provides a hierarchical early warning and fault diagnosis method and system for energy storage power stations, which solves problems such as false alarms and missed alarms of static thresholds in existing energy storage power stations, insufficient confidence in edge judgment, lack of deep reasoning and experience reuse in the cloud, lack of diagnostic self-evolution, and closed-loop optimization of cloud-edge models.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for graded early warning and fault diagnosis of energy storage power stations, characterized in that it includes:

[0007] Based on a dynamic multi-threshold table, the edge-side real-time monitoring module identifies the operation data of the energy storage power station; when the set operation and maintenance characteristics are detected to exceed the corresponding set threshold, an edge-side alarm is generated and the original time-series data within the abnormal time window is pushed to the edge node;

[0008] Based on the original time-series data, the edge node extracts fault features and generates edge warnings through the mechanism model and the lightweight time-series model. When the edge warning meets the preset triggering conditions, it initiates a deep inference request to the cloud self-evolving agent and uploads the original time-series data, the fault features and key data fragments to the cloud self-evolving agent.

[0009] The cloud-based self-evolving intelligent agent transforms the original time-series data into an LLM-friendly statistical indicator table. Using the statistical indicator table, fault characteristics, and key data fragments as input, and based on the ReAct framework and fault tree analysis logic, it iteratively updates the variable investigation state MIS using state log visitor technology until the fault tree termination condition is met, completing deep diagnosis and obtaining diagnostic results. When the diagnostic path benefit corresponding to the diagnostic result exceeds a preset success threshold, a diagnostic skill is generated or updated through a path benefit function. The diagnostic skill is stored in a knowledge base and the knowledge graph is updated. Based on the diagnostic results, the updated diagnostic skill, and the updated knowledge graph, a structured diagnostic report is generated, and a handling strategy is formed by matching the standard operating procedures in the knowledge graph.

[0010] The cloud-based self-evolving intelligent agent distributes the structured diagnostic report and the handling strategy to the edge nodes; the edge nodes perform alarm escalation or linkage control based on the structured diagnostic report and the handling strategy, forming a cloud-edge collaborative closed loop; the cloud-based self-evolving intelligent agent, based on accumulated historical data, the updated diagnostic skill, and the updated knowledge graph, periodically optimizes the mechanism model and the lightweight time series model on the edge side through a knowledge distillation strategy, thereby achieving continuous evolution of diagnostic capabilities.

[0011] As a preferred solution for graded early warning and fault diagnosis methods for energy storage power stations, the discrimination formula is as follows during the process of the end-side real-time monitoring module discriminating the operating data of the energy storage power station:

[0012]

[0013] In the formula, The real-time feature value of the i-th feature; The historical statistical mean of the i-th feature; Let i be the alarm coefficient for the i-th feature; Let be the historical statistical standard deviation of the i-th feature.

[0014] As a preferred embodiment of a graded early warning and fault diagnosis method for energy storage power stations, the edge early warning includes: anomaly type, fault level, location information, and confidence level; when the confidence level of the edge early warning is less than a preset confidence threshold or the fault level is greater than a high-level fault threshold, a deep inference request is initiated to the cloud-based self-evolving intelligent agent.

[0015] As a preferred solution for graded early warning and fault diagnosis methods for energy storage power stations, in the process of transforming the original time-series data into an LLM-friendly statistical index table by a cloud-based self-evolving intelligent agent, the original time-series data is normalized and integerized through a telemetry data preprocessing pipeline to obtain integerized time-series data. Based on the integerized time-series data, key data points are selected through a change significance function, and only significantly changing segments are retained for compression context processing to obtain compressed key data segments. Based on the compressed key data segments, a statistical index table containing mean, variance, extreme values, and rate of change is generated.

[0016] The expression for normalization and integerization is:

[0017]

[0018] In the formula, The parameter value is converted to an integer. These are the measured values ​​of the parameters; , These represent the minimum and maximum values ​​of the parameter's range, respectively; B is the number of quantization bits.

[0019] The expression for the compression context processing is:

[0020]

[0021] In the formula, The score represents the significance of the change at time t; Let p be the weight of the p-th parameter; Let p be the real-time value of the p-th parameter at time t; Let p be the real-time value of the p-th parameter at time t-1; Let be the historical standard deviation of the p-th parameter.

[0022] As a preferred embodiment of a graded early warning and fault diagnosis method for energy storage power stations, the expression for the variable survey state MIS is:

[0023]

[0024] In the formula, List of confirmed early signs of failure; This is the highest priority candidate root cause hypothesis. This is the set of logic gates in the fault tree that have not yet been satisfied. This represents the confidence vector for each hypothesis;

[0025] The expression for the diagnostic path is:

[0026]

[0027]

[0028] In the formula, For diagnostic pathways; For diagnostic purposes; The tools invoked for the i-th step of the diagnosis; This is the input data for the i-th step of the diagnosis; This is the output result of the diagnosis in step i;

[0029] The formula for calculating the benefits of the diagnostic pathway is as follows:

[0030]

[0031] In the formula, Benefits of the diagnostic pathway; For diagnostic accuracy; For computation / time cost; For manual feedback from operations and maintenance personnel; , , These are the weighting coefficients;

[0032] The structured diagnostic report includes: event summary, telemetry indicator table, root cause diagnostic thought chain, maintenance recommendations, and audit trail.

[0033] As a preferred solution for graded early warning and fault diagnosis methods for energy storage power stations, the cloud-based self-evolving intelligent agent, in the deep diagnosis process, retrieves historical diagnostic experience from long-term memory based on the current diagnostic task through similarity and uses it for reasoning.

[0034] The similarity calculation formula is as follows:

[0035]

[0036] In the formula, For comprehensive similarity; For the current diagnostic tasks to be processed; For the i-th historical diagnostic task in the knowledge base; To embed similarity-weighted fusion coefficients; For vector embedding similarity; This is the embedding vector for the current diagnostic task; For historical diagnostic tasks, the embedding vector is used. For the similarity of the map structure; This is the fault reasoning structure pattern corresponding to the current task; This is the fault reasoning structure pattern corresponding to historical tasks.

[0037] This invention also provides a graded early warning and fault diagnosis system for energy storage power stations, employing the above-mentioned graded early warning and fault diagnosis method for energy storage power stations, including:

[0038] The edge data processing unit is used to identify the operation data of the energy storage power station based on the dynamic multi-threshold table and the edge real-time monitoring module; when the set operation and maintenance characteristics are detected to exceed the corresponding set threshold, an edge alarm is generated and the original time series data within the abnormal time window is pushed to the edge node;

[0039] The edge node data processing unit is used to extract fault features and generate edge warnings based on the original time series data by the edge node through a mechanism model and a lightweight time series model; when the edge warning meets the preset triggering conditions, it initiates a deep inference request to the cloud self-evolving agent and uploads the original time series data, the fault features and key data fragments to the cloud self-evolving agent.

[0040] The cloud-based self-evolving intelligent agent diagnostic processing unit is used to transform the original time-series data into an LLM-friendly statistical indicator table. Taking the statistical indicator table, fault characteristics, and key data fragments as input, and based on the ReAct framework and fault tree analysis logic, it iteratively updates the variable investigation state MIS using state log visitor technology until the fault tree termination condition is met, completing deep diagnosis and obtaining diagnostic results. When the diagnostic path benefit corresponding to the diagnostic result exceeds a preset success threshold, a diagnostic skill is generated or updated through a path benefit function. The diagnostic skill is stored in a knowledge base and the knowledge graph is updated. Based on the diagnostic results, the updated diagnostic skill, and the updated knowledge graph, a structured diagnostic report is generated, and a handling strategy is formed by matching the standard operating procedures in the knowledge graph.

