Thermal power station real-time early warning system based on knowledge graph and particle swarm optimization

Abstract

The invention discloses a thermal power station real-time early warning system based on a knowledge graph and a particle swarm algorithm, and relates to the technical field of thermal power station real-time early warning. S2, time sequence feature extraction and significance screening; s3, constructing and optimizing a fault time sequence rule knowledge graph; s4, performing refined calculation on the equipment health degree; s5, dynamic grading early warning of working condition-health degree cooperative adjustment is carried out; and S6, carrying out operation and maintenance feedback driven model iteration and operation and maintenance closed-loop response. According to the invention, the limitation of a traditional single time parameter or single time sequence characteristic is broken through; meanwhile, a mutual information entropy and variance expansion factor combined screening method is adopted, saliency features strongly associated with the fault are reserved, multiple collinearity redundancy features are removed, the calculation efficiency is improved, the situation that redundant information covers weak signals in the early stage of the fault is avoided, and therefore the problem that early recognition precision is insufficient is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of real-time early warning technology for thermal power plants, specifically a real-time early warning system for thermal power plants based on knowledge graphs and particle swarm optimization algorithms. Background Technology

[0002] Thermal power generation is one of the core modes of energy supply in my country. Failures in critical equipment such as boiler tube leaks, turbine rotor imbalance, and generator insulation damage can easily lead to unplanned shutdowns, causing not only daily economic losses in the millions of yuan, but also seriously affecting the stable operation of the power system. Practice shows that the temporal evolution of critical equipment failures in thermal power plants exhibits characteristics of weak early-stage signals, accelerated evolution in the middle stage, and sudden outbreaks in the late stage. Furthermore, the health status of equipment dynamically declines with the years of operation and the quality of maintenance. Fluctuations in operating conditions such as sudden load changes and fuel replacements further exacerbate the failure evolution rate. This provides a critical time window and technical entry point for failure early warning. Existing early warning technologies for thermal power plants mostly rely on instantaneous values ​​of parameters at a single moment or single time-series features to trigger early warnings, failing to form a multi-dimensional feature system that comprehensively depicts the evolution of faults. At the same time, redundant features are not eliminated, resulting in low computational efficiency. Furthermore, weak early-stage fault signals are easily masked by redundant information, leading to insufficient early identification accuracy. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a real-time early warning system for thermal power plants based on knowledge graphs and particle swarm optimization, thus solving the problems mentioned in the background section.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a real-time early warning system for thermal power plants based on knowledge graphs and particle swarm optimization, comprising the following steps: S1. Multi-dimensional time-series data acquisition and preprocessing: Three types of core data are collected: operating status parameters, equipment physical parameters, and environmental and interference parameters. After outlier identification and completion, data smoothing, and dynamic calibration of benchmark values, standardized time-series data are obtained. S2. Temporal Feature Extraction and Significance Screening: Based on the adaptive sliding window of working conditions, a three-dimensional time-series feature composed of the rate of change, cumulative deviation, and trend matching degree is extracted. A joint screening method of mutual information entropy and variance inflation factor is used to obtain a concise and effective feature vector. In the joint screening method, the mutual information entropy between features and fault labels is used. Combined with variance inflation factor Perform the screening; S3. Construction and optimization of fault sequence rule knowledge graph: Construct a knowledge graph containing five types of nodes: time-series features, fault types, equipment, components, and operating conditions. Generate association rules composed of feature combinations, operating conditions, and faults, and dynamically update rule thresholds based on newly added fault data. S4. Refined Calculation of Equipment Health Status: A weighted evaluation model consisting of overall equipment health and key component health is adopted, and parameters such as equipment operating years, maintenance records, failure frequency, and component material characteristics are combined to calculate the overall equipment health. S5. Dynamic hierarchical early warning of coordinated adjustment between operating conditions and health status: A dynamic grading threshold is constructed that integrates health status, operating condition, and their synergistic effect. The threshold coefficient is optimized using a particle swarm optimization algorithm (the particle swarm optimization algorithm for threshold coefficients is an existing publicly available algorithm, which will not be elaborated here). That is, the dynamic grading threshold is iteratively optimized using the particle swarm optimization algorithm. , , By combining salient features with the knowledge graph matching results, a corresponding level of early warning is triggered; S6. Operation and maintenance feedback-driven model iteration and closed-loop response: The system collects four types of feedback data: early warning verification results, fault handling records, manual correction data, and equipment status update information. The model parameters and knowledge graph rules are updated iteratively on a monthly or quarterly basis.

