Typical discharge fault diagnosis method for high-voltage switch cabinet based on multi-source data fusion

Abstract

The invention discloses a typical discharge fault diagnosis method for a high-voltage switch cabinet based on multi-source data fusion. The typical discharge fault diagnosis method comprises the following steps: step 1, arranging a multi-mode sensing high-voltage switch cabinet insulation fault diagnosis device; 2, detecting and collecting a partial discharge signal generated when partial discharge occurs in the high-voltage switch cabinet; step 3, carrying out amplification and filtering processing on the partial discharge signal; 4, performing ADC data conversion on the received partial discharge signal, and transmitting the converted partial discharge signal to a PC upper computer through an edge computing gateway; and step 5, firstly performing information fusion by a power internet-of-things monitoring system loaded on the PC upper computer, then sending fused partial discharge data into the convolutional neural network for feature extraction and partial discharge mode identification, and finally outputting an identification result. According to the invention, the technical problems of limited sensitivity, insufficient anti-interference performance, low reliability of single-source detection and the like in the insulation fault detection of the existing high-voltage switch cabinet are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of partial discharge detection of high-voltage switchgear, and in particular relates to a method for diagnosing typical discharge faults of high-voltage switchgear based on multi-source data fusion. Background Art

[0002] High-voltage switchgear, a key node in the power distribution network, performs the core functions of receiving, distributing, and controlling electrical energy at voltage levels of 35kV and below. According to State Grid's 2022 operational statistics, the number of 10kV switchgear units in operation in my country has exceeded 6 million, with an average annual growth rate of 8.7%. Their operational reliability is directly related to the reliability of urban power grids and the quality of power delivered to users. In specialized applications such as rail transit and data centers, where power continuity is crucial, the insulation condition of switchgear becomes a key factor impacting power supply security.

[0003] However, latent insulation defects can occur at every stage of the equipment lifecycle: internal air gaps (mostly 0.1 to 3 mm) caused by poor workmanship during manufacturing, metal debris left over from assembly (typically 50 to 500 μm in size), epoxy resin cracks (growing at a rate of approximately 0.02 mm / year) caused by mechanical stress during transportation and installation, contact resistance degradation caused by thermal cycling during operation (typically increasing at a rate of 15% per thousand operations), and surface tracking of insulation components under condensation conditions (where the initial field strength drops to 1.2 kV / mm). These microscopic defects can trigger localized field strength concentrations with electric field distortion coefficients of 2.3 to 5.6 times, resulting in a discharge initiation probability exceeding 78%.

[0004] The development of partial discharge exhibits significant nonlinear characteristics: the initial discharge typically remains at 5-50pC, with a repetition rate of less than 10 times per minute. In the mid-stage, the discharge amplitude can surge to 200-800pC, accompanied by a 20dB increase in signal strength in the characteristic frequency band (300kHz-1.5MHz). Finally, a pre-breakdown discharge pulse with a single discharge exceeding 2000pC occurs. It is worth noting that organic insulating materials experience molecular chain scission under partial discharge (with a thermal decomposition temperature threshold of approximately 180°C), leading to an average annual increase in their dielectric loss factor (tanδ) of 0.05%, ultimately resulting in a 35-60% decrease in power frequency withstand voltage capability. The window for the formation of a through-breakdown channel is typically 3-8 years.

[0005] Therefore, partial discharge is not only a key factor in accelerating insulation material degradation but also the most effective indicator for evaluating insulation status. Accurately detecting partial discharge allows us to promptly identify potential equipment safety hazards, enabling real-time monitoring, fault diagnosis, and condition assessment of electrical equipment insulation. This provides a scientific basis for equipment maintenance and overhaul, effectively preventing and curbing power grid accidents.

[0006] Currently, the main methods for monitoring partial discharge (PD) in high-voltage switchgear include ultrasonic and pulse current methods. The pulse current method, which detects pulse current signals caused by partial discharge within the high-voltage switchgear through impedance detection, offers advantages such as 0.1pC-level detection sensitivity and quantitative detection capabilities. However, it is susceptible to wide-spectrum interference sources such as inverters and rectifiers. The ultrasonic method, on the other hand, utilizes sensors mounted on the outer casing of electrical equipment to detect ultrasonic signals generated by the violent collisions between electrons during internal defect discharges. While this method offers advantages such as high confidence and immunity to electromagnetic interference, it suffers from an acoustic attenuation rate of up to 30dB / cm for discharges within epoxy castings, resulting in a detection rate of less than 40% for deep-seated defects.

