Direct-current valve hall equipment discharge fault detection method and device and computer equipment

Abstract

The invention relates to a direct-current valve hall equipment discharge fault detection method and device and computer equipment, and the method comprises the steps: carrying out the image fusion processing of an original visible light image and an original ultraviolet image of direct-current valve hall equipment, and obtaining a plurality of fusion images of different frequency sub-bands; and inputting the fused images of the different frequency sub-bands into a pre-constructed fault detection model for fault detection to obtain a fault detection result corresponding to the target frequency sub-band. According to the invention, the visible light image can clearly present the macroscopic features of the device, and the ultraviolet image can reflect the potential discharge fault area. And after the two are fused, the fused image highlights feature information closely related to the discharge fault. And the fused image is input into the pre-constructed fault detection model for fault detection, and the model can accurately capture fault related characteristics, so that the misjudgment condition caused by insufficient single image information is avoided, and the accuracy of direct current valve hall equipment discharge fault detection is remarkably improved.

Description

Technical Field

[0001] The present application relates to the technical field of fault detection, and particularly to a method, device and computer equipment for detecting discharge faults of DC valve hall equipment. Background Art

[0002] As the core part of the high-voltage DC transmission system, the DC valve hall undertakes the key task of converting alternating current and direct current. The stable operation of its internal components plays a decisive role in the reliability of the entire transmission system. However, in practical applications, the insulation system in the DC valve hall faces many severe challenges. Due to the high-voltage environment and the complex electromagnetic environment, abnormal discharge phenomena frequently occur in the weak insulation areas of the DC valve hall components. Such abnormal discharge will cause a series of serious consequences.

[0003] Traditional fault detection methods mainly rely on regular maintenance inspections and off-line tests. However, it is difficult to capture early fault signals in a timely manner by this method, resulting in a serious lag in fault diagnosis and greatly increasing the risk of further fault expansion. Summary of the Invention

[0004] Based on this, it is necessary to provide a method, device and computer equipment for detecting discharge faults of DC valve hall equipment that can solve the above problems in view of the above technical problems.

[0005] In a first aspect, the present application provides a method for detecting discharge faults of DC valve hall equipment, the method comprising:

[0006] Performing image fusion processing on the pre-acquired original visible light image and original ultraviolet image of the DC valve hall equipment to obtain fusion images of multiple different frequency sub-bands;

[0007] Inputting the fusion images of each different frequency sub-band into a pre-constructed fault detection model for fault detection to obtain a fault detection result corresponding to the target frequency sub-band; the fault detection result is used to characterize whether there is a discharge fault in the DC valve hall equipment and the type and severity of the fault.

[0008] In one embodiment, the performing image fusion processing on the pre-acquired original visible light image and original ultraviolet image of the DC valve hall equipment to obtain fusion images of multiple different frequency sub-bands includes:

[0009] Performing feature extraction processing on the original visible light image and the original ultraviolet image respectively to obtain visible light feature data and ultraviolet feature data;

[0010] Performing feature relationship extraction processing on the visible light feature data and the ultraviolet feature data to obtain the feature change relationship between the original visible light image and the original ultraviolet image;

[0011] Enhance the visible light feature data and the ultraviolet feature data respectively to obtain enhanced visible light data and enhanced ultraviolet data;

[0012] Fuse the enhanced visible light data and the enhanced ultraviolet data according to the feature change relationship to obtain the fused images of each frequency sub-band.

[0013] In one embodiment, the above-mentioned enhancing the visible light feature data and the ultraviolet feature data respectively to obtain enhanced visible light data and enhanced ultraviolet data includes:

[0014] Determine the gray-scale data of the visible light feature data and the gray-scale data of the ultraviolet feature data respectively;

[0015] Determine the gray-scale distribution probability of the original visible light image according to the gray-scale data of the visible light feature data and the total number of pixels of the original visible light image, and determine the gray-scale distribution probability of the original ultraviolet image according to the gray-scale data of the ultraviolet feature data and the total number of pixels of the original ultraviolet image;

[0016] Determine the cumulative distribution rate of each gray-scale value in the original visible light image according to the gray-scale distribution probability of the original visible light image, and determine the cumulative distribution rate of each gray-scale value in the original ultraviolet image according to the gray-scale distribution probability of the original ultraviolet image;

[0017] Perform histogram equalization processing on the cumulative distribution rate of each gray-scale value in the original visible light image and the cumulative distribution rate of each gray-scale value in the original ultraviolet image respectively to obtain enhanced visible light data and enhanced ultraviolet data.

[0018] In one embodiment, the above-mentioned fusing the enhanced visible light data and the enhanced ultraviolet data according to the feature change relationship to obtain the fused images of each frequency sub-band includes:

[0019] Use the non-downsampling shearlet transform algorithm to decompose the enhanced visible light data and the enhanced ultraviolet data respectively to obtain visible light images of different frequency sub-bands and ultraviolet images of different frequency sub-bands;

[0020] Fuse the visible light images of each frequency sub-band with the ultraviolet images of the corresponding frequency sub-band to obtain fused image data of different frequency sub-bands.

[0021] In one embodiment, the above-mentioned fused image data of different frequency sub-bands includes low-frequency fused image data and high-frequency fused image data; fusing the visible light images of each frequency sub-band with the ultraviolet images of the corresponding frequency sub-band to obtain fused image data of different frequency sub-bands includes:

[0022] For the visible light image of the low-frequency sub-band and the ultraviolet image of the low-frequency sub-band, a Gaussian membership function is used to perform weighted fusion processing on the visible light image of the low-frequency sub-band and the ultraviolet image of the low-frequency sub-band to obtain low-frequency fusion image data;

[0023] For the visible light image of the high-frequency sub-band and the ultraviolet image of the high-frequency sub-band, an impulse coupled neural network is used to perform image fusion processing on the visible light image of the high-frequency sub-band and the ultraviolet image of the high-frequency sub-band to obtain high-frequency fusion image data.

[0024] In one embodiment, the above fault detection model includes a feature extraction layer, a feature fusion layer, and a classification layer; the fusion images of each different frequency sub-band are input into a pre-constructed fault detection model for fault detection to obtain the fault detection results corresponding to the target frequency sub-band, including:

[0025] The fusion images of each different frequency sub-band are input into the feature extraction layer for convolution and pooling processing to obtain the front-end feature map data of each different frequency sub-band;

[0026] The front-end feature map data of each different frequency sub-band are input into the feature fusion layer for cross-channel feature fusion to obtain the feature fusion data of each different frequency sub-band;

[0027] The feature fusion data of each different frequency sub-band are input into the classification layer for classification judgment and fault location processing to obtain the fault detection results corresponding to the target frequency sub-band.

[0028] In one embodiment, the training process of the above fault detection model includes:

[0029] The sample visible light images of different frequency sub-bands and the sample ultraviolet images of different frequency sub-bands obtained in advance are subjected to image fusion processing to obtain the sample fusion data of each different frequency sub-band, and a labeling algorithm is used to perform labeling processing on the sample fusion data of each different frequency sub-band to obtain training sample data;

[0030] The training sample data are input into the initial detection model for fault detection processing to obtain the initial detection results;

[0031] The initial detection results and the true labeling results in the training sample data are input into the loss function for loss calculation processing to obtain the loss value;

[0032] The parameters of the initial detection model are adjusted according to the loss value until the loss value meets the preset conditions and the training ends to obtain the fault detection model.

[0033] In one embodiment, the above method further includes:

[0034] For the visible light images, ultraviolet images, and fused image data in the test set respectively, determine the recognition accuracy and recall rate;

[0035] According to the recognition accuracy and recall rate, determine the recognition effect of the fault detection.

[0036] The above-mentioned direct current valve hall equipment discharge fault detection method, device, and computer equipment perform image fusion processing on the pre-acquired original visible light images and original ultraviolet images of the direct current valve hall equipment to obtain fused images of multiple different frequency sub-bands; input the fused images of each different frequency sub-band into the pre-constructed fault detection model for fault detection to obtain the fault detection results corresponding to the target frequency sub-band. In this application, the visible light image can clearly present the macroscopic features such as the appearance and structure of the equipment, while the ultraviolet image can reflect the potential discharge fault area. After fusing the two, the fused image highlights the feature information closely related to the discharge fault. Inputting the fused image into the pre-constructed fault detection model for fault detection again, the model can accurately capture the fault-related features, thus effectively avoiding misjudgment caused by insufficient single-image information and significantly improving the accuracy of the discharge fault detection of the direct current valve hall equipment. Description of the Drawings

[0037] Figure 1 It is an application environment diagram of the direct current valve hall equipment discharge fault detection method in an embodiment;

[0038] Figure 2 It is a flow schematic diagram of the direct current valve hall equipment discharge fault detection method in an embodiment;

[0039] Figure 3 It is a flow schematic diagram of obtaining fused images of multiple different frequency sub-bands in an embodiment;

[0040] Figure 4 It is a flow schematic diagram of obtaining visible light enhancement data and ultraviolet enhancement data in an embodiment;

[0041] Figure 5 It is a flow schematic diagram of obtaining the fused images of each frequency sub-band in an embodiment;

[0042] Figure 6 It is a flow schematic diagram of obtaining the fault detection results corresponding to the target frequency sub-band in an embodiment;

[0043] Figure 7 It is a flow schematic diagram of the training process of the fault detection model in an embodiment;

[0044] Figure 8 It is a flow schematic diagram of determining the recognition effect of the fault detection in an embodiment;

[0045] Figure 9 It is a structural block diagram of a discharge fault detection device for DC valve hall equipment in an embodiment;

[0046] Figure 10 It is an internal structure diagram of a computer device in an embodiment. Specific implementation manners

[0047] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0048] The DC valve hall equipment discharge fault detection method provided by the embodiments of the present application can be applied to, for example, Figure 1 the application environment shown in the figure. Among them, the terminal 102 communicates with the server 104 through a network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or placed in the cloud or other network servers. The server 104 first performs image fusion processing on the original visible light image and the original ultraviolet image of the DC valve hall equipment sent by the terminal 102 to obtain fusion images of multiple different frequency sub-bands; inputs the fusion images of each different frequency sub-band into a pre-constructed fault detection model for fault detection to obtain a fault detection result corresponding to the target frequency sub-band. Among them, the terminal 102 can be, but is not limited to, an ultraviolet-visible light imager, a camera, an image acquisition device, etc. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers.