[0005] Furthermore, in step S1: The operating status parameters include, but are not limited to, parameters such as main steam temperature and rotor vibration amplitude, with a sampling frequency of 1Hz. The physical parameters of the equipment include, but are not limited to, static parameters such as equipment material, rated power, and design life. The environmental and interference parameters include, but are not limited to, parameters such as fuel ash content, ambient temperature, and power grid load rate, and the sampling frequency is 5 minutes / time.

[0006] Furthermore, the preprocessing procedure in step S1 is as follows: Outlier identification and completion: using 3 Criteria identify outliers, i.e., those that meet the criteria. The parameter values ​​are identified as outliers; linear interpolation is used for imputation when the missing rate is high, and the KNN algorithm is used for imputation when the missing rate is between 5% and 20%. Indicates the first A monitoring data point is a specific value of a real-time monitoring indicator collected during the operation of a thermal power plant, including but not limited to temperature, pressure, and flow rate. Representing the The mean of each monitoring data point, that is, the average level of the monitoring indicator under normal operating conditions, calculated through statistical methods; For the first The standard deviation of each monitoring data point is used to measure the dispersion of the monitoring indicator data and reflect the fluctuation of the data relative to the mean. This formula compares individual monitoring data. and The absolute value of the difference and the standard deviation The magnitude relationship between three times and other values ​​is used to determine whether the current monitoring data is abnormal. If the data exceeds the normal fluctuation range at the time of establishment, it indicates that there is a potential risk in the operation of the thermal power plant; Data smoothing: A Savitzky-Golay filter is used to eliminate high-frequency noise. The window size is set to 11 points, the polynomial order is 2, and the filtering formula is: in, Indicates the first The predicted value or the processed estimated value at a given moment; It is a normalization coefficient used to ensure that the calculation results are on a reasonable scale; The width of the sliding window determines the range of historical data points involved in the calculation; These are weighting coefficients, reflecting the first... Time data in calculation The importance of time; Representing the The original data points at time t, where The range of values ​​is from arrive , including the first Before and after the moment One data point; Reference value calibration process: The normal reference values ​​of parameters are dynamically updated using the 1-hour moving average method. The calibration formula is: in, Indicates the first Each monitoring indicator Standard or reference values ​​under normal operating conditions at any given time; Indicates the first Each monitoring indicator The actual measured value at that moment; , For parameters Design safety range threshold, This is an indicator function that indicates when the parameter is within a certain range. Otherwise, it is 0.

[0007] Furthermore, in step S2, the adaptive sliding window... Based on load factor Dynamic adjustment: hour, ; hour, 300 seconds; hour, 600 seconds; The specific calculation method for three-dimensional temporal features is as follows: The rate of change, i.e., the slope of the sliding window, is: in, Indicates at time Next, the The rate of change of a monitoring parameter is used to measure how fast or slow the parameter changes over time. Indicates at time Next, the The actual values ​​of each monitoring parameter; Indicates at time Next, the The values ​​of each monitoring parameter; The cumulative deviation, or cumulative change, is: in, Indicates the first Each monitoring indicator The cumulative deviation metric at any given time is used to quantify the degree to which the indicator deviates from its normal state over a past period of time. Indicates the first Each monitoring indicator The actual measured value at that moment; Indicates the first Each monitoring indicator Standard or reference values ​​under normal operating conditions at any given time; The attenuation factor is a constant greater than 0, used to adjust the degree of influence of the deviation value on the cumulative measurement; For time indices in the time series, during summation calculations, from arrive The value is obtained by iterating through the historical time points used to calculate the cumulative deviation measure. The trend matching degree, or trend consistency coefficient, is: in, Indicates at time No. The real-time early warning coefficient corresponding to each monitoring indicator is used to quantify the degree of abnormality of the indicator in the current period. For at any time No. The actual measured values ​​of each monitoring indicator; For at any time No. The fault threshold for each monitoring indicator, which is related to time. And a certain time Changing parameters Related.

