Power supply fault early warning model design method for industrial high-risk loop

Abstract

The invention discloses a power supply fault early warning model design method for an industrial high-risk loop, and belongs to the technical field of electrical safety and intelligent operation and maintenance. The method comprises the steps that multivariate time sequence operation data are collected and preprocessed; constructing a comprehensive feature vector containing the voltage and current unbalance degree and a time sequence rolling statistical feature; carrying out training and real-time prediction by using a Transform deep learning model, and outputting an early risk probability; a large language model is innovatively integrated, the risk probability and key feature data are input into LLM, and a deep analysis report containing fault reason reasoning and disposal suggestions is automatically generated. According to the method, the problems of high false alarm rate, early warning lagging and lack of deep analysis capability of a traditional early warning method are solved, and accurate early warning and intelligent decision support of power supply faults are realized.

Description

Technical Field

[0001] This invention belongs to the field of electrical safety and intelligent operation and maintenance technology, specifically relating to a design method for an early warning model of power supply faults in high-risk industrial circuits. Background Technology

[0002] The continuity of power supply to high-risk industrial circuits (such as the main production line power distribution circuits in chemical and metallurgical enterprises) is crucial, as their failures can lead to significant economic losses and safety accidents. Traditional early warning methods have significant drawbacks: First, they rely on single electrical parameter thresholds (such as overcurrent) for judgment, only acting after a fault occurs and lacking pre-emptive warning capabilities; second, they use fixed thresholds, which cannot adapt to dynamic operating conditions such as load fluctuations, easily leading to false alarms or missed alarms; third, they lack comprehensive and coordinated analysis of multiple source parameters such as electrical, environmental, time, and equipment status; fourth, methods based on human experience, such as expert systems, have low levels of intelligence, struggle to handle complex nonlinear relationships, and have poor adaptive capabilities; fifth, they do not fully utilize time-series characteristics and fail to effectively capture trend changes in system operation. These problems result in low accuracy and high false alarm rates in existing early warning technologies, failing to meet the needs of modern industry for power supply safety and predictive maintenance. Summary of the Invention

[0003] The purpose of this invention is to overcome the aforementioned deficiencies of the prior art and provide a design method for an early warning model of power supply faults in high-risk industrial circuits. This method achieves accurate, early prediction and in-depth analysis of power supply faults through multi-source data fusion, intelligent feature engineering, and deep integration of Transformer models with large-scale language models.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A method for designing an early warning model for power supply failures in high-risk industrial circuits, characterized by the following steps:

[0006] Collect multi-dimensional time-series operating data of high-risk industrial circuits, including at least three-phase electrical parameters, environmental parameters, and equipment status parameters;

[0007] The collected data is preprocessed, and a comprehensive feature vector is constructed based on the preprocessed data. The comprehensive feature vector includes voltage imbalance, current imbalance, and time-series rolling statistical features.

[0008] The preprocessed data is used to train a deep learning model based on the Transformer architecture to obtain the early warning model for power supply failure.

[0009] The comprehensive feature vector constructed based on the real-time data is input into the trained early warning model to obtain the early risk probability of power supply failure;

[0010] The early risk probability and the key feature data leading to the risk are input into a large language model to generate an in-depth analysis report, and an early warning is issued accordingly.

[0011] Furthermore, the voltage imbalance and current imbalance are calculated using the three-phase electrical parameters. The voltage imbalance is calculated as the ratio of the standard deviation of the three-phase voltage to the average value of the three-phase voltage; the current imbalance is calculated as the ratio of the standard deviation of the three-phase current to the average value of the three-phase current.

[0012] Furthermore, the time-series rolling statistical features are obtained by performing rolling calculations on electrical parameters through a preset time window. The time-series rolling statistical features include: the rolling average value of the average current and the rolling standard deviation of the average voltage within the preset time window. As a preferred embodiment, the length of the preset time window is 60 consecutive sampling points. This window size effectively balances the feature response speed and stability. Those skilled in the art will understand that, depending on different sampling frequencies and application scenarios, the window size can be adjusted within the range of 30 to 100 consecutive sampling points, for example, 30, 45, 60, or 90 sampling points, all of which fall within the protection scope of this invention.

[0013] Furthermore, when constructing the comprehensive feature vector, time dimension features extracted from timestamps are also introduced, including: hourly features of the day, a flag indicating whether it is a workday, and seasonal features.

