Switch cabinet fault sensing and intelligent inspection method and system

Abstract

The invention belongs to the technical field of power equipment state sensing, and discloses a switch cabinet fault sensing and intelligent inspection method, which comprises the following steps: synchronously acquiring audio and video data of a switch cabinet through a microphone array and a video acquisition device on an intelligent inspection trolley; carrying out multi-source separation on the mixed audio by adopting an improved Mossform model, and outputting a pure fault audio; performing short-time Fourier transform on the separated fault audio to generate a spectrogram, and preprocessing a video frame; audio features and video features are extracted and fused through a lightweight VATT model, and sound source space probability distribution is generated; and superposing the sound source probability distribution on a video picture in a thermodynamic diagram form on edge equipment, and visually displaying the fault position and strength. The invention also discloses a system for realizing the method. The method has the advantages of being high in positioning precision, high in real-time performance, suitable for edge equipment deployment and the like, and the intelligence and operation and maintenance efficiency of switch cabinet inspection can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of power equipment status sensing technology, specifically to a method and system for fault sensing and intelligent inspection of switchgear. Background Technology

[0002] High-voltage switchgear, as a critical piece of equipment in power systems, operates in a complex environment with diverse internal fault types, including but not limited to partial discharge, mechanical loosening, arc discharge, insulation aging, and foreign object discharge. These faults are typically accompanied by characteristic abnormal acoustic signals in their early stages. With the continuous expansion of power systems, the operating status of switchgear, as an important power distribution device, directly affects grid security and power supply reliability.

[0003] Traditional inspection methods rely primarily on periodic manual checks, which suffers from low efficiency, high missed detection rates, and poor real-time performance. Furthermore, the sounds of switchgear malfunctions are often mixed with environmental noise, human voices, and the noise from the inspection cart, making fault identification difficult. Existing inspection systems struggle to accurately locate fault sound sources and fuse multimodal information, thus failing to meet the demands of intelligent operation and maintenance.

[0004] To address the shortcomings of existing technologies, this invention discloses a switchgear fault perception and intelligent inspection system and method. Through audio-visual multimodal information fusion, the system employs an improved Mossformer speech separation algorithm to separate the collected mixed audio from multiple sources, extracting the target fault audio while suppressing human voices, environmental noise, and inspection trolley operating noise. The separated fault audio is input into a multimodal feature extraction network based on the VATT model, and unified with video features for semantic alignment, achieving joint modeling of audio and video in image space and enabling high-precision localization of switchgear fault sound sources. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and system for fault detection and intelligent inspection of switchgear, which solves the problem of low recognition rate caused by the mixing of switchgear fault sound and environmental noise, and improves the recognition accuracy and efficiency of the system.

[0006] The technical solution adopted in this invention is to provide a switchgear fault perception and intelligent inspection system and method. The system includes a microphone array and a video acquisition device deployed on an intelligent inspection vehicle to collect audio and video data of the switchgear. An improved Mossformer model is used to separate the mixed audio from multiple sources, suppressing human voices, environmental noise and vehicle noise. A lightweight VATT model is used to extract audio features and fuse them with video features. An audio and video fusion positioning module generates a spatial probability distribution of sound sources and visualizes the fault location through a heat map. The system processes the data in real time on an edge device to achieve intelligent inspection.

[0007] In a preferred embodiment, the present invention provides an improved Mossformer audio separation model. This model collects mixed audio signals from the switchgear operating environment using a microphone array. After filtering and preprocessing the mixed signals, the signals are input into the improved Mossformer algorithm. A frequency-time bidirectional attention module is added to the algorithm, making the model more sensitive to the frequency patterns and transient characteristics of switchgear fault sounds. A spatial feature encoder extracts sound source direction information to improve sound source separation accuracy. A residual normalization strategy is introduced in each attention sub-layer to enhance the deep network's response to weak fault sounds. A dynamic noise gating layer is added at the output end to adaptively suppress background noise and trolley motion noise based on the ambient noise intensity.