[0008] Furthermore, in step S3, the node types of the knowledge graph include time-series feature nodes, fault type nodes, equipment nodes, component nodes, and operating condition nodes; the association edge types include causal edges composed of feature combinations and faults, attribution edges composed of faults and equipment / components, and adjustment edges composed of operating conditions and feature thresholds. The process for generating and optimizing association rules is as follows: Data annotation: For valid faults in the historical fault data, annotate the characteristic combination, corresponding operating conditions, fault type, and component location 1-3 hours before the fault occurred; Rule mining: The Apriori algorithm is used to mine frequent feature combinations, with a minimum support of 0.15 and a minimum confidence of 0.85 to generate an initial rule set; Manual verification: A verification team composed of power plant operation and maintenance experts and algorithm engineers eliminates rules with logical contradictions and low practicality with a support score of <0.05; Threshold dynamic update: The feature thresholds in the rules are updated quarterly based on newly added fault data using the moving average method; Knowledge graph queries use a rule-based index table to accelerate matching, mapping feature combination hashes to index values, with a query response time of ≤100ms.

[0009] Furthermore, the calculation model for the overall health of the equipment in step S4 is as follows: in, Indicates at time No. The comprehensive early warning indicator value of each monitored object; To comprehensively consider the global weight of all factors and reflect the overall impact of all factors on the comprehensive early warning indicator, and ; For at any time No. The basic early warning indicator value for each monitored object is calculated based on a comprehensive analysis of all factors; This indicates the number of factor categories affecting the early warning indicators; For the first The weights of influencing factors are used to measure the relative importance of each type of factor to the comprehensive early warning indicator, and must satisfy the following: and For at any time No. The monitoring object is based on the first Early warning indicator values ​​calculated based on various influencing factors; The calculation method is as follows: in, This is an adjustment coefficient used to adjust the rate of change of the function; The component material attenuation coefficient; For components Time since the last maintenance; The frequency of component failures over the past 3 years; Indicates from component The time interval from when the component is put into use is used to reflect the impact of the component's usage time on the indicator value; It is a constant term that provides a basic numerical offset for the formula; As coefficient, and , Together, they are used to quantify the material attenuation coefficient of components. and The degree of influence of the factors represented on the results; As coefficient, and Combined, used to measure components Since the last maintenance Impact on the results; As coefficient, and Related, used for quantification of components Failure frequency in the past 3 years Impact on the results.

[0010] Furthermore, in step S5, the warning level is divided into three levels: Level 1 warning indicates an emergency fault: triggers an audible and visual alarm, and simultaneously pushes a notification to the operations and maintenance manager and production plant manager via SMS or APP, requiring a response within 10 minutes, reducing the load to below 30%, and initiating the shutdown and maintenance process; Level 2 warning indicates a critical fault: triggers an audible and visual alarm, sends an SMS to the maintenance team leader, requiring a response within 30 minutes, reduces the load to below 70%, and arranges for dedicated monitoring; Level 3 warning indicates a potential fault: the message is pushed to maintenance personnel via the app, requiring a response within 2 hours, and the equipment is added to the daily key monitoring list; The formula for calculating the dynamic grading threshold is: in, These serve as the baseline thresholds for each level of early warning and are the basic reference values ​​for early warning classification. The health impact coefficient is used to measure the degree to which the health status of equipment affects the warning threshold. This is the operating condition influence coefficient, which reflects the magnitude of the effect of the current operating conditions on the early warning threshold. The operating condition-health synergy coefficient reflects the comprehensive impact of the interaction between operating conditions and equipment health on the early warning threshold, where the first-level early warning... Level II warning Level III early warning ; Indicates in Time device The health status of the device ranges from 0 to 1, with a value closer to 1 indicating a better health condition. For load factor, understand as The operating parameter values ​​at any given time represent quantitative indicators of the current operating conditions of the equipment. The maximum value of the operating condition parameter is used as the reference upper limit for the normalization process of the operating condition parameter; The formula for calculating the comprehensive time series feature values ​​using adaptive weight allocation is as follows: in, Indicates at time Next One comprehensive time series feature value; The total number of features; At any moment Next The first feature sequence in the nth feature sequence One eigenvalue; Indicates at time Below, eigenvalues With target variable The mutual information entropy between features is positively correlated with the weight of the feature; that is, the greater the mutual information entropy, the greater the weight of the corresponding feature in the comprehensive calculation. For all features and target variables The mutual information entropy is summed and used to normalize the weights, ensuring that the sum of the weights of each feature is 1.