[0049] In an exemplary embodiment, as Figure 2 shown in the figure, the present application provides a DC valve hall equipment discharge fault detection method, and taking the method applied to Figure 1 the server 104 in the figure as an example for description, it includes the following steps:

[0050] S201, perform image fusion processing on the original visible light image and the original ultraviolet image of the DC valve hall equipment obtained in advance to obtain fusion images of multiple different frequency sub-bands.

[0051] Among them, the original visible light image is an image obtained by a visible light imaging device (such as an ordinary camera, an industrial camera, etc.) photographing the DC valve hall equipment, which can present information such as the appearance, structure, color, and surface condition of the equipment under visible light conditions for the human eye. The original ultraviolet image is an image obtained by using a device with ultraviolet imaging function (such as an ultraviolet imager) to photograph the DC valve hall equipment, mainly used to capture the ultraviolet light situation generated by the discharge phenomenon during the operation of the equipment. Since ultraviolet light is radiated during the discharge process, the ultraviolet image can reflect the characteristics related to discharge faults such as potential discharge areas and discharge intensities of the equipment, and these information may not be obvious in the visible light image.

[0052] In the embodiment of the present application,

[0053] S202, input the fusion images of each different frequency sub-band into a pre-constructed fault detection model for fault detection to obtain the fault detection results corresponding to the target frequency sub-band.

[0054] Among them, the fault detection results are used to characterize whether there is a discharge fault in the DC valve hall equipment and the type and severity of the fault. The fault detection results are the judgment results output by the fault detection model on whether there is a discharge fault in the DC valve hall equipment, the specific fault type (such as partial discharge, surface discharge, etc.), and the severity of the fault (such as mild, moderate, severe levels).

[0055] In the embodiment of the present application, the server sequentially inputs the fused images of each different frequency sub-band obtained in the previous step into a pre-constructed fault detection model. The fault detection model contains multiple levels of network structures inside. For example, in the front-end feature extraction layer, it will extract features from the input fused image. Different convolutional kernels in the convolutional layer are used to scan the fused image to extract feature information related to equipment faults with different scales and directions, such as abnormal textures that may appear on the equipment surface, unique shapes of discharge areas, etc. Then, through the intermediate feature integration and optimization layer, operations such as integrating, denoising, and strengthening key features are performed on the extracted multi-dimensional features to make the feature representation more accurate and effective. Then, the processed features are passed to the back-end classification and localization layer. In the classification part, according to the feature patterns corresponding to different fault modes learned during previous training, it is judged whether there is a discharge fault in the DC valve hall equipment and the specific fault type is determined (for example, by judging whether the features conform to the feature pattern of partial discharge to determine whether it is a partial discharge fault); in the localization part, according to the position information contained in the features, the position range of the fault area in the fused image is roughly determined. Finally, combining information such as the fault type and position, and then according to the pre-set evaluation rules, the severity of the fault is judged, so as to obtain the fault detection result corresponding to the target frequency sub-band. For example, it is determined that there is a partial discharge fault in the equipment corresponding to the fused image of a certain high-frequency sub-band, the fault position is in a specific area of the image, and the severity of the fault is medium.

[0056] In another implementation, the fused images of different frequency sub-bands are input into a fault detection model with another architecture. This model uses a depthwise separable convolutional network to extract features of the fused image. Compared with the traditional convolutional network, it can extract features more efficiently while reducing the computational amount. After feature extraction, through an attention mechanism module, the model automatically focuses on the areas and features more critical for fault detection in the fused image to improve the effectiveness of the features. Then, a multi-classifier fusion method is used for fault classification, that is, multiple classifiers with different structures (such as decision tree classifiers, support vector machine classifiers, etc.) are used to judge the fault type of the fused image at the same time, and the accuracy of classification is improved by integrating the results of each classifier. For fault localization, a method based on feature matching is adopted. The features in the fused image are matched with known fault feature templates, and the position information corresponding to the most similar template is found as the fault position. Finally, according to the classification and localization results, the severity of the fault is determined according to the set fault severity evaluation criteria, so as to obtain the fault detection result corresponding to the target frequency sub-band.

[0057] The above method for detecting discharge faults in DC valve hall equipment performs image fusion processing on the original visible light image and the original ultraviolet image of the DC valve hall equipment obtained in advance to obtain fusion images of multiple different frequency sub-bands; inputs the fusion images of each different frequency sub-band into a pre-constructed fault detection model for fault detection, and obtains the fault detection results corresponding to the target frequency sub-band. In this application, the visible light image can clearly present macroscopic features such as the appearance and structure of the equipment, while the ultraviolet image can reflect potential discharge fault areas. After fusing the two, the fusion image highlights the feature information closely related to the discharge fault. Inputting the fusion image into the pre-constructed fault detection model for fault detection again, the model can accurately capture the fault-related features, thus effectively avoiding misjudgment caused by insufficient single-image information and significantly improving the accuracy of detecting discharge faults in DC valve hall equipment.

[0058] In an exemplary embodiment, based on the above embodiment, please refer to Figure 3 , the process of the embodiment of this application for performing image fusion processing on the original visible light image and the original ultraviolet image of the DC valve hall equipment obtained in advance to obtain fusion images of multiple different frequency sub-bands includes the following steps:

[0059] S301, respectively perform feature extraction processing on the original visible light image and the original ultraviolet image to obtain visible light feature data and ultraviolet feature data.

[0060] In the embodiment of this application, after the server receives the original visible light image and the original ultraviolet image of the DC valve hall equipment transmitted from the terminal device, it starts to perform feature extraction processing. For the original visible light image, the server uses the feature extraction layer in the convolutional neural network to mine the feature information therein. For example, by sliding convolutional kernels of different sizes on the image for convolution operations, the edge features of the equipment in the image are extracted, such as clear line information like the outer shell contour of the equipment and the boundary of the connecting components; at the same time, texture features can also be extracted, such as the texture pattern of the insulating material on the surface of the equipment; in addition, shape features are also obtained, such as the specific geometric shapes presented by different components, and these are combined to form the visible light feature data. For the original ultraviolet image, a similar convolutional neural network structure is also used. However, since the ultraviolet image mainly focuses on discharge-related features, the features that may reflect discharge phenomena are mainly extracted, such as the abnormally bright areas in the image (because discharge generates ultraviolet light, making the corresponding area have a different brightness performance from the surrounding area in the ultraviolet image), weak light-emitting points and other feature information, and after sorting, the ultraviolet feature data is obtained.

[0061] S302, perform feature relationship extraction processing on the visible light feature data and the ultraviolet feature data to obtain the feature change relationship between the visible light image and the ultraviolet light image.

[0062] In the embodiment of the present application, after the server obtains the visible light feature data and the ultraviolet feature data, it starts to perform feature relationship extraction processing. The server will conduct regional comparison and analysis on the two types of data. For example, for the same insulator component in the DC valve hall equipment, check the conventional features such as the appearance color and texture in the visible light feature data, and then check in the ultraviolet feature data whether there are ultraviolet light bright spots or local brightness change areas generated by discharge at the insulator part. Through comparison and analysis, it is found that in a certain edge area of the insulator, the texture is normal under visible light, but in the corresponding feature data of the ultraviolet image, it shows that there is a weak ultraviolet light enhancement phenomenon in this area, which indicates that there may be a potential discharge risk here, which is a feature change relationship. The server will traverse the feature data area corresponding to the entire device, sort out all similar feature associations, difference situations, etc., so as to obtain the comprehensive feature change relationship between the visible light image and the ultraviolet light image, such as which area's feature change means that a discharge fault may occur, and which changes are caused by normal environmental factors, etc. are all recorded.

[0063] S303. Respectively perform enhancement processing on the visible light feature data and the ultraviolet feature data to obtain visible light enhanced data and ultraviolet enhanced data.

[0064] In the embodiment of the present application, when the server performs enhancement processing on the visible light feature data, it first strengthens the features representing the key structures of the device in the data. For example, for those edge features reflecting the main contour of the device, by adjusting the weight of their feature values, make them more prominent in subsequent processing, so as to clearly identify the overall shape of the device. At the same time, for those texture features that are slightly blurred in the image but may be related to faults, use the method of local contrast enhancement to increase the contrast of the texture and make it easier to be distinguished, to obtain the visible light enhanced data. For the ultraviolet feature data, since many of its features reflecting discharge may be relatively weak, the server uses the method of signal amplification to appropriately amplify those feature data representing ultraviolet light bright spots, light-emitting areas, etc. related to discharge, and enhance their expressiveness in the whole data. In addition, noise reduction processing will also be performed on the ultraviolet feature data to remove noise features generated by environmental interference and other factors, so that the features truly related to the device discharge are more pure and prominent, and finally obtain the ultraviolet enhanced data.

[0065] S304. According to the feature change relationship, perform fusion processing on the visible light enhanced data and the ultraviolet enhanced data to obtain the fusion image of each frequency sub-band.

