Substation adaptive inspection method based on equipment health degree dynamic evaluation

Abstract

The invention provides a substation self-adaptive inspection method based on equipment health degree dynamic evaluation. The substation self-adaptive inspection method based on equipment health degree dynamic evaluation comprises the steps of S1, collecting electrical parameters, mechanical vibration parameters and environmental parameters of substation equipment in real time through a multi-source sensor network, and S2, performing data cleaning and time synchronization on the electrical parameters, the mechanical vibration parameters and the environmental parameters, and performing subset dynamic normalization processing. According to the substation self-adaptive inspection method based on equipment health degree dynamic assessment, electrical, mechanical and environmental parameters of the equipment are acquired in real time through the multi-source sensor network, and the real-time performance and accuracy of equipment health degree assessment are remarkably improved by combining dynamic normalization processing and real-time anomaly detection. And the LSTM neural network is used for analyzing the change trend of the equipment health index, so that the residual life of a key component can be predicted, and differentiated inspection and resource optimization allocation can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of power equipment, and in particular to a substation adaptive inspection method based on dynamic evaluation of equipment health. Background Art

[0002] Traditional substation operations and maintenance (O&M) involves routine monitoring, maintenance, and repair of traditional power substations. Substations are crucial nodes in the power system, responsible for converting high-voltage electricity into usable low-voltage energy and distributing it to users via the transmission network. The primary goals of traditional substation O&M are to ensure safe and stable substation operation, extend equipment life, and avoid power supply interruptions.

[0003] Traditional substation operation and maintenance mostly relies on manual regular inspections, using fixed cycles and unified thresholds, which makes it difficult to reflect the actual status of the equipment, leading to problems such as missed inspections of high-risk equipment and excessive inspections of healthy equipment. Although existing technologies have introduced online monitoring sensors, the ability to fuse multi-source data is insufficient. A single normalization method (such as Min-Max or Z-score) is often used to process all features, ignoring the differences in the physical characteristics of electrical, mechanical, and environmental parameters, resulting in low sensitivity in anomaly detection. In addition, weight allocation mostly relies on expert experience to preset static coefficients, which cannot adapt to dynamic scenarios such as equipment aging and sudden changes in working conditions. Threshold setting also lacks adaptability, resulting in frequent false alarms and missed alarms. In terms of inspection strategies, path planning usually aims at the shortest distance, without dynamically adjusting priorities based on equipment health, resulting in delayed responses to key defects. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a substation adaptive inspection method based on dynamic evaluation of equipment health, which solves the problems of insufficient data fusion, rigid weight distribution and delayed threshold setting in traditional inspection methods, as well as the problems of poor substation operation and maintenance efficiency and reliability and high labor costs.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a substation adaptive inspection method based on dynamic evaluation of equipment health, comprising:

[0006] S1. Real-time collection of electrical parameters, mechanical vibration parameters, and environmental parameters of substation equipment through a multi-source sensor network;

[0007] S2. After data cleaning and time synchronization of the electrical parameters, mechanical vibration parameters and environmental parameters, dynamic normalization of the subset is performed;

[0008] S3. The normalized feature vector is further optimized through two-dimensional weight allocation to form a closed-loop feedback loop.

[0009] S4. Integrate the normalized subset indicators with the dynamic weights to generate a comprehensive health index for the device, and classify the device status based on the health index value;

[0010] S5. Dynamically adjust the threshold based on the statistical mean and variance of the device health index within the time window;

[0011] S6. Dynamically generate differentiated inspection strategies based on the health index and preset thresholds, and optimize the inspection path;

[0012] S7. Analyze health index trends using an LSTM neural network to predict the remaining life of the equipment and generate pre-maintenance recommendations.

[0013] S8. Through the closed-loop feedback mechanism, the actual maintenance results are compared with the predicted results, and the health assessment model parameters and threshold settings are reversely optimized.

[0014] Preferably, the two-dimensional weight distribution is static historical weight and dynamic real-time contribution.

[0015] Preferably, the data cleaning includes: removing outliers using a sliding window dynamic 3σ criterion, wherein the window length matches the sensor sampling frequency, and the subset dynamic normalization processing includes the following steps:

[0016] a. Electrical signature subset: The comprehensive health index is calculated using the dynamic range compression method. The formula is:

[0017]

[0018] Among them, x Ei is the i-th electrical parameter, R Ei is the parameter rating, It is the maximum value of the comprehensive electrical parameters in the sliding time window.

[0019] b. Mechanical feature subset: Generate mechanical status index based on the joint normalization of working condition energy. The formula is:

[0020]

[0021] Among them, x Mj are parameters such as vibration energy entropy and frequency band energy ratio, w Mj is the initial weight, is the baseline energy value under the current working conditions.

