Substation inspection system and method based on image detection

Abstract

The invention provides a transformer substation inspection system and method based on image detection, and belongs to the technical field of transformer substation inspection. In order to solve the technical problems, the transformer substation inspection system hardware structure based on image detection and the improvement of the inspection method are provided, and the adopted technical scheme is that a row of advancing wheels are installed on each of the two sides of an inspection vehicle frame, and rotating shafts of the advancing wheels located at the ends are in transmission connection with an output shaft of a driving motor through a driving gear; a crawler belt is further arranged on the outer side of the advancing wheel; a control module, a power supply module, a communication module, a display module and a data acquisition module are mounted on the frame, the control module comprises a microcontroller packaged on a circuit board, the power supply module comprises a solar panel and a storage battery which are arranged on the frame, the output end of the solar panel is connected with the charging end of the storage battery, and the communication module is connected with the display module. The communication module comprises a wireless network card arranged on the frame; the method is applied to substation inspection.

Description

Technical Field

[0001] The invention provides a substation inspection system and method based on image detection, belonging to the technical field of substation inspection. Background Art

[0002] With the continuous expansion of investment and construction of power grid facilities, it is necessary to frequently inspect and maintain the existing power facilities to ensure that they can operate reliably and efficiently during peak power consumption periods. Among them, inspection of substations is one of the important maintenance tasks.

[0003] At present, the main inspection method for 35kV substations is regular manual inspections, which has three defects: first, there are gaps in manual inspections, and equipment hidden dangers cannot be discovered in a timely and accurate manner. In particular, manual inspections are more difficult in bad weather, and the inspection efficiency is seriously insufficient; second, manual inspections are heavily dependent on inspection experience, and relevant personnel are irreplaceable, with a long rotation cycle. In addition to the use of detection equipment, inspection efficiency still requires relevant personnel to identify hidden dangers, and it is impossible to quickly generate accurate inspection reports; third, manual inspections themselves have safety risks, which are particularly prominent in bad weather. Safety accidents are very likely to occur during personnel inspections, but this is also the time when substations are most likely to fail or have accidents.

[0004] At present, unmanned inspection and patrol methods are also adopted, but the relevant patrol stations are basically deployed in 110kV and above stations. The progress of intelligent deployment of 35kV stations can no longer meet the needs of reliable power supply. The operation and maintenance workload of grassroots power units is heavy, and daily patrol work can barely be coped with. However, during peak load periods, nighttime hours, and bad weather, the inherent problems such as insufficient operation and maintenance forces, poor patrol efficiency, and poor stability are particularly prominent, resulting in the inability to report the summary work after the inspection in a timely and rapid manner, and the fault is not handled in time, which in turn causes large-scale power outages and affects the power supply quality. Summary of the invention

[0005] In order to overcome the deficiencies in the prior art, the present invention aims to solve the technical problem of providing an improvement in the hardware structure and inspection method of a substation inspection system based on image detection.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a substation inspection system based on image detection, comprising an inspection vehicle, the inspection vehicle comprising a frame, a row of traveling wheels are installed on both sides of the frame, the rotating shaft of the traveling wheel located at the end is connected to the output shaft of the driving motor through a driving gear, and a crawler is also arranged on the outer side of the traveling wheel;

[0007] The frame is equipped with a control module, a power module, a communication module, a display module and a data acquisition module, wherein the control module includes a microcontroller packaged on a circuit board, the power module includes a solar panel and a battery arranged on the frame, the output end of the solar panel is connected to the charging end of the battery, the communication module includes a wireless network card arranged on the frame, the display module includes a display screen and an indicator light arranged on the frame, and the data acquisition module includes a pan-tilt mounted on the frame, the pan-tilt is connected to an operating table via a pan-tilt motor, and a camera is installed on the operating table;

[0008] The microcontroller is connected to the drive motor, the wireless network card, the display screen, the indicator light, the pan / tilt motor, and the camera through wires;

[0009] The power input terminal of the microcontroller is connected to the output terminal of the battery;

[0010] The inspection vehicle establishes a wireless connection with the inspection server through a wireless network card to exchange data.