[0008] In a preferred embodiment, the present invention provides a lightweight VATT audio-video fusion localization model. The target fault audio, after being separated by an improved Mossformer, is used to generate a spectrogram through a short-time Fourier transform (STFT). Video frames obtained from the inspection vehicle's camera are preprocessed, including resolution adjustment, normalization, and key region cropping. The spectrogram and video frames are then subjected to feature extraction and unified representation and semantic alignment. This data is input into the lightweight localization module, which outputs the probability distribution of the switchgear fault sound source in the image space, intuitively displaying the location and volume intensity of the fault sound source.

[0009] A method for fault detection and intelligent inspection of switchgear, characterized by comprising the following steps:

[0010] S1. Audio and video data of the switch cabinet are collected synchronously through the microphone array and video acquisition device on the intelligent inspection vehicle;

[0011] S2. An improved Mossformer model is used to separate the mixed audio from multiple sources, outputting clean fault audio; a short-time Fourier transform is performed on the separated fault audio to generate a spectrogram, and the video frames are preprocessed.

[0012] S3. Extract and fuse audio and video features using a lightweight VATT model to generate a spatial probability distribution of sound sources;

[0013] S4. On the edge device, the probability distribution of the sound source is overlaid on the video screen in the form of a heat map to visualize the location and intensity of the fault.

[0014] The above-described method for fault detection and intelligent inspection of switchgear is characterized in that, in step S2, the improved Mossformer model processing includes: preprocessing the mixed audio by filtering, enhancing the sensitivity to fault sound features through a frequency-time bidirectional attention module, extracting sound source direction information through a spatial feature encoder, introducing a residual normalization strategy in each attention sub-layer, and adaptively suppressing background noise at the output end through a dynamic noise gating layer.

[0015] The aforementioned switchgear fault detection and intelligent inspection method is characterized by the following implementation method of the dynamic noise gating layer: real-time monitoring of the ambient noise energy level, dynamic adjustment of the gating threshold according to the noise intensity, strengthening the suppression when strong noise is detected, relaxing the suppression when the noise is weak, and weighting the audio signal frame by frame through a nonlinear function to retain the target fault audio while effectively suppressing residual noise.

[0016] The aforementioned method for fault detection and intelligent inspection of switchgear is characterized in that, in step S3, the processing of the lightweight VATT model includes:

[0017] The faulty audio is converted into a spectrogram using a short-time Fourier transform; the video frames are then adjusted for resolution, pixel values ​​are normalized, and key regions are cropped.

[0018] Lightweight convolutional neural networks were used to extract spectrogram features and video frame features, respectively.

[0019] Align audio features with video features at the semantic level;

[0020] The input is fed into the lightweight localization module to generate the probability distribution of the sound source in the image space.

[0021] The aforementioned method for fault detection and intelligent inspection of switchgear is characterized in that, in step S4, the heat map visualization process includes: performing non-maximum suppression processing on the generated probability distribution; determining the coordinates of the most likely fault point; overlaying the heat map on the original video image, marking the fault point location, confidence level, and fault type; and automatically determining whether there is a potential fault and triggering an alarm mechanism based on a preset threshold.

[0022] A switchgear fault detection and intelligent inspection system is provided to implement the switchgear fault detection and intelligent inspection method described above. The system is characterized by including an improved Mossformer audio separation module, a lightweight VATT audio and video fusion positioning module, and an edge computing unit.

[0023] The improved Mossformer audio separation module is used to perform multi-source separation on the mixed audio collected by the microphone array, and to suppress human voice, environmental noise and vehicle noise.

[0024] The lightweight VATT audio-video fusion localization module is used to extract the separated audio features and fuse them with the video features to generate a spatial probability distribution of the sound source.

[0025] The edge computing unit is used to process audio and video data locally in real time and visualize the location of the fault through a heat map.

[0026] The aforementioned switchgear fault detection and intelligent inspection system is characterized in that the improved Mossformer audio separation module includes a frequency-time bidirectional attention module, a spatial feature encoder, a residual normalization layer, and a dynamic noise gating layer.