[0022] c. Environmental feature subset: The environmental risk score is calculated using the accumulation-mutation dual model, using the formula:

[0023]

[0024] Among them, λ lis the cumulative coefficient of environmental parameters; α and β are weight coefficients;

[0025] Finally, the standardized feature vector X′=[X' E ,X' M ,X' Env ].

[0026] Preferably, the two-dimensional weight allocation includes static historical weight and dynamic weight, dynamic real-time contribution, the static historical weight is set based on historical fault data statistics, the static historical weight includes electrical subset weight Mechanical subset Environment subset The dynamic weight is the dynamic real-time contribution, which is adjusted according to the real-time abnormal frequency of the subset. The formula is:

[0027]

[0028] The dynamic weight is updated every 5 minutes, and the data anomaly detection delay does not exceed 10 seconds.

[0029] Preferably, the calculation formula of the health index is:

[0030] HI(t)=w E (t)·X' E (t)+w M (t)·X' M (t)+w Env (t)·X' Env (t)

[0031] Preferably, the dynamically adjusted threshold in S5 is dynamically set based on the mean μ and standard deviation σ of the health index within a 72-hour sliding window, and the dynamically adjusted threshold is marked as μ+1.5σ for the health status threshold and μ-2σ for the abnormal threshold.

[0032] Preferably, the generation of the differentiated inspection strategy includes: formulating a differentiated inspection strategy based on the equipment health grading results, and the substation adaptive inspection method also includes: combining real-time monitoring data, analyzing the historical change trend of the equipment health index through long- and short-time neural networks, predicting the remaining life of key components and generating pre-maintenance recommendations.

[0033] Preferably, the adjustment of the weight coefficients α and β is divided into offline optimization and online adjustment:

[0034] The offline optimization is based on historical data and uses the NSGA-II algorithm to generate a Pareto optimal solution set and determine the initial weights;

[0035] The online adjustment combines the real-time health distribution and the environmental risk level, dynamically adjusts the α and β weights through fuzzy logic, and gives priority to covering high-risk equipment.

[0036] Preferably, the data cleaning stage includes: using a gradient mutation detection method to identify short-term data peaks.

[0037] Preferably, the long-short time neural network model includes an updating rule.

[0038] The present invention provides a substation adaptive inspection method based on dynamic equipment health assessment. It has the following beneficial effects:

[0039] This adaptive substation inspection method, based on dynamic equipment health assessment, uses a multi-source sensor network to collect real-time electrical, mechanical, and environmental parameters of equipment. Combined with dynamic normalization and real-time anomaly detection, it significantly improves the real-time and accuracy of equipment health assessment. Using an LSTM neural network to analyze trends in equipment health indices, it predicts the remaining life of key components and dynamically adjusts inspection strategies based on equipment health and environmental risks, enabling differentiated inspections and optimized resource allocation. This approach effectively improves inspection efficiency and reduces the risk of failures.

[0040] This method combines static historical data with real-time dynamic weights, performs offline optimization using the NSGA-II algorithm, and dynamically adjusts weights using fuzzy logic, prioritizing coverage for high-risk equipment. A closed-loop feedback mechanism ensures the comparison of actual maintenance results with predicted ones, providing reliable data support for continuous model optimization. This allows equipment inspection strategies to be continuously refined and adapted to different operating environments, thereby improving overall equipment operation and maintenance management and fault prevention capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a system architecture diagram of the present invention;

[0042] Figure 2 Schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] like Figure 1-2As shown, an embodiment of the present invention provides a substation adaptive inspection method based on dynamic evaluation of equipment health, which specifically includes the following steps:

[0045] S1. Comprehensive monitoring of substation equipment status is achieved by building a multi-dimensional perception network. The perception layer consists of a distributed sensor array, covering three functional modules: an electrical characteristics monitoring unit, a mechanical status perception unit, and an environmental parameter acquisition unit. The electrical characteristics monitoring unit integrates a partial discharge signal capture device and a dielectric loss angle detection module to provide early warning of electrical defects such as insulation degradation and poor contact. The mechanical status perception unit uses a three-axis accelerometer combined with an acoustic emission detection device to accurately capture mechanical anomalies such as stuck circuit breaker operating mechanisms and transformer winding deformation. The environmental parameter acquisition unit includes a temperature and humidity composite sensor and a gas concentration monitor, forming a foundation for multi-parameter coupled analysis of the equipment's operating microenvironment.