[0011] Furthermore, in step S6, the operation and maintenance feedback data includes four categories: early warning verification results, fault handling records, manual correction data, and equipment status update information; the model iteration is divided into monthly small iterations and quarterly large iterations: monthly iterations update feature weights and threshold coefficients, and quarterly iterations update knowledge graph rules and health calculation factors. The iterative process is as follows: the feedback data is cleaned and labeled, then the least squares method is used to correct the health calculation factor and the warning threshold coefficient, and then the knowledge graph rule set is updated. The iteration termination condition is that after three consecutive iterations, the improvement in early warning accuracy is ≤0.5% and the reduction in false alarm rate is ≤0.3%.

[0012] Furthermore, the component material attenuation coefficient The values ​​are: 0.08 for carbon steel, 0.03 for alloy steel, and 0.02 for stainless steel.

[0013] This invention provides a real-time early warning system for thermal power plants based on knowledge graphs and particle swarm optimization, which has the following advantages: 1. This real-time early warning system for thermal power plants, based on knowledge graphs and particle swarm optimization, constructs an adaptive sliding window to dynamically adapt to the differences in fault evolution rates under different loads. It extracts three-dimensional time-series features composed of change rate, deviation accumulation, and trend matching degree, forming a multi-dimensional feature system that comprehensively describes the fault evolution law, breaking through the limitations of traditional single-moment parameters or single-time-series features. At the same time, it adopts a joint screening method of mutual information entropy and variance inflation factor to retain significant features strongly correlated with faults and eliminate multicollinear redundant features. This not only improves computational efficiency but also avoids redundant information from obscuring weak early fault signals, thus effectively solving the problem of insufficient early identification accuracy.

[0014] 2. This real-time early warning system for thermal power plants, based on knowledge graphs and particle swarm optimization, achieves accurate early detection of weak fault signals through a collaborative mechanism of adaptive temporal feature mining and saliency screening, effectively overcoming the limitations of traditional early warning systems in fully depicting fault evolution patterns. Furthermore, relying on a dynamic hierarchical early warning mechanism that coordinates operating conditions and health status, it realizes differentiated and precise early warnings, solving the problem of low adaptability between traditional early warning hierarchies and actual operating scenarios, making the early warning response more aligned with the actual operational needs of thermal power plant equipment.

[0015] 3. This real-time early warning system for thermal power plants based on knowledge graphs and particle swarm optimization demonstrates superior stability and practicality. The model can continuously adapt to real-world scenarios such as equipment aging and changes in operating conditions, effectively avoiding the accuracy degradation problem that traditional models experience after long-term operation. Furthermore, the optimized feature selection and knowledge graph query mechanism ensures computational efficiency, meeting the timeliness requirements of real-time early warning for thermal power plants and providing reliable support for equipment lifecycle fault prevention and operation and maintenance scheduling. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the operating flow structure of the engine speed sensor in a real-time early warning system for thermal power plants based on knowledge graphs and particle swarm optimization algorithms, according to the present invention. Detailed Implementation