[0007] In summary, facing the problem that the current single partial discharge detection methods have their own advantages and disadvantages, it is impossible to achieve a comprehensive evaluation of the insulation fault of the high-voltage switchgear. Summary of the Invention

[0008] The purpose of the present invention is to provide a typical discharge fault diagnosis method for high-voltage switchgear based on multi-source data fusion, so as to solve the problems of insufficient sensitivity, weak anti-interference ability and limited reliability of single detection method in existing high-voltage switchgear insulation fault detection.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] A typical discharge fault diagnosis method for high-voltage switchgear based on multi-source data fusion includes the following steps:

[0011] Step 1: Arrange a multi-modal sensing high-voltage switchgear insulation fault diagnosis device;

[0012] The multimodal sensing high-voltage switchgear insulation fault diagnosis device includes a partial discharge signal detection module, a signal conditioning module, a data processing module, an edge computing gateway and a PC host computer; the partial discharge signal detection module includes an ultrasonic detection sensor and a pulse current detection sensor, and is respectively connected to the signal conditioning module via a radio frequency signal line; the signal conditioning module is connected to the data processing module via a radio frequency coaxial line; the data processing module and the edge computing gateway transmit data via a LoRa communication network; the edge computing gateway and the PC host computer transmit data via a 4G or 5G network;

[0013] Step 2: The ultrasonic detection sensor detects and collects ultrasonic signals generated by violent collisions between electrons when partial discharge occurs inside the high-voltage switchgear, and then transmits them to the signal conditioning module; the pulse current detection sensor detects and collects pulse current signals generated when partial discharge occurs inside the high-voltage switchgear, and then transmits them to the signal conditioning module;

[0014] Step 3: The signal conditioning module amplifies and filters the received ultrasonic signal and pulse current signal and then outputs them to the data processing module;

[0015] Step 4: The data processing module performs ADC data conversion on the received ultrasonic signal and pulse current signal respectively to obtain the partial discharge ultrasonic time domain signal and the partial discharge pulse current time domain signal, and then transmits them to the PC host computer via the edge computing gateway;

[0016] In step 5, the power IoT monitoring system installed on the PC host computer first fuses the received partial discharge ultrasonic time domain signal and partial discharge pulse current time domain signal, then feeds the fused partial discharge data into a convolutional neural network for feature extraction and partial discharge pattern recognition, and finally outputs the recognition result.

[0017] The present invention integrates the pulse current method and the ultrasonic detection method through heterogeneous sensing to construct a composite detection system and diagnostic method with time-frequency-space multi-dimensional feature extraction capabilities: on the one hand, the pulse current method is used to achieve the quantitative characterization of defects such as electrical tree discharge and surface discharge inside the insulating medium by utilizing its high sensitivity and ability to quantitatively detect the discharge amount; on the other hand, the ultrasonic detection method is combined with its advantages such as high confidence and immunity to electromagnetic interference to improve the insulation fault detection capability in strong electromagnetic interference environments.

[0018] The present invention is further described. In step 1, the ultrasonic detection sensor is a piezoelectric ultrasonic sensor and is attached to the cabinet of the high-voltage switchgear. The pulse current detection sensor includes a capacitor embedded in a ceramic insulator, a display unit, and a detection impedance. The ceramic insulator is connected to the main circuit of the high-voltage switchgear. The capacitor is connected in series with the display unit, and the detection impedance is connected in parallel with the display unit.

[0019] In step 2, the pulse current detection sensor utilizes the capacitor embedded in the ceramic insulator to provide a loop for the pulse current generated by partial discharge while avoiding the direct effect of the power frequency high voltage on the detection impedance, thereby realizing the pulse current signal detection of partial discharge.

[0020] In the present invention, the ultrasonic detection sensor can be a piezoelectric ultrasonic sensor commonly used in the field of partial discharge detection. It operates by utilizing the piezoelectric effect of piezoelectric materials. Piezoelectric ceramics are commonly used piezoelectric component materials. When the mechanical waves generated by partial discharge act on the piezoelectric ceramics, the mechanical waves are converted into pressure, causing the ceramic sheet to expand and contract. This generates charges of opposite polarity on both surfaces of the ceramic sheet. The more intense the expansion and contraction of the ceramic sheet, the more charges are generated. These charges are converted into voltage and output. They can be measured and recorded by devices such as oscilloscopes via coaxial cables, thus achieving acoustic-to-electric conversion, that is, converting the ultrasonic signal generated by the partial discharge into an electrical signal.

[0021] The present invention is further described. In step 1, the signal conditioning module is independent of the data processing module and includes an amplifying circuit module and a filtering circuit module.