[0046] The data communication network utilizes a layered, heterogeneous architecture, implementing differentiated transmission strategies tailored to the real-time requirements of different monitoring parameters. High-priority alarm signals are transmitted with millisecond-level latency via 5G network slicing technology. Routine monitoring data is periodically reported using the LoRa wide-area IoT network. A fiber-optic ring network is deployed in key equipment areas to ensure lossless transmission of high-bandwidth data such as vibration waveforms and partial discharge patterns. The communication protocol supports MQTT / CoAP dual-mode adaptive switching, ensuring data integrity and transmission reliability in complex electromagnetic environments.

[0047] S2. After data cleaning and time synchronization of electrical parameters, mechanical vibration parameters, and environmental parameters, perform dynamic subset normalization. Data cleaning includes: using a sliding window dynamic 3σ criterion to eliminate outliers, with the window length matching the sensor sampling frequency (e.g., a 1-second window for 10kHz vibration data); performing periodic online sensor calibration, and correcting drift errors based on the equipment's steady-state operating baseline data; aligning multi-source sensor timestamps using the PTP protocol, and addressing sampling frequency differences through sliding window interpolation: downsampling high-frequency data using the window mean, and upsampling low-frequency data using linear interpolation, ensuring synchronous fusion of multi-source data under a unified time base. Subset dynamic normalization includes the following steps:

[0048] a. Electrical signature subset: The comprehensive health index is calculated using the dynamic range compression method. The formula is:

[0049]

[0050] Among them, x Ei is the i-th electrical parameter, R Ei is the parameter rating, It is the maximum value of the comprehensive electrical parameters in the sliding time window.

[0051] b. Mechanical feature subset: Generate mechanical status index based on the joint normalization of working condition energy. The formula is:

[0052]

[0053] Among them, x Mj are parameters such as vibration energy entropy and frequency band energy ratio, w Mj is the initial weight, is the baseline energy value under the current working conditions.

[0054] c. Environmental feature subset: The environmental risk score is calculated using the accumulation-mutation dual model, using the formula:

[0055]

[0056] Among them, λ l is the cumulative coefficient of environmental parameters; α and β are weight coefficients;

[0057] Finally, the standardized feature vector X′=[X' E ,X' M ,X' Env The adjustment of weight coefficients α and β is divided into offline optimization and online adjustment:

[0058] Offline optimization is based on historical data and uses the NSGA-II algorithm to generate the Pareto optimal solution set and determine the initial weights;

[0059] Online adjustment combines real-time health distribution with environmental risk levels, dynamically adjusting α and β weights through fuzzy logic to prioritize high-risk devices. The data cleaning phase includes: Gradient mutation detection is used to identify short-term data peaks. If the parameter change rate within 10 seconds exceeds three times the historical average, the detected peak data is marked as "suspected interference." If the same sensor triggers a peak alarm three times in a row within five minutes, the following processing is initiated:

[0060] Electrical parameter peak value: replaced by the median value of the sliding window, with a sliding window of 1 hour;

[0061] Mechanical vibration peak: freeze the current weight calculation and give priority to the backup sensor data;

[0062] Sudden environmental changes such as sudden rainstorms: Delay HI calculation for 10 minutes and wait for environmental parameters to return to a steady state before re-evaluating.

[0063] S3. The normalized eigenvector is further optimized through two-dimensional weight allocation to form a closed-loop feedback. The two-dimensional weight allocation is static historical weight and dynamic real-time contribution. The two-dimensional weight allocation includes static historical weight and dynamic weight, dynamic real-time contribution. The static historical weight is based on the statistical setting of historical fault data. The static historical weight includes the electrical subset weight. Mechanical subset =0.3, environment subset The dynamic weight is the dynamic real-time contribution, which is adjusted according to the real-time abnormal frequency of the subset. The formula is:

[0064]

[0065] The dynamic weight is updated every 5 minutes, and the data anomaly detection delay does not exceed 10 seconds, ensuring the real-time statistics of subset anomalies and weight allocation. If an anomaly is detected in a subset three times in a row, such as the vibration subset exceeding the threshold, the number of anomalies of the subset in the dynamic weight calculation will only be counted as 50% of the total number of anomalies. At the same time, the data of backup sensors such as redundant vibration probes are activated to replace the abnormal source to avoid the excessive impact of single point failures on the health index.

