Defect identification method and system for photoelectric detection image of power transmission and transformation equipment

Abstract

The invention discloses a defect identification method and system for a photoelectric detection image of power transmission and transformation equipment, and relates to the technical field of defect identification, and the method comprises the steps: collecting photoelectric detection data of target power transmission and transformation equipment; carrying out coupling analysis in combination with an equipment topology connection relationship, and drawing up a feature matrix; meanwhile, identifying a defect type mode based on a device topology connection relation and a change trend of a gray scale deviation coefficient in a plurality of detection periods; based on the characteristic matrix, combining the defect type mode and the photoelectric detection data to configure a defect confidence factor; and determining a defect severity level by using the defect confidence factor and carrying out graded reminding. The technical problems that in the prior art, electric transmission and transformation equipment photoelectric detection data integration is insufficient, the defect recognition precision is low, and operation and maintenance response lacks pertinence are solved, and the technical effects that integration analysis of multiple types of photoelectric detection data of the electric transmission and transformation equipment is achieved, and the defect recognition accuracy and the operation and maintenance response pertinence are improved are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of defect identification, and in particular to a defect identification method and system for photoelectric detection images of power transmission and transformation equipment. BACKGROUND

[0002] In the field of power transmission and transformation equipment operation and maintenance, the equipment is exposed to complex outdoor environments for a long time, and various faults such as thermal defects, discharge defects, mechanical damage defects, and material aging defects may occur due to factors such as high temperature, corona discharge, mechanical stress, and material aging. If these faults are not accurately identified in a timely manner, they may cause equipment downtime or even power grid safety accidents. Currently, the detection of defects in power transmission and transformation equipment in the industry relies on single type of photoelectric detection data, and there is a lack of effective integration and analysis of multi-source data such as infrared thermal imaging images, ultraviolet discharge images, and visible light appearance images, resulting in incomplete defect feature extraction. At the same time, existing recognition methods often ignore the influence of device topology connection relationship on defect determination, and do not establish a strong correlation mechanism between defect features and recognition model parameters, making it difficult to accurately distinguish similar defects, and resulting in low defect recognition accuracy and high type misjudgment rate. In addition, traditional methods have poor adaptability to changes in the detection environment, and do not set a dynamic adjustment mechanism, which may lead to defect level determination deviation due to environmental interference, and thus the operation and maintenance work lacks pertinence, and cannot meet the actual needs of efficient and accurate operation and maintenance of power transmission and transformation equipment.

[0003] The existing technology has the technical problems of insufficient integration of photoelectric detection data of power transmission and transformation equipment, low defect recognition accuracy, and lack of pertinence in operation and maintenance response. SUMMARY

[0004] The present application provides a defect identification method and system for photoelectric detection images of power transmission and transformation equipment, which is used to solve the technical problems of insufficient integration of photoelectric detection data of power transmission and transformation equipment, low defect recognition accuracy, and lack of pertinence in operation and maintenance response in the prior art.

[0005] In view of the above problems, the present application provides a defect identification method and system for photoelectric detection images of power transmission and transformation equipment.

[0006] In a first aspect of the present application, a defect identification method for photoelectric detection images of power transmission and transformation equipment is provided, which comprises: The photoelectric detection data of the target power transmission and transformation equipment includes infrared thermal imaging images, ultraviolet discharge images and visible light appearance images; a feature matrix is drafted by coupling analysis according to the temperature gradient threshold corresponding to the equipment insulator, the discharge light spot critical area corresponding to the sleeve surface, the gray scale deviation coefficient corresponding to the conductor joint, in combination with the equipment topology connection relationship; meanwhile, a defect type mode is identified based on the equipment topology connection relationship and the change trend of the gray scale deviation coefficient in multiple detection cycles; a defect confidence factor is configured based on the feature matrix, in combination with the defect type mode and the photoelectric detection data, the defect confidence factor being strongly associated with the input layer weight of a defect recognition model; the defect confidence factor is used to determine a defect severity level and perform a hierarchical reminder.

[0007] In a second aspect of the present application, a defect recognition system for photoelectric detection images of power transmission and transformation equipment is provided, and the system comprises: A photoelectric detection data acquisition module is configured to acquire photoelectric detection data of the target power transmission and transformation equipment, including infrared thermal imaging images, ultraviolet discharge images and visible light appearance images; a feature matrix drafting module is configured to draft a feature matrix by coupling analysis according to the temperature gradient threshold corresponding to the equipment insulator, the discharge light spot critical area corresponding to the sleeve surface, the gray scale deviation coefficient corresponding to the conductor joint, in combination with the equipment topology connection relationship; a defect type mode identification module is configured to identify a defect type mode based on the equipment topology connection relationship and the change trend of the gray scale deviation coefficient in multiple detection cycles; a defect confidence factor configuration module is configured to configure a defect confidence factor based on the feature matrix, in combination with the defect type mode and the photoelectric detection data, the defect confidence factor being strongly associated with the input layer weight of a defect recognition model; and a hierarchical reminder module is configured to use the defect confidence factor to determine a defect severity level and perform a hierarchical reminder.

