Multi-source data fusion energy storage abnormity causal chain automatic generation method

Abstract

The invention discloses a multi-source data fusion energy storage abnormity causal chain automatic generation method, and relates to the technical field of large language models. The method comprises the following steps: firstly, carrying out synchronous preprocessing on voltage and current time sequence data, environment sensing data and BMS log data, and eliminating data heterogeneous interference through time alignment, missing measurement filling and de-noising normalization; positioning an abnormal interval based on a residual method, and extracting a threshold event, a peak event and a log event to construct a directed graph; quantifying causal edge weights among events by adopting an improved PCMCI algorithm, and searching an optimal path with the maximum average causal score; and inputting the optimal path into a large language model to generate a natural language causal chain, fusing the path average causal score and the model confidence to improve the result reliability, and performing local reasoning in the whole process. According to the method, the accurate and readable abnormal causal chain can be output, the industrial data safety and the generation real-time performance are guaranteed, the energy storage fault positioning efficiency is effectively improved, and powerful support is provided for operation and maintenance decision making.

Description

Technical Field

[0001] This invention relates to the field of large language model technology, specifically to a method for automatically generating causal chains of energy storage anomalies through multi-source data fusion. Background Technology

[0002] During the operation of energy storage systems, various safety accidents and abnormal faults are often accompanied by the coordinated changes of multi-dimensional signals, including electrical signals (such as voltage and current fluctuations), environmental signals (such as temperature rise, abnormal humidity, and excessive concentration of special gases), and log alarm codes generated by the battery management system (BMS). There are complex correlations and causal relationships between these signals, and cross-signal analysis is required to fully trace the logic of abnormal evolution.

[0003] Traditional threshold or single-source anomaly detection can only provide alarm scores, lacking explanations of causal chains across signals. Large Language Model (LLM) techniques alone do not possess accurate temporal statistical capabilities.

[0004] In view of the above, this application is hereby submitted. Summary of the Invention

[0005] The purpose of this invention is to provide a method for automatically generating causal chains of energy storage anomalies through multi-source data fusion, so as to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, the present invention provides a method for automatically generating causal chains of energy storage anomalies based on multi-source data fusion, comprising the following steps:

[0007] S1: Perform data synchronization processing. The input data includes three types: first, voltage and current timing data. ,in For discrete-time indexing, This represents the total number of time steps. The first is a multi-dimensional measured observation vector containing electrical quantities such as voltage and current of individual cells / clusters; the second is environmental sensor data. The data includes ambient temperature, humidity, and gas concentration; thirdly, BMS logs. , For the number of log entries, Includes timestamps and alarm codes. During synchronization, first... , Resampling is performed to a common time grid (e.g., 1 second), and the event stamps of the BMS logs are mapped to the most recent time grid. If clock skew exists, cross-correlation or hardware synchronization information is used for correction. For short-term missing data, spline or local linear interpolation is used to fill in the gaps, and long-term missing data is marked as "undeterminable" and masked during subsequent event extraction. High-frequency noise is then denoised using low-pass or wavelet thresholding, and the dimensions of each data channel are normalized to ultimately form a joint observation sequence on a unified time axis. ,in A binary or category flag indicating whether there is a log event at that moment.

[0008] S2: Conduct anomaly detection, using the predictive residual method to determine anomaly intervals. Define the residual. ,in Obtained based on LSTM model Predicted value Describes the 2-norm. That is, the residual magnitude of the measured observation vector and the predicted vector; The standard deviation of the residuals or its robust estimate (e.g.) , (This is the absolute deviation of the median). When an outlier occurs, that moment is marked as an anomaly. Adjacent or nearest-neighbor anomalies are merged along the timeline according to the minimum dwell time and maximum gap rules, forming an anomaly interval set. For example in At that time, the actual value of the single-cell voltage Predicted value Calculated ,because This moment is marked as an anomaly.