[0066] In the embodiments of the present application, the server performs fusion processing based on the feature change relationship between the previously obtained visible light image and the ultraviolet light image. For example, for a certain frequency sub-band (assuming it is a low-frequency sub-band, which mainly reflects the overall contour and general structural features of the device), according to the feature change relationship, it is known that there may be potential discharge features under ultraviolet light at some key connection parts of the device. Then, when fusing the visible light enhanced data and the ultraviolet enhanced data, the weight of the corresponding features in the ultraviolet enhanced data will be appropriately increased in the fusion of the low-frequency sub-band at this part, so that the fused low-frequency sub-band image can clearly present the overall contour of the device and also reflect the possible discharge-related features at the key parts. For the high-frequency sub-band (which focuses on reflecting image details and local changes), if it is found that the texture of a certain local area is normal under visible light but there are obvious brightness change features under ultraviolet light, it indicates that the detail changes here may be related to discharge. When fusing, more emphasis will be placed on integrating the high-frequency features of this area in the ultraviolet enhanced data. After performing fusion processing on each frequency sub-band in this way according to the feature change relationship, the fused images of each frequency sub-band are obtained.

[0067] Perform feature-based image approval on the visible light image and the ultraviolet light image. A two-dimensional image can be represented by a matrix. I 1 (x, y) represents the gray value of the image I at the pixel point (x, y). If there are two images, namely I 1 (x, y) and I 2 (x, y), where I 1 (x, y) is used as the reference image, then there is the following transformation relationship. Where g represents the transformation relationship between the two images. Here, a feature-based transformation relationship is adopted. Feature extraction includes feature elements such as edges, corners, points, and lines. The feature change relationship is obtained based on the feature elements of the two images.

[0068] I 2 (x, y) = g(I 1 (x, y)) (1)

[0069] In the embodiments of the present application, first, in the feature extraction stage, the key feature information contained in each of the two types of images can be accurately mined, providing a rich data basis for subsequent in-depth analysis of the device status and avoiding the problem of missing important details that may occur when directly processing the original images. Second, the extraction and processing of feature relationships enable the server to clarify the internal connection and change situation between the visible light image and the ultraviolet light image, which helps to integrate the advantages of the two types of images in a targeted manner during the fusion process, so that the fused image not only retains conventional information such as the device appearance but also highlights the key features related to discharge faults. Third, the enhancement processing of the feature data further optimizes the data quality, strengthening those useful features that may originally be unobvious or easily interfered, and improving the effectiveness of the data in subsequent fusion and fault detection. Finally, the fusion processing is carried out according to the feature change relationship to generate the fused images of each frequency sub-band. Such fused images comprehensively integrate information from different perspectives, from the overall contour to local details, and from the normal appearance to potential fault features, providing a more comprehensive, accurate, and valuable image basis for subsequent fault detection, thereby enhancing the performance and reliability of the entire discharge fault detection system for DC valve hall equipment, ensuring the safe and stable operation of the equipment, and assisting the maintenance personnel to carry out equipment maintenance and fault troubleshooting more efficiently.

[0070] In an exemplary embodiment, based on the above embodiment, please refer to Figure 4 , the embodiments of the present application are related to the process of respectively enhancing the visible light feature data and the ultraviolet feature data to obtain the visible light enhanced data and the ultraviolet enhanced data, including the following steps:

[0071] S401, respectively determine the grayscale data of the visible light feature data and the grayscale data of the ultraviolet feature data.

[0072] In the embodiments of the present application, after the server receives the visible light feature data and the ultraviolet feature data obtained from the previous processing steps, it starts to extract their respective grayscale data. For the visible light feature data, the server traverses all the pixel information contained therein and extracts the grayscale value corresponding to each pixel to form a complete set of grayscale values, that is, the grayscale data of the visible light feature data. For example, in the visible light feature data of the DC valve hall equipment, the pixel grayscale values representing the device shell part may be relatively high, presenting a brighter visual effect, while the pixel grayscale values in some darker areas inside the device are lower, and these grayscale values in different areas together constitute the grayscale data of the visible light feature data. Similarly, for the ultraviolet feature data, the server also extracts the corresponding grayscale values from the pixel information related to ultraviolet imaging contained therein in a similar manner to form the grayscale data of the ultraviolet feature data. For example, in the ultraviolet image, the pixel grayscale values corresponding to the discharge area may be significantly different from those of the surrounding normal areas, and these grayscale values are all included in this grayscale data.

[0073] S402. Determine the gray - level distribution probability of the original visible - light image based on the gray - level data of the visible - light feature data and the total number of pixels of the original visible - light image, and determine the gray - level distribution probability of the original ultraviolet image based on the gray - level data of the ultraviolet feature data and the total number of pixels of the original ultraviolet image.

[0074] In the embodiment of the present application, after obtaining the gray - level data of the visible - light feature data, the server combines the total number of pixels of the original visible - light image to determine its gray - level distribution probability. In this way, by successively counting the occurrence probabilities corresponding to each gray - level value, the gray - level distribution probability of the original visible - light image is obtained. Similarly, for the ultraviolet feature data, the server counts the occurrence times of each gray - level value in the original ultraviolet image according to its gray - level data and the total number of pixels of the original ultraviolet image, and then calculates the gray - level distribution probability corresponding to each gray - level value.

[0075] S403. Determine the cumulative distribution rate of each gray - level value in the original visible - light image according to the gray - level distribution probability of the original visible - light image, and determine the cumulative distribution rate of each gray - level value in the original ultraviolet image according to the gray - level distribution probability of the original ultraviolet image.

[0076] In the embodiment of the present application, after obtaining the gray - level distribution probability of the original visible - light image, the server starts to calculate the cumulative distribution rate of each gray - level value. For example, starting from gray - level value 0, first, the cumulative distribution rate of gray - level value 0 is its own gray - level distribution probability (assumed to be 0.01). Then, for gray - level value 1, its cumulative distribution rate is the sum of the gray - level distribution probabilities of gray - level value 0 and gray - level value 1 (assuming the gray - level distribution probability of gray - level value 1 is 0.02, then the cumulative distribution rate is 0.01 + 0.02 = 0.03). In this way, by successively adding the gray - level distribution probabilities of each gray - level value in ascending order, the server calculates the cumulative distribution rates corresponding to all gray - level values in the original visible - light image. Similarly, for the original ultraviolet image, also based on its determined gray - level distribution probability, according to the same addition method, the cumulative distribution rate of each gray - level value is calculated one by one. For example, the cumulative distribution rate of gray - level value 20 is the sum of the probabilities accumulated to it by the previous gray - level values, and so on, to obtain the cumulative distribution rates of all gray - level values in the complete original ultraviolet image.

[0077] S404. Perform histogram equalization processing on the cumulative distribution rates of each gray - level value in the original visible - light image and the cumulative distribution rates of each gray - level value in the original ultraviolet image respectively to obtain visible - light enhancement data and ultraviolet enhancement data.

[0078] In the embodiments of the present application, after the server obtains the cumulative distribution rates of the respective gray values in the original visible light image and the original ultraviolet image, it performs histogram equalization processing on them respectively. For the cumulative distribution rate of the original visible light image, the server remaps the gray value of each pixel according to its value. For example, a pixel with an original gray value of 30 may be remapped to a new gray value of 50 after the analysis of the cumulative distribution rate and the equalization processing rule (specifically, according to the correspondence between the cumulative distribution rate and the target equalized gray range), so that the gray values of the pixels in the originally darker areas of the image are adjusted, making the overall gray distribution more uniform, enhancing the contrast and detail expressiveness of the image. After such processing of all pixels, the visible light enhancement data is obtained. Similarly, for the cumulative distribution rate of the original ultraviolet image, the server adjusts the gray values of the respective pixels according to a similar remapping method. For example, some pixels with insufficient contrast in the discharge area originally have their gray values changed after processing, making the discharge-related features more prominent, and finally obtaining the ultraviolet enhancement data.

[0079] For the two approved images, enhance the images according to the processing of histogram equalization. Let the gray value of the image be f(x, y), the change range of the gray value of the image be [a, b], the transformed image be g(x, y), and the change range of the gray value of the transformed image be [c, d], h f represents the gray level histogram of the original image. First, find the gray distribution probability of the original image as p f , where N f is the total number of pixels in the image.

[0080]

[0081] Then calculate the cumulative distribution rate of each gray value.

[0082]

[0083] Among them, let p a (0)=0. Finally, perform histogram equalization calculation to obtain the processed pixel g(x, y).

[0084] g(x, y)=d×p a (k)(4)

[0085] The enhanced images are fused according to the fusion rules of the PCNN-NSSTR algorithm.

[0086] In the embodiments of the present application, first, the grayscale data is determined and the grayscale distribution probability and cumulative distribution rate are further analyzed, enabling the server to deeply understand the detailed features of the original visible light image and the original ultraviolet image at the grayscale level, grasp the overall brightness, contrast, etc. of the images, and provide an accurate data basis for subsequent enhancement processing. Second, based on the above analysis results, histogram equalization processing is used to adjust the grayscale distribution of the images in a targeted manner. For the original visible light image, it can highlight the appearance details, structural features, etc. of the device, making the blurred areas that may be caused by uneven illumination and other reasons clearly visible, enhancing the visibility of the image and the feature recognition ability during subsequent processing. For the original ultraviolet image, through equalization processing, the features related to discharge can be strengthened, such as making the brightness difference in the discharge area more obvious and the weak discharge signs more easily detectable, thereby improving the ability to capture fault-related information from the ultraviolet image.

[0087] In an exemplary embodiment, based on the above embodiment, please refer to Figure 5 , the embodiments of the present application relate to the process of fusing visible light enhancement data and ultraviolet enhancement data according to the feature change relationship to obtain the fused images of each frequency subband, including the following steps:

[0088] S501, respectively decompose the visible light enhancement data and the ultraviolet enhancement data by using the non-downsampling shearlet transform algorithm to obtain the visible light images of different frequency subbands and the ultraviolet images of different frequency subbands.