[0008] The one or more technical solutions provided in the present application have at least the following technical effects or advantages: The photoelectric detection data of the target power transmission and transformation equipment includes infrared thermal imaging images, ultraviolet discharge images and visible light appearance images; a feature matrix is drafted by coupling analysis according to the temperature gradient threshold corresponding to the equipment insulator, the discharge light spot critical area corresponding to the sleeve surface, the gray scale deviation coefficient corresponding to the conductor joint, in combination with the equipment topology connection relationship; a defect type mode is identified; a defect confidence factor is configured in combination with the defect type mode and the photoelectric detection data; and the defect confidence factor is used to determine a defect severity level and perform a hierarchical reminder. The technical effect of realizing integrated analysis of multiple types of photoelectric detection data of power transmission and transformation equipment is achieved, and the defect recognition accuracy and the operation and maintenance response pertinence are improved. BRIEF DESCRIPTION OF DRAWINGS

[0009] In order to make the technical solutions in the embodiments of the present application clearer, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the embodiment description are only some of the embodiments of the present application, and all other drawings obtained by those skilled in the art without creative effort based on the drawings are within the protection scope of the present application.

[0010] Figure 1 A defect recognition method flow chart of the photoelectric detection image of the power transmission and transformation equipment provided by the embodiment of the present application is shown in the figure. Figure 2 A defect recognition system structure diagram of the photoelectric detection image of the power transmission and transformation equipment provided by the embodiment of the present application is shown in the figure.

[0011] Legend: photoelectric detection data acquisition module 10, feature matrix preparation module 20, defect type pattern recognition module 30, defect confidence factor configuration module 40, and hierarchical prompting module 50. DETAILED DESCRIPTION

[0012] The present application provides a defect recognition method and system of the photoelectric detection image of the power transmission and transformation equipment, which is used to solve the technical problems of insufficient integration of the photoelectric detection data of the power transmission and transformation equipment, low defect recognition accuracy and lack of targeted operation and maintenance response in the prior art.

[0013] The technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present application.

[0014] Embodiment one, as shown in the figure, the present application provides a defect recognition method of the photoelectric detection image of the power transmission and transformation equipment, which comprises: Figure 1 Step S100: acquiring photoelectric detection data of the target power transmission and transformation equipment including infrared thermal imaging image, ultraviolet discharge image and visible light appearance image.

[0015] ​Specifically, through the configured photoelectric detection data collection module, multi-dimensional and full-coverage data collection operations are carried out on target power transmission and transformation equipment, such as core components of transformers, insulators, wire joints, and bushings. Among them, the infrared thermal imaging image collection focuses on the temperature distribution state of each component of the equipment. Through the infrared imaging equipment, the thermal radiation differences of different areas are captured, and the parts that may have thermal abnormalities, such as local overheating of insulators and high temperature of wire joints, are accurately recorded. The ultraviolet discharge image collection uses ultraviolet imaging equipment to monitor the surface of the equipment, especially the areas prone to discharge such as bushings and insulator sheds, to determine whether there is a corona discharge phenomenon. The position, shape, and distribution characteristics of the discharge light spot are clearly captured. The visible light appearance image collection uses a high-definition visible light camera to record the details of the equipment's appearance structure, including the connection state of the wire joint, the damage condition of the insulator, and the surface cracks of the bushing, to ensure that the mechanical form and appearance integrity of the equipment can be intuitively reflected. The three types of image data are synchronously collected and stored, and together they form a photoelectric detection data set covering the thermal state, electrical state, and physical form of the equipment.

[0016] Step S200: According to the temperature gradient threshold corresponding to the equipment insulator, the discharge light spot critical area corresponding to the bushing surface, and the gray scale deviation coefficient corresponding to the wire joint, coupled analysis is carried out combined with the equipment topology connection relationship to formulate a feature matrix.

