A non-electric quantity fault discrimination method for low-voltage switch cabinet of thermal power plant

Abstract

This invention provides a method for identifying non-electrical faults in low-voltage switchgear of thermal power plants, belonging to the field of power equipment fault diagnosis technology. It can at least partially solve the problems of poor real-time performance, easy omission of early-stage hidden faults, and insufficient linkage between early warning and location-based response in existing technologies that rely on manual inspection, single-parameter monitoring, and fixed threshold alarms. The method includes: deploying sensors in the switchgear busbar compartment, circuit breaker compartment, cable compartment, and / or operating mechanism compartment to collect signals of temperature, relative humidity, partial discharge, vibration, and gas concentration; performing outlier removal, noise reduction, and standardization preprocessing on the data; extracting time-domain, frequency-domain, and trend features and constructing feature vectors; using a random forest model and expert rule base to identify and output fault types and levels; triggering graded early warnings according to levels and outputting fault location information, forming a closed-loop linkage. This method achieves multi-parameter fusion and robust identification, improves early identification and response efficiency, and is engineering-feasible.

Description

Technical Field

[0001] This invention belongs to the field of power equipment fault diagnosis technology, specifically relating to a method for identifying non-electrical quantity faults in low-voltage switchgear of thermal power plants. Background Technology

[0002] As a crucial foundational power source in the power system, the safe and economical operation of thermal power plants heavily relies on the stable power supply of their auxiliary power distribution systems. These systems are responsible for providing electricity to auxiliary equipment such as boilers, turbines, electrostatic precipitators, desulfurization and denitrification systems, feedwater systems, and forced draft / induced draft systems, as well as public and ancillary systems. They are critical infrastructure ensuring continuous unit operation, and the rationality of load allocation directly impacts the energy consumption level, power supply reliability, and operational economy of thermal power plants. With the deepening of peak shaving and flexibility upgrades in thermal power units, and the gradual integration of new energy equipment such as photovoltaic, energy storage, and charging systems, auxiliary loads exhibit diverse types, frequent fluctuations, and high degree of operational coupling. Auxiliary loads not only include a large number of inductive motor loads and reactive power compensation devices, but also involve motor starting impact loads and bidirectional power fluctuations caused by on-site photovoltaic and energy storage systems. Their load demand changes significantly with unit load variations, start-up and shutdown processes, and abnormal operating conditions, placing higher demands on the dynamic adjustment capabilities of the auxiliary power distribution system.

[0003] Existing load allocation methods for thermal power plant auxiliary power distribution systems mostly employ fixed allocation coefficients or experience-based static allocation, making it difficult to dynamically adjust based on real-time unit operating conditions, equipment operating status, and load fluctuations. This can easily lead to bus overload or excessive energy consumption under high load or deep peak shaving conditions. Furthermore, existing methods often focus solely on reducing line losses, failing to comprehensively consider equipment losses, power supply reliability constraints, and fluctuations in renewable energy output, thus lacking multi-objective collaborative optimization capabilities. In addition, existing technologies lack sufficient modeling of the correlation between auxiliary loads and boilers, turbines, and environmental protection systems, and rely heavily on manual intervention or offline calculations, resulting in slow response times and difficulty in meeting the real-time load reconfiguration requirements under load surges or fault scenarios. Therefore, we propose a non-electrical fault identification method for low-voltage switchgear in thermal power plants. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art, and provides a method for identifying non-electrical faults in low-voltage switchgear of thermal power plants.

[0005] This invention provides a method for identifying non-electrical quantity faults in low-voltage switchgear of thermal power plants, comprising the following steps: S1: Collect temperature, relative humidity, partial discharge signal, vibration signal, and gas concentration signal inside the low-voltage switchgear of the thermal power plant; S2: The collected raw data are sequentially subjected to outlier removal, noise reduction, and standardization to obtain a preprocessed data sequence for fusion analysis; S3: Extract time-domain features, frequency-domain features, and trend features based on the preprocessed data sequence, and construct a feature vector for fault identification; S4: Input the feature vector into the random forest model and combine it with the expert rule base for mixed discrimination, and output the fault type and the corresponding fault level; S5: Trigger a graded early warning based on the fault level, and output fault location information based on the deployment location of the sensor.

[0006] Further, step S1 involves acquiring signals via sensors, including: a temperature sensor, a humidity sensor, and a partial discharge ultrasonic sensor installed in the busbar compartment of the low-voltage switchgear in the thermal power plant; a vibration sensor and a gas concentration sensor installed in the circuit breaker compartment of the low-voltage switchgear in the thermal power plant; a temperature sensor and a humidity sensor installed in the cable compartment of the low-voltage switchgear in the thermal power plant; and a temperature sensor and a vibration sensor installed in the operating mechanism compartment of the low-voltage switchgear in the thermal power plant.

[0007] Specifically, in step S1, the sequence of the original data is defined as ,in, i =1,2,...,5 correspond to temperature, relative humidity, partial discharge amplitude, vibration acceleration, and gas concentration, respectively; t This refers to the time of data collection.

[0008] Specifically, in step S2, the outlier removal adopts the Grubbs criterion and includes the following steps: The mean and standard deviation of the original data for each type are calculated, the Grubbs statistic is constructed, and outliers are identified based on the critical value corresponding to the significance level. The original data identified as outliers are filled by linear interpolation to maintain the continuity of the data sequence.