[0041] The periodic optimization and upgrade unit is used by the cloud-based self-evolving intelligent agent to distribute the structured diagnostic report and the handling strategy to the edge nodes; the edge nodes perform alarm upgrades or linkage control based on the structured diagnostic report and the handling strategy, forming a cloud-edge collaborative closed loop; the cloud-based self-evolving intelligent agent, based on accumulated historical data, the updated diagnostic skill, and the updated knowledge graph, periodically optimizes the mechanism model and the lightweight time series model on the edge side through a knowledge distillation strategy, so as to achieve continuous evolution of diagnostic capabilities.

[0042] As a preferred solution for a graded early warning and fault diagnosis system for energy storage power stations, in the end-side data processing unit, during the process of the end-side real-time monitoring module judging the operating data of the energy storage power station, the discrimination formula is as follows:

[0043]

[0044] In the formula, The real-time feature value of the i-th feature; The historical statistical mean of the i-th feature; Let i be the alarm coefficient for the i-th feature; Let be the historical statistical standard deviation of the i-th feature.

[0045] As a preferred solution for a graded early warning and fault diagnosis system for energy storage power stations, the edge node data processing unit includes the following: the edge early warning includes: anomaly type, fault level, location information, and confidence level; when the confidence level of the edge early warning is less than a preset confidence threshold or the fault level is greater than a high-level fault threshold, a deep inference request is initiated to the cloud-based self-evolving intelligent agent.

[0046] As a preferred solution for a graded early warning and fault diagnosis system for energy storage power stations, the cloud-based self-evolving intelligent agent diagnostic processing unit performs normalization and integerization on the original time-series data during the process of the cloud-based self-evolving intelligent agent converting the original time-series data into an LLM-friendly statistical index table. This is achieved through a telemetry data preprocessing pipeline to obtain integerized time-series data. Based on the integerized time-series data, key data points are selected using a change significance function, and only significantly changing segments are retained for compression context processing to obtain compressed key data segments. Based on the compressed key data segments, a statistical index table containing mean, variance, extreme values, and rate of change is generated.

[0047] The expression for normalization and integerization is:

[0048]

[0049] In the formula, The parameter value is converted to an integer. These are the measured values ​​of the parameters; , These represent the minimum and maximum values ​​of the parameter's range, respectively; B is the number of quantization bits.

[0050] The expression for the compression context processing is:

[0051]

[0052] In the formula, The score represents the significance of the change at time t; Let p be the weight of the p-th parameter; Let p be the real-time value of the p-th parameter at time t; Let p be the real-time value of the p-th parameter at time t-1; Let be the historical standard deviation of the p-th parameter.

[0053] As a preferred solution for a graded early warning and fault diagnosis system for energy storage power stations, the expression for the variable survey state MIS in the cloud-based self-evolving intelligent agent diagnostic processing unit is:

[0054]

[0055] In the formula, List of confirmed early signs of failure; This is the highest priority candidate root cause hypothesis. This is the set of logic gates in the fault tree that have not yet been satisfied. This represents the confidence vector for each hypothesis;

[0056] The expression for the diagnostic path is:

[0057]

[0058]

[0059] In the formula, For diagnostic pathways; For diagnostic purposes; The tools invoked for the i-th step of the diagnosis; This is the input data for the i-th step of the diagnosis; This is the output result of the diagnosis in step i;

[0060] The formula for calculating the benefits of the diagnostic pathway is as follows:

[0061]

[0062] In the formula, Benefits of the diagnostic pathway; For diagnostic accuracy; For computation / time cost; For manual feedback from operations and maintenance personnel; , , These are the weighting coefficients;

[0063] The structured diagnostic report includes: event summary, telemetry indicator table, root cause diagnostic thought chain, maintenance recommendations, and audit trail.

[0064] As a preferred solution for a graded early warning and fault diagnosis system for energy storage power stations, in the cloud-based self-evolving intelligent agent diagnostic processing unit, the cloud-based self-evolving intelligent agent, during the deep diagnosis process, retrieves historical diagnostic experience from long-term memory based on the current diagnostic task through similarity and uses it for the reasoning process.

[0065] The similarity calculation formula is as follows:

[0066]

[0067] In the formula, For comprehensive similarity; For the current diagnostic tasks to be processed; For the i-th historical diagnostic task in the knowledge base; To embed similarity-weighted fusion coefficients; For vector embedding similarity; This is the embedding vector for the current diagnostic task; For historical diagnostic tasks, the embedding vector is used. For the similarity of the map structure; This is the fault reasoning structure pattern corresponding to the current task; This is the fault reasoning structure pattern corresponding to historical tasks.

[0068] The present invention has the following advantages:

[0069] First, this invention uses a dynamic multi-threshold table to replace the traditional fixed threshold judgment method, which can adapt to the dynamic fluctuations of the operating conditions of energy storage power stations and the frequent load switching, avoid the alarm deviation problem caused by the poor adaptability of static thresholds, and improve the adaptability of end-side anomaly monitoring and primary alarms.

[0070] Second, the edge-side synchronous fusion mechanism model and lightweight time series model are used to extract fault features. Based on the dual judgment logic of early warning confidence and fault level, the cloud-based deep reasoning is triggered, which can cover both explicit conventional faults and identify hidden difficult faults caused by multi-factor coupling.

[0071] Third, the cloud-based self-evolving intelligent agent builds a complete telemetry data preprocessing pipeline, converting raw time-series data into statistical indicator tables adapted to large model inputs. Combined with fault tree logic deduction and historical diagnostic memory similarity retrieval, it improves the fault reasoning logic chain and enhances the diagnostic completeness in complex fault scenarios.

[0072] Fourth, by constructing standardized diagnostic paths and path benefit evaluation mechanisms, and completing the generation and updating of diagnostic skills and the iterative accumulation of knowledge graphs based on benefit thresholds, we can continuously accumulate experience based on on-site operation and maintenance data and achieve autonomous iterative optimization of diagnostic logic.

[0073] Fifth, relying on the structured diagnostic reports and disposal strategies distributed from the cloud to the edge to form an operation and maintenance closed loop, while using knowledge distillation to regularly optimize the edge-side mechanism model and lightweight time series model, so that the edge model can adapt to the evolution trend of power plant operating conditions, and improve the generalization ability and long-term operational adaptability of the entire early warning and diagnostic system. Attached Figure Description

[0074] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0075] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0076] Figure 1 This is a flowchart illustrating a graded early warning and fault diagnosis method for energy storage power stations provided in Embodiment 1 of the present invention;

[0077] Figure 2 This is a schematic diagram illustrating the specific implementation process of a graded early warning and fault diagnosis method for energy storage power stations provided in Embodiment 1 of the present invention;

[0078] Figure 3 This is a schematic diagram of the end-side threshold over-limit detection log in the graded early warning and fault diagnosis method for energy storage power stations provided in Embodiment 1 of the present invention;

[0079] Figure 4 This is a schematic diagram of edge-side feature extraction in a graded early warning and fault diagnosis method for energy storage power stations provided in Embodiment 1 of the present invention;

[0080] Figure 5 This is a schematic diagram of cloud-based inference in a graded early warning and fault diagnosis method for energy storage power stations provided in Embodiment 1 of the present invention;

[0081] Figure 6 This is a schematic diagram of the architecture of a graded early warning and fault diagnosis system for energy storage power stations provided in Embodiment 2 of the present invention. Detailed Implementation

[0082] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0083] Example 1

[0084] See Figure 1 and Figure 2 Embodiment 1 of the present invention provides a graded early warning and fault diagnosis method for energy storage power stations, comprising the following steps:

[0085] S1. Based on a dynamic multi-threshold table, the edge-side real-time monitoring module judges the operation data of the energy storage power station. When the set operation and maintenance characteristics are detected to exceed the corresponding set threshold, an edge-side alarm is generated and the original time-series data within the abnormal time window is pushed to the edge node.