[0017] Specifically, key decision indicators are determined for each core component of the equipment: for insulators, based on their thermal field distribution map, the temperature gradient threshold that can distinguish between normal and abnormal states is obtained through early infrared thermal imaging image analysis, which can accurately reflect whether the insulator has local overheating and other thermal defect risks; for bushings, the pixel ratio of surface discharge light spots is extracted, and the correlation function of the critical area of the discharge light spot is established combined with the actual operating voltage of the equipment to ensure that the judgment standard of the bushing discharge defect under different voltage levels is adaptive and accurate; for wire joints, the gray mean deviation in the visible light appearance image is calculated, and the distribution function of the gray deviation coefficient is defined to quantify the degree of gray abnormality of the wire joint appearance, providing a basis for defect judgment of mechanical damage or poor contact. Subsequently, the topological connection relationship of the equipment is introduced, that is, the connection logic, position association and functional dependence of components such as insulators, bushings and wire joints in the overall structure of the power transmission and transformation equipment are determined, and the key indicators of the above three types of components are coupled with the topological information for analysis, for example, combined with the connection path of the insulator and the wire joint, it is analyzed whether the abnormal temperature gradient of the insulator can affect the gray deviation of the wire joint, or whether the discharge of the bushing can interfere with the thermal state of the adjacent insulator. Finally, the multi-dimensional information after coupling analysis, including the key indicator values of each component, the topological association weight, the influence coefficient between components, etc., is structured and integrated, a matrix framework is constructed according to the preset dimensions, such as component type, defect judgment dimension, and topological correlation degree, and a feature matrix is formed after filling the corresponding data, which can comprehensively reflect the state of each component of the equipment and the mutual influence, and provide a structured analysis basis for subsequent defect type recognition and confidence factor configuration.

[0018] Step S300: At the same time, based on the device topological connection relationship and the change trend of the gray deviation coefficient in multiple detection cycles, the defect type mode is identified.

[0019] Specifically, based on the device topology connection relationship framework, the connection logic and functional association of components such as insulators, bushings, and conductor joints in power transmission and transformation equipment are clarified, which provides a basis for the spatial reference to determine the mutual influence range of different component defects and provides a basis for subsequent association analysis of defect types. Next, the numerical change data of the gray deviation coefficient corresponding to the conductor joint in multiple consecutive detection periods is extracted, the change trend is analyzed, and it is judged whether the coefficient is continuously increasing, periodically fluctuating, or remaining stable. Through the trend characteristics, it is preliminarily judged whether there is a potential defect and the development trend of the defect. Subsequently, combined with the multi-class photoelectric detection image features, the defect type mode is further subdivided: in the infrared thermal imaging image, the thermal defects and material aging defects are distinguished by the entropy change rate; in the ultraviolet discharge image, the discharge defects are identified according to the spot shape factor and the gray level co-occurrence matrix; at the same time, based on the device topology connection relationship, the visible light appearance image is decomposed into multiple image pyramid layers, the edge feature components related to the equipment component profile are extracted, and they are compared and analyzed with the mechanical damage defects in the defect type mode. If the continuous fracture length of the edge feature components in the preset detection area exceeds the fracture threshold and the profile fitting error increases sharply, the dynamic correction mechanism of the gray deviation coefficient corresponding to the conductor joint is triggered. Finally, by comprehensively considering the spatial analysis results of the device topology connection relationship, the periodic change trend of the gray deviation coefficient, and the judgment of the multi-class image features, the target power transmission and transformation equipment corresponding to the defect type mode is clarified from the four preset types of thermal defects, discharge defects, mechanical damage defects, and material aging defects.

[0020] Step S400: Based on the feature matrix, the defect type mode is combined with the photoelectric detection data to configure a defect confidence factor, and the defect confidence factor is strongly associated with the input layer weight of the defect recognition model.

[0021] Specifically, key quantitative indicators are extracted from the photoelectric detection data, and the infrared temperature peak in the infrared thermal imaging image, the ultraviolet spot energy in the ultraviolet discharge image, and the visible light edge gradient value in the visible light appearance image are taken as core input variables to provide multi-dimensional data support for the configuration of the defect confidence factor. Then, the above input variables are fused with the proposed feature matrix, and the support function of defect recognition is updated by integrating the information such as the insulator temperature gradient threshold, the sleeve discharge spot critical area, the conductor joint gray scale deviation coefficient, and the device topology connection relationship, so that the function can comprehensively reflect the matching degree of the current device state and the defect characteristics. Subsequently, the updated support function is used to perform correlation coefficient feedback verification on the defect confidence factor to ensure that the confidence factor can accurately quantify the reliability of the defect recognition result and avoid distortion of the confidence factor due to single data deviation. At the same time, a mapping relationship matrix is established between the support function and the input layer weight of the defect recognition model, and the influence weight vector of the defect confidence factor on each weight of the input layer is determined through matrix decomposition, so that the defect confidence factor and the input layer weight of the defect recognition model form a strong association. When the defect confidence factor changes, it can directly drive the dynamic adjustment of the model input layer weight, improving the recognition adaptability of the model to the current defect type. In addition, if the correlation coefficient of the defect confidence factor is lower than the preset correlation coefficient threshold for Q consecutive detection periods, the weight adaptive update mechanism will be triggered to further optimize the association between the defect confidence factor and the model input layer weight, ensuring the accuracy of subsequent defect severity level determination.

[0022] Step S500: using the defect confidence factor, determining the defect severity level and performing classification reminder.