[0027] The frequency-time bidirectional attention module is used to enhance the model's sensitivity to the frequency patterns and transient characteristics of switchgear fault sounds;

[0028] The spatial feature encoder is used to extract sound source direction information to improve sound source separation accuracy;

[0029] The residual normalization layer is used to enhance the response capability of deep networks to weak fault sounds;

[0030] The dynamic noise gating layer is used to adaptively suppress background noise based on the intensity of ambient noise.

[0031] The switchgear fault detection and intelligent inspection system described above is characterized in that the lightweight VATT audio and video fusion positioning module includes an STFT processing unit, a video preprocessing unit, a feature extraction unit, a semantic alignment unit, and a positioning output unit.

[0032] The STFT processing unit converts the separated fault audio into a spectrum.

[0033] The video preprocessing unit performs resolution adjustment, normalization, and key region cropping on video frames.

[0034] The feature extraction unit extracts audio spectrogram features and video frame features respectively;

[0035] The semantic alignment unit maps audio features and video features to a unified representation space;

[0036] The positioning output unit generates the probability distribution of the fault sound source in the image space.

[0037] The aforementioned switchgear fault perception and intelligent inspection system is characterized by further including a feature fusion module. The feature fusion module uses a cross-modal attention mechanism to map audio features and video features to the same low-dimensional shared embedding space, and uses a neural network to predict the spatial probability distribution of the sound source in the image coordinate system.

[0038] The aforementioned switchgear fault perception and intelligent inspection system is characterized in that the edge computing unit adopts an NVIDIA Jetson series processor and deploys a lightweight deep learning model to achieve real-time perception and location of switchgear partial discharge, mechanical loosening, arc discharge, insulation aging and foreign object discharge faults.

[0039] This invention has the advantages of high positioning accuracy, strong real-time performance, and suitability for edge device deployment, which can effectively improve the intelligence of switch cabinet inspection and operation and maintenance efficiency. Attached Figure Description

[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0041] Figure 1 This is a flowchart of the system operation of the present invention.

[0042] Figure 2 This is a flowchart of the improved Mossformer audio separation model of the present invention.

[0043] Figure 3 This is the overall structural diagram of the improved Mossformer audio separation model of this invention.

[0044] Figure 4 This is a flowchart of the lightweight VATT audio and video fusion positioning model of the present invention. Detailed Implementation

[0045] To better understand the purpose, system architecture, and functional implementation of this embodiment, the embodiments and features described herein can be combined with each other without conflict. The exemplary embodiments disclosed herein will be described below with reference to the accompanying drawings, including specific technical details disclosed to aid understanding; however, these details should be considered exemplary rather than restrictive. Therefore, those skilled in the art should understand that various improvements and adjustments can be made to the embodiments described herein without departing from the scope and core ideas of the invention. Similarly, for clarity, detailed descriptions of well-known technologies, functions, and structures are omitted in the following description.

[0046] Example 1: Figure 1 This is a flowchart of the system operation of the present invention.

[0047] like Figure 1 As shown, before the inspection, the vehicle-mounted control system calibrates the spatial layout of the microphone array, adjusts the camera's focal length and angle, and checks the communication module's connectivity. It then activates the Jetson TX2 edge computing unit, loads the pre-trained improved Mossformer model and the lightweight VATT model, and sets the inspection path to ensure the vehicle can cover all switchgear areas to be inspected. A linear or circular microphone array is positioned on the top of the vehicle, facing the switchgear, to collect a mixed audio signal containing equipment operating sounds, background noise, human voices, and the vehicle's own mechanical noise. The sampling rate is set to 44.1kHz or higher to ensure the capture of high-frequency bands generated by abnormal vibrations of the switchgear.