[0086] S2. Based on the original time-series data, the edge node extracts fault features and generates an edge warning through the mechanism model and the lightweight time-series model. When the edge warning meets the preset triggering conditions, it initiates a deep inference request to the cloud self-evolving agent and uploads the original time-series data, the fault features and key data fragments to the cloud self-evolving agent.

[0087] S3. The cloud-based self-evolving intelligent agent transforms the original time-series data into an LLM-friendly statistical indicator table. Using the statistical indicator table, the fault characteristics, and the key data fragments as input, and based on the ReAct framework and fault tree analysis logic, it iteratively updates the variable investigation state MIS using state log visitor technology until the fault tree termination condition is met, completing deep diagnosis and obtaining diagnostic results. When the diagnostic path benefit corresponding to the diagnostic result exceeds a preset success threshold, a diagnostic skill is generated or updated through a path benefit function. The diagnostic skill is stored in the knowledge base and the knowledge graph is updated. Based on the diagnostic results, the updated diagnostic skill, and the updated knowledge graph, a structured diagnostic report is generated, and a handling strategy is formed by matching the standard operating procedures in the knowledge graph.

[0088] S4. The cloud-based self-evolving intelligent agent distributes the structured diagnostic report and the handling strategy to the edge nodes; the edge nodes perform alarm escalation or linkage control based on the structured diagnostic report and the handling strategy to form a cloud-edge collaborative closed loop; the cloud-based self-evolving intelligent agent, based on accumulated historical data, the updated diagnostic Skill, and the updated knowledge graph, periodically optimizes the mechanism model and the lightweight time series model on the edge side through a knowledge distillation strategy to achieve continuous evolution of diagnostic capabilities.

[0089] In this embodiment, in step S1, the edge-side real-time monitoring module judges the operation data of the energy storage power station based on the dynamic multi-threshold table; when the set operation and maintenance characteristics are detected to exceed the corresponding set threshold, an edge-side alarm is generated and the original time series data within the abnormal time window is pushed to the edge node.

[0090] Specifically, the edge-side real-time monitoring module is deployed in the battery management system (BMS) and uses a dynamic multi-threshold table. Matching key operational features Achieve millisecond-level discrimination.

[0091] The criteria for judgment are as follows:

[0092]

[0093] In the formula, The real-time feature value of the i-th feature; The historical statistical mean of the i-th feature; Let i be the alarm coefficient for the i-th feature; Let be the historical statistical standard deviation of the i-th feature.

[0094] When any operational feature exceeds the corresponding threshold, an edge alarm is immediately generated, and the abnormal time window is automatically captured. All raw time series data within After data packaging is completed, it is pushed to edge nodes in real time to ensure the complete upload of abnormal data. For example... Figure 3 The image shows an example of the threshold over-limit detection log for this step. It visually displays the comparison results between the measured values ​​of parameters such as temperature and voltage monitored on the end side and the threshold, the alarm timestamp, the over-limit level, and other log information. It is the actual presentation of the threshold judgment and alarm generation in this step.

[0095] In this embodiment, in step S2, based on the original time-series data, the edge node extracts fault features and generates an edge warning through a mechanism model and a lightweight time-series model; when the edge warning meets the preset triggering conditions, it initiates a deep inference request to the cloud self-evolving agent and uploads the original time-series data, the fault features and key data fragments to the cloud self-evolving agent.

[0096] Specifically, edge nodes use a dual-model parallel detection channel to process the raw time-series data.

[0097] The mechanism model is an electrochemical-thermal coupling model combined with an extended Kalman filter (EKF). Taking internal short-circuit detection as an example, the voltage residual is calculated. Temperature residual and Pearson correlation coefficient Implement internal short-circuit fault detection.

[0098] The formula for calculating the voltage residual is as follows:

[0099]

[0100] In the formula, This is the voltage residual; This represents the measured terminal voltage of a single battery cell / cluster. To estimate the voltage value.

[0101] The formula for calculating the temperature residual is:

[0102]

[0103] In the formula, This refers to the temperature residual. This represents the measured terminal temperature of a single battery cell / cluster. To estimate the temperature.

[0104] like and In the time window The value continuously exceeded the limit, and its Pearson correlation coefficient was high. If so, it is determined to be an internal short circuit.

[0105] In this embodiment, the lightweight time series model is a pruned and quantized LSTM model, which takes a multi-dimensional time series consisting of voltage V, current I, temperature T, and state of charge (SOC) as input and outputs a predicted health degradation rate. .

[0106] The dual-model outputs are processed by the fusion decision module to generate anomaly type, fault level, location information, and confidence level. Edge warning .like confidence level or fault level This involves initiating a deep inference request in the cloud and simultaneously uploading the original time-series data and fault feature vectors. And key data snippets. For example... Figure 4The image shown is an example of edge-side feature extraction in this step. It displays the real-time monitoring and feature extraction interface for key parameters such as edge-side engineering configuration, algorithm point table, SOH, maximum SOH difference, remaining cycle lifetime, and abnormal capacity cells. It is the actual presentation of dual-model feature extraction and data processing in this step.

[0107] In this embodiment, in step S3, the cloud-based self-evolving intelligent agent transforms the original time-series data into an LLM-friendly statistical indicator table. Using the statistical indicator table, the fault characteristics, and the key data fragments as input, and based on the ReAct framework and fault tree analysis logic, it iteratively updates the variable investigation state MIS using state log visitor technology until the fault tree termination condition is met, completing deep diagnosis and obtaining diagnostic results. When the diagnostic path benefit corresponding to the diagnostic result exceeds a preset success threshold, a diagnostic skill is generated or updated through a path benefit function. The diagnostic skill is stored in the knowledge base and the knowledge graph is updated. Based on the diagnostic results, the updated diagnostic skill, and the updated knowledge graph, a structured diagnostic report is generated, and a handling strategy is formed by matching the standard operating procedures in the knowledge graph.

[0108] Specifically, a large-scale vertical model for energy storage safety is deployed in the cloud as the core intelligent agent, integrating knowledge graphs, long / short-term memory, and tool invocation, and performing reasoning based on the ReAct (Reasoning + Acting) framework. The core architectural components of the intelligent agent include:

[0109] Knowledge Graph (KG): Constructing a Knowledge Graph for Energy Storage Faults In this graph, node E represents the failure mode, cause, phenomenon, and remedial measures; edge R represents causal relationships, temporal relationships, etc.; and T is the timestamp or confidence level. The graph supports complex path queries.

[0110] Knowledge Base (KB): Stores structured "diagnostic skills" and operational case studies. Each skill is:

[0111] .

[0112] The memory module comprises short-term memory (STM) and long-term memory (LTM). STM stores the chain of thought, intermediate results, and tool call records for the current diagnostic session. LTM stores historical successful / failed diagnostic paths and user feedback corrections, using embedded vectors. Implement similarity-based retrieval.