[0017] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0018] like Figure 1 As shown, the present invention provides a technical solution: a real-time early warning system for thermal power plants based on knowledge graphs and particle swarm optimization. Specific embodiments are as follows. This embodiment uses a 300MW power plant unit (boiler model DG1025 / 18.2-Ⅱ10, turbine model N300-18.2 / 540 / 540) as the test object to verify the effectiveness of the present invention. Data Acquisition and Preprocessing Examples: Example of outlier handling: At a certain moment, the main steam temperature of the boiler was 620℃, exceeding the design safety range of 540~560℃. According to 3 According to the criteria (average main steam temperature 551℃, standard deviation 8℃), |620-551|=69>3×8=24, it is judged as an outlier. Using the linear interpolation method, based on the values ​​of 550℃ at the previous time and 552℃ at the next time, the outlier is completed to 551℃. Reference value calibration example: The 1-hour moving average baseline value of the main steam temperature at t=3600 seconds is calculated as follows: 3600 data points were collected from t=0 to t=3600 seconds. After removing 3 outliers exceeding the safe range, the average value of the remaining 3597 data points was calculated to be 551.2℃, i.e. ; Example of temporal feature extraction and filtering, with adaptive sliding window adjustment: When the grid load factor =90% (high load), ; =60% (medium load), 300 seconds; =40% (low load), 600 seconds; Feature calculation example: Under medium load The three-dimensional time-series characteristics of the main steam temperature at 300 seconds and t=3600 seconds are calculated as follows: Rate of change: Cumulative Deviation: Trend matching degree: The historical fault slope under the same operating conditions is 0.018; Significance screening results: The mutual information value can be obtained through calculation: , , All satisfy conditions; Variance inflation factor value: , , All satisfy Requirements; Based on the above calculation results, all features are retained, and the final determined feature vector is: ; Knowledge graph construction and rule examples, explanation of core nodes and associated edges: Time-series characteristic nodes: temperature change rate > 0.015℃ / s; cumulative deviation > 75; trend matching degree > 0.8; operating condition node: medium load; fault type node: high temperature corrosion of furnace tubes; Associated edge: Medium load + temperature change rate > 0.015℃ / s + cumulative deviation > 75 + trend matching degree > 0.8 → high temperature corrosion of furnace tubes. The support of this rule is 0.18 and the confidence level is 0.92. Rule update example: During the quarterly iteration, three new failure cases of high-temperature corrosion of furnace tubes were added, characterized by medium load, temperature change rate > 0.016℃ / s, and cumulative deviation > 78. The rule thresholds were updated using the moving average method to temperature change rate > 0.015 ± 0.001℃ / s and cumulative deviation > 75 ± 3, to ensure rule adaptability. Example of health score calculation: Overall equipment health calculation: The boiler has been in operation for 8 years, with the last overhaul 6 months ago, and has experienced 2 failures in the past 3 years; Health status calculation of key components: The furnace tube is made of alloy steel (k_m=0.03), has been in operation for 8 years, has been under maintenance for 6 months, and has experienced one failure in the past 3 years; Overall equipment health status: Example of dynamic hierarchical early warning: Current load factor Level 1 warning baseline threshold , , , ; Dynamic threshold calculation: Comprehensive time series eigenvalue calculation: Feature mutual information entropy weight The deviation from the cumulative normalized value is 0.783; Warning trigger result: Known , ,because No Level 1 warning was triggered; continue matching the Level 2 warning threshold. The result is If the Level 2 warning has not yet been triggered, then the Level 3 warning threshold will be matched. ,because If the fault type is identified as a potential risk of high-temperature corrosion of furnace tubes, it will be triggered by the knowledge graph and pushed to the maintenance personnel for inclusion in daily key monitoring. Experimental verification and result analysis: The experimental conditions are as follows: Test objects: 300MW and 600MW power plant units; Test period: 12 months (covering the break-in period of new equipment, the aging period of equipment, high / medium / low load conditions, and fuel replacement scenarios); Test data volume: 1.2TB of real-time operating data, 156 historical fault data, and 89 maintenance feedback data. Comparison methods: traditional fixed threshold method; existing time-series early warning method (no significant screening); static hierarchical early warning method (no collaborative adjustment). The final verification results are shown in the table below: The verification results show that the present invention is significantly superior to the existing technology in terms of key indicators such as early warning accuracy, early warning lead time, and operation and maintenance resource adaptability. Moreover, the model has good long-term operational stability, which fully verifies the technical advantages and engineering applicability of the present invention. Based on the above description, this invention constructs an adaptive sliding window to dynamically adapt to the differences in fault evolution rates under different loads. It extracts three-dimensional time-series features composed of change rate, deviation accumulation, and trend matching degree, forming a multi-dimensional feature system that comprehensively describes the fault evolution law, breaking through the limitations of traditional single-moment parameters or single time-series features. At the same time, it adopts a joint screening method of mutual information entropy and variance inflation factor to retain significant features strongly correlated with faults and eliminate multicollinear redundant features, which not only improves computational efficiency but also avoids redundant information from obscuring weak early fault signals, thus effectively solving the problem of insufficient early identification accuracy. In terms of fault identification and early warning adaptability, the collaborative mechanism of adaptive time-series feature mining and saliency screening can accurately capture weak early fault signals, effectively overcoming the limitation of insufficient characterization of fault evolution patterns in traditional early warning. Relying on the dynamic hierarchical early warning mechanism of coordinated adjustment of operating conditions and health, differentiated and accurate early warning is achieved, solving the problem of low adaptability between traditional early warning levels and actual operating scenarios, and making the early warning response more in line with the actual operating needs of thermal power plant equipment. In terms of model stability and practicality, the model can continuously adapt to actual scenarios such as equipment aging and changes in operating conditions, effectively avoiding the problem of accuracy decay after long-term operation of traditional models. At the same time, the optimized feature selection and knowledge graph query mechanism ensures the computational efficiency of the method, meets the timeliness requirements of real-time early warning for thermal power plants, and provides reliable support for fault prevention and operation and maintenance scheduling throughout the entire life cycle of equipment.