[0089] In an embodiment of the present application, after the server obtains the visible light enhancement data and ultraviolet enhancement data obtained in the previous processing steps, it starts to use the non-subsampled shear wave transform algorithm for decomposition processing. For the visible light enhancement data, the server starts the NSST algorithm, which gradually decomposes the visible light enhancement data into information of multiple different frequency sub-bands according to its built-in specific decomposition rules and parameter settings, and then generates corresponding visible light images of different frequency sub-bands. For example, the low-frequency sub-band visible light image clearly presents the overall outline of the DC valve hall equipment, and macroscopic features such as the general shape of the converter valve and the direction of the busbar can be intuitively seen; while the high-frequency sub-band visible light image can show some subtle texture changes on the surface of the equipment, such as fine lines on the surface of the insulator, tiny scratches on the equipment casing and other detailed features. Similarly, for the ultraviolet enhanced data, the server also uses the same non-subsampled shear wave transform algorithm to decompose it and obtain ultraviolet images of different frequency sub-bands. In the ultraviolet image of the low-frequency sub-band, the general contour changes of the device as a whole under ultraviolet light can be seen, especially those areas where contour feature changes may be related to discharge; the ultraviolet image of the high-frequency sub-band highlights the brightness changes in small local areas of the device under ultraviolet light, weak light points and other detailed features related to discharge. These ultraviolet images of different frequency sub-bands provide key discharge-related information for subsequent fusion.

[0090] S502 , fusing the visible light image of each frequency sub-band with the ultraviolet image of the corresponding frequency sub-band to obtain fused image data of different frequency sub-bands.

[0091] In the embodiments of the present application, after the server obtains the visible light images of each frequency sub-band and the ultraviolet images of the corresponding frequency sub-bands, it starts the fusion process. Taking the low-frequency sub-band as an example, when the server fuses the visible light image of the low-frequency sub-band and the ultraviolet image of the low-frequency sub-band, it will comprehensively consider the characteristic situations reflected by the two images in this frequency sub-band. For the contour features of the main body of the device, if there are some slight changes in this area in the ultraviolet image that may imply a discharge risk (such as a slight blur or abnormal brightness at the contour edge, etc.), the server will appropriately adjust the fusion weight during the fusion process, so that these discharge-related features in the ultraviolet image can be incorporated into the fusion image, while retaining the clear and accurate main body contour information in the visible light image, so that the fused low-frequency sub-band image can not only display the complete appearance contour of the device, but also reflect the potential discharge-related features. For the fusion of the high-frequency sub-band, the server will focus on those local areas with obvious brightness changes in the ultraviolet image. For example, a small area with a discharge phenomenon will be brighter than the surrounding area under ultraviolet light. The server extracts the high-frequency features of these areas from the ultraviolet image and fuses them with the detailed features of the corresponding areas in the visible light image of the high-frequency sub-band, so that the fused high-frequency sub-band image can more clearly present the local detailed changes of the device and the discharge-related features. By performing the fusion process for each frequency sub-band in turn like this, the fused image data of different frequency sub-bands are finally obtained. These fused image data comprehensively integrate the advantageous features of the visible light and ultraviolet images from different scales and angles, providing a more comprehensive image basis for subsequent operations such as fault detection.

[0092] In the embodiments of the present application, first, the use of the non-subsampled shearlet transform algorithm for decomposition processing can fully extract the characteristic information of the visible light enhanced data and the ultraviolet enhanced data at different frequency levels, refine the original complex image data according to different scales and characteristic dimensions, so that the subsequent fusion operation can more specifically integrate features at different levels, avoiding problems such as feature loss or poor fusion effect that may occur when directly fusing the original data. Second, fusing the visible light images of each frequency sub-band with the ultraviolet images of the corresponding frequency sub-bands can organically combine the features of the appearance, structure, and conventional textures of the device in the visible light image with the key features such as discharge faults reflected in the ultraviolet image. In the fused image data of different frequency sub-bands, whether it is the overall contour or local details, important information related to the device state is incorporated, providing a more comprehensive, accurate, and multi-dimensional image basis for subsequent fault detection.

[0093] In an exemplary embodiment, based on the above embodiment, the fused image data of different frequency subbands in the embodiments of the present application includes low-frequency fused image data and high-frequency fused image data; the embodiments of the present application relate to the process of fusing visible light images of each frequency subband with ultraviolet images of the corresponding frequency subband to obtain fused image data of different frequency subbands, including the following steps: for the visible light image of the low-frequency subband and the ultraviolet image of the low-frequency subband, a Gaussian membership function is used to perform weighted fusion processing on the visible light image of the low-frequency subband and the ultraviolet image of the low-frequency subband to obtain low-frequency fused image data; for the visible light image of the high-frequency subband and the ultraviolet image of the high-frequency subband, a pulse-coupled neural network is used to perform image fusion processing on the visible light image of the high-frequency subband and the ultraviolet image of the high-frequency subband to obtain high-frequency fused image data.

[0094] In the embodiments of the present application, after the server obtains the visible light image of the low-frequency subband and the ultraviolet image of the low-frequency subband, it starts to perform weighted fusion processing using the Gaussian membership function. First, the server analyzes the contour features of the DC valve hall equipment in the visible light image of the low-frequency subband. For example, for the main housing part of the equipment, in the area where the contour is clear and there are no obvious abnormal features under ultraviolet light, the server sets the weight of the visible light image in this area relatively high through the Gaussian membership function, such as setting it to 0.7 (this is just an example weight value for easy understanding). Correspondingly, the weight of the ultraviolet image in this area is set to 0.3, which means that more emphasis is placed on retaining the clear housing contour information in the visible light image during fusion. For some connection parts of the equipment, if there are slight blurs or weak brightness changes in the ultraviolet image that may indicate a discharge risk, the server dynamically adjusts the weight according to the characteristics of the Gaussian membership function based on the degree of these feature changes. It may increase the weight of the ultraviolet image in this area to 0.5 and reduce the weight of the visible light image to 0.5, so that the fused image can reflect both the original connection structure contour of the equipment and the potential discharge-related features reflected in the ultraviolet image. The server traverses the entire image area of the low-frequency subband in this way, determines the fusion weight of each area one by one through the Gaussian membership function according to the feature conditions of different areas, and then performs weighted fusion on the visible light image of the low-frequency subband and the ultraviolet image of the low-frequency subband according to the corresponding weights, finally obtaining the low-frequency fused image data. The image corresponding to this data can comprehensively display the overall contour of the equipment at the low-frequency level and potential discharge-related features.

[0095] After the server obtains the visible light image of the high-frequency sub-band and the ultraviolet image of the high-frequency sub-band, it starts the Pulse Coupled Neural Network (PCNN) for fusion processing. In the PCNN, each neuron corresponds to a pixel point in the image. For the pixel points in the visible light image of the high-frequency sub-band, their grayscale, brightness and other detailed features are used as the initial input information of the neuron. Similarly, the pixel features in the ultraviolet image of the high-frequency sub-band are also used as the input of the corresponding neuron. Then, based on the pulse coupling mechanism between the internal neurons of the PCNN, for example, when a neuron receives the pulse signal transmitted from the surrounding neurons, it will adjust its internal state according to its own state and the preset connection rules, etc., and then decide whether to generate a pulse output (i.e., fire). In this process, for the visible light image of the high-frequency sub-band and the ultraviolet image of the high-frequency sub-band, if in a certain local area, the pixels in the visible light image present normal texture details, while there is an obvious brightness change in this area in the ultraviolet image (possibly caused by discharge), then in the iteration process of the PCNN, the neurons corresponding to this area will automatically adjust the fusion method according to the pixel feature differences between the two images and the coupling relationship inside the network, so that in the final fusion, it will be more inclined to integrate the discharge-related detailed features reflected in the ultraviolet image into the fused image. After multiple such neuron interactions and network iterations, the server completes the fusion processing of the entire visible light image of the high-frequency sub-band and the ultraviolet image of the high-frequency sub-band through the PCNN, and obtains the high-frequency fusion image data. The image corresponding to this data can highlight the detailed changes of the device at the high-frequency level and the microscopic features related to discharge.

[0096] In the embodiments of the present application, first, for the acquisition of the low-frequency fusion image data, the Gaussian membership function is used for weighted fusion processing, which can flexibly and reasonably allocate the fusion weights according to the actual feature conditions of different regions in the visible light image of the low-frequency sub-band and the ultraviolet image of the low-frequency sub-band, so that the fused low-frequency fusion image data can not only retain the macroscopic features such as the clear contour and overall structure of the device in the visible light image, but also effectively integrate the potential discharge-related features reflected in the ultraviolet image, providing comprehensive and key image information for subsequent fault detection from a macroscopic perspective, and helping to improve the accuracy of judging the overall state of the device during fault detection. Secondly, the Pulse Coupled Neural Network (PCNN) is used to obtain the high-frequency fusion image data, which makes full use of the advantage of the PCNN that it can perform adaptive fusion according to the detailed features of the image pixels and the interaction between neurons, so that the fused high-frequency fusion image data can accurately integrate the microscopic features in the visible light image of the high-frequency sub-band and the ultraviolet image of the high-frequency sub-band, highlighting the detailed changes of the device locally and the features related to discharge, providing strong microscopic-level image support for more refined fault detection and fault location, etc., and improving the accuracy of fault detection.

[0097] In an exemplary embodiment, based on the above embodiment, please refer to Figure 6 , the fault detection model of the embodiment of the present application includes a feature extraction layer, a feature fusion layer, and a classification layer; the embodiment of the present application relates to the process of inputting the fusion images of different frequency sub-bands into a pre-constructed fault detection model for fault detection to obtain the fault detection results corresponding to the target frequency sub-band, including the following steps:

[0098] S601, input the fusion images of different frequency sub-bands into the feature extraction layer for convolution and pooling operations to obtain the front-end feature map data of different frequency sub-bands.