[0023] Specifically, relying on the previously configured defect confidence factor, combined with the identified defect type mode, i.e., thermal defect, discharge defect, mechanical damage defect, and material aging defect, the association rules between the confidence factor and the defect severity level are established. For example, for thermal defects, if the defect confidence factor is higher than the preset high threshold and the infrared temperature peak is much higher than the insulator temperature gradient threshold, it is determined as a severe defect; if the confidence factor is in the medium threshold interval and the temperature peak is slightly higher than the threshold, it is determined as a moderate defect; if the confidence factor is lower than the low threshold and the temperature is near the normal range, it is determined as a slight defect. Similarly, for discharge defects, mechanical damage defects, and material aging defects, the corresponding severity levels (slight, moderate, and severe) are determined by matching the features such as ultraviolet spot energy, visible light edge fracture degree, and entropy change rate with the defect confidence factor. At the same time, according to the determined defect severity level, a differentiated reminder mechanism is triggered to ensure that the maintenance personnel can obtain effective information in a timely manner. In addition, if the detection scene triggers the environmental light replacement condition or the humidity replacement condition, an incremental learning mechanism is automatically triggered for local parameter fine-tuning, further ensuring the accuracy of the defect severity level determination, so that the classification reminder can accurately match the actual defect condition of the equipment, providing clear guidance for the operation and maintenance of the power transmission and transformation equipment.

[0024] In one possible implementation manner, the step S200 further includes: Step S210: obtaining a temperature gradient threshold corresponding to the device insulator according to a thermal field distribution map of the key component of the device.

[0025] Step S220: extracting a discharge light spot pixel proportion of the sleeve surface, and setting an associated function of a discharge light spot critical area and a device operating voltage.

[0026] Step S230: defining a distribution function of a gray scale deviation coefficient through a gray scale mean value deviation of the wire joint.

[0027] Specifically, taking the device insulator as the core analysis object, based on the collected infrared thermal imaging image, a thermal field distribution map of the insulator is generated through a thermal field analysis technology. The map can directly present real-time temperature values, temperature distribution densities and temperature differences between regions of different regions of the insulator, and accurately mark the thermal state of key parts such as the insulator shed, flange and the like. Subsequently, combined with the material characteristics of the insulator, such as the heat resistance threshold of ceramic, composite insulating material, the rated operating temperature standard of the device and the thermal field data in the history of fault-free operation, the temperature fluctuation range of the insulator under normal working conditions is demarcated, and the temperature difference interval between adjacent regions, such as between the sheds and between the shed and the flange under normal conditions is further calculated. Finally, based on the temperature difference interval, the temperature gradient range of the insulator under normal working conditions, i.e. the temperature difference interval between adjacent regions, is analyzed, and the temperature gradient threshold that can distinguish between the “normal state” and the “thermal abnormal state” is determined. When the temperature gradient of a certain region of the insulator exceeds the threshold, it can be preliminarily determined that there is a risk of thermal defects.

[0028] For the bushing components of the target power transmission and transformation equipment, the acquired ultraviolet discharge images are retrieved. Image segmentation algorithms, such as threshold segmentation, are used to separate the discharge spot region on the bushing surface from the image background. This process requires eliminating irrelevant environmental interference pixels in the image to ensure that only the discharge spot pixels generated on the bushing surface are accurately extracted. Subsequently, the total number of pixels in the discharge spot region is counted, and its proportion to the total number of pixels in the overall imaging area of ​​the bushing in the ultraviolet discharge image is calculated. This yields the discharge spot pixel ratio on the bushing surface, providing basic data for subsequent quantification of discharge intensity. Based on this, the operating voltage parameters of the target power transmission and transformation equipment are collected, including key data such as the equipment's rated operating voltage and real-time monitored operating voltage. The characteristics of bushing discharge phenomena under different voltage levels are analyzed. Since the higher the operating voltage of the power transmission and transformation equipment, the greater the electric field strength on the bushing surface, the allowable weak discharge spot area and the critical spot area for abnormal discharge under normal insulation conditions will change accordingly. By combining the discharge allowable thresholds for bushings of different voltage levels in the equipment industry standards, and the corresponding relationship between "discharge spot area - operating voltage - fault occurrence probability" in historical fault data, a correlation function between the critical area of ​​the discharge spot and the operating voltage of the equipment is established. This function takes the operating voltage as the independent variable and the critical area of ​​the discharge spot as the dependent variable. Based on the real-time input operating voltage value of the equipment, it can automatically calculate the critical value of the spot area for judging whether there is abnormal discharge in the bushing at the corresponding voltage level. This ensures that the judgment criteria for bushing discharge defects under different voltage conditions are adaptable and accurate, and avoids misjudgment or omission due to voltage differences.