[0011] An electromagnetic sensor is also arranged at the bottom of the frame, and a signal output end of the electromagnetic sensor is connected to a microcontroller after passing through a signal processing circuit.

[0012] The battery is also provided with a temperature sensor, and the output end of the battery is also provided with a voltage sensor. The signal output ends of the temperature sensor and the voltage sensor are both connected to the microcontroller through wires.

[0013] The pan-tilt head is specifically a two-axis pan-tilt head, and the operation platform can be rotated up and down or left and right by controlling the actions of two pan-tilt motors.

[0014] The camera is specifically an infrared thermal imaging camera, or a common optical camera.

[0015] The driving motor is specifically a Hall-encoded DC motor.

[0016] A method for performing inspection by a substation inspection system based on image detection includes the following inspection steps:

[0017] Step 1: The inspection vehicle is remotely controlled through the inspection server to automatically inspect the substation in accordance with the electromagnetic inspection method. During the inspection process, the camera on the pan / tilt is controlled to collect images along the inspection route in real time, and the camera transmits the collected images wirelessly back to the inspection server.

[0018] Step 2: The inspection server pre-processes the collected image data:

[0019] Step 2.1: Use convolutional neural network to extract visual features from the image:

[0020] Step 2.2: Convert the original infrared image through the HSV color space transformation, convert the RGB pixel points into a pixel matrix in the HSV color space, judge each of the three HSV components of each pixel point according to the segmentation threshold, and present the judgment result in the form of a binary image;

[0021] The segmentation threshold is 10 < H < 150, 10 < S < 100, 200 < V < 255. If all three HSV components are within this range, then the pixel point is mapped to white in the binary image. Then search whether there are white pixel points around the white pixel point, calculate the size of the set of adjacent white pixel points. If the set exceeds the threshold, it is judged that there are defects in the power equipment in the infrared image. Otherwise, it is judged that the power equipment in the infrared image is operating normally;

[0022] Step 2.3: Adopt a defect area segmentation algorithm based on HSV to separate the defect area in the infrared image of the faulty power equipment from the background;

[0023] Step 3: After the preprocessing is completed, use the AlexNet model to perform convolution operations on the input image with multiple convolutional kernels, then introduce non-linear characteristics through the ReLU activation function, and reduce the spatial size of the feature map through max pooling or average pooling. After multiple convolutional and pooling operations, a set of high-dimensional feature maps are obtained. Finally, the judgment of defect hidden dangers is completed through training and testing;

[0024] Step 4: Input the visual feature vector into a multi-modal Transformer network. This network calculates the importance weights of each modal feature through the self-attention mechanism, dynamically adjusts these weights to highlight important information, and finally generates and sends an inspection report.

[0025] The specific method for converting the RGB pixel points into a pixel matrix in the HSV color space in Step 2.2 is as follows:

[0026] The conversion calculation formula for converting the RGB space of the image to the HSV space is:

[0027] V = max(R, G, B);

[0028]

[0029] To extract the defect area in the infrared image, by modifying the calculation formula of the H value in the conversion formula and performing normalization processing, the new calculation formula is obtained as:

[0030]

[0031] The segmentation thresholds of the three HSV components are set to 10 < H < 150, 10 < S < 100, and 200 < V < 255 for judgment.

[0032] The specific steps of the HSV-based defect area segmentation algorithm adopted in step 2.3 are as follows:

[0033] 1) Initialize the image clustering center points (l i , a i , b i , x i , y i ), T such that each clustering center point is evenly distributed in the image at a distance of S;

[0034] 2) Select the pixel point with the minimum gradient in each 3×3 neighborhood as the clustering center point in that neighborhood;

[0035] 3) Through the SLIC superpixel segmentation algorithm, calculate the distance metric from each pixel point to the clustering center point, calculate the distance metric and select the minimum distance metric to assign the pixel point to the corresponding clustering center point;

[0036] 4) Normalize the processed image and calculate the new H value according to the new formula for calculating the H value;

[0037] 5) Set the threshold. If the three HSV components satisfy 10 < H < 150, 10 < S < 100, and 200 < V < 255, separate the device defect area in the infrared image through this threshold;

[0038] 6) Use the OTSU threshold segmentation algorithm to segment the original image to obtain the binary image I b ;

[0039] 7) Connect each isolated defect point position to determine the defect area R S ;

[0040] 8) In the binary image I b , find each connected region R0. If in the defect area R0 ∩ R S > 0.8R S , then this area is the defective power equipment area to be extracted.