[0022] In the present invention, the signal conditioning module is primarily responsible for bandpass filtering and amplification of the partial discharge signal. The frequency range of the ultrasonic detection method for switchgear is 20kHz to 200kHz, while the frequency range of the pulse current method is 10kHz to 500kHz. The frequency range of the pulse current detection method encompasses that of the ultrasonic detection method, so the same signal conditioning circuit can be used to amplify and filter the partial discharge signal.

[0023] The present invention further illustrates that in step 1, the data processing module includes a high-speed signal acquisition circuit, an FPGA chip, a main control chip, a power module, a LoRa communication module, and an SD storage module. The high-speed signal acquisition circuit is connected to the main control chip via the FPGA chip; the main control chip is connected to the edge computing gateway via the LoRa communication module; and the power module and SD storage module are each connected to the main control chip. The high-speed signal acquisition circuit primarily comprises an ADC08D1000 chip, an AD8009 amplifier chip, an ADA4939 single-ended to differential converter chip, a high-voltage module, and a power chip. The main control chip utilizes an MSP430F6767 microcontroller.

[0024] In this invention, the core of the data processing module is an MSP430F6767 microprocessor, which is primarily responsible for ADC data conversion, data processing and analysis, data storage, data display, and data transmission. To ensure the accuracy of the sampled signal, the high-speed signal acquisition circuit is based on the ADC08D1000. Because the microprocessor's operating frequency cannot meet the high-speed signal acquisition circuit's requirements, an FPGA is first used to communicate with the high-speed signal acquisition circuit, and the FPGA then sends the collected data to the microprocessor for processing.

[0025] The present invention is further described. In step 5, the power IoT monitoring system loaded on the PC host computer first performs information fusion on the received partial discharge ultrasonic time domain signal and the partial discharge pulse current time domain signal. The specific steps are as follows:

[0026] 1) Converting the received partial discharge ultrasonic time domain signal and partial discharge pulse current time domain signal from one-dimensional time domain signal into two-dimensional time-frequency image by continuous wavelet transform formula;

[0027] The continuous wavelet transform formula is:

[0028]

[0029] Where f(t) is a one-dimensional time domain signal; is the wavelet basis function, * represents the conjugate complex number; s and τ are the expansion and translation factors of the wavelet transform. The expansion factor s is used to expand and contract the wavelet basis function, which is inversely proportional to the frequency. When s < 1, the waveform is compressed, corresponding to the increase in the fluctuation frequency of the wavelet basis function; when s > 1, the waveform is stretched, corresponding to the decrease in the fluctuation frequency of the mother wavelet function; the translation factor τ represents the length of the position shifted along the t-axis, corresponding to the time information;

[0030] The size of the two-dimensional time-frequency image is 100 pixels × 150 pixels × 3 channels (100 pixels wide × 150 pixels high × 3 channels);

[0031] 2) Use the cat(dim,A,B) function to merge two 100pixel×150pixel×3channels images into a 100pixel×150pixel×6channels image to achieve information fusion of the pulse current signal of the ultrasonic signal.

[0032] Furthermore, in step 5, the convolutional neural network includes: an input layer, three convolutional layers, three fully connected layers, and an output layer;

[0033] The input layer is a 100×150×6 time-frequency image, which is extracted through three convolutional layers. After the convolutional layers complete the feature extraction, three fully connected layers are responsible for implementing feature combination to reduce the number of feature parameters. Finally, the output layer is a 1×5 fully connected layer with a Softmax activation function to realize common tip discharge, internal discharge, suspended discharge, surface discharge, and normal pattern recognition.

[0034] The Maxpool layer of the convolutional layer is responsible for reducing the number of elements in the feature map and making the observation window of the consecutive convolutional layers larger and larger, thereby introducing a layer structure of spatial filters; the nonlinear activation function of the convolutional layer is selected as ReLU to maximize the screening ability of neurons; to avoid overfitting and improve the generalization ability of the model, the convolutional layer introduces Batch Normalization, which also improves the convergence speed of the model and reduces gradient explosion and gradient smallness.

[0035] The partial discharge pattern recognition model in the convolutional neural network can be trained by the following algorithm:

[0036] Build a partial discharge mode experimental platform to simulate typical insulation defect tests and collect corresponding partial discharge signals. The data within every 50 power frequency cycles is a set of data.

[0037] Manually label the partial discharge signals with reference to the power Internet of Things monitoring platform;

[0038] Perform CWT transformation on the partial discharge signal and perform information fusion to establish a data set for CNN model training;

[0039] The data set is divided into 80% for training and 20% for testing.