[0009] Preferably, in step S2, the denoising process includes wavelet thresholding and moving average filtering, wherein the window length of the moving average filtering is in the range of 3 to 6, and the formula for the moving average filtering is: ,in, This is the filtered data. For the first in the window k Wavelet-denoised data at each time point.

[0010] Specifically, in step S2, the standardization process includes: standardizing the denoised original data based on the mean and standard deviation, and then mapping the standardized original data to a preset normalization interval.

[0011] Furthermore, in step S3, the multi-dimensional feature extraction includes the following steps: The mean and peak factor of the temperature, relative humidity, partial discharge signal, vibration signal and gas concentration signal are extracted as time-domain features respectively. Perform Fourier transform on the vibration signal and the partial discharge signal and extract the dominant frequency as a frequency domain feature; The rate of change is extracted as a trend feature from the preprocessed data sequence based on a sliding window.

[0012] Furthermore, in step S3, the extracted time-domain features, frequency-domain features, and trend features are screened through correlation analysis to remove redundant features, and a multi-dimensional feature vector for fault identification is constructed accordingly. The multi-dimensional feature vector includes at least: time-domain features and trend features of temperature; time-domain features and trend features of relative humidity; time-domain features, frequency-domain features, and trend features of partial discharge signals; time-domain features and frequency-domain features of vibration signals; and trend features of gas concentration signals.

[0013] Furthermore, in step S4, the training of the random forest model includes the following steps: Construct a labeled dataset containing normal samples and faulty samples, and divide it into training and test sets; Set the parameters of the random forest model; The parameters of the random forest model are optimized using the out-of-bag (OOB) error, and training is stopped when the OOB error is no greater than 3%.

[0014] Specifically, in steps S4 and S5, the hybrid discrimination and the hierarchical early warning include: The output of the random forest model and the output of the expert rule base are fused using a weighted voting method. When the output of the random forest model and the output of the expert rule base are inconsistent, a secondary verification is performed based on the rate of change of the trend features to determine the final judgment result. The fault level obtained from the final judgment is used to trigger a graded early warning.

[0015] The beneficial effects of this invention are as follows: This invention employs zoned temperature, humidity, partial discharge, vibration, and gas concentration sensors to achieve continuous monitoring of various non-electrical anomalies in switchgear, reducing the risk of missed detections caused by single-parameter monitoring. It utilizes outlier removal, combined noise reduction, and standardization to improve monitoring data quality and mitigate the impact of electromagnetic interference and environmental fluctuations on the judgment results. Furthermore, it extracts key features from the time, frequency, and trend domains and removes redundancy to construct feature vectors, taking into account parameter levels, impact characteristics, frequency characteristics, and trends, thereby enhancing the ability to identify progressive early-stage faults. Attached Figure Description

[0016] Figure 1This is a flowchart illustrating the steps of a non-electrical quantity fault identification method for low-voltage switchgear in thermal power plants, according to a specific embodiment of the present invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown in the figure, a method for identifying non-electrical quantity faults in low-voltage switchgear of thermal power plants provided by a specific embodiment of the present invention includes the following steps: S1: Collect temperature, relative humidity, partial discharge signal, vibration signal, and gas concentration signal inside the low-voltage switchgear of the thermal power plant; S2: The collected raw data are sequentially subjected to outlier removal, noise reduction, and standardization to obtain a preprocessed data sequence for fusion analysis; S3: Extract time-domain features, frequency-domain features, and trend features from the preprocessed data sequence, and construct feature vectors for fault identification; S4: Input the feature vector into the random forest model and combine it with the expert rule base for mixed discrimination, and output the fault type and the corresponding fault level; S5: Triggers graded early warning based on fault level and outputs fault location information based on sensor deployment location, realizing closed-loop linkage of judgment, early warning and processing.

[0019] Specifically, in step S1, the sensor needs to be adapted to the harsh environment of high temperature and high electromagnetic interference in the thermal power plant. The selection and parameter definition are as follows: Temperature (T): unit ℃, platinum resistance sensor is selected, measurement range -20℃ to 120℃, accuracy ±0.5℃, to meet the monitoring requirements of normal equipment operation (20℃ to 60℃) and fault heating (up to 100℃ or more); Relative humidity (RH): Unit %RH, using SHT30 digital sensor, measurement range 0~100%RH, accuracy ±2%RH, can accurately capture humidity changes when insulation is damp; Partial discharge amplitude (PD): unit mV. PD-200 ultrasonic sensor is used, with a detection frequency band of 20kHz~200kHz and a sensitivity of ≥1mV. It can identify early weak partial discharge signals. Vibration acceleration (a): unit g (g=9.8m / s², gravitational acceleration), using ADXL355 MEMS sensor, measurement range 0.1g-50g, frequency response 10Hz-1kHz, capable of capturing abnormal vibrations of mechanical parts; Gas concentration (C): Unit ppm (volume concentration unit, 1ppm=10 -6The IR-600 infrared sensor is selected, with a detection limit of ≤10ppm and an accuracy of ±5%FS (FS is full scale), which can detect SF6 gas leaks in a timely manner.