[0099] In the embodiment of the present application, after the server obtains the fusion images of different frequency sub-bands, it inputs them into the feature extraction layer in sequence. Taking the fusion image of one of the low-frequency sub-bands as an example, in the convolution operation stage, multiple convolutional kernels in the feature extraction layer (these convolutional kernels have different sizes and parameter settings) start to slide on the image at a set stride. For example, when a convolutional kernel with a smaller size slides on the image, it can extract the subtle change features of the device contour edge in the fusion image, while a convolutional kernel with a larger size can capture more macroscopic features such as the overall shape of the device. After multiple convolution operations, a feature map containing various local features is obtained. Then, a pooling operation is performed. The server adopts the maximum pooling method, selects the maximum value in each small area of the feature map as the representative value of the area. Through such a downsampling process, the data volume of the feature map is reduced, but key information such as the most obvious boundary features of the device contour and the main texture features of key parts is retained, and the front-end feature map data of the low-frequency sub-band is obtained. For the fusion image of the high-frequency sub-band, similar convolution and pooling operations are also performed. However, since the high-frequency sub-band focuses on detail features, the convolutional kernels will more finely extract high-frequency detail features such as minute texture changes on the device surface and local brightness differences caused by discharge. After pooling, the front-end feature map data of the high-frequency sub-band is output. By analogy, the server completes the processing of the fusion images of all different frequency sub-bands and obtains the front-end feature map data of different frequency sub-bands.

[0100] S602, input the front-end feature map data of different frequency sub-bands into the feature fusion layer for cross-channel feature fusion to obtain the feature fusion data of different frequency sub-bands.

[0101] In the embodiments of the present application, the server inputs the front-end feature map data of each different frequency sub-band into the feature fusion layer for cross-channel feature fusion. The feature fusion layer first analyzes the channel structures of the front-end feature map data of different frequency sub-bands. For example, the front-end feature map data of the low-frequency sub-band has several channels, and each channel carries macroscopic feature information in different aspects. The front-end feature map data of the high-frequency sub-band also has multiple channels, corresponding to different microscopic detail features. Then, a cross-channel fusion method based on the attention mechanism is used to achieve feature fusion. For a channel representing the overall contour feature of the device in the low-frequency sub-band and a channel reflecting the local detail feature in the high-frequency sub-band, the feature fusion layer will automatically assign different weights according to the degree of association between the features they contain and the device failure. For example, if a certain local detail feature is more valuable for judging whether there is a discharge fault in the device, then a higher weight will be given to the feature corresponding to the high-frequency sub-band channel during fusion, so that it can be more fully integrated into the fused feature representation. By traversing the features of different frequency sub-bands and different channels in this way, and performing weight assignment and fusion operations, the features of different frequency sub-bands are interactively fused at the channel level, and finally the feature fusion data of each different frequency sub-band is obtained, so that the fused data contains both the macroscopic features of the whole device and highlights the key microscopic detail features related to the fault.

[0102] S603, input the feature fusion data of each different frequency sub-band into the classification layer for classification judgment and fault location processing to obtain the fault detection result corresponding to the target frequency sub-band.

[0103] In the embodiments of the present application, the server inputs the feature fusion data of each different frequency sub-band into the classification layer for classification judgment and fault location processing. In terms of classification judgment, the classification layer is pre-trained through a large number of labeled fusion image sample data of DC valve hall equipment (including annotation information such as whether there is a fault, fault type, and fault location) to form feature patterns corresponding to different fault modes. When the feature fusion data of each different frequency sub-band is input, the classification layer will compare the feature patterns of these data with the learned fault patterns to judge whether there is a discharge fault in the DC valve hall equipment.

[0104] After image decomposition, the image is fused according to different fusion rules, and the low-frequency coefficients are adaptively weighted using fuzzy logic. For the membership function of fuzzy logic, a Gaussian distribution is used for the low-frequency coefficient fusion as, where μ and σ are the center and width of the function respectively:

[0105]

[0106] For the adaptive weighting after the model logic, a Gaussian membership function is also used to determine the coefficient after image weighted fusion, where the weighting coefficient is:

[0107]

[0108] Among them, μ is the pixel mean of the low-frequency sub-band after NSSTR decomposition, σ represents the pixel variance, and k = 1.5. Thus, the low-frequency fusion expression is obtained, where η 1 (i,j) = 1 - η 0 (i,j), C A and C B are the low-frequency sub-band coefficients at the coordinate (i,j) in ultraviolet and visible light respectively, and C F is the corresponding coefficient after fusion:

[0109] C F (i,j) = η 0 (i,j)C A (i,j) + η 1 (i,j)C B (i,j) (7)

[0110] The high-frequency coefficients are fused using a PCNN pulse neural network. The high-frequency fusion rule formula is as follows:

[0111] Among them, n is the number of iterations, ij is the pixel position, F ij is the feed input at the coordinate point (i,j), I ij is the neuron input at this point, corresponding to the gray value at this pixel point; L ij is the connection input of the neuron, W ijkl is the matrix composed of the link weights of each neuron synapse, U ij represents the internal state information of the neuron at this moment, β is the link strength, Y ij is the external output signal of the neuron, α L , α θ are both delay constants, representing the decay of the coupling link domain L and the dynamic threshold θ over time, V L , V θ represent the amplitude gains of the coupling link domain L input and the dynamic threshold θ.

[0112]

[0113] In the embodiments of the present application, first, at the feature extraction layer, the front-end feature map data of each different frequency sub-band is obtained through convolution and pooling processing, which can fully extract the feature information of different scales and different levels in the fused image. From the macroscopic device contour to the microscopic detail changes, it provides a rich and targeted feature basis for subsequent fault detection, avoiding the situation of missed or misjudged faults caused by incomplete feature extraction, and improving the comprehensiveness and accuracy of fault detection. Secondly, the cross-channel feature fusion operation performed by the feature fusion layer effectively integrates the features of each channel in the front-end feature map data of different frequency sub-bands, enabling the features of different levels to complement and cooperate with each other, enhancing the overall expression ability of the features, establishing connections between the originally isolated features, and being more conducive to highlighting the key features related to device faults, further improving the recognition ability of the fault detection model for different fault situations, and creating good conditions for accurately judging the fault type and locating the fault position.

[0114] In an exemplary embodiment, based on the above embodiment, please refer to Figure 7 , the training process of the fault detection model in the embodiments of the present application includes the following steps:

[0115] S701, perform image fusion processing on the pre-acquired sample visible light images of different frequency sub-bands and the sample ultraviolet images of different frequency sub-bands to obtain the sample fusion data of each different frequency sub-band, and use the annotation algorithm to perform annotation processing on the sample fusion data of each different frequency sub-band to obtain the training sample data.

[0116] Among them, the sample fusion data is the data obtained by performing fusion operations on the sample visible light images of different frequency sub-bands and the corresponding sample ultraviolet images of different frequency sub-bands according to certain fusion rules and algorithms.

[0117] In the embodiment of the present application, the server first obtains the pre-collected sample visible light images of different frequency sub-bands and sample ultraviolet images of different frequency sub-bands from the database. These images are obtained after preliminary collection, sorting and processing by appropriate image decomposition algorithms, covering the image information of DC valve hall equipment in various different states. Then, the server uses the image fusion method mentioned above (for example, Gaussian membership function weighted fusion for low-frequency sub-bands, pulse coupled neural network fusion for high-frequency sub-bands, etc.) to perform image fusion processing on the sample visible light images and sample ultraviolet images of different frequency sub-bands. Taking the low-frequency sub-band as an example, the server reasonably allocates weights for fusion through the Gaussian membership function according to the equipment outline and potential discharge characteristics presented by the sample visible light image and the sample ultraviolet image at the low-frequency level, so that the fused low-frequency sub-band sample fusion data can not only reflect the clear overall outline of the equipment, but also integrate the characteristics that may indicate faults under ultraviolet light. For high-frequency sub-bands, pulse coupled neural networks are used to complete fusion based on the detailed features of pixels and the interaction between neurons, highlighting the local details of the equipment and discharge-related features. After the fusion is completed, the server uses a labeling algorithm to perform labeling processing and arranges professional operation and maintenance personnel or uses labeling tools to perform detailed labeling of the sample fusion data of each different frequency sub-band according to the actual fault situation of the equipment (determined by historical detection records, on-site verification, etc.) according to dimensions such as whether there is a fault, the type of fault, and the location of the fault. For example, it is marked that the equipment has a local discharge fault in the sample fusion data of a high-frequency sub-band, and the fault location is in a specific area of ​​the image. In this way, complete training sample data is obtained.

[0118] S702, input the training sample data into the initial detection model to perform fault detection processing to obtain an initial detection result.

[0119] In the embodiments of the present application, the server inputs the obtained training sample data into the initial detection model for fault detection processing. The initial detection model internally includes structures such as a feature extraction layer, a feature fusion layer, and a classification layer. In the feature extraction layer, for the sample fusion data of different frequency sub-bands input, convolution operations are performed using convolution kernels to extract features in each frequency sub-band. For example, features such as the outline and general structure of the device are extracted from the sample fusion data in the low-frequency sub-band, and features such as local details and brightness changes are extracted from the sample fusion data in the high-frequency sub-band. Then, through pooling operations, the data volume is reduced to obtain the front-end feature map data. Next, in the feature fusion layer, the front-end feature map data of each frequency sub-band are subjected to cross-channel feature fusion to integrate feature information at different levels and obtain the fused feature data. Finally, based on these fused feature data, the classification layer attempts to determine whether there is a fault, determine the fault type, locate the fault position, etc. according to the existing parameter settings, and output the initial detection result. For example, it is determined that there is no fault in the device corresponding to the sample fusion data of a certain low-frequency sub-band, or it is determined that there is a surface discharge fault in the device corresponding to the sample fusion data of a certain high-frequency sub-band and a corresponding approximate fault position range is given. However, since the model has not been fully trained, this initial detection result may be quite different from the true annotation result.