[0040] CNN model training, continuously adjusting model parameters during training to obtain the optimal network model;

[0041] After obtaining the optimal network model, it is tested on the test set to obtain the model's partial discharge pattern recognition results.

[0042] The TSNE visualization results of the Maxpool1 layer, Maxpool2 layer, Maxpool3 layer and SoftmaxLayer layer are used to analyze the effect of partial discharge pattern recognition.

[0043] Advantages of the present invention:

[0044] This invention builds a multi-source heterogeneous sensing fusion system combining pulse current and ultrasonic testing to form a composite detection system capable of extracting multidimensional features across time, frequency, and space. First, leveraging the high sensitivity of the pulse current method, it accurately quantifies the discharge of insulation defects such as electrical treeing and creeping discharge. Simultaneously, by combining the electromagnetic interference resistance of ultrasonic testing, a dual-channel verification mechanism is established for use in electromagnetic noise environments. This multimodal fusion strategy retains the quantitative detection capabilities of the pulse current method while enhancing diagnostic confidence through the spatial localization characteristics of ultrasonic testing, significantly improving the robustness of insulation fault detection under complex operating conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is the overall architecture of a multi-modal sensing high-voltage switchgear insulation fault diagnosis device according to an embodiment of the present invention;

[0046] Figure 2 This is a primary wiring diagram of the live display device;

[0047] Figure 3 The basic principle of partial discharge detection using pulse current method;

[0048] Figure 4 It is the amplifying circuit module in the signal conditioning module;

[0049] Figure 5 This is the frequency-gain curve of the INA129 instrumentation amplifier;

[0050] Figure 6 It is the filter circuit module in the signal conditioning module;

[0051] Figure 7 It is the overall structure of the high-speed signal acquisition circuit;

[0052] Figure 8 This is the communication topology diagram between FPGA and high-speed signal acquisition circuit;

[0053] Figure 9 It is the MSP430F6767 microcontroller and its peripheral circuits;

[0054] Figure 10 It is the LoRa wireless transmission module circuit;

[0055] Figure 11 Power supply circuit for LoRa;

[0056] Figure 12 Schematic diagram of partial discharge information fusion;

[0057] Figure 13 is the convolutional neural network architecture;

[0058] Figure 14 This is a flowchart of the multi-information fusion partial discharge pattern recognition algorithm;

[0059] Figure 15 It is a partial discharge mode experimental platform;

[0060] Figure 16 Visualize the model training process;

[0061] Figure 17 is the confusion matrix measured on site. DETAILED DESCRIPTION

[0062] The present invention will be further described below with reference to the accompanying drawings.

[0063] Example:

[0064] A typical discharge fault diagnosis method for high-voltage switchgear based on multi-source data fusion includes the following steps:

[0065] Step 1: Arrange a multi-modal sensing high-voltage switchgear insulation fault diagnosis device;

[0066] like Figure 1 As shown, the multimodal sensing high-voltage switchgear insulation fault diagnosis device includes a partial discharge signal detection module, a signal conditioning module, a data processing module, an edge computing gateway and a PC host computer; the partial discharge signal detection module includes an ultrasonic detection sensor and a pulse current detection sensor, and is respectively connected to the signal conditioning module through a radio frequency signal line; the signal conditioning module is connected to the data processing module through a radio frequency coaxial line; the data processing module and the edge computing gateway transmit data through the LoRa communication network; the edge computing gateway and the PC host computer transmit data through a 4G or 5G network;

[0067] Step 2: The ultrasonic detection sensor detects and collects ultrasonic signals generated by violent collisions between electrons when partial discharge occurs inside the high-voltage switchgear, and then transmits them to the signal conditioning module; the pulse current detection sensor detects and collects pulse current signals generated when partial discharge occurs inside the high-voltage switchgear, and then transmits them to the signal conditioning module;

[0068] Step 3: The signal conditioning module amplifies and filters the received ultrasonic signal and pulse current signal and then outputs them to the data processing module;

[0069] Step 4: The data processing module performs ADC data conversion on the received ultrasonic signal and pulse current signal respectively to obtain the partial discharge ultrasonic time domain signal and the partial discharge pulse current time domain signal, and then transmits them to the PC host computer via the edge computing gateway;

[0070] In step 5, the power IoT monitoring system installed on the PC host computer first fuses the received partial discharge ultrasonic time domain signal and partial discharge pulse current time domain signal, then feeds the fused partial discharge data into a convolutional neural network for feature extraction and partial discharge pattern recognition, and finally outputs the recognition result.