[0014] Furthermore, the Transformer architecture is an encoder model that includes a multi-head self-attention mechanism and a feedforward neural network to capture long-range dependencies in the integrated feature vector.

[0015] Furthermore, the preprocessing includes a dynamic standardization process, specifically: Z-score standardization of the input features based on the real-time calculated data mean and standard deviation, and incremental learning is used to dynamically update the mean and standard deviation as new data flows in.

[0016] Furthermore, the method also includes a model optimization step: periodically collecting new operational data and corresponding actual fault records, evaluating model performance based on F1 scores, and iteratively optimizing feature selection based on feature importance analysis results to update the early warning model.

[0017] Furthermore, the in-depth analysis report includes fault cause reasoning and handling suggestions, and its generation specifically includes:

[0018] Construct a structured prompt word template for the industrial fault diagnosis field. This template is not a simple question and answer, but pre-sets the insertion positions of risk probability, key electrical characteristics, environmental parameters and a standardized fault analysis framework.

[0019] The real-time risk probability value and the corresponding key feature value are filled into the template to form a complete prompt word;

[0020] The prompt words are input into a large language model, which is then driven to generate a structured natural language report containing technical risk identification, cause inference and actionable operation and maintenance suggestions. This transforms the model's numerical output into technical decision-making information that can directly guide on-site operation and maintenance.

[0021] Furthermore, the method processes real-time data through a streaming computing framework and provides risk probability query and in-depth analysis report retrieval functions to the outside world via a RESTful API service interface.

[0022] Furthermore, the method also includes an automated model management step: after the early warning model is deployed, its online performance indicators are continuously monitored; when the performance indicators are lower than a preset threshold, the model retraining process is automatically triggered to generate and verify candidate models using recent data; if the performance improvement of the candidate model exceeds a set range, it is updated to the current early warning model; otherwise, the original model is retained and a model performance alarm is issued.

[0023] The beneficial effects of this invention are that, compared with the prior art, this invention:

[0024] It solves the problems of low accuracy and high false alarm rate in early warning: By leveraging the powerful sequence modeling capabilities of the Transformer model to perform in-depth analysis of multi-dimensional features, it can accurately identify early signs of faults. Validated by actual data, compared with traditional threshold methods, this method can improve the prediction accuracy to over 95% and reduce the false alarm rate to below 5%.

[0025] It solves the problem of insufficient early warning capability: it can identify potential fault risks 30-60 minutes in advance, providing sufficient time window for emergency response and achieving true early warning.

[0026] The problem of poor adaptability of static fixed values ​​has been solved: the model has online learning capabilities and can dynamically adjust according to real-time data to adapt to different operating conditions and environmental changes.

[0027] It solves the problems of missing multi-parameter collaborative analysis and lack of in-depth analysis capabilities: It innovatively constructs a comprehensive feature vector that integrates electrical, environmental, time and equipment status parameters, and combines it with a large-scale language model to generate in-depth reports containing fault cause reasoning and handling suggestions, providing comprehensive decision support. Attached Figure Description

[0028] Figure 1 This is an overall flowchart of the method described in this invention;

[0029] Figure 2 This is a flowchart of the feature engineering module described in this invention;

[0030] Figure 3 This is a flowchart of the Transformer model training and prediction process described in this invention;

[0031] Figure 4 This is a flowchart of the LLM deep analysis and report generation described in this invention. Detailed Implementation

[0032] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings, so that those skilled in the art can better understand the essence and implementation of the present invention.

[0033] In the specific implementation process, the overall operation flow of the system is as follows: Figure 1 As shown, multi-dimensional time-series operational data are first collected from key monitoring nodes in high-risk industrial circuits. Three-phase electrical parameters, including those collected by voltage transformers, are acquired through a sensor network deployed on the power distribution circuits. , , Data collected by three-phase voltage and current transformers , , Three-phase current is sampled at a frequency of 4kHz to ensure complete capture of electrical characteristic waveforms; this parameter is derived from the technical requirements of this invention. Simultaneously, ambient temperature (T) and humidity (H) are acquired via digital temperature and humidity sensors, covering a range of -40℃ to +125℃ with an accuracy of ±0.5℃. Equipment status parameters, including the equipment's operating age, are read from the equipment management system. (in monthly units) and maintenance status indicators (Binary variable, 1 indicates maintenance within 3 months). All data are accompanied by timestamps accurate to milliseconds to ensure the integrity of time series information and establish a reliable data foundation for subsequent analysis.