[0066] S4. Integrate the normalized subset indicators and dynamic weights to generate a comprehensive health index for the device. Then, classify the device status based on the health index value. The health index calculation formula is:

[0067] HI(t)=w E (t)·X' E (t)+w M (t)·X' M (t)+w Env (t)·X' Env (t)

[0068] Dynamic weight w E ,w M ,w Env According to the real-time risk allocation, the allocation follows the following rules: When the environmental subset detects a short-term mutation risk such as a sudden increase in salt spray concentration, the electrical subset weight w E Increase by 10% to strengthen insulation status monitoring; if the mechanical subset is continuously abnormal, its dynamic weight w M The system decays by 50% until the anomaly is resolved. The health index calculation integrates the synergistic effect of multi-source data. For example, the baseline value of mechanical vibration in a high-temperature environment is dynamically adjusted as the temperature rises, and the equipment status is divided according to the HI value:

[0069] HI ≥ 0.8: healthy, perform routine inspections;

[0070] 0.6≤HI<0.8: Sub-health, potential risks exist, increase inspection frequency and detection dimensions;

[0071] HI<0.6: Abnormal, immediately triggering fault diagnosis and generating a maintenance work order.

[0072] S5. Dynamically adjust thresholds based on the statistical mean and variance of the device health index within the time window. The dynamically adjusted thresholds are dynamically set based on the mean μ and standard deviation σ of the health index within a 72-hour sliding window. The dynamic adjustment thresholds are marked as μ + 1.5σ for the health threshold and μ - 2σ for the abnormal threshold. Threshold adjustments are calculated and take effect within 2 seconds of the window update, and the abnormality determination delay is controlled within 15 seconds. If the health index data within the window contains a peak marked as "suspected interference," this data point is not included in the calculation of the mean μ and standard deviation σ. An additional 10% buffer, μ + 1.5σ × 1.1, is added to the health threshold μ + 1.5σ to prevent misjudgments of health status due to transient fluctuations. The 72-hour sliding window uses a rolling update mechanism, updating window data every 5 minutes. Data within the window is arranged in chronological order, and marginal data, including the first and last 5%, is smoothed using linear interpolation to avoid sudden interference. When the HI standard deviation σ within the window increases significantly, such as σ>0.15, the abnormal threshold is automatically lowered to μ-1.5σ to improve risk sensitivity and grade the results according to the equipment health.

[0073] S6. Dynamically generate differentiated inspection strategies based on the health index and preset thresholds, and optimize inspection routes. This includes developing differentiated inspection strategies based on the equipment health grading results: reducing inspection frequency for high-health equipment and increasing detection parameters and prioritizing low- and medium-health equipment. The adaptive substation inspection method also includes: combining real-time monitoring data with a long- and short-term neural network to analyze historical trends in the equipment health index, predict the remaining life of key components, and generate proactive maintenance recommendations. The LSTM model is trained based on the equipment's historical health index (HI) sequence, environmental operating condition data, and fault labels, requiring a predicted 72-hour HI with a MAE of ≤0.05. The model performs incremental training quarterly, integrating the latest operating data to optimize weights. A closed-loop feedback mechanism is established to compare actual maintenance results with predicted results. If the difference between the actual maintenance results and the predicted results exceeds 10%, parameter retraining of the health assessment model is triggered. Model retraining for high-risk equipment with an HI < 0.6 is completed within 30 minutes, and for sub-healthy equipment with an HI ≤ 0.6 < 0.8, it is completed within 2 hours. The optimized model parameters are deployed to the online assessment system within 1 minute after training. If the error does not exceed 10%, the weight distribution rules and threshold setting logic of the health assessment model are dynamically optimized through the back-propagation algorithm to continuously improve the accuracy of the status assessment. After the maintenance work order is executed, the system automatically traces the peak data records within 72 hours before the fault occurs. If it is found that the interference data that has not been effectively suppressed leads to misjudgment, the peak detection threshold gradient change rate threshold is optimized from 3 times to 5 times, and updated to the data cleaning rule base of all related devices. The long and short-term neural network models include update rules; the rules include integrating the newly generated health index sequence every quarter, the sequence includes 100,000 data, performing 8 hours of incremental training on the edge server, and updating the model weights; when the main transformer burns out or the salt spray season begins, a special training process is triggered to optimize the relevant feature weights in a targeted manner; the optimized model is compressed to 30% of its original size through the pruning algorithm to ensure real-time operation in embedded devices;

[0074] S7. Use LSTM neural network to analyze the health index change trend, predict the remaining life of the equipment and generate pre-maintenance recommendations.

[0075] S8. Through the closed-loop feedback mechanism, the actual maintenance results are compared with the predicted results, and the health assessment model parameters and threshold settings are reversely optimized.