[0019] The embodiments of the present invention are given for illustrative and descriptive purposes only, and are not intended to be exhaustive or to limit the invention to the forms disclosed. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of the invention, and to enable those skilled in the art to understand the invention and to design various embodiments with various modifications suitable for a particular purpose.

Claims

1. A real-time early warning system for thermal power plants based on knowledge graphs and particle swarm optimization, characterized in that: Includes the following steps: S1. Multi-dimensional time-series data acquisition and preprocessing: Three types of core data are collected: operating status parameters, equipment physical parameters, and environmental and interference parameters. After outlier identification and completion, data smoothing, and dynamic calibration of benchmark values, standardized time-series data are obtained. S2. Temporal Feature Extraction and Significance Screening: Based on the adaptive sliding window of working conditions, a three-dimensional time-series feature composed of the rate of change, cumulative deviation, and trend matching degree is extracted. A joint screening method of mutual information entropy and variance inflation factor is used to obtain a concise and effective feature vector. In the joint screening method, the mutual information entropy between features and fault labels is used. Combined with variance inflation factor Perform the screening; S3. Construction and optimization of fault sequence rule knowledge graph: Construct a knowledge graph containing five types of nodes: time-series features, fault types, equipment, components, and operating conditions. Generate association rules composed of feature combinations, operating conditions, and faults, and dynamically update rule thresholds based on newly added fault data. S4. Refined Calculation of Equipment Health Status: A weighted evaluation model consisting of overall equipment health and key component health is adopted, and parameters such as equipment operating years, maintenance records, failure frequency, and component material characteristics are combined to calculate the overall equipment health. S5. Dynamic hierarchical early warning of coordinated adjustment between operating conditions and health status: A dynamic grading threshold is constructed that integrates health status, working condition, and the synergistic effect of the two. The threshold coefficient is optimized by particle swarm optimization algorithm, and the corresponding level warning is triggered by combining the salience features and knowledge graph matching results. S6. Operation and maintenance feedback-driven model iteration and closed-loop response: The system collects four types of feedback data: early warning verification results, fault handling records, manual correction data, and equipment status update information. The model parameters and knowledge graph rules are updated iteratively on a monthly or quarterly basis.

2. The real-time early warning system for thermal power plants based on knowledge graphs and particle swarm optimization as described in claim 1, characterized in that: In step S1: The operating status parameters include, but are not limited to, parameters such as main steam temperature and rotor vibration amplitude, with a sampling frequency of 1Hz. The physical parameters of the equipment include, but are not limited to, static parameters such as equipment material, rated power, and design life. The environmental and interference parameters include, but are not limited to, parameters such as fuel ash content, ambient temperature, and power grid load rate, and the sampling frequency is 5 minutes / time.