[0034] After data acquisition is complete, the data preprocessing stage begins. The system first cleans the collected raw data, handling any missing or outlier values. Specifically, a data preprocessing-based approach is used. The outlier detection method follows the principle of cropping data that exceeds three times the standard deviation of the mean. Subsequently, a dynamic standardization process is implemented, applying Z-score standardization to the input features based on the real-time calculated mean and standard deviation. The calculation formula is as follows: ,in This is the original data. For standardized data, The mean of the data. The standard deviation of the data is the formula explicitly claimed in this invention. This standardization process employs an incremental learning mechanism, dynamically updating the statistics as new data continuously flows in. The update formula uses an exponentially weighted moving average method, and a learning rate is set. =0.01 enables the system to adapt to changes in data distribution, providing a stable and reliable data foundation for subsequent analysis.

[0035] Based on the preprocessed data, the system follows... Figure 2 The feature engineering process shown constructs a comprehensive feature vector, which innovatively integrates multi-dimensional information. The electrical features include voltage and current imbalances calculated from three-phase parameters. The calculation of these two imbalances strictly follows a mathematical formula, namely, voltage imbalance: and current imbalance: .

[0036] in Indicates standard deviation, and These represent the average values ​​of voltage and current, respectively, and these formulas are the core algorithms of this invention. The rolling statistical features are calculated through a preset time window, specifically calculating the rolling average of the average current and the rolling standard deviation of the average voltage within a window of 60 consecutive sampling points. Simultaneously, the feature vector also incorporates time-dimensional features extracted from timestamps, as described above.

[0037] Hourly characteristics: ,

[0038] Workday sign: ,

[0039] Seasonal characteristics: By combining environmental parameters and equipment status parameters, a 16-dimensional comprehensive feature vector is formed.

[0040] These 16-dimensional feature vectors specifically include: average voltage, voltage imbalance, average current, current imbalance, rolling average current, rolling standard deviation of voltage, ambient temperature, ambient humidity, equipment operating age, maintenance status indicator, hourly feature, working day indicator, seasonal feature, and three reserved extended feature bits. Specifically, the average voltage is calculated by dividing the sum of the three-phase voltages by three; the average current is calculated by dividing the sum of the three-phase currents by three; the rolling average current is calculated as a moving average of the average current from sixty consecutive sampling points; the rolling standard deviation of voltage is calculated as a moving standard deviation of the average voltage from sixty consecutive sampling points; the equipment operating age is recorded in months; the maintenance status indicator uses zero and one to indicate whether maintenance has been performed recently; the hourly feature is directly taken from the hours of the timestamp; the working day indicator is set to one for Monday to Friday and zero for weekends; the seasonal feature is divided into four seasons based on the months; and the three reserved extended feature bits are for adding new feature indicators later.

[0041] Using the constructed feature vectors and historical data, the system follows Figure 3 The model flow shown is used for training and prediction. The Transformer model employs an encoder structure, equipped with a multi-head self-attention mechanism and a feedforward neural network. In a preferred embodiment, the model parameters are set as follows: =64, =8, =3. It should be understood that these parameters can be adjusted according to the actual data size and complexity, for example... You can choose between 32 and 128. You can choose between 4 and 12. A value between 2 and 6 can be selected. The model effectively captures long-range dependencies and non-linear features in feature vectors. By setting the learning rate to 0.001, batch size to 32, and training epochs to 100, and using a weighted binary cross-entropy loss function (positive sample weight = 5.0) to address class imbalance, the model gradually learns early warning patterns for power supply failures. After training, the comprehensive feature vector formed from real-time data is input into the trained early warning model to obtain the early risk probability value of power supply failures, achieving a precise mapping from multi-source data to risk quantification indicators.