[0048] The original mixed audio is bandpass filtered to remove low-frequency interference below 50Hz and ultrasonic waves above 20kHz. After preliminary filtering, the signal is input into the improved Mossformer model. The model inference outputs a clean target fault audio signal, retaining only the abnormal sound components inside the switchgear. A high-definition camera transmits a real-time video stream to the edge computing unit at a frame rate of 30fps. The frames are stored sequentially as a frame sequence. Each frame is preprocessed and then used as visual input to the fusion model. The separated target fault audio is converted into a spectrogram using a short-time Fourier transform and used as audio feature input. The video frame is processed by a convolutional neural network to extract spatial features and generate a visual feature map. A cross-modal alignment mechanism is used to match the audio spectrogram with the video features at the semantic level. Context modeling is used to achieve accurate localization of the sound source in the image. A two-dimensional probability heatmap is output, representing the probability that each pixel is the location of the sound source. The darker the color in the heatmap, the higher the probability that there is a fault sound at that location. The system can automatically determine whether there is a potential fault based on a threshold. The heatmap is superimposed on the original video frame to form an audio-video fusion diagnostic view. It supports saving images and alarm records and uploading them to the backend management system for maintenance personnel to review.

[0049] Figure 2 This is a flowchart of the improved Mossformer audio separation model of the present invention.

[0050] like Figure 2 As shown, the ring microphone array on the intelligent inspection vehicle acquires multi-channel mixed audio signals from the switchgear operating environment, including equipment operating sounds, human voices, environmental noise, and the vehicle's own mechanical noise. A bandpass filter is applied to the original audio signal to retain typical switchgear fault frequency bands, and the output is a clean time-domain signal sequence used as model input. The model introduces a dual-dimensional attention mechanism based on the Transformer architecture, where the frequency dimension... By focusing on the importance of different frequency components, specific spectral patterns associated with the fault can be identified, as shown in the following formula. As shown.

[0051]

[0052] in, , and These are the query matrix, key matrix, and value matrix, respectively, representing the frequency dimension. The attention score is converted into a probability distribution by a normalization function. Given the input time-frequency representation matrix, The values ​​are taken in the real matrix space, where: For the real number field, For time frames, For frequency points, A scaling factor for the attention dimension to prevent the frequency dimension from becoming too large. Gradient vanishing.

[0053] The time dimension captures the temporal evolution of transient events, as shown in the following formula. As shown.

[0054]

[0055] in, , and The query matrix, key matrix, and value matrix are time-dimension related. Given the transpose of the input time-frequency representation, For the normalized function and the formula same, Scaling factor for the attention dimension and equation same.

[0056] Through bidirectional attention mechanism The model can more accurately focus on key segments of the faulty sound, ignoring irrelevant speech or persistent background noise, as shown in the following formula. As shown.

[0057]

[0058] in, To integrate weight parameters, For frequency-dimensional attention modules, For the temporal attention module, the importance of frequency and temporal attention is balanced. Meanwhile, residual connections are introduced after each attention sub-layer to directly pass the input to the output, alleviating the training difficulties of deep networks. Layer normalization is also used to stabilize the activation value distribution, enhancing the model's response to weak fault sounds and preventing it from being drowned out by other strong sound sources due to low signal strength.

[0059] Simultaneously, a learnable gating mechanism is set up in the bidirectional attention calculation module. Its weights are dynamically adjusted according to the noise intensity of the current input audio. By monitoring the ambient noise energy level, the gating threshold is automatically adjusted. When strong noise is detected, the gating layer strengthens the suppression, and when the noise is weak, the suppression is relaxed to retain more useful information. After the gating layer, the final noise suppression operation is performed. Smoothing is performed in combination with context information, and a nonlinear function is used to weight the audio signal frame by frame. While retaining the target fault audio, residual noise is effectively suppressed. The final output is a clean audio stream obtained after multi-stage processing, which mainly contains the sound generated by abnormal operation inside the switch cabinet. This audio can be used for audio-video fusion localization of the subsequent lightweight VATT model.

[0060] Improve the overall structure of the Mossformer model as follows: Figure 3 As shown, the encoder-decoder architecture is responsible for feature extraction and waveform reconstruction. The encoder, responsible for feature extraction, consists of a one-dimensional convolutional layer (Conv1D) and rectified linear units (ReLU), the latter constraining the encoded output to non-negative values. The decoder is a one-dimensional transposed convolutional layer that uses the same kernel size and stride as the encoder. The masking network performs a non-linear mapping from the encoder output to the mask. The main component of the masking network is the improved MossFormer module, which is based on an adaptive gated single-head transformer architecture with convolution-augmented joint self-attentions.