[0071] In step 1 of this embodiment, the ultrasonic detection sensor is a piezoelectric ultrasonic sensor and is attached to the cabinet of the high-voltage switchgear. The pulse current detection sensor includes a capacitor embedded in a ceramic insulator, a display unit, and a detection impedance. The ceramic insulator is connected to the main circuit of the high-voltage switchgear. The capacitor is connected in series with the display unit, and the detection impedance is connected in parallel with the display unit.

[0072] In step 2, the pulse current detection sensor utilizes the capacitor embedded in the ceramic insulator to provide a loop for the pulse current generated by partial discharge while avoiding the direct effect of the power frequency high voltage on the detection impedance, thereby realizing the pulse current signal detection of partial discharge.

[0073] In this embodiment, the ultrasonic detection sensor can be a piezoelectric ultrasonic sensor commonly used in the field of partial discharge detection. It operates by utilizing the piezoelectric effect of piezoelectric materials. Piezoelectric ceramics are a common piezoelectric component material. When the mechanical waves generated by partial discharge act on the piezoelectric ceramics, the mechanical waves are converted into pressure, causing the ceramic to expand and contract. This generates charges of opposite polarity on both surfaces of the ceramic. The more intense the expansion and contraction of the ceramic, the more charges are generated. These charges are converted into voltage and output. They can be measured and recorded by devices such as oscilloscopes via coaxial cables, thus achieving acoustic-to-electric conversion, converting the ultrasonic signals generated by partial discharge into electrical signals.

[0074] In this embodiment, the pulse current detection sensor mainly utilizes the ceramic insulator capacitor with a charge display device to prevent the power frequency high voltage from directly acting on the detection impedance while providing a loop for the pulse current generated by partial discharge to achieve partial discharge detection. The basic principle is as follows Figure 2 and Figure 3 shown. Figure 2 This is a primary wiring diagram of the energized display device. Figure 2 C k It is a capacitor embedded in the ceramic insulator. The capacitance value is generally 115pF. It presents high impedance to power frequency excitation, but low impedance to high-frequency pulse excitation generated by partial discharge. Therefore, the partial discharge signal in the high-voltage switchgear can be coupled through this capacitor. Figure 2 C ie and R ie The equivalent circuit of the indicator light can be detected by connecting the impedance in parallel at both ends of the indicator light, such as Figure 3 As shown: The pulse current acts on the detection impedance, which is manifested as a pulse voltage signal U d (t), after transmission, collection, amplification and display, the partial discharge of the equipment can be measured. When the equipment generates partial discharge, the capacitance C x There will be an instantaneous voltage change ΔU at both ends, and a pulse current i will be generated in the circuit, passing through the coupling capacitor (ceramic insulator capacitor) C k Coupled to the detection impedance Z d It is represented by a pulse voltage signal U d (t), through the processing of transmission, collection, amplification and display, some basic parameters of partial discharge generated by the equipment can be measured.

[0075] In step 1 of this embodiment, the signal conditioning module is independent of the data processing module and includes an amplifying circuit module and a filtering circuit module.

[0076] In this embodiment, the signal conditioning module is primarily responsible for bandpass filtering and amplification of the partial discharge signal. The frequency range of the switchgear ultrasonic detection method is 20kHz to 200kHz, while the frequency range of the pulse current method is 10kHz to 500kHz. The frequency range of the pulse current detection method encompasses that of the ultrasonic detection method, so the same signal conditioning circuit can be used to amplify and filter the partial discharge signal.

[0077] The amplifier circuit module uses the INA129 low-power precision instrumentation amplifier produced by Texas Instruments (TI) to design the amplifier circuit, which has the advantages of low offset voltage, low drift, low input bias current, and low noise. The INA129 op amp gain formula is:

[0078]

[0079] In the formula, G is the amplification factor of the INA129 instrumentation amplifier, and R = 49.4kΩ is the sum of the two internal amplifier feedback resistors, which has the advantages of high precision and low temperature drift. G It is an external resistor. Changing the resistance of the external resistor can change the amplification factor.

[0080] in accordance with Figure 5 The frequency-gain curve of the INA129 instrumentation amplifier is shown in the figure. As can be seen from the figure, considering that when the amplification factor G is too large, the frequency band of the amplified signal will be severely reduced, this embodiment selects the external resistor value as 750Ω, at which time the amplification factor is 67dB.

[0081] In order to ensure the signal integrity as much as possible, the filter circuit module uses a Butterworth bandpass filter, and the chip is the integrated operational amplifier NE5532 produced by Texas Instruments TI to design an active bandpass Butterworth filter. Figure 6 The resistor values of R1 to R8 and the capacitor values of C1 to C4 are as follows: According to the design manual, C1 = C2 = C3 = C4 = 1nF.

