A Fault Diagnosis Method for Ring Main Unit Integrating Multiple Types of Sensors

Abstract

The present invention discloses a fault diagnosis method for a ring main unit integrating multiple types of sensors, including: hybrid model construction: building a typical fault simulation experimental platform for the ring main unit to conduct fault simulation. During the fault simulation process, a UHF sensor, a smoke sensor, and a water immersion sensor are used for data acquisition; fault feature extraction is carried out, and model parameter training is carried out; the actual data of the ring main unit is collected; data preprocessing is carried out at the edge of the UHF sensor; the features extracted by the UHF, the data values uploaded by the smoke sensor, and the data values uploaded by the water immersion sensor are used as feature data and input into the trained hybrid model, and finally the fault type diagnosis value is output. The present invention adopts a hybrid model data processing and diagnosis method combining sensor edge and background, reduces the data transmission pressure, improves the online system fault diagnosis rate, and provides a theoretical basis for the operation status evaluation of the ring main unit.

Description

Technical Field

[0001] The present invention belongs to the technical field of insulation condition detection of power equipment, and particularly relates to a fault diagnosis method for ring main units integrating multiple types of sensors. Background Art

[0002] In the insulation system of electrical equipment, the electric field distribution is often not completely uniform. When the electric field strength in a local area reaches the breakdown field strength of that area, discharge occurs in that area, but the discharge does not penetrate between the two conductors to which the voltage is applied, and the phenomenon that the insulation system has not been broken down is local discharge. The occurrence of local discharge will lead to the continuous deterioration of the insulation system and may ultimately cause power outage accidents, etc.

[0003] Currently, the commonly used detection methods for local discharge in ring main units include the pulse current method, the TEV (Transmission Electro-Voice) method, the UHF (Ultra High Frequency) method, and the ultrasonic method. The local discharge detection of ring main units involves the measurement technologies of multiple sensors in multiple fields of electrical and non-electrical quantities. The traditional single local discharge detection method is vulnerable to external signal interference, and its application has certain limitations. By adopting the fusion technology of multiple sensors, the advantages of the common operation of multiple sensors are exerted, the limitations of single or a small number of sensors are eliminated, and the effectiveness and accuracy of local discharge detection are greatly improved. Insulation faults mainly include internal insulation faults, external insulation flashover breakdown to the ground, phase-to-phase insulation flashover breakdown, porcelain insulator flashover breakdown and explosion, CT flashover breakdown and explosion, flashover breakdown caused by overvoltage, etc. It can be seen that detecting the insulation defects existing in the ring main unit and eliminating potential hazards at an early stage are of great significance for improving the operation stability of the power system. Summary of the Invention

[0004] The present invention proposes a fault diagnosis method for ring main units integrating multiple types of sensors, which is applicable to the online monitoring of ring main units in different environments.

[0005] To achieve the above object, the present invention adopts the following technical solutions: A fault diagnosis method for ring main units integrating multiple types of sensors, including a laboratory model construction module, a sensor data edge processing module, and a background service fusion center processing module; the hybrid model constructed by the laboratory model construction module is used for the background service fusion center, and the data processed by the sensor data edge processing module is wirelessly transmitted to the background service fusion center; the steps are as follows:

[0006] S1. Hybrid model construction: Build a typical fault simulation experimental platform for ring main units to simulate faults. During the fault simulation, use UHF sensors, smoke sensors, and water immersion sensors to collect data; extract fault features based on the partial discharge information, smoke sensor values, and water immersion sensor values collected from typical faults; train a hybrid model based on Stacking: the basic models include BP neural network, probabilistic neural network, and LSTM network, and a linear regression model is selected as the secondary learner for the final result diagnosis; the process of training the hybrid model is as follows:

[0007] a. Divide the collected data into a training set and a test set, and divide the training set into a fixed number of parts;

[0008] b. Use the cross-validation method to train the basic models. Select the training set as the validation set in turn to train the basic models, and then make predictions on the test set. In this way, new data predictions with the same number of parts as the training set and a set of prediction values on the test set will be obtained;

[0009] c. Calculate the average value of the data obtained from each basic model to obtain a new training set Train;

[0010] d. Input the training set Train and the obtained prediction values into the linear regression model for model training to establish a linear regression model;