[0020] Furthermore, in step S2, the outliers are mainly caused by transient sensor malfunctions and electromagnetic pulse interference, which need to be accurately identified and appropriately filled. Calculate statistical parameters: for each type of original data sequence Calculate its mean and standard deviation , where the mean Reflects the average level and standard deviation of the data. Reflects the degree of dispersion of the data; Constructing the Grubbs statistic: used to quantify the deviation of individual data points from the overall data, the formula is as follows: (1) In the formula, For the first t Grubbs statistic for time-matter data; Outlier identification: Select significance level (Industry-standard confidence level), obtained from the Grubbs critical value table. ( N (for the sample size), if Then determine This is an outlier; Outlier imputation: Linear interpolation is used to imput outliers to avoid missing data affecting subsequent analysis. The formula is as follows: (2) In the formula, The data for the locations of outliers after imputation. , This method uses normal data at times adjacent to outliers to ensure the continuity and smoothness of the data sequence.

[0021] Furthermore, in step S3, two types of core temporal features are extracted for the five types of standardized data sequences: mean ( ): Reflects the average level of parameters and can indicate the overall operating status of the equipment. For example, a continuous increase in the average temperature may indicate an overheating fault in the equipment. Peak factor ( ): Reflects the impact characteristics of the parameter and is sensitive to transient signals such as vibration and partial discharge. The formula is as follows: (3) In the formula, The peak value (maximum value) of the parameter sequence is given by the denominator, which is the effective value (RMS) of the parameter sequence. The larger the peak value, the stronger the signal impact. For example, loose mechanical parts can cause the peak value of the vibration signal to increase significantly.

[0022] Based on the above basic implementation method, step S1 involves acquiring signals through sensors, including: a temperature sensor, a humidity sensor, and a partial discharge ultrasonic sensor installed in the busbar compartment of the low-voltage switchgear in the thermal power plant; a vibration sensor and a gas concentration sensor installed in the circuit breaker compartment of the low-voltage switchgear in the thermal power plant; a temperature sensor and a humidity sensor installed in the cable compartment of the low-voltage switchgear in the thermal power plant; and a temperature sensor and a vibration sensor installed in the operating mechanism compartment of the low-voltage switchgear in the thermal power plant. In step S1, the sequence of the raw data is defined as... ,in, i =1,2,...,5 correspond to temperature, relative humidity, partial discharge amplitude, vibration acceleration, and gas concentration, respectively; t At the time of data acquisition, the sampling frequency of various raw data is 1Hz; the sensor communicates with the data processing unit via RS485 bus, and the data processing unit uploads the acquired data to the background monitoring system via industrial Ethernet.

[0023] Specifically, based on the structural characteristics of low-voltage switchgear and areas prone to failure, sensors are deployed in functional areas to ensure comprehensive monitoring: Busbar compartment: As the core of current transmission, it is prone to problems such as busbar joint overheating, partial discharge of insulation, and insulation moisture. Three types of sensors are deployed: ① Temperature sensor (monitors joint overheating); ② Humidity sensor (monitors the risk of insulation moisture); ③ Partial discharge ultrasonic sensor (monitors partial discharge phenomenon in insulation). Circuit breaker compartment: Circuit breaker operating mechanisms are prone to wear and loosening, and gas-insulated switchgear poses a risk of SF6 leakage. Two types of sensors are deployed: ① Vibration sensors (to monitor the mechanical condition of the operating mechanism); ② SF6 gas concentration sensors (to monitor for leakage of insulating gas). Cable compartment: Cable joints are prone to overheating and insulation is susceptible to moisture. Two types of sensors are deployed: ① Temperature sensor (to monitor cable joint temperature); ② Humidity sensor (to monitor cable insulation moisture). Operating mechanism room: The operating mechanism coil is prone to overheating and the mechanical parts are prone to wear. Two types of sensors are deployed: ① Temperature sensor (to monitor coil temperature); ② Vibration sensor (to monitor the vibration status of mechanical parts).

[0024] In another specific embodiment, in step S2, outlier removal adopts the Grubbs criterion, including the following steps: The mean and standard deviation of various raw data are calculated, the Grubbs statistic is constructed, and outliers are identified based on the critical values ​​corresponding to the significance level. Linear interpolation is used to impute the outliers to maintain the continuity of the data sequence. In step S2, denoising includes wavelet thresholding and moving average filtering. Wavelet thresholding uses the db4 wavelet as the wavelet basis and performs a 3-level decomposition. The window length for moving average filtering ranges from 3 to 6, and the formula for moving average filtering is: (4) in, This is the filtered data. For the first in the window k The wavelet denoising data at each time point; in step S2, the standardization process includes: standardizing the denoised original data based on the mean and standard deviation, and then mapping the standardized original data to a preset normalization interval for use in multi-parameter fusion analysis.

[0025] Specifically, the wavelet basis is selected as db4 wavelet (which balances denoising effect and computational efficiency and is suitable for power equipment monitoring data), and the number of decomposition layers is set to 3 (which can effectively separate signal and noise). Adaptive threshold calculation: threshold The formula is adaptively adjusted based on the data standard deviation and sample size, as follows: (5) In the formula, The standard deviation of the original data sequence. N The number of samples; Wavelet coefficient processing: The decomposed wavelet coefficients are processed with soft thresholding—coefficients greater than the threshold are retained and corrected, while coefficients less than the threshold are set to zero. The signal is then reconstructed through inverse wavelet transform to achieve high-frequency noise suppression.