[0082]

[0083] Where, f cH is the high-pass cutoff frequency, which is 500kHz, so R3=R4=820Ω; f cL is the high-pass cutoff frequency, which is 10kHz, R7=R8=820Ω; Q is the filter quality factor, which is 0.8; A vpis the gain of a single-stage op amp; furthermore, it can be determined that R1 = 5kΩ, R2 = 3kΩ, R5 = 47kΩ, and R6 = 27kΩ.

[0084] After the partial discharge signal is collected, amplified and filtered, the signal is sent to the data processing module via the radio frequency coaxial cable.

[0085] In step 1 of this embodiment, the data processing module includes a high-speed signal acquisition circuit, an FPGA chip, a main control chip, a power module, a LoRa communication module, and an SD storage module. The high-speed signal acquisition circuit is connected to the main control chip via the FPGA chip; the main control chip is connected to the edge computing gateway via the LoRa communication module; and the power module and SD storage module are each connected to the main control chip. The high-speed signal acquisition circuit primarily consists of an ADC08D1000 chip, an AD8009 amplifier chip, an ADA4939 single-ended to differential converter chip, a high-voltage module, and a power chip. The main control chip utilizes an MSP430F6767 microcontroller.

[0086] In this embodiment, the core of the data processing module is the microprocessor MSP430F6767 single-chip microcomputer, which is mainly responsible for completing ADC data conversion, data processing and analysis, data storage, data storage and data transmission, etc. In order to ensure the accuracy of the sampling signal, the high-speed signal acquisition circuit is based on ADC08D1000. Since the operating frequency of the microprocessor cannot meet the high-speed signal acquisition circuit, FPGA is first used to communicate with the high-speed signal acquisition circuit, and then the FPGA sends the collected data to the microprocessor for processing (the circuit topology of this part is shown in Figure 1). Figure 8 shown).

[0087] Figure 9 In the example, the crystal oscillator of the main control chip is 32.768kHz, the system power supply voltage is 3.3V, the reference voltage of the ADC is 1.65V, and the system power supply voltage is monitored by VCC-DETECT, which can monitor whether the system voltage is normal at any time.

[0088] This embodiment is equipped with LoRa wireless communication technology to transmit data to the edge computing gateway. Among them, the LoRa wireless transmission module circuit is as follows Figure 10 In addition, since the power consumption of the LoRa wireless data transmission module is about 150mW when sending data and the standby power consumption is no more than 10μW, the power consumption of the LoRa module is relatively large when working. In order to achieve low power consumption, it is necessary to design a power supply circuit for the LoRa module. The power supply is turned on when the monitoring device needs to upload data and turned off after the data transmission is completed. Figure 11 shown.

[0089] After LoRa uploads the partial discharge data to the edge computing gateway, the gateway transmits it to the "Power Internet of Things Monitoring System" via 4G, thereby realizing the diagnosis of the partial discharge type.

[0090] In step 5 of this embodiment, Figure 12 As shown, the specific steps of the power IoT monitoring system loaded on the PC host computer firstly fusing the received partial discharge ultrasonic time domain signal and the partial discharge pulse current time domain signal are as follows:

[0091] 1) Converting the received partial discharge ultrasonic time domain signal and partial discharge pulse current time domain signal from one-dimensional time domain signal into two-dimensional time-frequency image by continuous wavelet transform formula;

[0092] The continuous wavelet transform formula is:

[0093]

[0094] Where f(t) is a one-dimensional time domain signal; is the wavelet basis function, * represents the conjugate complex number; s and τ are the expansion and translation factors of the wavelet transform. The expansion factor s is used to expand and contract the wavelet basis function, which is inversely proportional to the frequency. When s < 1, the waveform is compressed, corresponding to the increase in the fluctuation frequency of the wavelet basis function; when s > 1, the waveform is stretched, corresponding to the decrease in the fluctuation frequency of the mother wavelet function; the translation factor τ represents the length of the position shifted along the t-axis, corresponding to the time information;

[0095] The size of the two-dimensional time-frequency image is 100 pixels × 150 pixels × 3 channels (100 pixels wide × 150 pixels high × 3 channels);

[0096] 2) Use the cat(dim,A,B) function to merge two 100pixel×150pixel×3channels images into a 100pixel×150pixel×6channels image to achieve information fusion of the pulse current signal of the ultrasonic signal.

[0097] Furthermore, in step five, if Figure 13 As shown, the convolutional neural network includes: an input layer, three convolutional layers, three fully connected layers and an output layer;

[0098] The input layer is a 100×150×6 time-frequency image, which is extracted through three convolutional layers. After the convolutional layers complete the feature extraction, three fully connected layers are responsible for implementing feature combination to reduce the number of feature parameters. Finally, the output layer is a 1×5 fully connected layer with a Softmax activation function to realize common tip discharge, internal discharge, suspended discharge, surface discharge, and normal pattern recognition.