[0009] S3: Construct an event graph, including the exception interval. Three types of key events are extracted: first, threshold events, such as single-unit voltage below the safety threshold, temperature above the set threshold, and differential pressure exceeding the limit; second, peak events, such as charging / discharging current peaks, temperature rise peaks, and differential pressure peaks; and third, log events, i.e., alarm / protection codes (including code values ​​and text descriptions) from the BMS logs. These events are then analyzed based on their temporal relationship and causal delay windows. ( (To set the maximum causal delay), the earlier-occurring events are directed to the later-occurring events, forming a directed graph. ,in For a set of nodes, This is a set of causal edges between events. The temporal order is used to suppress the generation of reverse edges. If multiple events exist at the same time, the direction of the edges is determined according to the priority rule of "threshold event > peak event > log event". For example, if a "voltage imbalance" node is detected... The subsequent "rapid temperature rise" node Finally, the "Current Collector Corrosion Alarm (0x57E2)" node ,but , This forms a directed graph. .

[0010] S4: Perform causal scoring and compute the directed graph using the improved PCMCI algorithm. The causal strength of the causal edges is used to obtain the causal score for each edge. , The larger the value, the more significant the event. Regarding the event The stronger the causal strength, the better. The specific process is as follows: First, determine the candidate lag set. ( (for each pair of nodes, the maximum lag value is set). and each ,test (event exist (Observation of time) and (event exist The conditional independence of observations at different times is determined; then, a stepwise selection-removal strategy is used to select the condition set, controlling for spurious associations, and global significance is controlled by the FDR (false detection rate); finally, the significance statistic obtained from the test is... -value transformation to The interval is obtained For example, regarding edges Calculated ; targeting the edge Calculated .

[0011] S5: Perform optimal chain search in the directed graph. The path with the highest average causal score in the search is the optimal causal path. The path average causal score is defined as follows: ,in For path The number of sides, For path The sum of causal scores of all causal edges within the path; the optimal path satisfies .like Given a directed acyclic graph, dynamic programming is used to solve it in polynomial time; if If a cycle exists, first transform the graph into a directed acyclic graph according to the temporal topological relationships and the minimum cycle breaking rule, then perform dynamic programming search. For example, if there are three candidate paths: ( ), ( ), ( ), final choice As .

[0012] S6: Generate Chinese causal chains using LLM, and select the optimal causal path. The sequenced structured prompts, presented as "cause → evolution → consequence," serve as the input prompt for the LLM. To mitigate the risk of "illusion" in the LLM output, a predefined glossary is used, allowing the use of timestamps and thresholds corresponding to referenced nodes, with additional terminology definitions added when necessary. Finally, the LLM generates a Chinese causal chain that meets operational readability requirements, for example, for... The LLM outputs "Voltage imbalance → Temperature rise → Current collector corrosion alarm (0x57E2)".

[0013] S7: Perform confidence fusion, fusing the average causal score of the optimal path with the confidence score of the LLM output to obtain the comprehensive confidence score. The calculation formula is: ,in for Average causal score The internal confidence level for the LLM output is set to 0.6 and 0.4, which are preset weights and can be adjusted according to the on-site safety level. For example... , hour, .

[0014] S8: Output the report. The system generates a report containing structured information, including Markdown tables, thumbnails, and Chinese text explanations. The Markdown tables record the event name, the corresponding timestamp, and the relevant causal score (e.g., "voltage imbalance" corresponds to 120s, 0.92; "temperature rise" corresponds to 135s, 0.89; "current collector corrosion alarm" corresponds to 150s, no score). The thumbnails are visualized in the form of black and white arrow links. It intuitively presents the cause-and-effect relationship; the Chinese text explanation explains the causal chain and... The system provides explanations, such as "The system detected an abnormal imbalance in the voltage of a single cell, which was followed by a rapid increase in temperature, eventually triggering a BMS alarm (0x57E2). The confidence level of this causal chain is 0.849. It is recommended that maintenance personnel review the report." The system also supports exporting the report to PDF or Word format with one click for easy archiving and review.

[0015] Compared with the prior art, the beneficial effects of the present invention are:

[0016] 1. Multi-source data synergy and fusion, breaking through the limitations of single source: By aligning voltage and current time series data, environmental sensor data and BMS logs with timestamps, correcting clock offsets and filling missing measurements, and denoising and normalizing, a unified time axis fusion of multi-source heterogeneous data is achieved. This solves the problem of abnormal omissions and misjudgments caused by data fragmentation in traditional threshold methods or single-source detection, and lays the data foundation for accurate causal analysis.