[0029] For the conductor joint part of the target power transmission and transformation equipment, the collected visible light appearance image is called, the imaging area of the conductor joint is accurately framed by image cropping and region positioning technology, the interference of other components and background environment such as conductor body and insulator in the image is excluded, and the appearance gray scale features of the conductor joint are focused. Subsequently, the imaging area of the conductor joint is grayed, the average value of the gray scale values of all pixel points in the area is calculated, and the actual gray scale average value of the conductor joint is obtained; at the same time, the visible light image data of the normal running conductor joint under the same type and working condition is called, the gray scale average value is calculated and used as the reference gray scale average value, the difference between the actual gray scale average value and the reference gray scale average value is defined as the gray scale average deviation, and the abnormal degree of the appearance gray scale of the conductor joint is quantified. On this basis, combined with the material characteristics of the conductor joint, such as the normal gray scale range of copper and aluminum materials, the influence of surface treatment process such as plating and oxidation layer on gray scale, and historical defect data such as the probability of mechanical damage and poor contact corresponding to different gray scale average deviation, the defect risk frequency and distribution rule corresponding to multiple groups of different gray scale average deviation are counted. Further, taking the gray scale average deviation as the independent variable and the defect risk probability or the gray scale abnormal degree as the dependent variable, the distribution function of the gray scale deviation coefficient is defined, which needs to define the gray scale deviation coefficient value corresponding to different gray scale average deviation intervals, for example, when the gray scale average deviation is in the minimum range and close to the reference value, the gray scale deviation coefficient tends to 0, representing that the appearance of the conductor joint has no obvious abnormality; when the gray scale average deviation exceeds the normal threshold, the gray scale deviation coefficient increases with the deviation, and the increasing rate matches the defect risk increasing rate, so as to realize the quantitative representation of the gray scale abnormality of the conductor joint through the distribution function.

[0030] In a possible implementation manner, step S300 further includes: Step S310: the classification of the defect type mode is a thermal defect, a discharge defect, a mechanical damage defect and a material aging defect.

[0031] Step S320: in the infrared thermal imaging image, the thermal defect and the material aging defect in the defect type mode are distinguished by the entropy value change rate.

[0032] Step S330: at the same time, the discharge defect in the defect type mode is identified according to the spot shape factor and the gray level co-occurrence matrix of the ultraviolet discharge image.

[0033] Specifically, combined with the operation characteristics, core component functions and common fault modes of power transmission and transformation equipment, the classification basis of defect type mode is determined, and the possible defects of the equipment are divided into four categories based on the causes, forms and effects of the defects. Among them, the thermal defect is mainly caused by abnormal heating of the equipment components, which is often related to poor component contact and high load; the discharge defect is caused by the decline of the insulation performance of the equipment, often accompanied by corona, creeping and other phenomena; the mechanical damage defect refers to the physical structure damage of the equipment components caused by external force or installation error; the material aging defect is the deterioration problem of the component material caused by long-term environmental and voltage factors. Through this classification framework, the attribute category of different defects is clearly defined, which provides clear category guidance for subsequent targeted defect recognition based on multi-class photoelectric detection images, ensuring that the subsequent recognition process can match the corresponding detection data and analysis methods according to the defect type, improving the accuracy and efficiency of defect recognition.

[0034] The collected infrared thermal imaging image of the target power transmission and transformation equipment is retrieved, which can fully present the temperature distribution state of each component of the equipment. Then, the infrared thermal imaging image is regionally divided, focusing on the key component regions of insulators, wire joints and other components prone to thermal defects and material aging defects, and the temperature entropy values of each region at different detection time points are calculated. The entropy value can reflect the degree of disorder of the temperature distribution in the region, and the greater the temperature distribution difference, the higher the entropy value. On this basis, the change amplitude of the temperature entropy value of each key component region with the detection time, i.e. the entropy value change rate, is calculated. For the region corresponding to the thermal defect, since the thermal defect is often caused by poor component contact, short circuit and other problems, it shows a sudden rise in local temperature, and the temperature distribution changes significantly in a short time, so its entropy value increases rapidly, and the entropy value change rate remains at a high level. The region corresponding to the material aging defect, because the aging process is a long-term slow material degradation, the temperature rising trend is gentle and the temperature distribution is relatively uniform, the entropy value only increases slowly, and the entropy value change rate is in a low interval. By comparing the differences in entropy value change rates of different regions, the judgment rule of "high entropy value change rate corresponding to thermal defect, low entropy value change rate corresponding to material aging defect" is established, so as to accurately distinguish the two types of defects with temperature abnormalities, avoid misjudgment of defect attributes due to single temperature index, and provide accurate defect type basis for subsequent defect confidence factor configuration and severity level judgment.