[0099] The Maxpool layer of the convolutional layer is responsible for reducing the number of elements in the feature map and making the observation window of the consecutive convolutional layers larger and larger, thereby introducing a layer structure of spatial filters; the nonlinear activation function of the convolutional layer is selected as ReLU to maximize the screening ability of neurons; to avoid overfitting and improve the generalization ability of the model, the convolutional layer introduces Batch Normalization, which also improves the convergence speed of the model and reduces gradient explosion and gradient smallness.

[0100] The local discharge pattern recognition model in the convolutional neural network can be obtained by training the following algorithm (the algorithm process is as follows Figure 14 shown):

[0101] Build as Figure 15 The partial discharge mode experimental platform shown simulates a typical insulation defect test and collects the corresponding partial discharge signal. The data within every 50 power frequency cycles constitutes a set of data.

[0102] Manually label the partial discharge signals with reference to the power Internet of Things monitoring platform;

[0103] Perform CWT transformation on the partial discharge signal and perform information fusion to establish a data set for CNN model training;

[0104] The data set is divided into 80% for training and 20% for testing.

[0105] CNN model training, continuously adjusting model parameters during training to obtain the optimal network model;

[0106] After obtaining the optimal network model, it is tested on the test set to obtain the model's partial discharge pattern recognition results.

[0107] like Figure 16 Figure 1 shows the TSNE visualization results of the Maxpool1, Maxpool2, Maxpool3, and SoftmaxLayer layers, enabling analysis of the effectiveness of partial discharge pattern recognition. As the network deepens, the distance between the features of partial discharge signals from different insulation defects increases, while the distance between features of the same category decreases, further demonstrating the strong partial discharge pattern recognition capabilities of the proposed model.

[0108] Figure 17 This is a confusion matrix measured on-site. The accuracy of identifying normal and tip discharges reached 100%. Of 30 suspended discharges, one was identified as a tip discharge; of 30 surface discharges, two were identified as internal discharges; and of 30 internal discharges, one was identified as a surface discharge. The overall recognition accuracy reached 97.3%.

[0109] Obviously, the above embodiments are merely examples for the purpose of clearly illustrating the present invention and are not intended to limit the implementation of the present invention. Those skilled in the art will readily appreciate that other variations or modifications may be made based on the above description. It is not necessary and impossible to enumerate all possible implementations here. Obvious variations or modifications derived therefrom remain within the scope of protection of the present invention.

Claims

1. A typical discharge fault diagnosis method for high-voltage switchgear based on multi-source data fusion, characterized by The following steps are involved: Step 1: Arrange a multi-modal sensing high-voltage switchgear insulation fault diagnosis device; The multimodal sensing high-voltage switchgear insulation fault diagnosis device includes a partial discharge signal detection module, a signal conditioning module, a data processing module, an edge computing gateway and a PC host computer; the partial discharge signal detection module includes an ultrasonic detection sensor and a pulse current detection sensor, and is respectively connected to the signal conditioning module via a radio frequency signal line; the signal conditioning module is connected to the data processing module via a radio frequency coaxial line; the data processing module and the edge computing gateway transmit data via a LoRa communication network; the edge computing gateway and the PC host computer transmit data via a 4G or 5G network; Step 2: The ultrasonic detection sensor detects and collects ultrasonic signals generated by violent collisions between electrons when partial discharge occurs inside the high-voltage switchgear, and then transmits them to the signal conditioning module; the pulse current detection sensor detects and collects pulse current signals generated when partial discharge occurs inside the high-voltage switchgear, and then transmits them to the signal conditioning module; Step 3: The signal conditioning module amplifies and filters the received ultrasonic signal and pulse current signal and then outputs them to the data processing module; Step 4: The data processing module performs ADC data conversion on the received ultrasonic signal and pulse current signal respectively to obtain the partial discharge ultrasonic time domain signal and the partial discharge pulse current time domain signal, and then transmits them to the PC host computer via the edge computing gateway; In step 5, the power IoT monitoring system installed on the PC host computer first fuses the received partial discharge ultrasonic time domain signal and partial discharge pulse current time domain signal, then feeds the fused partial discharge data into a convolutional neural network for feature extraction and partial discharge pattern recognition, and finally outputs the recognition result.