[0113] Tools: Encapsulates callable external capabilities, including: Tool_KG_Query (spectral query), Tool_Data_Analyzer (time series data analysis), Tool_Simulation (electrochemical model simulation), Tool_Report_Gen (report generation), etc.

[0114] In this embodiment, the cloud-based self-evolving agent first converts the raw time-series data into an LLM-friendly format through the telemetry data preprocessing pipeline. Then, the raw time-series data is normalized and integerized using the telemetry data preprocessing pipeline to obtain integerized time-series data.

[0115] The expression for normalization and integerization is:

[0116]

[0117] In the formula, The parameter value is converted to an integer. These are the measured values ​​of the parameters; , These represent the minimum and maximum values ​​of the parameter range, respectively; B is the number of quantization bits. The length of the sequence remains unchanged after integerization, but the token representation of each value is reduced from multiple numbers to one integer token, achieving a compression rate of over 60%.

[0118] Based on the integerized time series data, key data points are filtered using a change significance function, and only fragments with significant changes are retained for compression context processing to obtain compressed key data fragments.

[0119] The expression for the compression context processing is:

[0120]

[0121] In the formula, The score represents the significance of the change at time t; Let p be the weight of the p-th parameter; Let p be the real-time value of the p-th parameter at time t; Let p be the real-time value of the p-th parameter at time t-1; Let be the historical standard deviation of the p-th parameter. Only retain . The compression ratio is dynamically adjusted, typically ranging from 5:1 to 20:1, while preserving the abrupt changes in fault precursors. Data points at the specified time and their immediate vicinity are included, while duplicate sampling points within stable segments are discarded.

[0122] Based on the compressed key data fragments, a statistical index table containing mean, variance, extreme values, and rate of change is generated.

[0123] The calculation formulas for each indicator are as follows:

[0124] Mean:

[0125]

[0126] Standard deviation:

[0127]

[0128] Range:

[0129]

[0130] Skewness:

[0131]

[0132] In the above formula, This represents the mean of the operation and maintenance data sample. This represents the total number of sample data. Let i be the i-th sample data point; Standard deviation; The range is 0. The maximum value in the sample data; It is the minimum value in the sample data; This refers to the skewness.

[0133] An example of a statistical indicator table is shown in Table 1:

[0134] Voltage 3.215 V 0.042 V 0.135 V 1.23 -0.008 V / s yes temperature 32.5 °C 2.1 °C 7.2 °C 0.45 +0.15 °C / s yes ... ... ... ... ... ... ...

[0135] Table 1 Examples of Statistical Indicators

[0136] In this embodiment, to reduce the dependence of single inference on the context window and enhance the logicality of causal backtracking, the present invention proposes a Stateful Log-walker technique. The agent holds a Mutable Investigation State (MIS) object, which is recursively passed between telemetry log blocks. The MIS is updated each time a fault precursor is confirmed, until the logic gate conditions in the fault tree are met.

[0137] The expression for the variable survey state MIS is:

[0138]

[0139] In the formula, List of confirmed early signs of failure; This is the highest priority candidate root cause hypothesis. This is the set of logic gates in the fault tree that have not yet been satisfied. This is the confidence vector for each hypothesis.

[0140] In this embodiment, statistical indicator tables, fault characteristics, and key data fragments are used as inputs. Based on the ReAct framework and combined with fault tree analysis (FTA), the variable investigation state (MIS) is iteratively updated through state log visitor technology. The fault tree logic gates are verified one by one until all are satisfied, thus completing the deep diagnosis.

[0141] In the Thought phase, the agent first reads the current MIS to determine the logic gates that need to be verified next. For example:

[0142] Thought: The current MIS contains "voltage sag" and "abnormal temperature rise" in Confirmed_Facts, and the remaining logic gates are "internal short circuit → (voltage sag AND abnormal temperature rise AND residual correlation > 0.7)". Both conditions are met, and the residual correlation needs to be verified.

[0143] Action: Call Tool_Data_Analyzer to calculate the Pearson correlation coefficient between the voltage residual and the temperature residual within the most recent 10-second window.

[0144] The agent stores the computation results in the MIS and updates Confirmed_Facts. Inference terminates and a diagnostic conclusion is output when Remaining_Gates is empty and the maximum value of Confidence_Vector exceeds a threshold (e.g., 0.85). This method breaks down long sequence inference into multiple short steps, significantly reducing the requirement for large model context lengths.

[0145] In this embodiment, the diagnostic process is abstracted into a diagnostic path. Single-step diagnostic action ,in, The tools invoked for the i-th step of the diagnosis; This is the input data for the i-th step of the diagnosis; This is the output of the diagnosis at step i. The path reward is calculated using the path reward function. When the reward exceeds a preset success threshold, a diagnostic skill is generated / updated, and the knowledge graph is updated synchronously.

[0146] The formula for calculating the benefit of the diagnostic pathway is as follows:

[0147]

[0148] In the formula, Benefits of the diagnostic pathway; For diagnostic accuracy; For computation / time cost; For manual feedback from operations and maintenance personnel; , , These are the weighting coefficients.

[0149] During the process of generating / updating diagnostic skills, when a diagnosis... When the success threshold is reached, the evolution module is activated:

[0150] Abstraction: from Extract general patterns and generalize specific input and output parameters into variables.

[0151] Skill Generation: Generates a new Skill Or update the weights of existing skills. Skills are stored in the knowledge base. middle.

[0152] The knowledge graph is constructed using failure modes, causes, and remedial measures as nodes, and causal and temporal relationships as edges. Knowledge graph updates involve updating the causal relationships confirmed in this diagnosis. As a new fact triplet Add it to the knowledge graph G and update its confidence T.

[0153] In this embodiment, historical diagnostic experience is reused through task similarity retrieval during the reasoning process, and finally a structured diagnostic report conforming to the NFPA 855 / EPRI standard is generated, which includes event summary, telemetry index table, root cause thought chain, maintenance suggestions, audit trail, and matches the operation and maintenance standard operating procedures in the knowledge graph to form a handling strategy.

[0154] The similarity calculation formula is as follows:

[0155]

[0156] In the formula, For comprehensive similarity; For the current diagnostic tasks to be processed; For the i-th historical diagnostic task in the knowledge base; To embed similarity-weighted fusion coefficients; For vector embedding similarity; This is the embedding vector for the current diagnostic task; For historical diagnostic tasks, the embedding vector is used. For the similarity of the map structure; This is the fault reasoning structure pattern corresponding to the current task; This is the fault reasoning structure pattern corresponding to historical tasks.

[0157] The structured diagnostic report template and LLM processing strategy are shown in Table 2:

[0158] Event Summary Time of failure, failure level (emergency / critical / warning), and scope of impact (cell / module / cluster) Natural language description is generated directly based on the final MIS state. Telemetry Index Table Statistical characteristics of key parameters within abnormal time windows Presented in structured Markdown table format, the data comes from the feature summary output of S3.2. Root cause diagnosis (RCA) Physical mechanism analysis and logical reasoning chain Using the chain-of-thought template, list out the reasoning steps one by one. Maintenance Recommendations Specific measures Search the pre-defined maintenance SOP knowledge base and match the most relevant SOP segment. Audit trail Tool call sequence, execution time, inference steps, and credibility score Directly export the agent's execution log (Action History).