[0026] Furthermore, the five types of non-electrical parameters have different dimensions (such as temperature in °C and humidity in %RH), making direct fusion analysis impossible. Standardization is needed to eliminate these dimensional differences, as shown in the following formula: (6) In the formula, For standardized data, This represents the mean of the filtered data. This represents the standard deviation of the filtered data. The standardized data is mapped to the [0,1] interval to ensure that all parameters have equal weight in subsequent feature extraction and model training.

[0027] In one specific implementation, step S3, multi-dimensional feature extraction includes the following steps: The mean and peak factor of temperature, relative humidity, partial discharge signal, vibration signal, and gas concentration signal are extracted as time-domain features, respectively; Fourier transform is performed on vibration signal and partial discharge signal and the main frequency is extracted as frequency-domain features; the rate of change is extracted as trend feature based on sliding window for preprocessed data sequence. In step S3, the extracted time-domain features, frequency-domain features, and trend features are screened through correlation analysis to remove redundant features, and a multi-dimensional feature vector for fault identification is constructed accordingly. The multi-dimensional feature vector includes at least: time-domain features and trend features of temperature; time-domain features and trend features of relative humidity; time-domain features, frequency-domain features, and trend features of partial discharge signals; time-domain features and frequency-domain features of vibration signals; and trend features of gas concentration signals.

[0028] In this embodiment, for vibration signals and partial discharge ultrasonic signals (both types of signals have high frequency characteristics and are highly correlated with faults), a fast Fourier transform (FFT) is used to convert them to the frequency domain, and one type of core features is extracted: main frequency ( ( ): This reflects the frequency where signal energy is most concentrated, and can effectively identify the fault condition of mechanical parts. The formula is as follows: (7) In the formula, The frequency domain amplitude after FFT transformation. The effective frequency range of the sensor. This refers to the frequency variable that corresponds to the maximum value.

[0029] Specifically, for the five types of standardized data sequences, one core trend feature is extracted: Sliding window rate of change ( r ( ): Reflects the average change intensity of parameters over a period of time, and can capture the development trend of faults. The formula is as follows: (8) In the formula, 30 is the size of the sliding window (in seconds, balancing trend capture and real-time performance). For the first t The first-order difference slope at time (reflecting the instantaneous rate of change). This represents the mean slope within the window. A larger rate of change (r) indicates more drastic parameter changes and a faster fault development rate. Redundant features were removed through correlation analysis (e.g., if the variance and mean are highly correlated, only the mean was retained), ultimately constructing a 10-dimensional feature vector. ,in: f_3-f_4 represent the time-domain trend features of temperature, f_5-f_7 represent the time-domain and frequency-domain trend features of partial discharge, f_8-f_9 represent the time-domain and frequency-domain features of vibration, and f_{10} represents the trend features of gas concentration. The feature vectors are both comprehensive and concise, which can effectively improve the discrimination efficiency of subsequent models.

[0030] In another specific embodiment, in step S4, the random forest model training includes the following steps: A labeled dataset containing normal and faulty samples is constructed and divided into training and test sets. The parameters of the random forest model are set as follows: the number of decision trees is 50, the maximum depth is 10, and the minimum number of samples for node splitting is 5. The parameters of the random forest model are optimized using the out-of-bag (OOB) error, and training is stopped when the OOB error is no greater than 3% to obtain the random forest model. In steps S4 and S5, the hybrid discrimination and hierarchical early warning includes: A weighted voting method is used to integrate the output of the random forest model and the expert rule base, with the random forest model having a weight of 0.7 and the expert rule base having a weight of 0.3. When the output of the random forest model and the expert rule base are inconsistent, a secondary verification is performed based on the rate of change of the trend feature to determine the final judgment result. The fault level obtained from the final judgment is used to trigger a graded early warning. For the third-level fault in the graded early warning, on-site audible and visual alarms, information push and background pop-up alarms are triggered, and the tripping circuit of the low-voltage switchgear of the thermal power plant is linked to cut off the power supply to the fault area. At the same time, the fault location information of the low-voltage switchgear of the thermal power plant is output.