[0041] Compared with the prior art, the invention has the following beneficial effects: the inspection vehicle provided by the invention can perform inspection tasks uninterruptedly at all times without manual operation, saving a lot of human resources and time costs. The inspection vehicle is equipped with a variety of sensors and cameras, which can monitor the operating status and environmental conditions of 35kV substation equipment in real time. Through the remote monitoring system, operation and maintenance personnel can view equipment data at any time, discover abnormal conditions in time and take corresponding measures to solve the problems of insufficient manpower, low efficiency and quality in the traditional inspection method of 35kV substations in grassroots units; the inspection system provided by the invention can provide a judgment basis for the operation and maintenance management of substations by collecting and analyzing a large amount of inspection data, combining infrared thermal imaging, video monitoring and other multimodal data analysis, especially in the process of collecting and processing inspection images, judging whether there are defects in the power equipment in the infrared image through the corresponding algorithm, and completing the judgment of defects and hidden dangers through the obtained high-dimensional feature map after training and testing; the invention is not only applicable to 35kV substations, but also can be applied to drone inspections, security inspections, remote monitoring and inspection of plant equipment in the petroleum industry and coal enterprises after fine-tuning and deployment by training a large language model by replacing the data set. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The present invention will be further described below in conjunction with the accompanying drawings:

[0043] Figure 1 It is a structural schematic diagram of the inspection vehicle of the present invention;

[0044] Figure 2 It is a schematic diagram of the circuit structure of the present invention;

[0045] Figure 3 Schematic diagram of RGB pixels in an embodiment of the present invention;

[0046] Figure 4 Schematic diagram of a pixel matrix in the HSV color space according to an embodiment of the present invention;

[0047] Figure 5 This is a schematic diagram of the structure of the AlexNet network used in the embodiment of the present invention;

[0048] In the figure: 1 is a frame, 2 is a driving motor, 3 is a crawler, 4 is a microcontroller, 5 is a solar panel, 6 is a battery, 7 is a wireless network card, 8 is a display screen, 9 is an indicator light, 10 is a pan-tilt head, 11 is a pan-tilt head motor, 12 is an operating table, 13 is a camera, 20 is an inspection server, 21 is an electromagnetic sensor, and 22 is a signal processing circuit. DETAILED DESCRIPTION

[0049] like Figure 1 and Figure 2As shown, the present invention provides a remote inspection system for 35kV substations. The system integrates functions such as infrared thermal imaging, remote control, and real-time video feedback. It can complete equipment inspections at a relatively low cost during key periods such as nighttime or peak load periods. It also has autonomous line inspection and automatic recharging (power supply) functions, which can ensure continuous operation and improve the inspection efficiency and safety of substations.

[0050] The inspection system includes multiple functional modules, including travel module, automatic line inspection module, camera module, PTZ control module, power module, detection module, etc.

[0051] The traveling module adopts a crawler chassis. Since the inspection vehicle is heavy and has a low center of gravity, the crawler track can better cross obstacles and adapt to the more complex terrain of the substation, making it have better stability and balance when inspecting the substation. The wireless network card installed on the frame can realize remote control operation function, which improves the operational convenience and safety of the inspection system.