[0159] Table 2. Structured Diagnostic Report Templates and LLM Processing Strategies

[0160] like Figure 5 The diagram shown is an example of this cloud-based inference system, displaying the cloud-based fault diagnosis operation interface, capacity anomaly attenuation early warning information, large model in-depth diagnosis results, automated maintenance suggestions, and diagnostic report display interface. It is the actual presentation of in-depth diagnosis, report generation, and strategy output in this step.

[0161] In this embodiment, in step S4, the cloud-based self-evolving intelligent agent distributes the structured diagnostic report and the handling strategy to the edge nodes; the edge nodes perform alarm escalation or linkage control based on the structured diagnostic report and the handling strategy, forming a cloud-edge collaborative closed loop; the cloud-based self-evolving intelligent agent, based on accumulated historical data, the updated diagnostic skill, and the updated knowledge graph, periodically optimizes the mechanism model and the lightweight time series model on the edge side through a knowledge distillation strategy, thereby achieving continuous evolution of diagnostic capabilities.

[0162] Specifically, the cloud-based self-evolving intelligent agent distributes structured diagnostic reports and handling strategies to edge nodes. The edge nodes then execute coordinated control operations such as graded alarm escalation, fault isolation, equipment start-up and shutdown, and cluster-level switch-out based on the report content and strategy instructions, completing a closed-loop process of end-edge-cloud data, instructions, and strategies. The cloud, based on the full historical operating data of the power station, the updated diagnostic skill library, and the knowledge graph, uses a knowledge distillation strategy to iterate parameters, optimize the structure, and improve the generalization ability of the edge-side electrochemical-thermal coupling mechanism model and the lightweight LSTM time series model. This enables the edge model to adapt to scenarios such as the evolution of energy storage power station operating conditions, battery aging, and fault type updates, achieving autonomous iteration and continuous evolution of graded early warning and fault diagnosis capabilities.

[0163] In one possible implementation, a verification example is provided as follows:

[0164] I. Application Scenarios

[0165] A 100MWh grid-side energy storage power station consists of 20 battery clusters and 6,400 lithium iron phosphate cells, employing a cloud-edge collaborative hierarchical early warning architecture. During full-power discharge operation, the BMS detected an abnormal voltage drop in cells #12 and #16 of cluster #, triggering a full-process fault diagnosis and handling process.

[0166] II. Tiered Early Warning and Fault Diagnosis Steps

[0167] T1, Real-time monitoring and alarm push from the device side

[0168] The edge-side BMS collects real-time voltage data for each cell and performs millisecond-level discrimination based on a dynamic multi-threshold table. Monitoring revealed that the voltage difference between cell #16 and the average cell voltage within the cluster reached 300mV, exceeding the preset consistency alarm threshold of 220mV. The edge-side immediately generated a "cell voltage consistency over-limit alarm," automatically extracting raw time-series data for 5 seconds before and after the alarm (a total of 10 seconds), including four types of parameters: cell voltage, current, temperature, and SOC, with a sampling rate of 100Hz, forming an abnormal data segment of 4000 floating-point numbers, which was synchronously pushed to the edge nodes.

[0169] T2, Edge-side Feature Extraction and Cloud Request Triggering

[0170] Edge nodes employ a dual-channel parallel processing approach, combining a mechanistic model and a lightweight time-series model, to process uploaded data.

[0171] Mechanism model channel: Based on the electrochemical-thermal coupling model combined with the extended Kalman filter (EKF), the theoretical voltage value of cell #16 was calculated and compared with the measured voltage to obtain a voltage residual of 120mV; at the same time, the difference between the measured value and the model estimate of the cell temperature was calculated to obtain a temperature residual of +2.5℃; further, the Pearson correlation coefficient between the voltage residual and the temperature residual within a 10-second window was calculated to be 0.85, which exceeds the internal short circuit judgment threshold of 0.7, and is judged as "severe internal short circuit suspected".

[0172] Timing Model Channel: Based on the pruned and quantized LSTM model, the input is a multi-dimensional timing sequence of voltage, current, temperature, and SOC, and the output is a predicted health degradation rate of -1.2% / 100 cycles, indicating abnormal cell aging rate.

[0173] By fusing the outputs of the two models, an edge warning is generated, labeled with the anomaly type as "voltage drop + abnormal temperature rise", the fault level as "severe", the location as "#16 cell", and the confidence level as 0.92. Since the fault level meets the "severe" trigger condition, the edge node immediately initiates a deep inference request to the cloud-based self-evolving agent, and simultaneously uploads the fault feature vector and key data fragments.

[0174] T3, Cloud-based Self-evolving Intelligent Agent Deep Reasoning and Knowledge Evolution

[0175] T31, Data Preprocessing Stage

[0176] The cloud-based self-evolving intelligent agent calls the Tool_Preprocess tool to receive raw data uploaded from the edge and runs data analysis scripts to obtain statistical values ​​such as mean, standard deviation, skewness, range, and skewness.

[0177] T32, FTA-guided MIS iterative inference

[0178] Based on statistical indicator tables, fault characteristics, and key data fragments, the agent initiates fault tree analysis under the ReAct framework:

[0179] Initialize MIS: Confirmed_Facts={"Voltage sag", "Abnormal temperature rise"}, Hypotheses={"Internal short circuit":0.7", "External short circuit":0.2", "Sensor failure":0.1}, Remaining_Gates={"Internal short circuit requires residual correlation > 0.7"};

[0180] Step 1: Verify residual correlation: Thought: "The core conditions for internal short circuit determination need to be confirmed", Action: Call Tool_Data_Analyzer to calculate. and The correlation coefficient; Observation: 0.85, updated MIS: Confirmed_Facts added "Residual Correlation 0.85", Remaining_Gates cleared, and the confidence of "Internal Short Circuit" in Confidence_Vector increased to 0.95;

[0181] Step 2: Trace historical causes: Thought: "Common causes of internal short circuits include lithium plating due to overcharging, which requires querying historical operating data", Action: Call Tool_KG_Query to query the knowledge graph; Observation: Returns "#16 Cell T-30 days have 3 overcharge events", Update MIS: Confirmed_Facts adds "historical overcharge records", Hypotheses adds "lithium plating leading to internal short circuit": 0.8;

[0182] Step 3: Verify Lithium Plating Characteristics: Thought: "Lithium plating failure will cause characteristic peaks in the voltage relaxation curve, which needs to be verified by combining similar cases and curve analysis." Action: Call Tool_LTM_Retrieve to retrieve historical cases and Tool_Data_Analyzer to calculate the second derivative of the voltage relaxation curve. Observation: Two highly similar successfully diagnosed cases were found, and the current cell voltage relaxation curve shows characteristic peaks of lithium plating. Update MIS: Confirmed_Facts adds "characteristic peaks of voltage relaxation curve", and the confidence of "lithium plating causes internal short circuit" in Confidence_Vector increases to 0.95.

[0183] The fault tree termination condition is met, and the final diagnostic result is output: #16 cell suffered a severe internal short circuit fault due to lithium plating caused by historical overcharging.

[0184] T33, Structured Report Generation

[0185] The agent invokes the Tool_Report_Gen tool to generate a structured diagnostic report according to the NFPA 855 standard template. This report includes an event summary, telemetry metrics table, root cause diagnostic thought process, maintenance recommendations, and audit trail information, clearly indicating the location of the faulty cell, the fault mechanism, the risk level, and recommended handling. Maintenance personnel confirm the report's accuracy and provide feedback (Feedbackuser=+1).