[0031] Furthermore, the steps for training a random forest model are as follows: Sample dataset construction: Historical fault data: Five years of operation data of 10 low-voltage switchgear in a thermal power plant were collected, including 200 sets of actual fault samples (35 sets of abnormal temperature, 40 sets of excessive humidity, 30 sets of partial discharge, 50 sets of abnormal vibration, and 25 sets of gas leakage) and 200 sets of normal state samples. Simulated fault data: A low-voltage switchgear fault simulation platform was built in the laboratory to simulate different fault levels (minor, moderate, severe) and environmental interference scenarios (electromagnetic interference, temperature fluctuation), generating 8,000 sets of simulated fault samples and 1,300 sets of normal state samples. Sample label coding: Fault type coding (0-normal, 1-abnormal temperature, 2-excessive humidity, 3-partial discharge, 4-abnormal vibration, 5-gas leak); Fault level coding (1-general fault, 2-moderate fault, 3-serious fault). Dataset partitioning: The dataset is divided into a training set (70%, used for model training) and a test set (30%, used for model validation) in a 7:3 ratio, with a total of ≥10,000 samples to ensure the model's generalization ability; Model parameter settings and optimization: The number of decision trees is set to 50 (to balance model accuracy and computational efficiency; too few decision trees will lead to underfitting, while too many will increase computational cost); the maximum depth of each decision tree is set to 10 (to avoid overfitting and ensure the model can capture the core patterns of the data); the minimum number of samples for node splitting is set to 5 (to ensure the statistical significance of nodes); model optimization: out-of-bag (OOB) data is used to validate the optimized parameters. The OOB error is calculated as "number of OOB errors / total number of samples". Training is stopped when the OOB error is ≤3%, resulting in the final random forest model. The model's input feature vector... Afterwards, the fault diagnosis result can be output. (Coded fault type + level); The steps for building an expert rule base are as follows: Based on the "Operating Procedures for Low-Voltage Power Distribution Equipment in Thermal Power Plants", equipment factory technical specifications, and non-electrical fault mechanisms, an expert rule base was established. The core rules are shown in the table below: ; In step S4, the steps for mixed discrimination are as follows: Weighted voting fusion: Considering that data-driven approaches better align with the actual operating patterns of equipment, and that mechanistic analysis can ensure the reliability of judgments, the following fusion weights are set: Random Forest model weights. Expert rule base weight Fusion results Calculate using the following formula (after converting the encoded result to a numerical value, perform a weighted sum, and then decode it into fault type and level): (9) Two-factor verification mechanism: when At the same time, combined with trend characteristics r (Sliding window rate of change) is used for secondary verification to avoid the influence of bias from a single model on the results: like r A value ≥0.1 (a preset rate of change threshold, determined through extensive sample validation) indicates drastic parameter changes and a high risk of failure, based on the results from the expert rule base. Subject to; like r A value <0.1 indicates that the parameter changes gradually, and the data patterns are more reliable, as shown by the results of random forest. As the standard.

[0033] Furthermore, in step S5, a differentiated early warning mechanism is triggered based on the fault level output by the hybrid discrimination model. Simultaneously, the fault is accurately located based on the sensor deployment location, realizing a closed-loop linkage of "discrimination-early warning-processing". The specific operation is as follows: Based on the safety operation requirements of thermal power plants, the fault levels are divided into three levels, and the risk level of each level of fault is clearly defined: Level 1 (General Fault): The parameters deviate slightly from the threshold, with no immediate safety risk. For example, if the humidity is 86%RH for 10 minutes, it may only affect the insulation performance of the equipment, and there is no risk of equipment damage in the short term. Level 2 (Moderate Fault): Parameters deviate significantly from the threshold, posing potential risks. For example, a partial discharge amplitude of 550mV, if left untreated for a long time, will lead to accelerated insulation aging and may cause a short circuit fault. Level 3 (Severe Fault): Parameters are severely out of control, which may cause equipment damage or power outages, such as temperature 90℃ or SF6 gas concentration 60ppm. Immediate action is required. Different types of early warning signals are triggered based on the fault level to ensure rapid response by maintenance personnel, while also outputting fault location information (based on sensor deployment location, such as "busbar compartment temperature sensor T1 detected abnormal temperature"): Level 1 fault: Only triggers "background pop-up warning" - the fault parameter value, fault location and occurrence time are displayed on the monitoring system in the central control room of the thermal power plant, reminding the operation and maintenance personnel to check and handle it during the next inspection (within 1 week), without the need for shutdown; Level 2 fault: Triggering "audio-visual warning + background pop-up warning + information push" - The audio-visual alarm of the on-site switch cabinet is activated (red light flashing + intermittent buzzing), the background monitoring system pops up to display detailed fault information, and at the same time pushes a warning text message to the mobile phone of the operation and maintenance person in charge. It is recommended to arrange for handling within 24 hours. Before handling, parameter changes need to be closely monitored. Level 3 fault: Triggers "audio-visual warning + SMS push + background pop-up warning + emergency shutdown linkage" - The on-site audio-visual alarm continuously alarms (red light stays on + continuous buzzing), pushes emergency SMS to the operation and maintenance manager and the central control room duty officer, the background system pops up an emergency fault window, and at the same time, it links the low-voltage switch cabinet trip circuit to cut off the power supply to the fault area to prevent the fault from spreading to other areas. The background monitoring system records the fault diagnosis results, warning time, personnel involved, handling measures, and parameter changes after handling, forming a fault handling ledger to provide data support for subsequent equipment maintenance and fault pattern analysis.

[0034] In another specific embodiment, a sensor and data processing system is deployed on the low-voltage switchgear (model GGD-2) of a 300MW unit in a thermal power plant. The specific steps are as follows: Sensor deployment: According to the above planning method, PT100 temperature sensors, SHT30 humidity sensors, PD-200 partial discharge sensors, ADXL355 vibration sensors, and IR-600SF6 gas concentration sensors are deployed in the busbar room, circuit breaker room, cable room, and operating mechanism room, respectively. A total of 8 sensors are deployed in each switch cabinet (the sensor models are the same in different areas, and they are numbered T1-T3, RH1-RH2, PD1, a1-a2, and C1 according to their location). The sensor is connected to the STM32F407 data processing unit via an RS485 bus. The data processing unit uploads data to the LabVIEW background monitoring system in the central control room via industrial Ethernet (PROFINET protocol), with a sampling frequency of 1Hz. The system operated continuously for 30 days, collecting a total of 2,592,000 sets of raw data (30 days × 24 hours × 3,600 seconds × 1 set / second). The data integrity rate reached 99.8%, with no sensor disconnections or data loss.