2. The method for diagnosing typical discharge faults of high-voltage switchgear based on multi-source data fusion according to claim 1 is characterized in that: In step 1, the ultrasonic detection sensor is a piezoelectric ultrasonic sensor and is attached to the cabinet of the high-voltage switchgear. The pulse current detection sensor includes a capacitor embedded in a ceramic insulator, a display unit, and a detection impedance. The ceramic insulator is connected to the main circuit of the high-voltage switchgear. The capacitor is connected in series with the display unit, and the detection impedance is connected in parallel with the display unit. In step 2, the pulse current detection sensor utilizes the capacitor embedded in the ceramic insulator to provide a loop for the pulse current generated by partial discharge while avoiding the direct effect of the power frequency high voltage on the detection impedance, thereby realizing the pulse current signal detection of partial discharge.

3. The method for diagnosing typical discharge faults of high-voltage switchgear based on multi-source data fusion according to claim 1 is characterized in that: In step 1, the signal conditioning module is independent of the data processing module and includes an amplifying circuit module and a filtering circuit module.

4. The method for diagnosing typical discharge faults of high-voltage switchgear based on multi-source data fusion according to claim 1 is characterized in that: In step one, the data processing module includes a high-speed signal acquisition circuit, an FPGA chip, a main control chip, a power module, a LoRa communication module and an SD storage module; the high-speed signal acquisition circuit is connected to the main control chip through the FPGA chip; the main control chip is connected to the edge computing gateway through the LoRa communication module; the power module and the SD storage module are respectively connected to the main control chip.

5. The method for diagnosing typical discharge faults of high-voltage switchgear based on multi-source data fusion according to claim 4 is characterized in that: The high-speed signal acquisition circuit is mainly composed of an ADC08D1000 chip, an amplifier circuit AD8009 chip, a single-ended to differential converter ADA4939 chip, a high-voltage module and a power supply chip.

6. The method for diagnosing typical discharge faults of high-voltage switchgear based on multi-source data fusion according to claim 4 is characterized in that: The main control chip adopts the microprocessor MSP430F6767 single-chip computer.

7. The method for diagnosing typical discharge faults of high-voltage switchgear based on multi-source data fusion according to claim 1 is characterized in that: In step 5, the power IoT monitoring system loaded on the PC host computer first fuses the received partial discharge ultrasonic time domain signal and partial discharge pulse current time domain signal. The specific steps are as follows: 1) Converting the received partial discharge ultrasonic time domain signal and partial discharge pulse current time domain signal from one-dimensional time domain signal into two-dimensional time-frequency image by continuous wavelet transform formula; The continuous wavelet transform formula is: Where f(t) is a one-dimensional time domain signal; is the wavelet basis function, * represents the conjugate complex number; s and τ are the expansion and translation factors of the wavelet transform. The expansion factor s is used to expand and contract the wavelet basis function, which is inversely proportional to the frequency. When s < 1, the waveform is compressed, corresponding to the increase in the fluctuation frequency of the wavelet basis function; when s > 1, the waveform is stretched, corresponding to the decrease in the fluctuation frequency of the mother wavelet function; the translation factor τ represents the length of the position shifted along the t-axis, corresponding to the time information; The size of the two-dimensional time-frequency image is 100 pixels × 150 pixels × 3 channels; 2) Use the cat(dim,A,B) function to merge two 100pixel×150pixel×3channels images into a 100pixel×150pixel×6channels image to achieve information fusion of the pulse current signal of the ultrasonic signal.

8. The method for diagnosing typical discharge faults of high-voltage switchgear based on multi-source data fusion according to claim 7 is characterized in that: In step 5, the convolutional neural network includes: an input layer, three convolutional layers, three fully connected layers and an output layer; The input layer is a 100pixel×150pixel×6channels time-frequency image, which is extracted through three convolutional layers. After the convolutional layers complete the feature extraction, three fully connected layers are responsible for implementing feature combination to reduce the number of feature parameters. Finally, the output layer is a 1×5 fully connected layer with a Softmax activation function to realize common tip discharge, internal discharge, suspended discharge, surface discharge, and normal pattern recognition.

9. The method for diagnosing typical discharge faults of high-voltage switchgear based on multi-source data fusion according to claim 8 is characterized in that: The Maxpool layer of the convolutional layer is responsible for reducing the number of elements in the feature map and making the observation window of the consecutive convolutional layers larger and larger, thereby introducing a layer structure of spatial filters; the nonlinear activation function of the convolutional layer is selected as ReLU to maximize the screening ability of neurons; to avoid overfitting and improve the generalization ability of the model, the convolutional layer introduces Batch Normalization, which also improves the convergence speed of the model and reduces gradient explosion and gradient smallness.

Images