[0017] 2. Synergistic Quantification of Causality and Language Generation, Balancing Accuracy and Readability: An improved PCMCI algorithm is employed to quantify the causal strength of events through conditional independence testing, condition set optimization, and score normalization. Simultaneously, the optimal causal path is used as a structured prompt input to the large language model, limiting the generation of natural language causal chains using specialized terminology. This combination overcomes the limitations of traditional statistical methods, such as lack of explanation and the susceptibility of single language models to illusions, ensuring causal accuracy while reducing operational and comprehension costs.

[0018] 3. Confidence-weighted fusion to reduce output risk: The innovative approach combines the average causal score of the optimal path with the internal confidence score of the large language model to balance the objectivity of statistical causality with the reliability of the language model. This effectively suppresses the risk of false causality due to reliance on the language model alone and improves the practical credibility of the causal chain.

[0019] 4. Accurately locate anomaly intervals and improve analysis efficiency: The prediction residual method is used to calculate the deviation between the measured value and the model prediction value. Combined with the anomaly point merging rule, the complete anomaly interval is locked, avoiding misjudgment of isolated anomalies or omission of key nodes. This ensures that subsequent analysis focuses on the effective range, reduces invalid interference, and improves the efficiency of causal chain generation.

[0020] 5. Localized deployment to meet industrial needs: Supports local inference, all data processing, causal analysis and report generation are completed locally, without the need to upload sensitive operational data to external sources, meeting industrial data privacy and security requirements, avoiding data leakage, and ensuring stable operation in complex network environments.

[0021] 6. Enhanced Practicality of Structured Reports: Generates structured reports with tables, thumbnails, and text explanations, supports exporting in common formats, transforms abstract causal relationships into intuitive information, adapts to existing operation and maintenance archiving and review processes, and improves the engineering implementation capabilities of technical solutions. Attached Figure Description

[0022] Figure 1 The overall architecture diagram of the method for automatically generating causal chains of energy storage anomalies through multi-source data fusion. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0024] Please see Figure 1 This invention provides a technical solution: an automatic generation method for causal chains of energy storage anomalies based on multi-source data fusion. This specific embodiment details the method for automatically generating causal chains of energy storage anomalies based on multi-source data fusion. Using the operation logs, voltage and current monitoring data, and environmental sensor data (including temperature, humidity, and gas concentration) of a 50MW / 100MWh energy storage power station as a basis, and combining it with a fault scenario of "rapid temperature rise within 15 seconds, increased differential pressure between individual cells, and triggering BMS log code 0x57E2," the method's practical application process is fully presented according to the flow of "data synchronization → anomaly detection → event graph construction → causal scoring → optimal chain search → LLM generation → confidence fusion → report output." The following will first define the symbols and abbreviations used in the implementation, and then elaborate on the specific details step by step.

[0025] I. Conventions for Symbols and Abbreviations

[0026] To ensure clarity in the implementation process, the following symbols are used (consistent with the context of the technical solution of this invention): Discrete-time index; Sampling period, in this embodiment ; : Multidimensional measured observation vector, including electrical quantities such as voltage, current, and temperature of individual cells / clusters; : Based on the LSTM prediction model Predicted vector; Residual modulus length This represents the 2-norm, used to quantify the deviation between measured and predicted values; The standard deviation of the residuals or its robust estimate is used in this embodiment. ( (This is the absolute deviation of the median). : The set of abnormal intervals, that is, those that satisfy A collection consisting of all time indices; Timing data such as voltage and current; Environmental sensor data (temperature, humidity, gas concentration); BMS log event sequence, including event timestamps and alarm / protection codes; : A directed graph after extracting events within the abnormal interval; It is a set of nodes (including threshold event, peak event, and log event nodes). It is a set of directed edges (representing causal relationships between events); Improve the causal edge weights calculated in the PCMCI algorithm; larger values ​​represent events. Regarding the event The stronger the causal strength; A directed path in a directed graph; :path Number of edges contained; :path The average causality score; The optimal causal path is the path with the highest average causal score. Internal confidence of the output text of a Large Language Model (LLM); The overall confidence level is set to 0.6 and 0.4, which are preset weights and can be adjusted according to the on-site safety level.