[0035] Furthermore, taking a set of monitoring data simulating a "partial discharge fault" (acquired by partial discharge sensor PD1, i=3) as an example, the data preprocessing and feature extraction process is demonstrated: Raw data: {480mV, 520mV, 490mV, 1500mV, 510mV, ...} (1000 sets in total, the 4th set is an outlier); Outlier removal: Calculate the mean =502mV, standard deviation =35mV; Grubbs statistic for the fourth set of data By looking up the critical value table, we can obtain... Since 28.51 > 3.29, it is determined to be an outlier; Linear interpolation fill: After filling, the data sequence is smooth and continuous; Noise suppression: Wavelet thresholding After performing soft thresholding on the wavelet coefficients, an inverse transform is performed to remove high-frequency electromagnetic interference. 5-point moving average filtering: The moving average value is calculated on the denoised data to obtain the filtered data {482mV, 518mV, 495mV, 500mV, 508mV, ...}, and low-frequency environmental noise is effectively suppressed; Standardization process: Filtered mean =501mV, standard deviation =32mV, standardized data Mapped to the interval [0,1]; Feature extraction: Time-domain characteristics: mean =0.48, peak factor =2.1; Frequency domain characteristics: dominant frequency =100kHz (consistent with the characteristic frequency band of partial discharge signals); Trend characteristics: Sliding window rate of change r =0.08; The eigenvector corresponding to this partial discharge-related feature f 5- f The 7 dimensions, combined with features from other parameters, form a multidimensional feature vector. .

[0036] Furthermore, multidimensional feature vectors Input the hybrid discriminant model for fault identification and early warning: Random Forest model output: (Partial discharge - moderate fault), the numerical code is 3×10+2=32; Expert rule base output: Partial discharge amplitude PD = 500mV ≥ 500mV, trigger rule R3. =[3,2], numerical code 32; Hybridization results: =0.7×32+0.3×32=32, decoded as [3,2] (partial discharge - moderate fault); Early warning linkage: Triggering a Level 2 fault warning - the on-site audible and visual alarm is activated (red light flashing + intermittent buzzer), and the background monitoring system pops up a window displaying "PD1 sensor in busbar room detected partial discharge abnormality, amplitude 500mV, fault level: moderate", and at the same time pushes a warning text message to the mobile phone of the operation and maintenance manager, suggesting that it be handled within 24 hours; Outcome: Maintenance personnel arrived at the site within 12 hours, inspected the insulation components of the busbar compartment, and found that the insulation sheath of the busbar joint was damaged. After timely replacement, the partial discharge signal disappeared and the equipment returned to normal operation. Historical fault data (200 fault samples and 200 normal samples) from 10 low-voltage switchgear units in a thermal power plant and laboratory simulation data (8000 fault samples and 1300 normal samples) were selected to compare the performance of the method of this invention with the traditional threshold method and the single random forest method. The verification results are shown in the table below: ;

[0037] The verification results show that the method of the present invention is significantly better than the existing methods in terms of discrimination accuracy, false judgment rate, response time and early fault identification rate. It also has strong engineering applicability and can effectively meet the non-electrical quantity fault discrimination requirements of low-voltage switchgear in thermal power plants.

[0038] To aid in a better understanding of the present invention, a more comprehensive and specific embodiment is described, in which the present invention provides a method for identifying non-electrical quantity faults in low-voltage switchgear of thermal power plants, comprising the following steps: S1: Collect temperature, relative humidity, partial discharge signal, vibration signal, and gas concentration signal inside the low-voltage switchgear of the thermal power plant; S2: The collected raw data are sequentially subjected to outlier removal, noise reduction, and standardization to obtain a preprocessed data sequence for fusion analysis; S3: Extract time-domain features, frequency-domain features, and trend features from the preprocessed data sequence, and construct feature vectors for fault identification; S4: Input the feature vector into the random forest model and combine it with the expert rule base for mixed discrimination, and output the fault type and the corresponding fault level; S5: Triggers graded early warning based on fault level and outputs fault location information based on sensor deployment location.

[0039] In this embodiment, step S1 involves acquiring signals using sensors, including: a temperature sensor, a humidity sensor, and a partial discharge ultrasonic sensor installed in the busbar compartment of the low-voltage switchgear in a thermal power plant; a vibration sensor and a gas concentration sensor installed in the circuit breaker compartment of the low-voltage switchgear in a thermal power plant; a temperature sensor and a humidity sensor installed in the cable compartment of the low-voltage switchgear in a thermal power plant; and a temperature sensor and a vibration sensor installed in the operating mechanism compartment of the low-voltage switchgear in a thermal power plant. In step S1, the sequence of the raw data is defined as follows: ,in, i =1,2,...,5 correspond to temperature, relative humidity, partial discharge amplitude, vibration acceleration, and gas concentration, respectively; t The data collection time is specified in step S2. Outlier removal employs the Grubbs criterion, including the following steps: calculating the mean and standard deviation of each type of raw data; constructing the Grubbs statistic and determining outliers based on the critical value corresponding to the significance level; and imputing outliers using linear interpolation to maintain data sequence continuity. In step S2, denoising includes wavelet thresholding and moving average filtering. Wavelet thresholding uses the db4 wavelet as the wavelet basis and performs a 3-level decomposition. The window length for moving average filtering ranges from 3 to 6, and the formula for moving average filtering is: ,in, This is the filtered data. For the first in the window kThe wavelet denoising data at each time point; in step S2, the standardization process includes: standardizing the denoised original data based on the mean and standard deviation, and then mapping the standardized original data to a preset normalization interval for use in multi-parameter fusion analysis.