[0011] S2. Monitoring data collection: UHF sensors, smoke sensors, and water immersion sensors collect data of ring main units;

[0012] S3. Data preprocessing: Perform data preprocessing at the edge of UHF sensors. First, conduct a preliminary diagnosis of data anomalies based on information divergence; divide the data initially judged as abnormal by data length and extract features based on the long short-term memory artificial neural network (LSTM); take the maximum values of the data from the smoke sensors and water immersion sensors respectively;

[0013] S4. Diagnose the fault type through the hybrid model: Take the features extracted by UHF, the data values uploaded by the smoke sensor, and the data values uploaded by the water immersion sensor as feature data, input them into the trained hybrid model, and finally output the fault type diagnosis value.

[0014] Furthermore, the process of constructing the hybrid model in step S1 includes: S11. Build a typical fault simulation experiment for ring main units in the laboratory, and collect the partial discharge signal U during typical faults std (i), smoke sensor values v std1 , water immersion sensor values v std2 ; S12. Calculate the data U std(i)'s information divergence factor, calculate α based on typical faults std Range distribution, determine range parameters K ; S13. Extract fault features according to the method described in step S12, train each model parameter, input the typical fault data of the laboratory into the trained model respectively, and obtain the fault accuracy of each model; the models include: BP neural network, probabilistic neural network, LSTM network to obtain the training parameters of each model.

[0015] The present invention further illustrates that the step S3 includes: S31. Calculate data U (i)'s information divergence factor α, perform signal anomaly diagnosis; S32. If it is an abnormal signal, divide the signal into fixed-length windows to generate a data set Y s , based on the long short-term memory artificial neural network (LSTM) for the data set Y s Perform feature extraction to obtain a feature set h l ; S33. Take the maximum value of the smoke sensor for the data of the smoke sensor v 1. Take the maximum value of the water immersion sensor for the data of the water immersion sensor v 2.

[0016] Furthermore, in step S4, the feature set h l , the value collected by the smoke sensor v 1, the value collected by the water immersion sensor v 2 are combined to form a new feature interval T = { h l , v 1, v 2}; Use each single trained model to diagnose the collected feature set T, and then calculate the average weighting of the output results of each model to obtain the final fault diagnosis result, and the fault types are typical fault types of switchgear: tip discharge, surface discharge, internal discharge, floating discharge, particle discharge.

[0017] Furthermore, in step S4, use the collected feature set T to predict based on the trained basic model to obtain three predicted values B1, B2, B3, use these three predicted values to construct three feature values (B1, B2, B3), and input them into the trained linear regression model for prediction to obtain the final fault diagnosis result.

[0018] Advantages of the present invention: By adopting a hybrid model data processing and diagnosis method combining the sensor edge and the background, it realizes the reasonable selection of data diagnosis models for different types of data, reduces the data transmission pressure, improves the online system fault diagnosis rate, and provides a theoretical basis for the operation status evaluation of the ring main unit. Description of the Drawings

[0019] Figure 1 is the flowchart of the present invention. Detailed Embodiments

[0020] The present invention will be further described below with reference to the accompanying drawings.

[0021] A ring main unit fault diagnosis method integrating multiple types of sensors includes a laboratory model construction module, a sensor data edge processing module, and a background service fusion center processing module; the hybrid model constructed by the laboratory model construction module is used for the background service fusion center, and the data processed by the sensor data edge processing module is wirelessly transmitted to the background service fusion center; as Figure 1 shown, it includes the following steps:

[0022] S1. Hybrid model construction: Build a typical fault simulation experimental platform for the ring main unit to conduct fault simulation. During the fault simulation process, use UHF sensors, smoke sensors, and water immersion sensors to collect data; extract fault features based on the partial discharge information, smoke sensor values, and water immersion sensor values collected from typical faults, extract fault features based on the partial discharge information, smoke sensor values, and water immersion sensor values collected from typical faults, and train a hybrid model based on Stacking; the basic models include: BP neural network, probabilistic neural network, and LSTM network to obtain the hybrid model, and the secondary learner selects a linear regression model for the final result diagnosis.

[0023] Specifically:

[0024] a. Divide the collected data into a training set and a test set, and divide the training set into fixed parts; in this embodiment, the training set is divided into 5 parts, namely train1, train2, trian3, train4, and train5.