[0042] After obtaining early risk probabilities, the system further integrates a large-scale language model for in-depth analysis, namely LLM (Limited Language Model) feature analysis. To achieve the technological transformation from numerical early warning to precise operational decision-making, the prompt word engineering in the LLM deep analysis module is implemented through a specially designed function. This function takes risk probability values, a key feature dictionary, and timestamps as input to construct a highly structured technical analysis prompt word template. The template first defines the AI ​​role as a professional electrical engineer and fault analysis expert to limit its output scope and expertise. Then, it embeds real-time monitoring data, including specific timestamps, formatted risk probability values, and specific values ​​of key anomalies. Next, through explicit instructions, it constrains the large-scale language model to generate a professional analysis report according to a five-part structure, including risk level assessment, main fault assumptions, root cause analysis, emergency response suggestions, and preventative maintenance suggestions. Finally, it requires responses using professional and accurate language. This carefully designed, structured prompt word template is a technical means to ensure that the large-scale language model can generate professional technical analyses with a stable format and practical operational guidance based on specific monitoring data, effectively solving the technical deficiency of traditional early warning systems lacking in-depth analysis capabilities.

[0043] The prompt word engineering implementation in the LLM deep analysis module is as follows:

[0044] Constructing LLM analysis prompts:

[0045] You are a professional electrical engineer and troubleshooting expert. Please reply using professional and accurate language:

[0046] (1) Please analyze the following monitoring data:

[0047] ① Risk Probability

[0048] ②Key abnormal characteristics: Phase A current, equipment temperature, voltage imbalance, and current imbalance.

[0049] (2) Please generate a professional analysis report according to the following structure:

[0050] ① Risk level assessment: [High, medium, and low risk levels are assessed based on risk probability];

[0051] ② Main failure assumptions: [List 2-3 most likely failure types];

[0052] ③ Root cause analysis: [Analyze the possible causes of the abnormality];

[0053] ④ Emergency Response Recommendations: [Provide specific operational suggestions];

[0054] ⑤ Preventive maintenance recommendations: [Provide long-term improvement measures].

[0055] Example 1: When the system detects that the early risk probability of a core power distribution circuit in a chemical plant reaches 0.87, a high-risk warning is immediately triggered because it exceeds the set high-risk threshold of 0.7. For example... Figure 4 As shown, the system inputs the risk probability and key characteristic data (including phase A current of 285A, equipment temperature of 95℃, and voltage imbalance of 0.15) into a large-scale language model, and generates an in-depth analysis report using preset prompt word templates. The report clearly indicates a high-risk level, proposes fault hypotheses such as loose phase A line connections and exacerbated load imbalance, and provides emergency handling suggestions such as immediately checking wiring terminals and measuring contact resistance. It also recommends optimizing the three-phase load distribution as a preventative maintenance measure. Based on the report, staff took timely measures, successfully avoiding a potential power outage.

[0056] Example 2: The system detected an early risk probability of 0.45 in a rolling mill production line circuit of a metallurgical enterprise, falling within the set medium risk level of 0.3-0.7. Although the risk probability did not reach an emergency level, the system still initiated a deep analysis process. After inputting the risk probability and key characteristic data (including current imbalance of 0.12, temperature of 78℃, and significant voltage fluctuations) into a large-scale language model, the generated report indicated a medium risk level. The analysis suggested a potential fault caused by local insulation aging or slight loosening of connection points, recommending a detailed inspection during the next planned maintenance period and increased frequency of temperature monitoring in the area. This tiered response mechanism ensures safety while optimizing the allocation of maintenance resources.

[0057] The system's model optimization mechanism, namely the optimization features, runs continuously and periodically evaluates model performance based on new operational data and actual fault records, continuously optimizing feature selection through feature importance analysis. Simultaneously, the system processes real-time data through a streaming computing framework and provides services externally via a RESTful API interface. The entire system is specifically designed for high-risk industrial circuits, including critical power distribution circuits in the main production lines of large chemical and metallurgical enterprises, where uninterrupted power supply is crucial for continuous production processes. Through the organic combination of the above technical solutions, it provides strong technical support for industrial power supply safety.