[0186] T34. Diagnostic Pathway Benefit Assessment and Knowledge Evolution

[0187] The path reward function calculates a reward of 0.95, exceeding the preset success threshold. The agent generates a new diagnostic Skill Sli_dendrite, stores it in the knowledge base, and updates the knowledge graph by adding a causal edge for "lithium plating → high confidence leading to → internal short circuit," thus completing the knowledge accumulation.

[0188] T4, Cloud-Edge Collaborative Closed Loop and Model Optimization

[0189] The cloud-based self-evolving intelligent agent distributes the structured diagnostic report and the "shutdown to replace #16 battery cell module" handling strategy to the edge nodes, escalates the alarm, and pushes a maintenance work order to the maintenance platform. Subsequently, based on the fault data, the updated diagnostic skills, and the knowledge graph, the cloud uses a knowledge distillation strategy to optimize the parameters of the edge-side mechanism model and the lightweight LSTM model, improving the model's ability to identify lithium-plated internal short-circuit faults and achieving continuous evolution of hierarchical early warning and fault diagnosis capabilities.

[0190] III. Effect Verification

[0191] In this embodiment, the root cause identification is completed in just 3 steps by FTA-guided MIS inference, with a diagnostic accuracy of over 95%. The cloud-edge collaborative closed loop enables rapid fault handling and avoids fault propagation. At the same time, knowledge accumulation and model optimization are completed, providing support for the diagnosis of similar faults in the future. This fully demonstrates the engineering practicality and technical advantages of this invention in complex energy storage power station scenarios.

[0192] The application scenarios of this invention are as follows:

[0193] In high-power charging and discharging scenarios where large-scale grid-side energy storage power stations of hundreds of MWh participate in grid frequency regulation, peak shaving, and reactive power support, this invention achieves hierarchical monitoring and deep reasoning through an edge-cloud collaborative architecture, which can stably identify early anomalies at the cell level and support the safe operation of the power station and the reliable execution of dispatch instructions.

[0194] In the intermittent operation scenarios where photovoltaic and wind power supporting energy storage power stations face severe power fluctuations and frequent operating condition switches, the present invention can adapt to dynamic load changes. Relying on the edge dual models and the cloud self-evolving agents, it can timely detect abnormal trends in voltage, temperature, and SOC and give disposal strategies.

[0195] In the scenarios where user-side energy storage power stations supporting industrial parks, data centers, and large commercial buildings need to ensure uninterrupted power supply, the present invention can quickly complete fault location and root cause analysis, form an executable operation and maintenance plan, reduce the power outage risk and lower the on-site maintenance cost.

[0196] In the scenarios of unified operation and maintenance of multiple sites with distributed energy storage clusters, hybrid layouts of household and industrial and commercial energy storage, the present invention supports cloud centralized deep diagnosis and edge local rapid response, realizing cross-site fault comparison, experience reuse, and model iteration optimization.

[0197] In the scenarios of full-life cycle operation and maintenance of energy storage power stations with long-term operation, gradually decaying battery performance, and increasing fault types, the present invention continuously improves the recognition and reasoning capabilities for complex faults such as aging, lithium plating, and internal short circuits through the update of diagnostic skills and the evolution of knowledge graphs.

[0198] In the scenarios of compliance operation and maintenance where energy storage power stations need to meet safety specifications and issue standardized fault analysis reports, the present invention can automatically generate structured diagnostic files that meet industry requirements, retain the reasoning process and audit trails, and support safety inspections and accident tracing. [[ID=1A]]

[0199] It should be noted that the method of the embodiments of the present disclosure can be executed by a single device, such as a computer or a server, etc. The method of this embodiment can also be applied to a distributed scenario and be completed by multiple devices cooperating with each other. In such a distributed scenario, one of these multiple devices can only execute one or more steps of the method of the embodiments of the present disclosure, and these multiple devices will interact with each other to complete the described method.

[0200] It should be noted that some embodiments of the present disclosure have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be executed in a different order from that in the above embodiments and still achieve the desired results. Additionally, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0201] Embodiment 2

[0202] See Figure 6 , the embodiment 2 of the present invention also provides an energy storage power station hierarchical early warning and fault diagnosis system, including:

[0203] The edge data processing unit 001 is used to identify the operation data of the energy storage power station based on the dynamic multi-threshold table and the edge real-time monitoring module; when the set operation and maintenance characteristics are detected to exceed the corresponding set threshold, an edge alarm is generated and the original time series data within the abnormal time window is pushed to the edge node.

[0204] The edge node data processing unit 002 is used to extract fault features and generate edge warnings based on the original time series data by the edge node through a mechanism model and a lightweight time series model; when the edge warning meets the preset triggering conditions, it initiates a deep inference request to the cloud self-evolving agent and uploads the original time series data, the fault features and key data fragments to the cloud self-evolving agent.

[0205] The cloud-based self-evolving intelligent agent diagnostic processing unit 003 is used to transform the original time-series data into an LLM-friendly statistical indicator table; taking the statistical indicator table, the fault characteristics, and the key data fragments as input, and based on the ReAct framework and fault tree analysis logic, it iteratively updates the variable investigation state MIS using state log visitor technology until the fault tree termination condition is met, completing deep diagnosis and obtaining diagnostic results; when the diagnostic path benefit corresponding to the diagnostic result exceeds a preset success threshold, a diagnostic skill is generated or updated through the path benefit function abstraction; the diagnostic skill is stored in the knowledge base and the knowledge graph is updated; based on the diagnostic results, the updated diagnostic skill, and the updated knowledge graph, a structured diagnostic report is generated, and a handling strategy is formed by matching the operation and maintenance standard operating procedures in the knowledge graph;

[0206] The periodic optimization and upgrade unit 004 is used for the cloud-based self-evolving intelligent agent to distribute the structured diagnostic report and the handling strategy to the edge nodes; the edge nodes perform alarm upgrades or linkage control according to the structured diagnostic report and the handling strategy, forming a cloud-edge collaborative closed loop; the cloud-based self-evolving intelligent agent, based on accumulated historical data, the updated diagnostic skill and the updated knowledge graph, periodically optimizes the mechanism model and the lightweight time series model on the edge side through a knowledge distillation strategy, so as to achieve continuous evolution of diagnostic capabilities.

[0207] In this embodiment, in the end-side data processing unit 001, during the process of the end-side real-time monitoring module judging the operating data of the energy storage power station, the judgment formula is as follows:

[0208]

[0209] In the formula, The real-time feature value of the i-th feature; The historical statistical mean of the i-th feature; Let i be the alarm coefficient for the i-th feature; Let be the historical statistical standard deviation of the i-th feature.

[0210] In this embodiment, the edge node data processing unit 002 includes the following edge warning: anomaly type, fault level, location information, and confidence level. When the confidence level of the edge warning is less than a preset confidence threshold or the fault level is greater than a high-level fault threshold, a deep inference request is initiated to the cloud-based self-evolving intelligent agent.

[0211] In this embodiment, in the cloud-based self-evolving agent diagnostic processing unit 003, during the process of the cloud-based self-evolving agent converting the original time-series data into the LLM-friendly statistical index table, the original time-series data is normalized and integerized through the telemetry data preprocessing pipeline to obtain integerized time-series data; based on the integerized time-series data, key data points are filtered through a change significance function, and only significantly changing segments are retained for compression context processing to obtain compressed key data segments; based on the compressed key data segments, the statistical index table containing mean, variance, extreme values, and rate of change is generated;

[0212] The expression for normalization and integerization is:

[0213]

[0214] In the formula, The parameter value is converted to an integer. These are the measured values ​​of the parameters; , These represent the minimum and maximum values ​​of the parameter's range, respectively; B is the number of quantization bits.