[0035] The ultraviolet discharge image of the target power transmission and transformation equipment collected in the early stage is called, and the key component areas prone to discharge phenomenon such as bushings and insulators are focused on. Through image preprocessing techniques such as denoising, enhancement and optimization of image quality, interference factors such as environmental stray light and equipment reflection are excluded to ensure that the discharge spot area is clear and distinguishable. Then, the shape factor of the spot in the ultraviolet discharge image is calculated: first, the contour boundary of the spot is extracted by edge detection algorithm, the perimeter and area of the spot are counted, and then the specific value is obtained according to the shape factor calculation formula, which is 4π × area / perimeter². The shape factor can quantify the regularity of the spot contour. Normal weak discharge, such as the light spot corresponding to the tiny corona when the equipment is running normally, is usually regular circular or elliptical, and the shape factor is close to 1. Abnormal discharge, such as the discharge caused by the surface creepage of the bushing and the partial breakdown of the insulator, usually presents irregular shape due to irregular discharge path, and the shape factor deviates from 1 significantly. In this way, it is preliminarily judged whether the light spot has abnormal characteristics. At the same time, the gray level co-occurrence matrix of the light spot area is constructed: the gray level image in the light spot area is selected, and reasonable distance and angle parameters are set, such as distance 1 and angle 0° / 45° / 90° / 135°. The co-occurrence probability of different gray level pixels in the matrix under certain spatial relationship is calculated, and the feature parameters such as energy, contrast and correlation are extracted through the matrix. The gray distribution of the light spot corresponding to the discharge defect is usually uneven, and the gray level co-occurrence matrix will show the characteristics of low energy and high contrast. The gray distribution of the normal weak discharge area or the no discharge area is relatively uniform, and the matrix characteristics are high energy and low contrast. Finally, the analysis results of the shape factor of the light spot and the gray level co-occurrence matrix are combined: if the shape factor deviates from the normal range and the gray level co-occurrence matrix shows the typical characteristics of discharge defects, it is determined that there are discharge defects in the equipment; if both of them meet the characteristics of normal weak discharge, it is determined that there is no abnormal discharge. In this way, the precise recognition of the discharge defect in the defect type mode is realized, and reliable discharge characteristics are provided for the subsequent defect confidence factor configuration and severity level judgment.

[0036] In one possible implementation, step S300 further includes: Step S340: based on the device topology connection relationship, the visible light appearance image is decomposed into multiple image pyramid layers, and the edge feature component related to the device component contour is extracted.

[0037] Step S350: based on the mechanical damage defect in the defect type mode, the edge feature component is compared and analyzed.

[0038] Step S360: if the continuity breaking length of the edge feature component in the preset detection area exceeds the breaking threshold and the contour fitting error increases sharply, the dynamic correction mechanism of the gray scale deviation coefficient corresponding to the wire joint is triggered.

[0039] Specifically, taking the device topology connection relationship as the core reference basis, the physical connection logic, spatial position distribution and mutual correlation of key components such as wire joints, insulators and bushings in power transmission and transformation equipment are determined, so as to delimit the key analysis area of each component in the visible light appearance image and avoid the deviation of subsequent feature extraction caused by the confusion of component positions. Subsequently, the image pyramid construction technology is used to process the visible light appearance image in layers: starting from the original resolution image, through step-by-step downsampling, the image size is reduced to 1 / 2 of the original size each time to generate multiple image layers of different scales, forming an image pyramid containing bottom high-resolution, middle middle-resolution and high low-resolution. In the layering process, the images in each layer are simultaneously subjected to smoothing filter processing to reduce the image distortion caused by downsampling and ensure the integrity of the component outlines in each layer of images. Finally, for each layer of the image pyramid, the outline area of the device components such as the connection part of the wire joint, the shed outline of the insulator and the shell edge of the bushing is focused on, and the edge detection algorithm such as Canny edge detection is used to extract the gray gradient, edge direction and continuous length of the outline and other feature information, and these information is integrated to form an edge feature component directly related to the outline of the device component. This component can not only reflect the microscopic details of the component outline, but also reflect the macroscopic structure of the outline, such as whether the overall shape of the component is complete, providing accurate outline feature support for the subsequent comparative analysis of mechanical damage defects.