[0052] The automatic line patrol module adopts the electromagnetic line patrol method. By laying a wire with alternating current under the path, an alternating electromagnetic field is generated. After the electromagnetic sensor installed on the inspection vehicle detects the electromagnetic field, the signal processing circuit converts the magnetic field change signal into an electrical signal that can be processed by the microcontroller, thereby realizing the tracking of the path; the electromagnetic line patrol method is suitable for various environments and terrains, and is not affected by external weather, with high reliability and adaptability; the laying of electromagnetic line patrol does not require digging the ground, so the 35kV substation does not need to be reconstructed and renovated, avoiding the safety risks in the construction process; the electromagnetic line patrol module can be equipped with an advanced software system to realize real-time recording of detection results, which is convenient for subsequent inspection analysis and reporting.

[0053] The camera module consists of an infrared sensor and a conventional camera. The infrared sensor can detect heat sources very sensitively and can realize non-contact real-time temperature detection within a certain distance, which can meet the temperature measurement needs of the substation; the conventional camera can transmit the captured images back to the control end in real time, which can meet the needs of operation and maintenance personnel to assist in manual inspections.

[0054] A two-axis gimbal is used for camera installation and fixed support equipment. Specifically, it is a gimbal that can achieve horizontal (left and right) and pitch (up and down) movements. The gimbal has a simple structure, which can meet the shooting needs of substation inspections and also reduces maintenance and repair costs.

[0055] The power modules installed in the inspection vehicle are specifically batteries and solar panels, which can directly convert solar energy into electrical energy, help save energy and reduce emissions, and provide a clean and efficient power source for the entire inspection system.

[0056] The detection module provided by the present invention includes corresponding patrol software, which includes a monitoring interface and can realize electromagnetic line patrol function, manual remote control function, recording and video recording function, infrared thermal imaging function, pan-tilt control function, etc.; the software further has a built-in image processing algorithm, which can analyze the detection picture based on preset judgment rules and complete the corresponding patrol tasks.

[0057] Furthermore, the above-mentioned automatic recharge function enables the inspection vehicle to have an automatic return and charging control mechanism. The specific process is as follows: the STM32 microcontroller will continuously monitor the battery voltage and reduce the battery voltage to an acceptable range of the ADC through a voltage divider resistor network. At the same time, a digital temperature sensor is used to monitor the battery temperature and read the data through the I2C interface. The STM32 initializes the ADC module, timer and UART interface, and sets the appropriate sampling rate, baud rate and other communication parameters. In the timer interrupt service program, the STM32 periodically reads the battery voltage and temperature data, and packages these data into a fixed format message, and periodically reports to the Raspberry Pi through the UART interface. When it is detected that the battery voltage is lower than the set threshold, the STM32 immediately stops the current task, triggers the return charging task, and sends a low battery alarm and a notification of returning to the charging point to the Raspberry Pi through the UART. The UART interface is initialized by the Raspberry Pi, the same communication parameters are set to match the STM332, and data from the STM32 is continuously received through the monitoring thread or process. The Raspberry Pi parses the received data packets, extracts the battery voltage and temperature information, updates the status display or log file, and records the event and may trigger other emergency measures when a low battery alarm is detected. Throughout the process, the inspection system ensures effective monitoring of the battery status and automatic recharging functions, thereby ensuring that the inspection vehicle can continue to operate without being affected.

[0058] Furthermore, the present invention is based on multimodal data fusion and deep learning technology to achieve efficient inspection task management and anomaly detection. The specific process is: first, the system collects various types of data such as images, videos and temperature through devices such as infrared thermal imaging cameras, ordinary cameras and sensors. Operators can interact with the system through natural language, issue inspection tasks or conduct queries, and process these multimodal data in the preprocessing stage. Specifically, the collected image and video data are subjected to preprocessing steps such as denoising, enhancement and standardization.

[0059] During preprocessing, a convolutional neural network (CNN) is specifically used to extract visual features from images and videos. First, for video data, key frames are uniformly extracted (e.g., 5 frames per second), and each frame is subjected to the following image preprocessing steps: infrared thermal imaging data is preprocessed, including normalizing pixel values ​​to between 0 and 1 and adjusting the image size to a uniform size (e.g., 224x224 pixels).