[0027] II. Specific Implementation Steps

[0028] Step 1: Data Synchronization and Preprocessing (S1)

[0029] This step aims to... , and To unify the data onto the same timeline and eliminate interference from sampling frequency differences, clock drift, and missing data on subsequent causal inference, the specific steps are as follows:

[0030] 1. Time alignment: The original sampling frequency was 0.5 s / time. The original sampling frequency is 2 seconds / time. First, the two types of data are resampled to a common time grid. ); BMS logs Event stamps such as "0x57E2 alarm triggered at 120.3s" and "temperature over-limit record at 135.6s" are mapped to the most recent 1-second time grid (120s and 136s respectively); cross-correlation analysis reveals... and There is a clock offset of 0.8s. The synchronization signal of the energy storage system hardware clock module is used for correction to ensure that the time base of the three types of data is consistent.

[0031] 2. Filling in missing measurements: The single-unit voltage data at time 121s was missing for a short period (1 sampling point), and spline interpolation was used to fill the gap. Humidity data for the period from 140 to 145 seconds was missing for an extended period (6 sampling points), marked as "undeterminable", and the generation of environmental data nodes for this time period was masked during subsequent event extraction.

[0032] 3. Noise reduction and normalization: for The 50Hz power frequency high-frequency noise of the medium current signal is removed using a low-pass filter with a cutoff frequency of 10Hz; for data with different dimensions such as voltage (V), current (A), and temperature (°C), min-max normalization is used for mapping to... This range ensures consistency in subsequent residual threshold settings.

[0033] 4. Unified Output: The final aligned multi-source joint sequence is obtained. (Analysis of data from 1 to 150 seconds), where Use a binary flag (1 indicates a BMS log event at that moment, 0 indicates no event), such as at 150s. (Corresponds to alarm 0x57E2).

[0034] Step 2: Anomaly detection (S2)

[0035] This step uses the prediction residual method to determine outlier intervals. The specific operation is as follows:

[0036] 1. Prediction Model Training: An LSTM model is selected as the predictor, with the input being the values ​​from the previous 10 time steps. (Right now The output is The model was trained based on the steady-state operation data (100,000 samples) of the power station over the past 6 months, and the root mean square error (RMSE) of the prediction on the validation set was less than 0.05 (after normalization).

[0037] 2. Residual Calculation and Outlier Marking: For joint sequences from 1 to 150 seconds, according to... Calculate the residual magnitude at each time step; obtained from historical steady-state data. (After normalization) then ;when The time is marked as an outlier, such as 120 seconds. (Single cell voltage 2.75V, normalized to 0.35) (Predicted voltage 3.00V, normalized to 0.42), calculated as follows Marked as an exception.

[0038] 3. Abnormal Interval Merging: Abnormal points are merged according to the rule of "minimum dwell time 2s, maximum gap 1s". There are 3 consecutive abnormal points from 120 to 122s (meeting the minimum dwell time), 123s is normal (gap 1s), and 124s is abnormal again, so they are merged into an abnormal segment from 120 to 124s; the final integrated abnormal interval set is obtained. (The residual exceeds the limit most of the time within this interval) (and meets the merger rules).

[0039] Step 3: Event Extraction and Event Graph Construction (S3)

[0040] This step is within the abnormal range. Extract key nodes from the graph and construct a directed graph based on temporal sequence and causal delay relationships. The specific steps are as follows:

[0041] 1. Key node extraction:

[0042] Threshold event node: Sets the individual cell voltage safety threshold (Normalized to 0.28), the single-cell voltage at 125s is 2.48V (lower than 0.28). Extract the "voltage imbalance / differential voltage exceedance" node. (Corresponds to 125s);

[0043] Peak event node: Calculate the slope of temperature change from 120 to 150 seconds. From 130 to 135 seconds, the temperature rises from 46.2℃ to 49.5℃, with a slope of 0.66℃ / s (the maximum value within the interval). Extract the peak node of "rapid temperature rise". (Corresponds to 135s);

[0044] Log event node: The BMS log triggers an alarm "0x57E2" (described as "current collector corrosion alarm") within 150 seconds. Extract the "current collector corrosion alarm" node. (Corresponds to 150s).