[0040] According to the definition and characteristics of mechanical damage defects in the defect type mode, a typical edge feature library of mechanical damage defects is constructed, which covers the damage morphological features of components prone to mechanical damage of power transmission and transformation equipment, such as conductor joints, insulator sheds and bushing shells, including the edge discontinuity features of fracture type damage, the elongated edge features of crack type damage, and the profile offset features of deformation type damage. At the same time, the gray scale gradient range, continuous length threshold and direction consistency parameters of various mechanical damage edges are labeled to provide standard reference for comparative analysis. Then, the extracted edge feature components related to the component profile are retrieved, which contain detailed information of the component profile in different image pyramid layers, such as the tiny crack edges of the bottom layer image and the overall deformation profile of the high layer image. For the edge feature components of key components such as conductor joints and insulators, they are classified and arranged according to the component type and high-incidence area of damage, to ensure that the edge features of each component are accurately matched with the mechanical damage edge features of the corresponding components in the feature library. Finally, dimensional comparative analysis is carried out: on the one hand, the continuity of the edge feature components is compared with the discontinuity features of the mechanical damage edges in the feature library to determine whether there are damage signs such as fracture and gap; on the other hand, the shape of the edge features is compared, such as whether it is in the form of elongated crack, whether there is profile distortion, and the degree of agreement with the typical shape of mechanical damage, and the differences between the edge gray scale gradient and direction parameters and the feature library indicators are quantified. Through multi-dimensional comparison, if the edge feature components and the typical edge features of mechanical damage defects are consistent beyond the preset matching threshold, it is preliminarily determined that the component has a risk of mechanical damage defects; if the degree of agreement is below the threshold, it is determined that the edge features have no obvious mechanical damage correlation, which provides a basis for the subsequent accurate judgment of mechanical damage defects and correction of gray deviation coefficient.

[0041] The range of the preset detection area is determined, the topological position of the conductor joint in the power transmission and transformation equipment and the high-incidence position of mechanical damage, such as the crimping position of the conductor joint and the bolt connection position, are combined to determine the area in the visible light appearance image that needs to be monitored, and the edge continuity breaking threshold is set according to the specification parameters of the conductor joint, such as the diameter, length and industry operation standard, for example, the breaking length is not more than 1 / 5 of the diameter of the conductor joint, and the contour fitting error threshold is, for example, the fitting error is not more than 0.1 mm. Subsequently, the conductor joint edge feature components extracted in the early stage are analyzed in detail: on the one hand, the continuity breaking of the edge feature components in the preset detection area is counted, the actual breaking length is obtained through pixel distance conversion, and is compared with the preset breaking threshold; on the other hand, the least square method or the Bezier curve fitting algorithm is used to fit and calculate the ideal contour of the conductor joint and the actual extracted edge contour, and the contour fitting error value is obtained, and whether the error suddenly increases is observed, for example, the error value suddenly increases from 0.08 mm to 0.2 mm. If the analysis results of the above two items both meet the determination condition, that is, the edge continuity breaking length exceeds the preset breaking threshold, and the contour fitting error suddenly increases, it is indicated that the conductor joint is highly likely to have mechanical damage, such as breaking and deformation, and the mechanical damage may cause the abnormal distribution of the gray scale on the surface of the conductor joint, thereby affecting the accuracy of the gray scale deviation coefficient calculated in the early stage. At this time, the gray scale deviation coefficient dynamic correction mechanism corresponding to the conductor joint is automatically triggered: the gray scale image data of the mechanical damage area is called, the actual gray scale mean value of the area is recalculated, the reference gray scale mean value is adaptively adjusted according to the damage degree, and the gray scale deviation coefficient is recalculated based on the corrected actual gray scale mean value and the reference gray scale mean value, so as to eliminate the interference of the mechanical damage on the gray scale deviation coefficient and ensure the accuracy and reliability of the subsequent defect analysis and determination results based on the gray scale deviation coefficient.

[0042] In one possible implementation manner, the step S400 further includes: Step S410: taking the infrared temperature peak value, the ultraviolet spot energy and the visible light edge gradient value as input variables.

[0043] Step S420: based on the input variables, fusing the feature matrix and updating the support function of defect recognition.

[0044] Step S430: using the support function to perform correlation coefficient feedback checking on the defect confidence factor.

[0045] Specifically, for the three types of core photoelectric detection images collected, key quantitative indicators that can accurately represent the state of equipment defects are extracted. From the infrared thermal imaging image, focus on the areas of components such as insulators and wire joints that are prone to thermal anomalies. Through a temperature extreme detection algorithm, the highest temperature point in each region is located, and the highest temperature value is defined as the infrared temperature peak value, which quantifies the heat intensity of thermal defects. From the ultraviolet discharge image, first, through image segmentation and energy conversion model, combined with the pixel area, gray mean value of the discharge spot, and the ultraviolet light radiation intensity calibration parameter, the energy value of the discharge spot, i.e., the ultraviolet spot energy, is calculated, which is used to measure the discharge intensity of discharge defects. From the visible light appearance image, the gradient of the outline edge of equipment components such as wire joints and insulators is calculated, and the gradient value of the pixel point corresponding to the most severe change in the gray value of the edge is selected as the visible light edge gradient value, which reflects the abnormal degree of the outline edge caused by mechanical damage defects. Finally, the extracted infrared temperature peak value, ultraviolet spot energy, and visible light edge gradient value are integrated and used as the core input variables for subsequent defect recognition and analysis, providing multi-dimensional and quantitative data support for the configuration of defect confidence factors and the update of support functions.