[0060] As Figure 3 and Figure 4 shown, the original infrared image is then converted through the HSV color space, and the RGB pixel points are converted into a pixel matrix in the HSV color space; by judging the three HSV components of each pixel point according to the segmentation threshold, and presenting the judgment result in the form of a binary image. The segmentation threshold is 10 < H < 150, 10 < S < 100, 200 < V < 255. If all three HSV components are within this range, then this pixel point is mapped to white in the binary image; search whether there are white pixel points around the white pixel point, calculate the size of the set of adjacent white pixel points. If this set exceeds the threshold, it is determined that there are defects in the power equipment in this infrared image. Otherwise, it is determined that the power equipment in this infrared image is operating normally.

[0061] The conversion calculation formula for converting the RGB space of the traditional image to the HSV space is as follows:

[0062] V = max(R, G, B);

[0063]

[0064] In order to extract the defective area in the infrared image, the calculation formula of the H value in the traditional conversion formula can be modified and normalized to obtain a new calculation formula as follows:

[0065]

[0066] Thus, after setting the segmentation threshold of the three HSV components to 10 < H < 150, 10 < S < 100, 200 < V < 255, the judgment is made.

[0067] In order to improve the accuracy of the subsequent image classification algorithm, the present invention also uses the defect area segmentation algorithm based on HSV to separate the defect area in the infrared image of the faulty power equipment from the background. Since simply analyzing the fault area cannot determine its defect type, it is necessary to perform segmentation based on mathematical morphology according to the location of the defect area. Through this method, the defective power equipment and the background area in the infrared image can be segmented, so as to reduce the interference caused by the background area in the infrared image to the subsequent defect type classification.

[0068] The steps of the defect area extraction algorithm based on HSV are as follows:

[0069] 1) Initialize the image clustering center points (l i , a i , b i , x i , y i ) T, making each cluster center point evenly distributed in the image at a distance of S.

[0070] 2) Select the pixel point with the minimum gradient in each 3×3 neighborhood as the cluster center point within this neighborhood.

[0071] 3) Calculate the distance metric by the method of calculating the distance metric from each pixel point to the cluster center point in the SLIC superpixel segmentation algorithm and select the minimum distance metric to assign the pixel point to the corresponding cluster center point.

[0072] 4) Normalize the processed image and calculate the new H value according to the new formula for calculating the H value.

[0073] 5) Set a threshold. If the three HSV components satisfy 10 < H < 150, 10 < S < 100, 200 < V < 255, separate the device defect area in this infrared image through this threshold.

[0074] 6) Use the OTSU threshold segmentation algorithm to segment the original image to obtain a binary image I b .

[0075] 7) Since there may be one or several defect points in the segmented binary image, and analyzing a single defect point cannot determine the defect type of the power device defect area, and there are obvious differences between the device defect area and the background area in the infrared image, so connect each isolated defect point according to the mathematical morphology closing operation to determine the defect area R S .

[0076] 8) Find each connected region R0 in the binary image I b . If in the defect area R0∩R S > 0.8R S , then this area is the defect power device area to be extracted.

[0077] After determining the extraction area, further analyze the defect hidden danger of the area image based on the AlexNet model defect classification algorithm. The network structure of AlexNet is as Figure 5 shown:

[0078] The whole network includes eight network layers, which are composed of five convolutional layers and three fully connected layers. There is a local response normalization layer, namely LRN, after the first two convolutional layers. The LRN layer can create a competition mechanism for the activities of local neurons and enhance the generalization ability of the model. There is a RELU activation function after each network layer, which solves the problem of gradient diffusion caused by Sigmoid when the network is deep. On the basis of the LeNet model, the average pooling layer is changed to the maximum pooling layer, which avoids the blurring effect caused by average pooling, and makes the step size smaller than the size of the pooling kernel, which increases the richness of features. During training, the DropOut mechanism is added to randomly ignore a part of the neurons to avoid overfitting of the model. DropOut is used in the last few fully connected layers. In order to reduce the overfitting phenomenon of the model and improve the generalization ability of the model, a data enhancement mechanism is used when inputting model data, that is, a 224×224 area is randomly cut from the original 256×256 input image instead of the original image as the new input of the model. This measure is equivalent to increasing the amount of data by 2048 times.