[0045] 2. Directed Graph Construction: Setting Causal Delay Windows (The time difference between the previous event and the next event must be ≤20s), determine the direction of the edges according to the order of time: (125s) and (135s) Time difference 10s≤20s, construct edge ; (135s) and (150s) Time difference 15s≤20s, construct edges Suppress edges with inverse time relationships (e.g.) Generates the final set of nodes. Directed edge set To form a directed graph .

[0046] Step 4: Causal scoring module (S4, improved PCMCI)

[0047] This step utilizes an improved PCMCI algorithm to compute directed graphs. Causal scores of each side The specific steps are as follows:

[0048] 1. Candidate lag set setting: Set the maximum lag value (Considering the impact of events within the previous 5 seconds on the current event), candidate lag set .

[0049] 2. Conditional independence test:

[0050] opposite side :test exist time( Observations (i.e., 124 to 120 seconds) and The conditional independence between observations at 135 s was determined; through a stepwise selection-removal strategy, "ambient humidity" and "current value" were selected as the condition set (excluding the weakly correlated "gas concentration"); the significance was calculated using the Fisher-Z test. -value=0.002.

[0051] opposite side :test exist time( Observations (i.e., from 134 to 130 s) and Conditional independence between observations at 150 s; selecting "rate of change of voltage" as the condition set; Fisher-Z test results. -value=0.005.

[0052] 3. Scoring normalization: ... -value is mapped to interval ( The smaller the -value, the better. (The larger), the mapping formula is ; calculated , .

[0053] Step 5: Optimal Chain Search (S5)

[0054] This step searches for the optimal causal path with the highest average causal score in the directed graph G. The specific steps are as follows:

[0055] 1. Candidate path enumeration and score calculation:

[0056] path Number of sides The sum of causal scores is The average causal score was ;

[0057] path Number of sides The sum of causal scores is Average causal score ;

[0058] path Number of sides Calculated by PCMCI C, average causality score .

[0059] 2. Optimal path determination: according to the formula Compare the average scores of the three groups to determine .

[0060] Step 6: LLM Text Generation (S6)

[0061] This step will determine the optimal causal path. The serialization process is as follows: Serialize the input Prompt into a structured LLM file to generate a Chinese causal chain that is readable by operations and maintenance personnel.

[0062] 1. Prompt Construction: The Prompt content is "Optimal causal path includes the following nodes: 1. Voltage imbalance / differential pressure exceeding limit (125s, causal score 0.92); 2. Rapid temperature rise (135s, causal score 0.89); 3. Current collector corrosion alarm (0x57E2, 150s). Please generate a Chinese causal chain according to the 'cause → evolution → consequence' logic, limiting the use of the terms 'voltage imbalance', 'temperature rise', and 'current collector corrosion alarm (0x57E2)', and referencing the node timestamps."

[0063] 2. LLM output control: The Llama 2-7B model with the energy storage field terminology is used (irrelevant terms are disabled). The output result is: "voltage imbalance (125s) → temperature rise (135s) → current collector corrosion alarm (0x57E2, 150s)", which conforms to the logic of "cause → evolution → consequence" and has no "illusion" content.

[0064] Step 7: Confidence fusion and gating (S7)

[0065] This step combines the average causal score of the optimal path with the LLM confidence level to calculate the overall confidence level. And execute the gating policy, the specific operation is as follows:

[0066] 1. Confidence level acquisition: LLM outputs confidence level through its internal probability module. (Indicates the level of confidence in the output text); Optimal path Average causal score .

[0067] 2. Overall confidence level calculation: according to the formula Substituting the values ​​into the equations is worthwhile. .

[0068] 3. Gating strategy execution: The on-site "audit threshold" is set to 0.85 ( Output directly. Marked "Pending Review"); because The report marks the causal chain as "slightly low and requires verification," but the generated results are retained normally.

[0069] Step 8: Report Output and Implementation (S8)

[0070] This step generates a structured report and supports format export. The specific steps are as follows:

[0071] 1. Report content generation:

[0072] Markdown table: Recording events, timestamps, and causal ratings, as shown in the table below:

[0073]

[0074] Thumbnail: A directed graph visualization of the link is drawn using Matplotlib, with black and white circles marking nodes, black and white arrows marking edges, and causal scores (0.92, 0.89) next to the arrows.