[0040] Specifically, in step S3, the multi-dimensional feature extraction includes the following steps: The mean and peak factor of temperature, relative humidity, partial discharge signal, vibration signal, and gas concentration signal are extracted as time-domain features, respectively; Fourier transform is performed on vibration signal and partial discharge signal and the main frequency is extracted as frequency-domain features; the rate of change is extracted as trend feature based on sliding window for preprocessed data sequence. In step S3, the extracted time-domain features, frequency-domain features, and trend features are screened through correlation analysis to remove redundant features, and a multi-dimensional feature vector for fault identification is constructed accordingly. The multi-dimensional feature vector includes at least: time-domain features and trend features of temperature; time-domain features and trend features of relative humidity; time-domain features, frequency-domain features, and trend features of partial discharge signals; time-domain features and frequency-domain features of vibration signals; and trend features of gas concentration signals. In step S4, and in steps S4 and S5, the hybrid discrimination and hierarchical early warning includes: A labeled dataset containing normal and faulty samples is constructed and divided into training and test sets. The parameters of the random forest model are set as follows: number of decision trees: 50; maximum depth: 10; minimum number of samples for node splits: 5. The parameters of the random forest model are optimized using out-of-bag (OOB) error, and training is stopped when the OOB error is no greater than 3%. When training each decision tree, a training subset of that decision tree is formed by sampling with replacement from the training sample set using the bootstrap method. Samples not included in this training subset are defined as out-of-bag samples of that decision tree. For any sample in the training sample set, the outputs of all decision trees using that sample as an out-of-bag sample are collected, and the out-of-bag prediction result for that sample is obtained through voting (classification task) or averaging (regression task). This out-of-bag prediction result is compared with the true label of the sample to determine whether the out-of-bag prediction is correct. The ratio of the number of samples with incorrect out-of-bag predictions in the training sample set to the total number of training samples is used to obtain the out-of-bag error, expressed as: OOB error = Number of out-of-bag prediction errors / Total number of samples.

[0041] Based on the OOB error, parameters such as the number of decision trees, maximum depth, and minimum number of split samples in the random forest model can be optimized, and the final model parameters can be determined when the OOB error reaches a preset threshold or no longer decreases significantly with the increase of the number of decision trees. The steps of hybrid discrimination and graded early warning include: A weighted voting method is used to fuse the output of the random forest model and the output of the expert rule base, with the random forest model having a weight of 0.7 and the expert rule base having a weight of 0.3. When the output of the random forest model is inconsistent with the output of the expert rule base, a secondary verification is performed based on the rate of change of the trend feature to determine the final judgment result, and a graded early warning is triggered based on the fault level obtained from the final judgment.

[0042] In summary, the embodiments disclosed herein have at least the following technical effects: By arranging temperature, humidity, partial discharge, vibration and gas concentration sensors in the busbar room, circuit breaker room, cable room and operating mechanism room, multi-dimensional continuous acquisition of non-electrical states of low-voltage switchgear can be achieved. It can simultaneously cover multiple types of hidden faults such as overheating, moisture, partial discharge, abnormal vibration of mechanism and gas leakage, reducing the risk of missed detection due to monitoring of a single parameter. The Grubbs criterion is used for outlier identification and interpolation repair. Wavelet threshold denoising and moving average filtering are combined to suppress electromagnetic interference and environmental fluctuations. Then, standardization / normalization is used to eliminate the influence of different dimensions, providing a stable and fusionable data foundation for subsequent feature extraction and model discrimination, thereby improving the reliability and availability of online monitoring. Features such as mean, peak factor, main frequency and sliding window change rate are extracted from the time domain, frequency domain and trend domain. Redundant features are eliminated through correlation analysis to construct feature vectors, which can characterize parameter levels and fluctuation impact characteristics, as well as reflect fault development trends and frequency characteristics. This is helpful for identifying early minor and progressive non-electrical faults. Random forest models can learn complex nonlinear correlations in historical and simulated data; expert rule bases introduce mechanistic and procedural threshold constraints. By weighted fusion and combining trend change rate for secondary verification, these two methods can improve the robustness of fault type and severity discrimination under complex operating conditions and environmental disturbances, and reduce the probability of false alarms and false negatives caused by a single model or single threshold method. Differentiated handling strategies are triggered based on the fault level, such as background pop-ups, on-site audio and visual effects, information push, and tripping linkage. Based on the sensor deployment location, specific fault location information is output, which can shorten the link time from anomaly detection to location and handling, improve operation and maintenance response efficiency, and prevent the fault from escalating. The acquisition frequency is 1Hz and the algorithm mainly consists of statistical testing, lightweight filtering, feature extraction and random forest inference, with relatively controllable computational complexity. The sensor is connected to the data processing unit via RS485 and accessed to the back-end system via industrial Ethernet, which is convenient for installation and deployment on existing switch cabinets with small modification requirements and good engineering promotion and operation and maintenance operability.