[0079] There is an LRN layer after the convolutional layer and pooling layer of the first two layers of the network, that is, local response normalization. The specific calculation formula is:

[0080]

[0081] In the formula, a represents the output result after the convolution operation and the pooling operation. The structure of the result is a four-dimensional array, namely [batch, height, width, channel], where batch is the number of batches of images for batch training, height and width are the height and width of the image, and channel is the number of channels of the image, which can also be understood as the number of neurons a output by a certain image in the batch after the convolution operation and the pooling operation. i x,y Represents a position [a, b, c, d] of the image in this output structure, corresponding to the four-dimensional array above, which is a point at a certain width and a certain height under a certain channel in a certain image, where N is the number of channels after convolution and pooling operations;

[0082] In the formula The direction of accumulation is along the channel direction, that is, the square sum is accumulated along the third dimension of the output result.

[0083] The present invention adopts the AlexNet model to perform convolution operations on the input image using multiple convolution kernels, then introduces nonlinear characteristics through the ReLU activation function, and reduces the spatial size of the feature map through maximum pooling or average pooling. After multiple layers of convolution and pooling operations, a set of high-dimensional feature maps is obtained, and finally the defect hidden danger judgment work is completed after training and testing.

[0084] The present invention finally inputs the visual feature vector into a multimodal Transformer network, which calculates the importance weight of each modal feature through the self-attention mechanism and dynamically adjusts these weights to highlight the most important information. The self-attention mechanism allows the model to focus on the most relevant part when processing the data of each modality, thereby improving the fusion effect. The fused data is used to generate a detailed inspection report, which contains key image evidence, text descriptions, anomaly detection results and recommended measures, and is sent to relevant personnel through the system interface or email.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A substation inspection system based on image detection, comprising an inspection vehicle, characterized in that: The inspection vehicle comprises a frame (1), a row of traveling wheels are mounted on both sides of the frame (1), the rotating shafts of the traveling wheels at the ends are connected to the output shafts of the driving motors (2) via driving gears, and tracks (3) are also arranged on the outsides of the traveling wheels; The frame (1) is equipped with a control module, a power module, a communication module, a display module and a data acquisition module, wherein the control module comprises a microcontroller (4) packaged on a circuit board, the power module comprises a solar panel (5) and a storage battery (6) arranged on the frame (1), the output end of the solar panel (5) is connected to the charging end of the storage battery (6), the communication module comprises a wireless network card (7) arranged on the frame (1), the display module comprises a display screen (8) and an indicator light (9) arranged on the frame (1), and the data acquisition module comprises a pan / tilt platform (10) installed on the frame (1), the pan / tilt platform (10) is connected to an operating table (12) via a pan / tilt platform motor (11), and a camera (13) is installed on the operating table (12); The microcontroller (4) is connected to the drive motor (2), the wireless network card (7), the display screen (8), the indicator light (9), the pan / tilt motor (11), and the camera (13) respectively through wires; The power input terminal of the microcontroller (4) is connected to the output terminal of the storage battery (6); The inspection vehicle exchanges data after establishing a wireless connection with the inspection server (20) via a wireless network card (7).

2. A substation inspection system based on image detection according to claim 1, characterized in that: An electromagnetic sensor (21) is also provided at the bottom of the vehicle frame (1), and a signal output end of the electromagnetic sensor (21) is connected to a microcontroller (4) after passing through a signal processing circuit (22).

3. The substation inspection system based on image detection according to claim 1 is characterized in that: A temperature sensor is also provided on the storage battery (6), and a voltage sensor is also provided at the output end of the storage battery (6). The signal output ends of the temperature sensor and the voltage sensor are both connected to the microcontroller (4) via wires.

4. The substation inspection system based on image detection according to claim 1 is characterized in that: The pan / tilt platform (10) is specifically a two-axis pan / tilt platform, and by controlling the actions of two pan / tilt motors (11), the operating platform (12) can be rotated up and down or left and right.