3. The real-time early warning system for thermal power plants based on knowledge graphs and particle swarm optimization as described in claim 1, characterized in that: The preprocessing procedure in step S1 is as follows: Outlier identification and completion: using 3 Criteria identify outliers, i.e., those that meet the criteria. The parameter values ​​are identified as outliers; linear interpolation is used for imputation when the missing rate is high, and the KNN algorithm is used for imputation when the missing rate is between 5% and 20%. Indicates the first A monitoring data point is a specific value of a real-time monitoring indicator collected during the operation of a thermal power plant, including but not limited to temperature, pressure, and flow rate. Representing the The mean of each monitoring data point, that is, the average level of the monitoring indicator under normal operating conditions, calculated through statistical methods; For the first The standard deviation of each monitoring data point is used to measure the dispersion of the monitoring indicator data and reflect the fluctuation of the data relative to the mean. This formula compares individual monitoring data. and The absolute value of the difference and the standard deviation The magnitude relationship between three times and other values ​​is used to determine whether the current monitoring data is abnormal. If the data exceeds the normal fluctuation range at the time of establishment, it indicates that there is a potential risk in the operation of the thermal power plant; Data smoothing: A Savitzky-Golay filter is used to eliminate high-frequency noise. The window size is set to 11 points, the polynomial order is 2, and the filtering formula is: in, Indicates the first The predicted value or the processed estimated value at a given moment; It is a normalization coefficient used to ensure that the calculation results are on a reasonable scale; The width of the sliding window determines the range of historical data points involved in the calculation; These are weighting coefficients, reflecting the first... Time data in calculation The importance of time; Representing the The original data points at time t, where The range of values ​​is from arrive , including the first Before and after the moment One data point; Reference value calibration process: The normal reference values ​​of parameters are dynamically updated using the 1-hour moving average method. The calibration formula is: in, Indicates the first Each monitoring indicator Standard or reference values ​​under normal operating conditions at any given time; Indicates the first Each monitoring indicator The actual measured value at that moment; , For parameters Design safety range threshold, This is an indicator function that indicates when the parameter is within a certain range. Otherwise, it is 0.

4. The real-time early warning system for thermal power plants based on knowledge graphs and particle swarm optimization as described in claim 1, characterized in that: In step S2, the adaptive sliding window Based on load factor Dynamic adjustment: hour, ; hour, 300 seconds; hour, 600 seconds; The specific calculation method for three-dimensional temporal features is as follows: The rate of change, i.e., the slope of the sliding window, is: in, Indicates at time Next, the The rate of change of a monitoring parameter is used to measure how fast or slow the parameter changes over time. Indicates at time Next, the The actual values ​​of each monitoring parameter; Indicates at time Next, the The values ​​of each monitoring parameter; The cumulative deviation, or cumulative change, is: in, Indicates the first Each monitoring indicator The cumulative deviation metric at any given time is used to quantify the degree to which the indicator deviates from its normal state over a past period of time. Indicates the first Each monitoring indicator The actual measured value at that moment; Indicates the first Each monitoring indicator Standard or reference values ​​under normal operating conditions at any given time; The attenuation factor is a constant greater than 0, used to adjust the degree of influence of the deviation value on the cumulative measurement; For time indices in the time series, during summation calculations, from arrive The value is obtained by iterating through the historical time points used to calculate the cumulative deviation measure. The trend matching degree, or trend consistency coefficient, is: in, Indicates at time No. The real-time early warning coefficient corresponding to each monitoring indicator is used to quantify the degree of abnormality of the indicator in the current period. For at any time No. The actual measured values ​​of each monitoring indicator; For at any time No. The fault threshold for each monitoring indicator, which is related to time. And a certain time Changing parameters Related.

5. A real-time early warning system for thermal power plants based on knowledge graphs and particle swarm optimization as described in claim 1, characterized in that: In step S3, the node types of the knowledge graph include time-series feature nodes, fault type nodes, equipment nodes, component nodes, and operating condition nodes; the association edge types include causal edges composed of feature combinations and faults, attribution edges composed of faults and equipment / components, and adjustment edges composed of operating conditions and feature thresholds. The process for generating and optimizing association rules is as follows: Data annotation: For valid faults in the historical fault data, annotate the characteristic combination, corresponding operating conditions, fault type, and component location 1-3 hours before the fault occurred; Rule mining: The Apriori algorithm is used to mine frequent feature combinations, with a minimum support of 0.15 and a minimum confidence of 0.85 to generate an initial rule set; Manual verification: A verification team composed of power plant operation and maintenance experts and algorithm engineers eliminates rules with logical contradictions and low practicality with a support score of <0.05; Threshold dynamic update: The feature thresholds in the rules are updated quarterly based on newly added fault data using the moving average method; Knowledge graph queries use a rule-based index table to accelerate matching, mapping feature combination hashes to index values, with a query response time of ≤100ms.