[0075] The Chinese text explains: "The system detected the following abnormalities within the 120-150s interval: voltage imbalance at 125s, temperature rise at 135s, and BMS alarm (0x57E2, current collector corrosion alarm) triggered at 150s; the overall confidence level of this causal chain is 0.849, slightly lower than the review threshold of 0.85. It is recommended that maintenance personnel review the voltage and temperature monitoring data and investigate the risk of current collector corrosion."

[0076] 2. Report Export: Using the python-docx and reportlab libraries, the above content is integrated into Word and PDF documents, which can be downloaded by maintenance personnel with one click for archiving (storing to the power plant maintenance database) and review (submitting to the maintenance supervisor for review).

[0077] Implementation effect verification

[0078] Sample source: Operation logs and monitoring data from a 50MW / 100MWh energy storage power station; Based on 120 historical fault samples, this method (multi-source synchronization + causal scoring + LLM + confidence fusion) was compared with the traditional threshold method (based on a single voltage / temperature threshold alarm) and the single-source anomaly detection + LLM method to verify whether the system performance meets the invention objectives stated in the document (causal chain Top-1 hit rate ≥88%, average generation time ≤4s, readability ≥4.5 / 5). The experimental results are as follows:

[0079]

[0080] The results show that the system's Top-1 causal chain hit rate in this embodiment is 88.3%, the average generation time is 3.7s, and the readability score is 4.6 / 5, all of which meet the invention objectives. At the same time, it supports local deployment without uploading data to the cloud, which meets the requirements of industrial data security and has strong engineering feasibility.

Claims

1. A method for automatically generating causal chains of energy storage anomalies based on multi-source data fusion, characterized in that: Includes the following steps: S1: Perform timestamp alignment processing on voltage and current time series data, environmental sensor data, and BMS log data, and merge the voltage and current time series data, environmental sensor data, and BMS log data into a unified time axis to form a joint observation sequence; S2: The residual between the measured observation vector and the predicted value obtained based on the prediction model is calculated using the prediction residual method. When the calculated residual exceeds three times the standard deviation, the time is marked as an abnormal time. Adjacent or nearby abnormal times on the time axis are merged according to the set rules to form an abnormal interval. S3: Extract three types of key nodes within the abnormal interval: threshold events, peak events, and log events. Construct a directed graph based on the temporal relationship of the three types of key nodes. This directed graph contains a set of nodes and a set of causal edges between events. S4: Calculate the causal edges in the directed graph using the improved PCMCI algorithm to obtain the causal score corresponding to each causal edge. The causal score ranges from 0 to 1 and is used to represent the causal strength of the previous event on the next event. S5: Search for the path with the largest average causal score in the directed graph. During the search, the average causal score of all causal edges in the path is used as the criterion. The path with the largest average causal score is determined as the optimal causal path. S6: Input the optimal causal path as a prompt into the large language model, and the large language model generates a Chinese causal chain that conforms to the logic of "cause → evolution → consequence"; S7: Calculate the average causal score of the optimal causal path, integrate the average causal score with the confidence score output by the large language model to obtain the comprehensive confidence score, and finally output the Chinese causal chain in the form of "cause → evolution → consequence" generated by the large language model.

2. The method for automatically generating causal chains of energy storage anomalies by multi-source data fusion as described in claim 1, characterized in that: The following steps are also included: S8: Generate a structured report. The structured report includes Markdown tables, thumbnails, and Chinese text explanations. The Markdown tables record events, the timestamps corresponding to the events, and the causal scores related to the events. The thumbnails visualize the optimal causal path and present the causal relationship in the form of black and white arrow links. The Chinese text explanations describe the generated causal chain for reference by the duty system and the work order system. It supports one-click export of structured reports to PDF or Word format, facilitating report archiving and review.

3. The method for automatically generating causal chains of energy storage anomalies based on multi-source data fusion as described in claim 1, characterized in that: In S3, threshold events include individual cell voltage below the safety threshold, temperature above the threshold, and differential pressure exceeding the limit; peak events include charge / discharge current peak, temperature rise peak, and differential pressure peak; and log events include BMS alarm codes and BMS protection codes.

Images