[0076] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A substation adaptive inspection method based on dynamic evaluation of equipment health, characterized by: The following steps are involved: S1. Real-time collection of electrical parameters, mechanical vibration parameters, and environmental parameters of substation equipment through a multi-source sensor network; S2. After data cleaning and time synchronization of the electrical parameters, mechanical vibration parameters and environmental parameters, dynamic normalization of the subset is performed; S3. The normalized feature vector is further optimized through two-dimensional weight allocation to form a closed-loop feedback loop. S4. Integrate the normalized subset indicators with the dynamic weights to generate a comprehensive health index for the device, and classify the device status based on the health index value; S5. Dynamically adjust the threshold based on the statistical mean and variance of the device health index within the time window; S6. Dynamically generate differentiated inspection strategies based on the health index and preset thresholds, and optimize the inspection path; S7. Analyze health index trends using an LSTM neural network to predict the remaining life of the equipment and generate pre-maintenance recommendations. S8. Through the closed-loop feedback mechanism, the actual maintenance results are compared with the predicted results, and the health assessment model parameters and threshold settings are reversely optimized.

2. The adaptive inspection method for substations based on dynamic equipment health evaluation according to claim 1 is characterized by: The two-dimensional weight distribution is static historical weight and dynamic real-time contribution.

3. The adaptive inspection method for substations based on dynamic equipment health evaluation according to claim 1 is characterized by: The data cleaning includes: using a sliding window dynamic 3σ criterion to eliminate outliers, the window length matches the sensor sampling frequency, and the subset dynamic normalization processing includes the following steps: a. Electrical signature subset: The comprehensive health index is calculated using the dynamic range compression method. The formula is: Among them, x Ei is the i-th electrical parameter, R Ei is the parameter rating, It is the maximum value of the comprehensive electrical parameters in the sliding time window. b. Mechanical feature subset: Generate mechanical status index based on the joint normalization of working condition energy. The formula is: Among them, x Mj are parameters such as vibration energy entropy and frequency band energy ratio, w Mj is the initial weight, is the baseline energy value under the current working conditions. c. Environmental feature subset: The environmental risk score is calculated using the accumulation-mutation dual model, using the formula: Among them, λ l is the cumulative coefficient of environmental parameters; α and β are weight coefficients; Finally, the standardized feature vector X′=[X′ E ,X' M ,X' Env ].

4. The adaptive inspection method for substations based on dynamic equipment health evaluation according to claim 1 is characterized by: The two-dimensional weight distribution includes static historical weight and dynamic weight, dynamic real-time contribution, the static historical weight is set based on historical fault data statistics, the static historical weight includes electrical subset weight Mechanical subset Environment subset The dynamic weight is the dynamic real-time contribution, which is adjusted according to the real-time abnormal frequency of the subset. The formula is: The dynamic weight is updated every 5 minutes, and the data anomaly detection delay does not exceed 10 seconds.

5. The adaptive inspection method for substations based on dynamic equipment health evaluation according to claim 1 is characterized by: The calculation formula of the health index is: HI(t)=w E (t)·X' E (t)+w M (t)·X' M (t)+w Env (t)·X' Env (t) Dynamic weight w E ,w M ,w Env Based on real-time risk allocation.

6. The adaptive inspection method for substations based on dynamic equipment health evaluation according to claim 1, characterized in that: The dynamic adjustment threshold in S5 is dynamically set based on the mean μ and standard deviation σ of the health index within a 72-hour sliding window. The dynamic adjustment threshold is marked as the health status threshold as μ+1.5σ and the abnormal threshold as μ-2σ.

7. The adaptive inspection method for substations based on dynamic equipment health evaluation according to claim 1 is characterized by: The generation of the differentiated inspection strategy includes: formulating a differentiated inspection strategy based on the equipment health grading results. The substation adaptive inspection method also includes: combining real-time monitoring data, analyzing the historical change trend of the equipment health index through long- and short-time neural networks, predicting the remaining life of key components and generating pre-maintenance recommendations.

8. The adaptive inspection method for substations based on dynamic equipment health evaluation according to claim 3 is characterized by: The adjustment of the weight coefficients α and β is divided into offline optimization and online adjustment: The offline optimization is based on historical data and uses the NSGA-II algorithm to generate a Pareto optimal solution set and determine the initial weights; The online adjustment combines the real-time health distribution and the environmental risk level, dynamically adjusts the α and β weights through fuzzy logic, and gives priority to covering high-risk equipment.

9. The adaptive substation inspection method based on dynamic equipment health evaluation according to claim 3 is characterized by: The data cleaning stage includes: using a gradient mutation detection method to identify short-term data peaks.

10. The adaptive inspection method for substations based on dynamic evaluation of equipment health according to claim 7, characterized in that: The long-short time neural network model includes an updating rule.

Images