6. A real-time early warning system for thermal power plants based on knowledge graphs and particle swarm optimization as described in claim 1, characterized in that: The calculation model for the overall health of the equipment in step S4 is as follows: in, Indicates at time No. The comprehensive early warning indicator value of each monitored object; To comprehensively consider the global weight of all factors and reflect the overall impact of all factors on the comprehensive early warning indicator, and ; For at any time No. The basic early warning indicator value for each monitored object is calculated based on a comprehensive analysis of all factors; This indicates the number of factor categories affecting the early warning indicators; For the first The weights of influencing factors are used to measure the relative importance of each type of factor to the comprehensive early warning indicator, and must satisfy the following: and For at any time No. The monitoring object is based on the first Early warning indicator values ​​calculated based on various influencing factors; The calculation method is as follows: in, This is an adjustment coefficient used to adjust the rate of change of the function; The component material attenuation coefficient; For components Time since the last maintenance; The frequency of component failures over the past 3 years; Indicates from component The time interval from when the component is put into use is used to reflect the impact of the component's usage time on the indicator value; It is a constant term that provides a basic numerical offset for the formula; As coefficient, and , Together, they are used to quantify the material attenuation coefficient of components. and The degree of influence of the factors represented on the results; As coefficient, and Combined, used to measure components Since the last maintenance Impact on the results; As coefficient, and Related, used for quantification of components Failure frequency in the past 3 years Impact on the results.

7. A real-time early warning system for thermal power plants based on knowledge graphs and particle swarm optimization as described in claim 1, characterized in that: In step S5, the warning level is divided into three levels: Level 1 warning indicates an emergency fault: triggers an audible and visual alarm, and simultaneously pushes a notification to the operations and maintenance manager and production plant manager via SMS or APP, requiring a response within 10 minutes, reducing the load to below 30%, and initiating the shutdown and maintenance process; Level 2 warning indicates a critical fault: triggers an audible and visual alarm, sends an SMS to the maintenance team leader, requiring a response within 30 minutes, reduces the load to below 70%, and arranges for dedicated monitoring; Level 3 warning indicates a potential fault: the message is pushed to maintenance personnel via the app, requiring a response within 2 hours, and the equipment is added to the daily key monitoring list; The formula for calculating the dynamic grading threshold is: in, These serve as the baseline thresholds for each level of early warning and are the basic reference values ​​for early warning classification. The health impact coefficient is used to measure the degree to which the health status of equipment affects the warning threshold. This is the operating condition influence coefficient, which reflects the magnitude of the effect of the current operating conditions on the early warning threshold. The operating condition-health synergy coefficient reflects the comprehensive impact of the interaction between operating conditions and equipment health on the early warning threshold, where the first-level early warning... Level II warning Level III early warning ; Indicates in Time device The health status of the device ranges from 0 to 1, with a value closer to 1 indicating a better health condition. For load factor, understand as The operating parameter values ​​at any given time represent quantitative indicators of the current operating conditions of the equipment. The maximum value of the operating condition parameter is used as the reference upper limit for the normalization process of the operating condition parameter; The formula for calculating the comprehensive time series feature values ​​using adaptive weight allocation is as follows: in, Indicates at time Next One comprehensive time series feature value; The total number of features; At any moment Next The first feature sequence in the nth feature sequence One eigenvalue; Indicates at time Below, eigenvalues With target variable The mutual information entropy between features is positively correlated with the weight of the feature; that is, the greater the mutual information entropy, the greater the weight of the corresponding feature in the comprehensive calculation. For all features and target variables The mutual information entropy is summed and used to normalize the weights, ensuring that the sum of the weights of each feature is 1.

8. A real-time early warning system for thermal power plants based on knowledge graphs and particle swarm optimization as described in claim 1, characterized in that: In step S6, the operation and maintenance feedback data includes four categories: early warning verification results, fault handling records, manual correction data, and equipment status update information; the model iteration is divided into monthly small iterations and quarterly large iterations: monthly iterations update feature weights and threshold coefficients, and quarterly iterations update knowledge graph rules and health calculation factors. The iterative process is as follows: the feedback data is cleaned and labeled, then the least squares method is used to correct the health calculation factor and the warning threshold coefficient, and then the knowledge graph rule set is updated. The iteration termination condition is that after three consecutive iterations, the improvement in early warning accuracy is ≤0.5% and the reduction in false alarm rate is ≤0.3%.

9. A real-time early warning system for thermal power plants based on knowledge graphs and particle swarm optimization as described in claim 6, characterized in that: The component material attenuation coefficient The values ​​are: 0.08 for carbon steel, 0.03 for alloy steel, and 0.02 for stainless steel.

Images