[0120] S703, input the initial detection result and the true annotation result in the training sample data into the loss function for loss calculation processing to obtain a loss value.

[0121] In the embodiments of the present application, the server inputs the initial detection result and the true annotation result in the training sample data into the loss function for loss calculation processing together. For example, for the fault type judgment task, if the initial detection result determines that the device is a certain fault type, while the true annotation result is another fault type, then in the loss function part corresponding to the classification task, a corresponding loss value component will be calculated according to this difference situation; for the fault position localization task, if there is a deviation between the fault position given by the initial detection result and the truly annotated position, then the loss function part corresponding to the localization task will also calculate a loss value component according to the degree of position deviation. Then, through an appropriate method (such as adding the classification loss and the localization loss according to a certain weight, which specifically depends on the model design), these loss value components corresponding to different tasks are integrated, and finally an overall loss value is obtained. This loss value intuitively reflects the overall difference degree between the initial detection result and the true annotation result.

[0122] S704, adjust the parameters of the initial detection model according to the loss value until the loss value meets the preset conditions, and then end the training to obtain the fault detection model.

[0123] In the embodiments of the present application, the server adjusts the parameters of the initial detection model according to the obtained loss value. For example, if the loss value is large, it indicates that there is a significant gap between the prediction result of the model and the actual situation. The server will start from the last layer (classification layer) of the model and backpropagate the loss value to the previous feature fusion layer, feature extraction layer, etc. according to certain rules. Based on the influence degree of the parameters of each layer on the loss value (measured by calculating the gradient), the parameters of each layer are adjusted. For example, the parameters of the convolution kernel are appropriately changed, and the weights during feature fusion are adjusted, etc., so that when the model processes the training sample data again, it can output a prediction closer to the true annotation result. The server will continuously repeat the processes of inputting the training sample data, obtaining the initial detection result, calculating the loss value, and adjusting the parameters, and continuously observe the change of the loss value until the loss value meets the preset conditions, and then end the training to obtain a well-trained fault detection model, which can accurately detect faults in the fused images of different frequency subbands.

[0124] In the embodiments of the present application, first, in the sample preparation stage, by fusing and annotating the sample visible light images and ultraviolet images of different frequency subbands, the advantageous information of the two images can be fully integrated to generate training sample data containing rich fault features, enabling the model to learn different feature representations in the normal and fault states of the device from multiple angles, laying a solid foundation for subsequent accurate fault detection, and avoiding the problem of poor model learning effect caused by single sample data or incomplete features. Second, inputting the training sample data into the initial detection model for fault detection processing and subsequent loss calculation can enable the model to continuously discover the difference between its own prediction and the actual situation based on the true annotation result, clarify the direction of improvement. Through such an iterative feedback mechanism, the model gradually adjusts its own parameters and structure, learns the accurate corresponding relationship between different features and faults, and improves the ability to judge fault situations, whether it is judging whether there is a fault, determining the fault type, or locating the fault position, it can be more and more accurate.

[0125] In an exemplary embodiment, based on the above embodiment, please refer to Figure 8 , the method of the embodiments of the present application further includes the following steps:

[0126] S801, respectively determine the recognition accuracy rate and recall rate for the visible light images, ultraviolet images, and fused image data in the test set.

[0127] Among them, the test set is a part of the data set specifically reserved for evaluating the model performance in the whole data set division, except for the training set used to train the model. It contains multiple groups of visible light images, ultraviolet images of DC valve hall equipment, and fused image data obtained through the previous process.

[0128] In the embodiment of the present application, after the server completes the training of the fault detection model, it obtains a prepared test set, which includes multiple groups of visible light images, ultraviolet images, and fused image data of DC valve hall equipment, and each group of image data has corresponding true annotation information.

[0129] For the visible light images in the test set, the server sequentially inputs these images into the trained fault detection model for fault detection processing. The model will output the fault detection results (whether there is a fault, fault type, fault location, etc.) corresponding to each visible light image. Then, the server compares the detection results output by the model with the true annotation results corresponding to the visible light image. For example, there are 200 visible light images in the test set. After the server detects and compares each of them one by one, it is found that the detection results of 160 images are completely consistent with the true annotation results (that is, the model correctly judges whether there is a fault, fault type, fault location, etc.). Then, the recognition accuracy rate for visible light images is 160÷200 = 80%.

[0130] Next, calculate the recall rate. The server first counts the number of visible light images with actual faults in the test set, assuming it is 100, and then checks the number of visible light images with faults accurately identified by the model, such as 80. Then, the recall rate for visible light images is 80÷100 = 80%.

[0131] Similarly, for the ultraviolet images in the test set, the server also follows the above process, inputs the ultraviolet images into the fault detection model, compares the model output results with the true annotation results, and calculates the recognition accuracy rate and recall rate. For example, in a test set of 150 ultraviolet images, the model correctly judges 120, so the recognition accuracy rate is 120÷150 = 80%; if there are 80 ultraviolet images with actual faults and the model accurately identifies 64 of them, the recall rate is 64÷80 = 80%.

[0132] For the fused image data, the server also performs the above operations. It inputs each group of fused image data into the fault detection model, and after detection, compares it with the true annotation. Suppose there are 300 groups of fused image data and the model accurately judges 240 groups. Then, the recognition accuracy rate is 240÷300 = 80%; if there are 150 groups of fused image data with actual faults and the model accurately identifies 120 groups, the recall rate is 120÷150 = 80%.

[0133] S802, determine the recognition effect of fault detection according to the recognition accuracy rate and recall rate.

[0134] In the embodiments of the present application, after the server obtains the recognition accuracy and recall rate corresponding to the visible light images, ultraviolet images, and fused image data in the test set respectively, it begins to comprehensively analyze to determine the recognition effect of fault detection.

[0135] If the recognition accuracy and recall rate of the visible light images, ultraviolet images, and fused image data are all at a relatively high and similar level (for example, all around 80%), it indicates that the model can have a relatively stable and good performance whether it is based on individual visible light images, ultraviolet images, or fused image data for fault detection. The overall fault detection recognition effect is good, and it can detect the fault conditions of DC valve hall equipment more accurately and comprehensively.

[0136] However, if it is found that the recognition accuracy and recall rate of a certain type of image (such as ultraviolet images) are significantly lower than the corresponding indicators of the other two types of images, it means that there may be some problems when the model uses ultraviolet images for fault detection. For example, the extraction and analysis of discharge-related features reflected in the ultraviolet images may not be accurate enough, resulting in more deviations in fault judgment. In this case, the recognition effect of fault detection needs to be further improved, and the model or related image preprocessing, feature extraction, etc. need to be optimized and improved for the case of ultraviolet images to improve the overall fault detection performance and enable the model to have a more balanced and excellent recognition effect on different types of image data.

[0137] In the embodiments of the present application, first, calculating the two indicators of recognition accuracy and recall rate can quantitatively evaluate the performance of the fault detection model when processing visible light images, ultraviolet images, and fused image data from different perspectives. The accuracy reflects the ability of the model to make correct judgments, and the recall rate reflects the ability to avoid missed detections. By analyzing the numerical values of these two indicators on different types of images respectively, it is possible to clearly understand the accuracy and comprehensiveness of the model in fault detection under various data inputs, accurately locate the possible advantages or deficiencies of the model in processing which image types, and provide a clear direction for subsequent optimization and improvement. Second, comprehensively considering the recognition accuracy and recall rate to determine the recognition effect of fault detection, whether through simple comparative analysis or introducing comprehensive evaluation indicators such as the F1 value, helps to more comprehensively and objectively evaluate the overall performance of the model.

[0138] In an exemplary embodiment, based on the above embodiment, the method of the embodiments of the present application further includes the following steps:

[0139] Step 1: Perform image fusion processing on the pre-acquired sample visible light images of different frequency sub-bands and the sample ultraviolet images of different frequency sub-bands to obtain the sample fusion data of each different frequency sub-band, and use the annotation algorithm to perform annotation processing on the sample fusion data of each different frequency sub-band to obtain the training sample data;

[0140] Step 2: Input the training sample data into the initial detection model for fault detection processing to obtain the initial detection result; input the initial detection result and the true annotation result in the training sample data into the loss function for loss calculation processing to obtain the loss value; adjust the parameters of the initial detection model according to the loss value until the loss value meets the preset condition and then end the training to obtain the fault detection model;

[0141] Step 3: Perform feature extraction processing on the original visible light image and the original ultraviolet image respectively to obtain visible light feature data and ultraviolet feature data; perform feature relationship extraction processing on the visible light feature data and the ultraviolet feature data to obtain the feature change relationship between the visible light image and the ultraviolet image;

[0142] Step 4: Determine the gray data of the visible light feature data and the gray data of the ultraviolet feature data respectively;

[0143] Determine the gray distribution probability of the original visible light image according to the gray data of the visible light feature data and the total number of pixels of the original visible light image, and determine the gray distribution probability of the original ultraviolet image according to the gray data of the ultraviolet feature data and the total number of pixels of the original ultraviolet image; determine the cumulative distribution rate of each gray value in the original visible light image according to the gray distribution probability of the original visible light image, and determine the cumulative distribution rate of each gray value in the original ultraviolet image according to the gray distribution probability of the original ultraviolet image; perform histogram equalization processing on the cumulative distribution rate of each gray value in the original visible light image and the cumulative distribution rate of each gray value in the original ultraviolet image respectively to obtain visible light enhancement data and ultraviolet enhancement data;

[0144] Step 14: Use the non-sub-sampled shearlet transform algorithm to decompose the visible light enhancement data and the ultraviolet enhancement data respectively to obtain visible light images of different frequency sub-bands and ultraviolet images of different frequency sub-bands;

[0145] Step 6: For the visible light image of the low-frequency sub-band and the ultraviolet image of the low-frequency sub-band, use the Gaussian membership function to perform weighted fusion processing on the visible light image of the low-frequency sub-band and the ultraviolet image of the low-frequency sub-band to obtain the low-frequency fusion image data in the fusion image data of different frequency sub-bands; for the visible light image of the high-frequency sub-band and the ultraviolet image of the high-frequency sub-band, use the pulse coupled neural network to perform image fusion processing on the visible light image of the high-frequency sub-band and the ultraviolet image of the high-frequency sub-band to obtain the high-frequency fusion image data in the fusion image data of different frequency sub-bands;

[0146] Step 7: Input the fused images of different frequency sub-bands into the feature extraction layer for convolution and pooling processing to obtain the front-end feature map data of different frequency sub-bands; input the front-end feature map data of different frequency sub-bands into the feature fusion layer for cross-channel feature fusion to obtain the feature fusion data of different frequency sub-bands; input the feature fusion data of different frequency sub-bands into the classification layer for classification judgment and fault location processing to obtain the fault detection result corresponding to the target frequency sub-band.