5. The substation inspection system based on image detection according to claim 1 is characterized in that: The camera (13) is specifically an infrared thermal imaging camera, or a common optical camera.

6. The substation inspection system based on image detection according to claim 1 is characterized in that: The driving motor (2) is specifically a Hall-encoded DC motor.

7. The method for performing inspection by a substation inspection system based on image detection according to claim 1, characterized in that: The inspection steps include the following: Step 1: The inspection vehicle is remotely controlled through the inspection server (20) to automatically inspect the substation in accordance with the electromagnetic inspection method. During the inspection process, the camera on the pan / tilt is controlled to collect images along the inspection route in real time, and the camera transmits the collected images wirelessly back to the inspection server (20); Step 2: The inspection server (20) pre-processes the collected image data: Step 2.1: Use convolutional neural network to extract visual features from the image: Step 2.2: The original infrared image is converted into HSV color space, and the RGB pixels are converted into a pixel matrix in HSV color space. The three HSV components of each pixel are judged according to the segmentation threshold, and the judgment result is presented in the form of a binary image; The segmentation threshold is 10 < H < 150, 10 < S < 100, 200 < V < 255. If all three HSV components are within this range, then the pixel is mapped to white in the binary image. Then, search for white pixels around the white pixel, calculate the size of the set of adjacent white pixels. If the set exceeds the threshold, it is determined that there is a defect in the power equipment in the infrared image; otherwise, it is determined that the power equipment in the infrared image is operating normally. Step 2.3: Use the HSV-based defect area segmentation algorithm to separate the defect area from the background in the infrared image of the faulty power equipment. Step 3: After preprocessing, use the AlexNet model to perform convolution operations on the input image using multiple convolutional kernels, then introduce non-linearity through the ReLU activation function, and reduce the spatial size of the feature map through max pooling or average pooling. After multiple convolutional and pooling operations, a set of high-dimensional feature maps is obtained. Finally, the determination of defect hazards is completed through training and testing. Step 4: Input the visual feature vector into a multi-modal Transformer network. This network calculates the importance weights of each modal feature through the self-attention mechanism, dynamically adjusts these weights to highlight important information, and finally generates and sends an inspection report.

8. A substation inspection method based on image detection according to claim 7, characterized in that: The specific method for converting RGB pixels to a pixel matrix in the HSV color space in Step 2.2 is as follows: The conversion calculation formula for converting the RGB space of the image to the HSV space is: V = max((R, G, B To extract the defect area in the infrared image, by modifying the calculation formula of the H value in the conversion formula and performing normalization processing, the new calculation formula is obtained as: Therefore, after setting the segmentation threshold of the three HSV components to 10 < H < 150, 10 < S < 100, 200 < V < 255, the judgment is made.

9. The substation inspection method based on image detection according to claim 7 is characterized in that: The specific steps of the HSV-based defect area segmentation algorithm used in Step 2.3 are as follows: 1) Initialize the image cluster center point (l i ,a i ,b i ,x i ,y i ) T , so that each cluster center point is evenly distributed in the image with a distance S; 2) Select the pixel point with the minimum gradient in each 3×3 neighborhood as the clustering center point in this neighborhood. 3) Using the SLIC superpixel segmentation algorithm, calculate the distance metric from each pixel to the cluster center point and calculate the distance metric And select the minimum distance metric to assign pixels to the corresponding cluster center points; 4) Perform normalization processing on the processed image and calculate the new H value according to the new formula for calculating the H value. 5) Set the threshold. If the three HSV components satisfy 10 < H < 150, 10 < S < 100, 200 < V < 255, separate the equipment defect area in the infrared image through this threshold. 6) Use the OTSU threshold segmentation algorithm to segment the original image and obtain a binary image I b ; 7) Connect each isolated defect point to determine the defect area R S ; 8) In the binary image I b Find each connected region R0 in the defect region. If R0∩R S >0.8R S , then this area is the defective power equipment area that needs to be extracted.

Images