[0061] Specifically, a MossFormer module consists of four convolutional modules, scaling and offset operations, joint local and global single-head self-attention, and three gating operations, responsible for processing long sequences. Within the MossFormer module, the sequence is processed by the convolutional modules and a bidirectional attention gating mechanism. The convolutional modules use linear projection and depthwise convolution to process the sequence. The bidirectional attention gating mechanism performs joint frequency and temporal self-attention as well as adaptive gating operations. An improved MossFormer module learns only the residual and applies skip connections from the input to the output to improve training efficiency.

[0062] Figure 4 This is a flowchart of the lightweight VATT audio and video fusion positioning model of the present invention.

[0063] like Figure 4 As shown, the left branch input contains the separated fault audio, including abnormal sound components inside the switchgear, such as typical fault sounds generated by partial discharge, mechanical vibration, and poor contact. A short-time Fourier transform is performed on the separated fault audio to convert it from the time domain to the time-frequency domain, generating a two-dimensional spectrogram. The spectrogram clearly reflects the frequency component changes of the fault sound over time. A lightweight feature extraction network, MobileNet, is used to extract features from the spectrogram. The extraction process focuses on features such as energy distribution and transient change patterns in key frequency bands to form high-dimensional feature vectors. These feature vectors retain the structured information related to the fault in the audio.

[0064] The right branch consists of a high-definition video stream captured in real-time by an onboard camera. Each frame corresponds to the current appearance of the switchgear, including visual information such as the equipment casing, wiring terminals, and indicator lights. The video frames serve as visual input to help determine the location of the sound source in physical space. The resolution of the original video frame images is adjusted to a uniform size to reduce computational complexity, and pixel values ​​are normalized to the range of [0,1]. Key areas are automatically cropped according to the switchgear outline to focus on the target area. A CNN model is used to extract features from the preprocessed video frame images, and the output is a visual feature vector with the same dimension as the audio features, ensuring the feasibility of subsequent fusion.

[0065] Map audio features and video features to the same low-dimensional shared embedding space. As shown in the following formula As shown.

[0066]

[0067] in, This is the transpose of the audio shared feature vector. To share feature vectors between videos, To share spatial dimensions, This refers to the number of key areas in the video. For the first Visual shared features of each region are calculated through shared spatial mapping and attention weights. It is the natural logarithm function.

[0068] By utilizing a cross-modal attention mechanism, a semantic association is established between "specific frequency events" in audio and "corresponding physical locations" in video. When the audio detects a discharge sound of a certain frequency, the system can match it with flickering or heating signs in a certain junction area in the video, as shown in the following formula. As shown.

[0069]

[0070] in, For normalization function, Using a classic Transformer network, this approach receives aligned audio and video features and predicts the spatial probability distribution of the sound source in the image coordinate system based on the learned joint representation. The output is a two-dimensional heatmap, the same size as the input video frame. Each pixel in the heatmap represents the probability that the location is a sound source; a higher probability indicates a more likely fault location. Non-maximum suppression is applied to the heatmap to determine the most probable fault point coordinates (x, y), and its specific location in the image is output. Historical data is used to determine whether it is a recurring fault or a new anomaly. The heatmap is overlaid on the original video frame to form a fused audio and video view. Fault point locations, confidence levels, and fault type suggestions are annotated in the image. The view can be saved as an image file or video clip and uploaded to the backend management system for maintenance personnel to view.

[0071] The specific embodiments described above do not constitute a limitation on the scope of protection disclosed in this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A switchgear fault perception and intelligent patrol method, characterized in that, The method comprises the following steps: S1, synchronously collecting audio and video data of the switch cabinet through a microphone array and a video acquisition device on an intelligent inspection trolley; S2, using an improved Mossformer model to perform multi-source separation on the mixed audio and output pure fault audio, performing short-time Fourier transform on the separated fault audio to generate a frequency spectrum, and preprocessing video frames; S3, extracting and fusing audio features and video features through a lightweight VATT model to generate a spatial probability distribution of the sound source; S4, superimposing the spatial probability distribution of the sound source on the video picture in the form of a heat map on an edge device to visually display the fault position and intensity.