[0058] To further ensure the reliability of the early warning system during long-term operation, this invention also designs an automated model management mechanism. This mechanism runs continuously after the early warning model is deployed, calculating performance metrics such as the F1 score of the model's prediction results in real time and comparing them with a preset safety threshold (e.g., 0.85). When the performance metrics fall below this threshold for several consecutive periods, the system automatically triggers a model retraining process, collecting data from the past period (e.g., 30 days) to generate candidate models. These candidate models are evaluated on an independent validation set. Only when their performance improvement compared to the original model exceeds a set value (e.g., 5%) will the system perform a model switch, updating the candidate model to the current early warning model. If the candidate model's performance fails to meet the standard, the system maintains the original model and simultaneously sends an alarm message to maintenance personnel indicating model performance degradation, prompting manual intervention. This mechanism effectively solves the problem of model performance degradation caused by equipment aging and changes in operating conditions, ensuring the high reliability of the early warning system throughout its entire lifecycle.

[0059] Although the present invention has been described in detail above with reference to specific embodiments, those skilled in the art will recognize that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the invention. All such changes should be included within the scope of protection defined by the appended claims.

[0060] Different embodiments of the present invention can be implemented individually or in combination; the parts of the present invention not described in detail are techniques known or commonly used by those skilled in the art.

[0061] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of the invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A design method for an early warning model of power supply faults in high-risk industrial circuits, characterized in that, Includes the following steps: Collect multi-dimensional time-series operating data of high-risk industrial circuits, including at least three-phase electrical parameters, environmental parameters, and equipment status parameters; The collected data is preprocessed, and a comprehensive feature vector is constructed based on the preprocessed data. The comprehensive feature vector includes voltage imbalance, current imbalance, and time-series rolling statistical features. The preprocessed data is used to train a deep learning model based on the Transformer architecture to obtain the early warning model for power supply failure. The comprehensive feature vector constructed based on the real-time data is input into the trained early warning model to obtain the early risk probability of power supply failure; The early risk probability and the key feature data leading to the risk are input into a large language model to generate an in-depth analysis report, and an early warning is issued accordingly.

2. The method according to claim 1, characterized in that, The voltage imbalance and current imbalance are calculated using the three-phase electrical parameters. The voltage imbalance is calculated as the ratio of the standard deviation of the three-phase voltage to the average value of the three-phase voltage; the current imbalance is calculated as the ratio of the standard deviation of the three-phase current to the average value of the three-phase current.

3. The method according to claim 1, characterized in that, The time-series rolling statistical features are obtained by performing rolling calculations on electrical parameters through a preset time window. The time-series rolling statistical features include: the rolling average value of the average current and the rolling standard deviation of the average voltage within the preset time window.

4. The method according to claim 1, characterized in that, When constructing the comprehensive feature vector, time dimension features extracted from timestamps are also introduced. These time dimension features include: hourly features of a day, a flag indicating whether it is a workday, and seasonal features.

5. The method according to claim 1, characterized in that, The Transformer architecture described is an encoder model that includes a multi-head self-attention mechanism and a feedforward neural network to capture long-range dependencies in the integrated feature vector.

6. The method according to claim 1, characterized in that, The preprocessing includes a dynamic standardization process, specifically: Z-score standardization of the input features is performed based on the real-time calculated data mean and standard deviation, and incremental learning is used to dynamically update the mean and standard deviation as new data flows in.

7. The method according to claim 1, characterized in that, The method also includes a model optimization step: periodically collecting new operating data and corresponding actual fault records, evaluating model performance based on F1 scores, and iteratively optimizing feature selection based on feature importance analysis results to update the early warning model.

8. The method according to claim 1, characterized in that, The in-depth analysis report includes fault cause reasoning and handling suggestions, and its generation specifically includes: Construct a structured prompt word template, which pre-defines the insertion positions of risk probabilities, key electrical characteristics, environmental parameters, and analysis framework; The real-time risk probability value and the corresponding key feature value are filled into the template to form a complete prompt word; The prompt words are input into a large language model, which drives it to generate a structured natural language report containing risk identification, cause inference, and operation and maintenance suggestions.

9. The method according to claim 1, characterized in that, The method processes real-time data through a streaming computing framework and provides risk probability query and in-depth analysis report retrieval functions to the outside world through a RESTful API service interface.

10. The method according to claim 1, characterized in that, The method also includes an automated model management step: after the early warning model is deployed, its online performance indicators are continuously monitored; when the performance indicators are lower than a preset threshold, the model retraining process is automatically triggered to generate and verify candidate models using recent data; if the performance improvement of the candidate model exceeds a set range, it is updated to the current early warning model; otherwise, the original model is retained and a model performance alarm is issued.

Images