[0215] The expression for the compression context processing is:

[0216]

[0217] In the formula, The score represents the significance of the change at time t; Let p be the weight of the p-th parameter; Let p be the real-time value of the p-th parameter at time t; Let p be the real-time value of the p-th parameter at time t-1; Let be the historical standard deviation of the p-th parameter.

[0218] In this embodiment, the expression for the variable survey state MIS in the cloud-based self-evolving intelligent agent diagnostic processing unit 003 is:

[0219]

[0220] In the formula, List of confirmed early signs of failure; This is the highest priority candidate root cause hypothesis. This is the set of logic gates in the fault tree that have not yet been satisfied. This represents the confidence vector for each hypothesis;

[0221] The expression for the diagnostic path is:

[0222]

[0223]

[0224] In the formula, For diagnostic pathways; For diagnostic purposes; The tools invoked for the i-th step of the diagnosis; This is the input data for the i-th step of the diagnosis; This is the output result of the diagnosis in step i;

[0225] The formula for calculating the benefits of the diagnostic pathway is as follows:

[0226]

[0227] In the formula, Benefits of the diagnostic pathway; For diagnostic accuracy; For computation / time cost; For manual feedback from operations and maintenance personnel; , , These are the weighting coefficients;

[0228] The structured diagnostic report includes: event summary, telemetry indicator table, root cause diagnostic thought chain, maintenance recommendations, and audit trail.

[0229] In this embodiment, in the cloud-based self-evolving intelligent agent diagnostic processing unit 003, the cloud-based self-evolving intelligent agent, during the deep diagnostic process, retrieves historical diagnostic experience from long-term memory based on the current diagnostic task through similarity and uses it for the reasoning process;

[0230] The similarity calculation formula is as follows:

[0231]

[0232] In the formula, For comprehensive similarity; For the current diagnostic tasks to be processed; For the i-th historical diagnostic task in the knowledge base; To embed similarity-weighted fusion coefficients; #imgpt236# represents the embedding similarity of vectors; #imgpt237# represents the embedding vector of the current diagnostic task; #imgpt238# represents the embedding vector of the historical diagnostic task; #imgpt239# represents the graph structure similarity; #imgpt239# represents the fault reasoning structure pattern corresponding to the current task; #imgpt240# represents the fault reasoning structure pattern corresponding to the historical task.

[0233] It should be noted that the information interaction and execution process between the modules of the above system are based on the same concept as the method embodiment in Embodiment 1 of this application, and the resulting technical effects are the same as those in the method embodiment of this application. For details, please refer to the description in the method embodiment shown above in this application, and it will not be repeated here.

[0234] Example 3

[0235] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium storing program code for a graded early warning and fault diagnosis method for an energy storage power station. The program code includes instructions for executing the graded early warning and fault diagnosis method for an energy storage power station according to Embodiment 1 or any possible implementation thereof.

[0236] Computer-readable storage media can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0237] Example 4

[0238] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;

[0239] The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor can execute the graded early warning and fault diagnosis method for energy storage power stations according to Embodiment 1 or any possible implementation thereof.

[0240] Specifically, a processor can be implemented in hardware or software. When implemented in hardware, the processor can be a logic circuit, an integrated circuit, etc. When implemented in software, the processor can be a general-purpose processor that reads software code stored in memory. This memory can be integrated into the processor or located outside the processor and exist independently.

[0241] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable system. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0242] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing systems. They can be centralized on a single computing system or distributed across a network of multiple computing systems. Optionally, they can be implemented using program code executable by a computing system, thereby storing them in a storage system for execution by the computing system. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0243] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for graded early warning and fault diagnosis of energy storage power stations, characterized in that, include: Based on a dynamic multi-threshold table, the edge-side real-time monitoring module identifies the operation data of the energy storage power station; when the set operation and maintenance characteristics are detected to exceed the corresponding set threshold, an edge-side alarm is generated and the original time-series data within the abnormal time window is pushed to the edge node; Based on the original time-series data, edge nodes extract fault features and generate edge early warnings through a mechanism model and a lightweight time-series model. When the edge warning meets the preset triggering conditions, a deep inference request is initiated to the cloud self-evolving agent, and the original time series data, the fault characteristics and key data fragments are uploaded to the cloud self-evolving agent. The cloud-based self-evolving intelligent agent transforms the original time-series data into an LLM-friendly statistical indicator table. Using the statistical indicator table, the fault characteristics, and the key data fragments as input, and based on the ReAct framework and fault tree analysis logic, it iteratively updates the variable investigation state MIS using state log visitor technology until the fault tree termination condition is met, completing deep diagnosis and obtaining diagnostic results. When the diagnostic path benefit corresponding to the diagnostic result exceeds a preset success threshold, a diagnostic skill is generated or updated through a path benefit function. The diagnostic skill is stored in the knowledge base and the knowledge graph is updated. Based on the diagnostic results, the updated diagnostic skills, and the updated knowledge graph, a structured diagnostic report is generated, and a handling strategy is formed by matching the standard operating procedures in the knowledge graph. The cloud-based self-evolving intelligent agent distributes the structured diagnostic report and the handling strategy to the edge nodes; the edge nodes perform alarm escalation or linkage control based on the structured diagnostic report and the handling strategy, forming a cloud-edge collaborative closed loop. The cloud-based self-evolving intelligent agent, based on accumulated historical data, the updated diagnostic skill, and the updated knowledge graph, periodically optimizes the mechanism model and the lightweight time-series model on the edge side through a knowledge distillation strategy, thereby achieving continuous evolution of diagnostic capabilities.

2. The method for graded early warning and fault diagnosis of energy storage power stations according to claim 1, characterized in that, During the process of the end-side real-time monitoring module discriminating the operating data of the energy storage power station, the discrimination formula is as follows: In the formula, The real-time feature value of the i-th feature; The historical statistical mean of the i-th feature; Let i be the alarm coefficient for the i-th feature; Let be the historical statistical standard deviation of the i-th feature.

3. The method for graded early warning and fault diagnosis of energy storage power stations according to claim 2, characterized in that, The edge warning includes: anomaly type, fault level, location information, and confidence level; when the confidence level of the edge warning is less than a preset confidence threshold or the fault level is greater than a high-level fault threshold, a deep inference request is initiated to the cloud-based self-evolving intelligent agent.

4. The method for graded early warning and fault diagnosis of energy storage power stations according to claim 3, characterized in that, During the process of transforming the original time series data into the LLM-friendly statistical index table by the cloud-based self-evolving intelligent agent, the original time series data is normalized and integerized through the telemetry data preprocessing pipeline to obtain integerized time series data. Based on the integerized time series data, key data points are filtered using a change significance function, and only segments with significant changes are retained for compression context processing to obtain compressed key data segments; based on the compressed key data segments, a statistical index table containing mean, variance, extreme values, and rate of change is generated; The expression for normalization and integerization is: In the formula, The parameter value is converted to an integer. These are the measured values ​​of the parameters; , These represent the minimum and maximum values ​​of the parameter's range, respectively; B is the number of quantization bits. The expression for the compression context processing is: In the formula, The score represents the significance of the change at time t; Let p be the weight of the p-th parameter; Let p be the real-time value of the p-th parameter at time t; Let p be the real-time value of the p-th parameter at time t-1; Let be the historical standard deviation of the p-th parameter.