2. The method of claim 1, wherein the method further comprises: In step S2, the processing process of the improved Mossformer model comprises: filtering and preprocessing the mixed audio, enhancing the sensitivity to fault audio features through a frequency-time bidirectional attention module, extracting sound source direction information through a spatial feature encoder, introducing a residual normalization strategy at each attention sublayer, and adaptively suppressing background noise through a dynamic noise gating layer at the output end.

3. The method of claim 2, wherein the method further comprises: The implementation of the dynamic noise gating layer is: real-time monitoring of the environmental noise energy level, dynamically adjusting the gating threshold according to the noise intensity, enhancing the suppression when strong noise is detected, relaxing the suppression when the noise is small, and performing frame-by-frame weighting on the audio signal through a nonlinear function to retain the target fault audio while effectively suppressing residual noise.

4. The method of claim 1, wherein the method further comprises: In step S3, the processing process of the lightweight VATT model comprises: Converting the fault audio into a frequency spectrum through short-time Fourier transform; adjusting the resolution of the video frame, normalizing the pixel value, and cropping the key area; using a lightweight convolutional neural network to extract the frequency spectrum features and video frame features respectively; aligning the audio features and video features at the semantic level; inputting into a lightweight positioning module to generate the probability distribution of the sound source in the image space.

5. The method of claim 1, wherein the method further comprises: In step S4, the heat map visualization process comprises: performing non-maximum suppression on the generated probability distribution; determining the most likely fault point coordinates; superimposing the heat map on the original video picture to mark the fault point position, confidence and fault type; automatically determining whether there is a potential fault according to a preset threshold and triggering an alarm mechanism.

6. A switchgear fault perception and intelligent patrol system for implementing the switchgear fault perception and intelligent patrol method of any one of claims 1 to 5, characterized in that, It comprises an improved Mossformer audio separation module, a lightweight VATT audio-video fusion positioning module and an edge computing unit; The improved Mossformer audio separation module is used for multi-source separation of mixed audio collected by the microphone array, and for suppressing human voice, environmental noise and trolley noise; The lightweight VATT audio-video fusion positioning module is used for extracting audio features after separation and fusing with video features to generate a spatial probability distribution of the sound source; The edge computing unit is used for real-time processing of audio and video data locally and visualizing the fault position through a heat map.

7. A switchgear fault perception and intelligent inspection system as claimed in claim 6, wherein, The improved Mossformer audio separation module comprises a frequency-time bidirectional attention module, a spatial feature encoder, a residual normalization layer and a dynamic noise gating layer; The frequency-time bidirectional attention module is used to enhance the sensitivity of the model to the frequency mode and transient characteristics of the switch cabinet fault sound; The spatial feature encoder is used to extract sound source direction information to improve the sound source separation accuracy; The residual normalization layer is used to enhance the response capability of the deep network to weak fault sound; The dynamic noise gating layer is used to adaptively suppress background noise according to the intensity of environmental noise; The improved Mossformer audio separation model is as follows: a ring microphone array on an intelligent inspection trolley acquires multi-channel mixed audio signals in the operating environment of a switch cabinet, including equipment operation sound, human voice, environmental noise and trolley self mechanical noise; a band-pass filter is applied to the original audio signal to retain typical switch cabinet fault frequency bands, and a clean time domain signal sequence is output as the model input; the model introduces a double dimension attention mechanism on the basis of the Transformer architecture, wherein the frequency dimension focuses on the importance of different frequency components, and identifies specific spectral patterns related to faults, as follows: wherein, , and are the query matrix, key matrix and value matrix of the frequency dimension respectively, is a normalization function for converting attention scores into probability distribution, is an input time-frequency representation matrix, takes values in a real matrix space, wherein is a real number field, is the number of time frames, is the number of frequency points, is a scaling factor of the attention dimension to prevent the frequency dimension from being too large to cause gradient disappearance; The time dimension captures the temporal evolution of the transient event as follows: wherein, , and are the query matrix, key matrix and value matrix of the time dimension, is the transpose matrix of the input time-frequency representation, is a normalization function, is a scaling factor of the attention dimension; Through a bidirectional attention mechanism The model can accurately focus on the key segment of the fault sound and ignore irrelevant speech or persistent background noise as follows: wherein, is a fusion weight parameter, is a frequency dimension attention module, is a time dimension attention module, balancing the importance of frequency and time attention; at the same time, a residual connection is introduced after each attention sublayer, which directly transmits the input to the output end, alleviates the difficulty problem of deep network training, and adopts layer normalization to stabilize the activation value distribution, thereby enhancing the response ability of the model to weak fault sound and avoiding being submerged in other strong sound sources due to low signal intensity.