[0043] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for identifying non-electrical quantity faults in low-voltage switchgear of thermal power plants, characterized in that, Includes the following steps: S1: Collect temperature, relative humidity, partial discharge signal, vibration signal, and gas concentration signal inside the low-voltage switchgear of the thermal power plant; S2: The collected raw data are sequentially subjected to outlier removal, noise reduction, and standardization to obtain a preprocessed data sequence for fusion analysis; S3: Extract time-domain features, frequency-domain features, and trend features based on the preprocessed data sequence, and construct a feature vector for fault identification; S4: Input the feature vector into the random forest model and combine it with the expert rule base for mixed discrimination, and output the fault type and the corresponding fault level; S5: Trigger a graded early warning based on the fault level, and output fault location information based on the deployment location of the sensor.

2. The method for identifying non-electrical quantity faults in low-voltage switchgear of thermal power plants according to claim 1, characterized in that, Step S1 involves acquiring signals through sensors, including: a temperature sensor, a humidity sensor, and a partial discharge ultrasonic sensor installed in the busbar compartment of the low-voltage switchgear in the thermal power plant; a vibration sensor and a gas concentration sensor installed in the circuit breaker compartment of the low-voltage switchgear in the thermal power plant; a temperature sensor and a humidity sensor installed in the cable compartment of the low-voltage switchgear in the thermal power plant; and a temperature sensor and a vibration sensor installed in the operating mechanism compartment of the low-voltage switchgear in the thermal power plant.

3. The method for identifying non-electrical quantity faults in low-voltage switchgear of thermal power plants according to claim 1, characterized in that, In step S1, the sequence of the original data is defined as ,in, i =1,2,...,5 correspond to temperature, relative humidity, partial discharge amplitude, vibration acceleration, and gas concentration, respectively; t This refers to the time of data collection.

4. The method for identifying non-electrical quantity faults in low-voltage switchgear of thermal power plants according to claim 1, characterized in that, In step S2, the outlier removal includes the following steps: The mean and standard deviation of the original data for each type are calculated, the Grubbs statistic is constructed, and outliers are identified based on the critical value corresponding to the significance level. The original data identified as outliers are filled by linear interpolation to maintain the sequence continuity of the original data.

5. The method for identifying non-electrical quantity faults in low-voltage switchgear of thermal power plants according to claim 1, characterized in that, In step S2, the denoising process includes wavelet thresholding and moving average filtering. The window length of the moving average filtering is in the range of 3 to 6, and the formula for the moving average filtering is: ,in, This is the filtered data. For the first in the window k Wavelet-denoised data at each time point.

6. The method for identifying non-electrical quantity faults in low-voltage switchgear of thermal power plants according to claim 1, characterized in that, In step S2, the standardization process includes: standardizing the denoised original data based on the mean and standard deviation, and then mapping the standardized original data to a preset normalization interval.

7. The method for identifying non-electrical quantity faults in low-voltage switchgear of thermal power plants according to claim 1, characterized in that, In step S3, the multi-dimensional feature extraction includes the following steps: The mean and peak factor of the temperature, relative humidity, partial discharge signal, vibration signal, and gas concentration signal are extracted as time-domain features, respectively; Fourier transform is performed on the vibration signal and the partial discharge signal and the main frequency is extracted as frequency-domain features; and the rate of change of the preprocessed data sequence is extracted as trend features based on a sliding window.

8. The method for identifying non-electrical quantity faults in low-voltage switchgear of thermal power plants according to claim 7, characterized in that, In step S3, the extracted time-domain features, frequency-domain features, and trend features are screened through correlation analysis to remove redundant features, and a multi-dimensional feature vector for fault identification is constructed accordingly. The multi-dimensional feature vector includes at least: time-domain features and trend features of temperature; time-domain features and trend features of relative humidity; time-domain features, frequency-domain features, and trend features of partial discharge signals; time-domain features and frequency-domain features of vibration signals; and trend features of gas concentration signals.

9. The method for identifying non-electrical quantity faults in low-voltage switchgear of thermal power plants according to claim 1, characterized in that, In step S4, the training of the random forest model includes the following steps: Construct a labeled dataset containing normal samples and faulty samples, and divide it into training and test sets; Set the parameters of the random forest model; The parameters of the random forest model are optimized using the out-of-bag (OOB) error, and training is stopped when the OOB error is no greater than 3%.

10. The method for identifying non-electrical faults in low-voltage switchgear of thermal power plants according to any one of claims 1 to 9, characterized in that, In steps S4 and S5, the hybrid discrimination and the hierarchical early warning include: The output of the random forest model and the output of the expert rule base are fused using a weighted voting method. When the output of the random forest model and the output of the expert rule base are inconsistent, a secondary verification is performed based on the rate of change of the trend feature to determine the final judgment result. The graded early warning is triggered based on the fault level obtained from the final judgment.

Images