[0025] b. Use the cross-validation method to train the basic model. Use train1, train2, trian3, train4, and train5 as validation sets in turn, and the remaining 4 sets as training sets. Perform cross-validation to train the model, and then make predictions on the test set. In this way, 5 predictions data can be obtained when training the BP neural network, and one prediction value B1 can be obtained on the test set. The probabilistic neural network and LSTM data network can also obtain 5 predictions data respectively. These 5 data are stacked vertically and recorded as A1.

[0026] c. After the basic model is trained, the prediction values of the basic model on the training set are used as new training sets Train={A1,A2,A3};

[0027] d. Input the training set Train and the obtained predicted value into the linear regression model for model training and establish a linear regression model;

[0028] S11. The laboratory sets up a typical fault simulation experiment for the ring main unit and collects the partial discharge signal U during a typical fault. std (i) Smoke sensor value v std1 , water sensor value v std2 ; S12. Calculate data U std (i) The information divergence factor is calculated based on typical faults. std Range distribution, determine range parameters K , providing a parameter range for the subsequent data anomaly diagnosis in S3; S13. Fault feature extraction is performed based on a fixed-length window and LSTM method, and each model parameter is trained respectively. The trained model is tested based on the laboratory typical fault data input to obtain the diagnostic accuracy of each model. The model includes: BP neural network, probabilistic neural network, and LSTM network to obtain the training parameters of each model. In this embodiment, the diagnostic accuracy of the BP neural network is α1=0.85, the diagnostic accuracy of the probabilistic neural network is α2=0.9, and the diagnostic accuracy of the LSTM network is α3=0.8.

[0029] S2. Monitoring data collection: In actual use, the sensor collects 1s data U (i); The sensor is a UHF sensor, a smoke sensor, and a water sensor.

[0030] S3, Data preprocessing: Data preprocessing is performed at the edge of the UHF sensor, and preliminary diagnosis of data abnormalities is performed based on information divergence. Data that is initially judged to be abnormal is divided into data lengths and features are extracted based on the long short-term memory artificial neural network (LSTM).

[0031] Specifically, in the actual test, for the collected data, first calculate the data U (i) the information divergence factor α, and perform signal anomaly diagnosis based on the parameter K determined in S12. In this embodiment, the value of K is 1.1. If α < 1.1, the signal is considered abnormal;

[0032] S32. If it is an abnormal signal, divide the signal data with a fixed-length window. Select the window as M (M < N), then the new sequence obtained after data division is Y s , and based on the long short-term memory artificial neural network (LSTM), extract features from the data set Y to obtain the feature set h l ; The selected activation function is: , where W xf 、W hf is the weighting coefficient of the input signal at x t , x t-1 , and b f is the bias term; otherwise, directly upload the extracted partial discharge amplitude without feature extraction and model prediction, and the fault diagnosis result is normal;

[0033] S33. For the data of the smoke sensor, take the maximum value of the smoke sensor v 1. For the data of the water immersion sensor, take the maximum value of the water immersion sensor v 2.

[0034] S4. Diagnose the fault type through the hybrid model: Use the features extracted by UHF, the data values uploaded by the smoke sensor, and the data values uploaded by the water immersion sensor as feature data, input them into the trained hybrid model, and finally output the fault type diagnosis value.

[0035] Specifically, for the extracted features h l upload them to the background processing center, and at the same time, upload the values collected by the smoke sensor v 1 and the values collected by the water immersion sensor v 2 to the background processing center, and combine them to form a new feature interval T = { h l , v 1, v 2}.

[0036] Specifically, the background processing center makes predictions based on the trained basic model of the collected feature set T, obtaining three predicted values B1, B2, and B3. Using these three predicted values to construct three feature values (B1, B2, B3), and inputting them into the trained linear regression model for prediction to obtain the final fault diagnosis result.

[0037] Obviously, the above embodiments are merely examples given for clearly illustrating the present invention and are not limitations on the implementation of the present invention. For those of ordinary skill in the art to which the present invention pertains, other different forms of changes or modifications can be made based on the above description; it is not necessary and impossible to enumerate all the implementation manners here; and the obvious changes or modifications derived therefrom still fall within the protection scope of the present invention.