8. A switchgear fault perception and intelligent inspection system as claimed in claim 6, wherein, The lightweight VATT audio-video fusion positioning module comprises an STFT processing unit, a video preprocessing unit, a feature extraction unit, a semantic alignment unit and a positioning output unit; The STFT processing unit converts the separated fault audio into a frequency spectrum graph; The video preprocessing unit adjusts the resolution, normalizes and crops the key area of the video frame; The feature extraction unit extracts audio spectrum graph features and video frame features respectively; The semantic alignment unit maps the audio features and the video features to a unified representation space; The positioning output unit generates the probability distribution of the fault sound source in the image space; The lightweight VATT audio-video fusion positioning model is: adjusting the resolution of the original video frame image to a unified size, thereby reducing the calculation complexity, normalizing the pixel value to the range of [0, 1], automatically cropping the key area according to the outline of the switch cabinet, focusing on the target area, using a CNN model to extract features from the preprocessed video frame image, and outputting a visual feature vector consistent with the dimension of the audio feature, ensuring the feasibility of subsequent fusion; Map audio features and video features to the same low-dimensional shared embedding space. As shown in the following formula: ,in, This is the transpose of the audio shared feature vector. To share feature vectors between videos, To share spatial dimensions, This refers to the number of key areas in the video. For the first Visual shared features of each region are calculated through shared spatial mapping and attention weights. It is the natural logarithm function; With the cross-modal attention mechanism, the "specific frequency event" in the audio is semantically associated with the "corresponding physical position" in the video. When the audio detects a discharge sound of a certain frequency, the system can match to the flicker or heating sign of a certain contact area in the video, as follows: wherein, is a normalization function, is a classic Transformer network, which receives the aligned audio and video features through the Transformer network, predicts the spatial probability distribution of the sound source in the image coordinate system based on the learned joint representation, and outputs a two-dimensional heat map with the same size as the input video frame. The value of each pixel point in the heat map represents the probability of the position being the sound source. The higher the probability, the more likely it is the location of the fault occurrence. Non-maximum suppression is performed on the heat map to determine the most likely fault point coordinates (x, y) and output its specific position in the image. Combined with historical data, it is determined whether it is a repeated fault or a new abnormality. The heat map is superimposed on the original video screen to form an audio and video fusion view, which labels the fault point position, confidence and fault type suggestion in the image, supports saving as an image file or a video segment, and uploading to the background management system for maintenance personnel to view.

9. A switchgear fault perception and intelligent patrol system according to claim 8, characterized in that, It also comprises a feature fusion module, which uses a cross-modal attention mechanism to map the audio features and the video features to the same low-dimensional shared embedding space, and predicts the spatial probability distribution of the sound source in the image coordinate system through a neural network.

10. A switchgear fault perception and intelligent inspection system as claimed in claim 6, wherein, The edge computing unit adopts an NVIDIA Jetson series processor, deploys a lightweight deep learning model, and realizes real-time sensing and positioning of switch cabinet partial discharge, mechanical looseness, arc discharge, insulation aging and foreign object discharge faults.

Images