[0147] Step 8: Respectively determine the recognition accuracy rate and recall rate for the visible light images, ultraviolet images, and fused image data in the test set; determine the recognition effect of the fault detection according to the recognition accuracy rate and recall rate.

[0148] It should be understood that although each step in the flowcharts involved in the above-described embodiments is shown in sequence according to the indication of the arrows, these steps do not necessarily have to be executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages do not necessarily have to be executed at the same time, but can be executed at different times. The execution order of these steps or stages does not necessarily have to be sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0149] Based on the same inventive concept, the embodiment of the present application also provides a direct current valve hall equipment discharge fault detection device for implementing the direct current valve hall equipment discharge fault detection method described above. The implementation solution provided by this device to solve the problem is similar to the implementation solution described in the above method. Therefore, the specific limitations in one or more embodiments of the direct current valve hall equipment discharge fault detection device provided below can refer to the limitations on the direct current valve hall equipment discharge fault detection method in the above text, and will not be repeated here.

[0150] In one embodiment, as Figure 9 shown, a direct current valve hall equipment discharge fault detection device 900 is provided, including:

[0151] An image fusion module 901, configured to perform image fusion processing on the pre-acquired original visible light image and original ultraviolet image of the direct current valve hall equipment to obtain fused images of multiple different frequency sub-bands;

[0152] A fault detection module 902, configured to input the fused images of different frequency sub-bands into a pre-constructed fault detection model for fault detection, so as to obtain the fault detection results corresponding to the target frequency sub-bands; the fault detection results are used to characterize whether there is a discharge fault in the DC valve hall equipment and the type and severity of the fault.

[0153] In one embodiment, the above image fusion module includes:

[0154] A feature extraction unit, configured to perform feature extraction processing on the original visible light image and the original ultraviolet image respectively to obtain visible light feature data and ultraviolet feature data;

[0155] A feature relationship extraction unit, configured to perform feature relationship extraction processing on the visible light feature data and the ultraviolet feature data to obtain the feature change relationship between the visible light image and the ultraviolet light image;

[0156] A feature enhancement unit, configured to perform enhancement processing on the visible light feature data and the ultraviolet feature data respectively to obtain visible light enhanced data and ultraviolet enhanced data;

[0157] A feature fusion unit, configured to perform fusion processing on the visible light enhanced data and the ultraviolet enhanced data according to the feature change relationship to obtain the fused images of each frequency sub-band.

[0158] In one embodiment, the above feature enhancement unit includes:

[0159] A gray level determination sub-unit, configured to determine the gray level data of the visible light feature data and the gray level data of the ultraviolet feature data respectively;

[0160] A probability determination sub-unit, configured to determine the gray level distribution probability of the original visible light image according to the gray level data of the visible light feature data and the total number of pixels of the original visible light image, and determine the gray level distribution probability of the original ultraviolet image according to the gray level data of the ultraviolet feature data and the total number of pixels of the original ultraviolet image;

[0161] A distribution determination sub-unit, configured to determine the cumulative distribution rate of each gray level value in the original visible light image according to the gray level distribution probability of the original visible light image, and determine the cumulative distribution rate of each gray level value in the original ultraviolet image according to the gray level distribution probability of the original ultraviolet image;

[0162] An equalization sub-unit, configured to perform histogram equalization processing on the cumulative distribution rate of each gray level value in the original visible light image and the cumulative distribution rate of each gray level value in the original ultraviolet image respectively to obtain visible light enhanced data and ultraviolet enhanced data.

[0163] In one embodiment, the above root feature fusion unit includes:

[0164] A data decomposition subunit, which is used to decompose and process the visible light enhanced data and the ultraviolet enhanced data respectively by using the non-downsampling shearlet transform algorithm to obtain the visible light images of different frequency subbands and the ultraviolet images of different frequency subbands;

[0165] A feature fusion subunit, which is used to fuse the visible light images of each frequency subband with the ultraviolet images of the corresponding frequency subband to obtain the fused image data of different frequency subbands.

[0166] In one embodiment, the above-mentioned fused image data of different frequency subbands includes low-frequency fused image data and high-frequency fused image data; the feature fusion subunit is specifically used for: for the visible light image of the low-frequency subband and the ultraviolet image of the low-frequency subband, using the Gaussian membership function to perform weighted fusion processing on the visible light image of the low-frequency subband and the ultraviolet image of the low-frequency subband to obtain the low-frequency fused image data; for the visible light image of the high-frequency subband and the ultraviolet image of the high-frequency subband, using the pulse coupled neural network to perform image fusion processing on the visible light image of the high-frequency subband and the ultraviolet image of the high-frequency subband to obtain the high-frequency fused image data.

[0167] In one embodiment, the above-mentioned fault detection model includes a feature extraction layer, a feature fusion layer and a classification layer; the fault detection module includes:

[0168] A convolution pooling unit, which is used to input the fused images of each different frequency subband into the feature extraction layer for convolution and pooling processing to obtain the front-end feature map data of each different frequency subband;

[0169] A feature fusion unit, which is used to input the front-end feature map data of each different frequency subband into the feature fusion layer for cross-channel feature fusion to obtain the feature fusion data of each different frequency subband;

[0170] A positioning unit, which is used to input the feature fusion data of each different frequency subband into the classification layer for classification judgment and fault location positioning processing to obtain the fault detection result corresponding to the target frequency subband.

[0171] In one embodiment, the above-mentioned device includes:

[0172] An image fusion module, which is used to perform image fusion processing on the sample visible light images of different frequency subbands and the sample ultraviolet images of different frequency subbands obtained in advance to obtain the sample fusion data of each different frequency subband, and use the annotation algorithm to perform annotation processing on the sample fusion data of each different frequency subband to obtain the training sample data;

[0173] A fault detection module, which is used to input the training sample data into the initial detection model for fault detection processing to obtain the initial detection result;

[0174] A loss calculation module for inputting the initial detection result and the true annotation result in the training sample data into a loss function for loss calculation processing to obtain a loss value;

[0175] A parameter adjustment module for adjusting the parameters of the initial detection model according to the loss value until the training ends when the loss value meets a preset condition, and obtaining a fault detection model.

[0176] In one embodiment, the above device further includes:

[0177] A data calculation module for determining the recognition accuracy rate and recall rate respectively for visible light images, ultraviolet images, and fused image data in a test set;

[0178] An effect determination module for determining the recognition effect of fault detection according to the recognition accuracy rate and recall rate.

[0179] Each module in the above direct current valve hall equipment discharge fault detection device can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in a processor in a computer device in hardware form or be independent of it, or can be stored in a memory in a computer device in software form, so that the processor can call and execute the operations corresponding to the above respective modules.

[0180] In one embodiment, a computer device is provided. The computer device can be a server, and its internal structure diagram can be as Figure 10 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used for the processor to exchange information with external devices. The communication interface of the computer device is used for communicating with an external terminal through a network connection. When the computer program is executed by the processor, it implements a method for detecting discharge faults of direct current valve hall equipment.

[0181] Those skilled in the art can understand that Figure 10 the structure shown in

[0182] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory. When the processor executes the computer program, the following steps are implemented:

[0183] Perform image fusion processing on the pre-acquired original visible light image and original ultraviolet image of the DC valve hall equipment to obtain fusion images of multiple different frequency sub-bands;

[0184] Input the fusion images of each different frequency sub-band into a pre-constructed fault detection model for fault detection to obtain a fault detection result corresponding to the target frequency sub-band; the fault detection result is used to characterize whether there is a discharge fault in the DC valve hall equipment and the type and severity of the fault.

[0185] In one embodiment, when the processor executes the computer program, the following steps are also implemented:

[0186] Respectively perform feature extraction processing on the original visible light image and the original ultraviolet image to obtain visible light feature data and ultraviolet feature data;

[0187] Perform feature relationship extraction processing on the visible light feature data and the ultraviolet feature data to obtain the feature change relationship between the visible light image and the ultraviolet light image;

[0188] Respectively perform enhancement processing on the visible light feature data and the ultraviolet feature data to obtain visible light enhanced data and ultraviolet enhanced data;

[0189] Perform fusion processing on the visible light enhanced data and the ultraviolet enhanced data according to the feature change relationship to obtain fusion images of each frequency sub-band.

[0190] In one embodiment, when the processor executes the computer program, the following steps are also implemented:

[0191] Respectively determine the gray-scale data of the visible light feature data and the gray-scale data of the ultraviolet feature data;

[0192] According to the gray-scale data of the visible light feature data and the total number of pixels of the original visible light image, determine the gray-scale distribution probability of the original visible light image, and according to the gray-scale data of the ultraviolet feature data and the total number of pixels of the original ultraviolet image, determine the gray-scale distribution probability of the original ultraviolet image;

[0193] According to the gray-scale distribution probability of the original visible light image, determine the cumulative distribution rate of each gray-scale value in the original visible light image, and according to the gray-scale distribution probability of the original ultraviolet image, determine the cumulative distribution rate of each gray-scale value in the original ultraviolet image;

[0194] Histogram equalization is performed on the cumulative distribution rates of each gray value in the original visible light image and the cumulative distribution rates of each gray value in the original ultraviolet image respectively to obtain visible light enhancement data and ultraviolet enhancement data.