Claims

1. A fault diagnosis method for ring main unit integrating multiple types of sensors, characterized in that, It includes a laboratory model construction module, a sensor data edge processing module, and a background service fusion center processing module; the hybrid model constructed by the laboratory model construction module is used for the background service fusion center, and the data processed by the sensor data edge processing module is wirelessly transmitted to the background service fusion center; the steps are as follows: S1. Hybrid model construction: Build a typical fault simulation experimental platform for ring main unit to conduct fault simulation. During the fault simulation, use UHF sensors, smoke sensors, and water immersion sensors to collect data; Extract fault features based on the partial discharge information, smoke sensor values, and water immersion sensor values collected from typical faults; Train the hybrid model based on Stacking: The basic models include BP neural network, probabilistic neural network, and LSTM network, and the hybrid model is trained. The secondary learner selects a linear regression model to conduct the final result diagnosis; The hybrid model training process is as follows: a. Divide the collected data into a training set and a test set, and divide the training set into fixed parts; b. Use the cross-validation method to train the basic model. Select the training set as the validation set in turn to train the basic model, and then make predictions according to the test set. In this way, new data predictions with the same number of parts as the training set and a prediction value on the test set will be obtained; c. Calculate the average value of the data obtained by each basic model to obtain a new training set Train; d. Input the new training set Train and the obtained prediction values into the linear regression model for model training to establish a linear regression model; S2. Monitoring data collection: Use UHF sensors, smoke sensors, and water immersion sensors to collect ring main unit data to obtain 1s of data U(i); S3. Data preprocessing: Perform data preprocessing at the edge of the UHF sensor. First, conduct a preliminary diagnosis of data anomalies based on information divergence; Divide the data initially judged to be abnormal according to data length and extract features based on the LSTM network; The smoke sensor and the water immersion sensor respectively take the maximum value of the data and transmit it to the background service fusion center through the wireless network; S4. Diagnose the fault type through the hybrid model: Use the features extracted by the UHF, the data values uploaded by the smoke sensor, and the data values uploaded by the water immersion sensor as feature data, input them into the trained hybrid model, and finally output the fault type diagnosis value.

2. The fault diagnosis method for a ring main unit integrating multiple types of sensors according to claim 1, wherein, The process of constructing the hybrid model in step S1 includes: S11. Build a typical fault simulation experiment of a ring main unit in the laboratory and collect the partial discharge time-domain signal U during typical faults. std (i), the value v of the smoke sensor std1 , the value v of the water immersion sensor std2 ; S12. Calculate the information divergence factor of the partial discharge time-domain signal U std (i), and determine the range parameter K according to the α std range distribution calculated based on typical faults; S13. Extract fault features based on the fixed-length window and the LSTM method, train each model parameter respectively, and input the typical fault data in the laboratory into the trained model for testing to obtain the diagnostic accuracy of each model.

3. The fault diagnosis method for a ring main unit integrating multiple types of sensors according to claim 1, wherein, Step S3 includes: S31. Calculating the information divergence factor of data U(i) for signal anomaly diagnosis; S32. If it is an abnormal signal, performing fixed-length window partitioning on the signal to generate a data set Y s , and based on the LSTM network for the data set Y s performing feature extraction to obtain a feature set h l ; S33. Taking the maximum value v1 of the smoke sensor for the data of the smoke sensor and the maximum value v2 of the water immersion sensor for the data of the water immersion sensor.

4. A fault diagnosis method for a ring main unit integrating multiple types of sensors according to claim 1, characterized in that, In step S4, the feature set h l , the maximum value v1 of the smoke sensor, and the maximum value v2 of the water immersion sensor are combined to form a new feature interval T = {h l , v1, v2}; the collected feature interval T is used to diagnose the feature set using each single trained model, and then the output results of each model are calculated and averaged and weighted to obtain the final fault diagnosis result, and the fault type is the typical fault type of the switch cabinet: tip discharge, surface discharge, internal discharge, floating discharge, particle discharge.

5. The fault diagnosis method for a ring main unit integrating multiple types of sensors according to claim 4, characterized in that, In step S4, use the collected feature interval T to predict based on the trained basic model to obtain three prediction values B1, B2, and B3. Use these three prediction values to construct three feature values (B1, B2, B3), input them into the trained linear regression model for prediction, and obtain the final fault diagnosis result.

Images