5. The method for graded early warning and fault diagnosis of energy storage power stations according to claim 4, characterized in that, The expression for the variable survey state MIS is: In the formula, List of confirmed early signs of failure; This is the highest priority candidate root cause hypothesis. This is the set of logic gates in the fault tree that have not yet been satisfied. This represents the confidence vector for each hypothesis; The expression for the diagnostic path is: In the formula, For diagnostic pathways; For diagnostic purposes; The tools invoked for the i-th step of the diagnosis; This is the input data for the i-th step of the diagnosis; This is the output result of the diagnosis in step i; The formula for calculating the benefits of the diagnostic pathway is as follows: In the formula, Benefits of the diagnostic pathway; For diagnostic accuracy; For computation / time cost; For manual feedback from operations and maintenance personnel; , , These are the weighting coefficients; The structured diagnostic report includes: event summary, telemetry indicator table, root cause diagnostic thought chain, maintenance recommendations, and audit trail.

6. The method for graded early warning and fault diagnosis of energy storage power stations according to claim 5, characterized in that, In the process of deep diagnosis, the cloud-based self-evolving intelligent agent retrieves historical diagnostic experience from long-term memory based on the current diagnostic task and uses it for reasoning. The similarity calculation formula is as follows: In the formula, For comprehensive similarity; For the current diagnostic tasks to be processed; For the i-th historical diagnostic task in the knowledge base; To embed similarity-weighted fusion coefficients; For vector embedding similarity; This is the embedding vector for the current diagnostic task; For historical diagnostic tasks, the embedding vector is used. For the similarity of the map structure; This is the fault reasoning structure pattern corresponding to the current task; This is the fault reasoning structure pattern corresponding to historical tasks.

7. A graded early warning and fault diagnosis system for energy storage power stations, employing the graded early warning and fault diagnosis method for energy storage power stations as described in any one of claims 1-6, characterized in that, include: The edge data processing unit is used to identify the operation data of the energy storage power station based on the dynamic multi-threshold table and the edge real-time monitoring module; when the set operation and maintenance characteristics are detected to exceed the corresponding set threshold, an edge alarm is generated and the original time series data within the abnormal time window is pushed to the edge node; The edge node data processing unit is used to extract fault features and generate edge early warning based on the original time series data by the edge node through the mechanism model and the lightweight time series model. When the edge warning meets the preset triggering conditions, a deep inference request is initiated to the cloud self-evolving agent, and the original time series data, the fault characteristics and key data fragments are uploaded to the cloud self-evolving agent. The cloud-based self-evolving intelligent agent diagnostic processing unit is used to transform the raw time-series data into an LLM-friendly statistical indicator table. Taking the statistical indicator table, fault characteristics, and key data fragments as input, and based on the ReAct framework and fault tree analysis logic, it iteratively updates the variable investigation state MIS using state log visitor technology until the fault tree termination condition is met, completing deep diagnosis and obtaining diagnostic results. When the diagnostic path benefit corresponding to the diagnostic result exceeds a preset success threshold, a diagnostic skill is generated or updated through a path benefit function. The diagnostic skill is stored in the knowledge base and the knowledge graph is updated. Based on the diagnostic results, the updated diagnostic skills, and the updated knowledge graph, a structured diagnostic report is generated, and a handling strategy is formed by matching the standard operating procedures in the knowledge graph. The regularly optimized and upgraded unit is used by the cloud-based self-evolving intelligent agent to distribute the structured diagnostic report and the handling strategy to the edge nodes; the edge nodes perform alarm upgrades or linkage control based on the structured diagnostic report and the handling strategy, forming a cloud-edge collaborative closed loop; The cloud-based self-evolving intelligent agent, based on accumulated historical data, the updated diagnostic skill, and the updated knowledge graph, periodically optimizes the mechanism model and the lightweight time-series model on the edge side through a knowledge distillation strategy, thereby achieving continuous evolution of diagnostic capabilities.

8. The graded early warning and fault diagnosis system for energy storage power stations according to claim 7, characterized in that, In the terminal data processing unit, during the process of the terminal real-time monitoring module judging the operation data of the energy storage power station, the judgment formula is as follows: In the formula, The real-time feature value of the i-th feature; The historical statistical mean of the i-th feature; Let i be the alarm coefficient for the i-th feature; Let be the historical statistical standard deviation of the i-th feature.

9. The graded early warning and fault diagnosis system for energy storage power stations according to claim 8, characterized in that, In the edge node data processing unit, the edge early warning includes: anomaly type, fault level, location information, and confidence level; when the confidence level of the edge early warning is less than a preset confidence threshold or the fault level is greater than a high-level fault threshold, a deep inference request is initiated to the cloud-based self-evolving intelligent agent.

10. A graded early warning and fault diagnosis system for energy storage power stations according to claim 9, characterized in that, In the cloud-based self-evolving intelligent agent diagnostic processing unit, during the process of the cloud-based self-evolving intelligent agent converting the original time-series data into the LLM-friendly statistical indicator table, the original time-series data is normalized and integerized through the telemetry data preprocessing pipeline to obtain integerized time-series data. Based on the integerized time series data, key data points are filtered using a change significance function, and only segments with significant changes are retained for compression context processing to obtain compressed key data segments; based on the compressed key data segments, a statistical index table containing mean, variance, extreme values, and rate of change is generated; The expression for normalization and integerization is: In the formula, The parameter value is converted to an integer. These are the measured values ​​of the parameters; , These represent the minimum and maximum values ​​of the parameter's range, respectively; B is the number of quantization bits. The expression for the compression context processing is: In the formula, The score represents the significance of the change at time t; Let p be the weight of the p-th parameter; Let p be the real-time value of the p-th parameter at time t; Let p be the real-time value of the p-th parameter at time t-1; Let p be the historical standard deviation of the p-th parameter; In the cloud-based self-evolving intelligent agent diagnostic processing unit, the expression for the variable survey state MIS is: In the formula, List of confirmed early signs of failure; This is the highest priority candidate root cause hypothesis. This is the set of logic gates in the fault tree that have not yet been satisfied. This represents the confidence vector for each hypothesis; The expression for the diagnostic path is: In the formula, For diagnostic pathways; For diagnostic purposes; The tools invoked for the i-th step of the diagnosis; This is the input data for the i-th step of the diagnosis; This is the output result of the diagnosis in step i; The formula for calculating the benefits of the diagnostic pathway is as follows: In the formula, Benefits of the diagnostic pathway; For diagnostic accuracy; For computation / time cost; For manual feedback from operations and maintenance personnel; , , These are the weighting coefficients; The structured diagnostic report includes: event summary, telemetry indicator table, root cause diagnostic thought chain, maintenance recommendations, and audit trail; In the cloud-based self-evolving intelligent agent diagnostic processing unit, during the deep diagnostic process, the cloud-based self-evolving intelligent agent, based on the current diagnostic task, retrieves historical diagnostic experience from long-term memory through similarity and uses it for the reasoning process. The similarity calculation formula is as follows: In the formula, For comprehensive similarity; For the current diagnostic tasks to be processed; For the i-th historical diagnostic task in the knowledge base; To embed similarity-weighted fusion coefficients; For vector embedding similarity; This is the embedding vector for the current diagnostic task; For historical diagnostic tasks, the embedding vector is used. For the similarity of the map structure; This is the fault reasoning structure pattern corresponding to the current task; This is the fault reasoning structure pattern corresponding to historical tasks.

Images