[0195] In one embodiment, when the processor executes the computer program, the following steps are also implemented:

[0196] The non - subsampled shearlet transform algorithm is used to decompose the visible light enhancement data and the ultraviolet enhancement data respectively to obtain visible light images of different frequency sub - bands and ultraviolet images of different frequency sub - bands;

[0197] The visible light images of each frequency sub - band are fused with the ultraviolet images of the corresponding frequency sub - band to obtain fused image data of different frequency sub - bands.

[0198] In one embodiment, when the processor executes the computer program, the following steps are also implemented:

[0199] For the visible light image of the low - frequency sub - band and the ultraviolet image of the low - frequency sub - band, a Gaussian membership function is used to perform weighted fusion processing on the visible light image of the low - frequency sub - band and the ultraviolet image of the low - frequency sub - band to obtain low - frequency fused image data;

[0200] For the visible light image of the high - frequency sub - band and the ultraviolet image of the high - frequency sub - band, a pulse - coupled neural network is used to perform image fusion processing on the visible light image of the high - frequency sub - band and the ultraviolet image of the high - frequency sub - band to obtain high - frequency fused image data.

[0201] In one embodiment, when the processor executes the computer program, the following steps are also implemented:

[0202] The fused images of each different frequency sub - band are input into the feature extraction layer for convolution and pooling processing to obtain front - end feature map data of each different frequency sub - band;

[0203] The front - end feature map data of each different frequency sub - band are input into the feature fusion layer for cross - channel feature fusion to obtain feature fusion data of each different frequency sub - band;

[0204] The feature fusion data of each different frequency sub - band are input into the classification layer for classification judgment and fault location processing to obtain the fault detection results corresponding to the target frequency sub - band.

[0205] In one embodiment, when the processor executes the computer program, the following steps are also implemented:

[0206] The sample visible light images of different frequency sub - bands and the sample ultraviolet images of different frequency sub - bands obtained in advance are subjected to image fusion processing to obtain sample fusion data of each different frequency sub - band, and a labeling algorithm is used to perform labeling processing on the sample fusion data of each different frequency sub - band to obtain training sample data;

[0207] Input the training sample data into the initial detection model for fault detection processing to obtain the initial detection result;

[0208] Input the initial detection result and the true annotation result in the training sample data into the loss function for loss calculation processing to obtain the loss value;

[0209] Adjust the parameters of the initial detection model according to the loss value until the loss value meets the preset conditions, and then end the training to obtain the fault detection model.

[0210] In one embodiment, when the processor executes the computer program, the following steps are also implemented:

[0211] Respectively determine the recognition accuracy and recall rate for the visible light images, ultraviolet images, and fused image data in the test set;

[0212] Determine the recognition effect of the fault detection according to the recognition accuracy and recall rate.

[0213] According to some embodiments of the present application, a computer program product is also provided. When the computer program is executed by a processor, the above method can be implemented. The computer program product includes one or more computer instructions. When these computer instructions are loaded and executed on a computer, part or all of the above method can be implemented in accordance with the process or function described in the embodiments of the present application.

[0214] According to some embodiments of the present application, a non - temporary computer - readable storage medium including instructions is also provided, such as a memory including instructions. The above instructions can be executed by the processor of an electronic device to complete the above method. For example, the non - temporary computer - readable storage medium can be ROM, random access memory (RAM), CD - ROM, magnetic tape, floppy disk, and optical data storage devices, etc.

[0215] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions.

[0216] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memories can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memories can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., without limitation.

[0217] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0218] The above-described embodiments merely represent several implementation manners of the present application. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for detecting discharge faults of DC valve hall equipment, characterized in that: The method comprises: Performing image fusion processing on the original visible light image and the original ultraviolet image of the DC valve hall device acquired in advance to obtain fused images of multiple different frequency sub-bands; The fused images of the different frequency sub-bands are input into a pre-built fault detection model for fault detection to obtain a fault detection result corresponding to the target frequency sub-band; the fault detection result is used to characterize whether there is a discharge fault in the DC valve hall equipment and the type and severity of the fault.

2. The method according to claim 1, characterized in that The image fusion processing is performed on the original visible light image and the original ultraviolet image of the DC valve hall device acquired in advance to obtain a plurality of fused images of different frequency sub-bands, including: Performing feature extraction processing on the original visible light image and the original ultraviolet image respectively to obtain visible light feature data and ultraviolet feature data; Performing feature relationship extraction processing on the visible light feature data and the ultraviolet feature data to obtain a feature change relationship between the original visible light image and the original ultraviolet image; Respectively performing enhancement processing on the visible light characteristic data and the ultraviolet characteristic data to obtain visible light enhanced data and ultraviolet enhanced data; The visible light enhancement data and the ultraviolet enhancement data are fused according to the characteristic change relationship to obtain a fused image of each frequency sub-band.

3. The method according to claim 2, characterized in that The enhancing process is performed on the visible light characteristic data and the ultraviolet characteristic data respectively to obtain the visible light enhanced data and the ultraviolet enhanced data, including: respectively determining the grayscale data of the visible light characteristic data and the grayscale data of the ultraviolet characteristic data; Determining the grayscale distribution probability of the original visible light image according to the grayscale data of the visible light feature data and the total pixels of the original visible light image, and determining the grayscale distribution probability of the original ultraviolet image according to the grayscale data of the ultraviolet feature data and the total pixels of the original ultraviolet image; Determining the cumulative distribution rate of each grayscale value in the original visible light image according to the grayscale distribution probability of the original visible light image, and determining the cumulative distribution rate of each grayscale value in the original ultraviolet image according to the grayscale distribution probability of the original ultraviolet image; The cumulative distribution rate of each gray value in the original visible light image and the cumulative distribution rate of each gray value in the original ultraviolet image are respectively subjected to histogram equalization processing to obtain the visible light enhanced data and the ultraviolet enhanced data.

4. The method according to claim 2, characterized in that: The fusing the visible light enhancement data and the ultraviolet enhancement data according to the characteristic change relationship to obtain a fused image of each frequency sub-band includes: Using a non-subsampled shear wave transform algorithm to decompose the visible light enhanced data and the ultraviolet enhanced data respectively, to obtain visible light images of different frequency sub-bands and ultraviolet images of different frequency sub-bands; The visible light image of each frequency sub-band is fused with the ultraviolet image of the corresponding frequency sub-band to obtain fused image data of the different frequency sub-bands.

5. The method according to claim 4, characterized in that The fused image data of different frequency sub-bands includes low-frequency fused image data and high-frequency fused image data; the fusion processing of the visible light image of each frequency sub-band with the ultraviolet image of the corresponding frequency sub-band to obtain the fused image data of different frequency sub-bands includes: For the visible light image of the low frequency sub-band and the ultraviolet image of the low frequency sub-band, a Gaussian membership function is used to perform weighted fusion processing on the visible light image of the low frequency sub-band and the ultraviolet image of the low frequency sub-band to obtain the low frequency fused image data; For the visible light image of the high frequency sub-band and the ultraviolet image of the high frequency sub-band, a pulse coupled neural network is used to perform image fusion processing on the visible light image of the high frequency sub-band and the ultraviolet image of the high frequency sub-band to obtain the high frequency fused image data.

6. The method according to claim 1, characterized in that The fault detection model includes a feature extraction layer, a feature fusion layer and a classification layer; the fusion images of the different frequency sub-bands are input into the pre-built fault detection model for fault detection to obtain the fault detection result corresponding to the target frequency sub-band, including: Inputting the fused images of the different frequency sub-bands into the feature extraction layer for convolution and pooling processing to obtain the front-end feature map data of the different frequency sub-bands; Inputting the front-end feature map data of each of the different frequency sub-bands into the feature fusion layer for cross-channel feature fusion to obtain feature fusion data of each of the different frequency sub-bands; The feature fusion data of each of the different frequency sub-bands are input into the classification layer for classification judgment and fault location processing to obtain the fault detection result corresponding to the target frequency sub-band.

7. The method according to claim 6, characterized in that The training process of the fault detection model includes: Performing image fusion processing on sample visible light images of different frequency sub-bands and sample ultraviolet images of different frequency sub-bands acquired in advance to obtain sample fusion data of each of the different frequency sub-bands, and performing labeling processing on the sample fusion data of each of the different frequency sub-bands using a labeling algorithm to obtain training sample data; Inputting the training sample data into the initial detection model to perform fault detection processing to obtain an initial detection result; Inputting the initial detection result and the real labeling result in the training sample data into the loss function to perform loss calculation processing to obtain a loss value; The parameters of the initial detection model are adjusted according to the loss value, and the training is terminated when the loss value meets the preset conditions to obtain the fault detection model.

8. The method according to claim 7, characterized in that The method further comprises: Determine the recognition accuracy and recall rate for the visible light image, ultraviolet image and fused image data in the test set respectively; The recognition effect of the fault detection is determined according to the recognition accuracy and the recall rate.

9. A DC valve hall equipment discharge fault detection device, characterized in that: The device comprises: An image fusion module is used to perform image fusion processing on the original visible light image and the original ultraviolet image of the DC valve hall device acquired in advance to obtain fused images of multiple different frequency sub-bands; The fault detection module is used to input the fusion image of the target frequency sub-band into a pre-built fault detection model for fault detection, and obtain the fault detection result corresponding to the target frequency sub-band; the fault detection result is used to characterize whether there is a discharge fault in the DC valve hall equipment and the type and severity of the fault.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images