[0046] The mapping relationship between the input variables and the structured information in the feature matrix, such as the insulator temperature gradient threshold, the critical area of the bushing discharge light spot, the conductor joint gray deviation coefficient, and the device topology connection relationship, needs to be established. For example, the infrared temperature peak value is calculated by difference with the insulator temperature gradient threshold to quantify the degree of temperature anomaly; the ratio analysis of the ultraviolet light spot energy and the critical area of the bushing discharge light spot is performed to determine whether the discharge intensity is excessive; the visible light edge gradient value is associated with the conductor joint gray deviation coefficient to reflect the influence of mechanical damage on the appearance gray scale. Subsequently, the multi-source data fusion algorithm, such as weighted fusion, is used to integrate the association results of the input variables and the feature matrix: according to the weight proportion of different defect types (thermal defect, discharge defect, and mechanical damage defect), the fusion weight of each association result is allocated, for example, the weight of the association result of the infrared temperature peak value and the insulator temperature gradient threshold is increased in thermal defect analysis, the weight of the association result of the ultraviolet light spot energy and the critical area of the bushing discharge light spot is increased in discharge defect analysis, to ensure that the fused data can highlight the core features of different defect types. Finally, based on the fused multi-dimensional data, combined with the corresponding rules of "input variables-feature matrix-defect existence probability" in the device historical defect database, the original defect recognition support function is updated: the coefficients of each parameter in the function are adjusted, so that the support value output by the function can more accurately match the corresponding relationship between the current device state and the existence of defects, for example, when the infrared temperature peak value far exceeds the insulator temperature gradient threshold and the fused data shows strong association, the support function outputs a high support value, indicating a high possibility of thermal defects; otherwise, a low support value is output, thereby updating the support function to improve the accuracy of defect state representation, providing a reliable tool for subsequent defect confidence factor verification.

[0047] The updated defect recognition support function is called, and the support value of the corresponding defect under the current equipment state is calculated by combining the photoelectric detection data of the current equipment and the recognized defect type mode. The value is used to quantify the matching degree of the current data characteristics and the existence of the defect. The higher the value, the greater the possibility of the existence of the corresponding defect under the current state. Subsequently, the preliminary configured defect confidence factor is called, which is used to represent the reliability of the defect recognition result and is strongly related to the input layer weight of the defect recognition model. By statistical methods such as Pearson correlation coefficient calculation method, a correlation model between the support value output by the support function and the defect confidence factor is established, and the correlation coefficient between the two is calculated. The value range of the correlation coefficient is [-1, 1], and the closer to 1, the higher the matching degree, and the closer to -1 or in the lower interval, the lower the matching degree, and there may be a deviation in the configuration of the confidence factor. Finally, according to the preset correlation coefficient threshold, such as 0.7, the calculation result is judged: if the correlation coefficient is higher than the threshold, it means that the defect confidence factor can accurately reflect the actual matching degree of the current equipment defect, and the confidence factor does not need to be adjusted and can be directly used for subsequent determination of the defect severity level; if the correlation coefficient is lower than the threshold, it means that the matching degree of the defect confidence factor and the current equipment state and the support function is insufficient, which may be caused by detection environment interference, data error or initial configuration deviation. At this time, the output result of the support function is taken as the feedback basis to modify and adjust the defect confidence factor, and the confidence factor value is fine-tuned by the correlation coefficient deviation ratio. The correlation coefficient is recalculated after adjustment until the coefficient meets the threshold requirement, so as to ensure the accuracy of the defect confidence factor through closed loop verification and provide reliable guarantee for subsequent defect severity level determination.

[0048] In one possible implementation, step S400 further includes: Step S440: setting a mapping relationship matrix of the support function and the input layer weight of the defect recognition model, and decomposing to determine the influence weight vector of the defect confidence factor.

[0049] Step S450: when the correlation coefficient of the defect confidence factor is lower than the correlation coefficient threshold for Q consecutive detection periods, triggering the weight self-adaptive updating mechanism.

[0050] Specifically, the association dimension between the support function and the input layer weight of the defect recognition model is determined, the support function can output the quantitative results reflecting the matching degree of the current device defect, covering the infrared temperature feature, the ultraviolet spot feature, the visible light edge feature and other defect-related dimensions, and the input layer weight of the defect recognition model corresponds to the importance of each feature in the model, such as the infrared temperature feature weight, the ultraviolet spot energy feature weight, the visible light edge gradient feature weight, etc. Based on these dimensions, a mapping relationship matrix is constructed, the rows of the matrix correspond to the output feature dimensions of the support function, the columns correspond to the weight parameters of the input layer of the model, and each element in the matrix is determined according to the historical detection data and the defect recognition result, and the influence coefficient of the output feature of the support function on the weight of the input layer is determined by statistical analysis, so as to establish the direct association between the two. Subsequently, a matrix decomposition algorithm such as singular value decomposition is used to decompose the mapping relationship matrix constructed, and the parameter component related to the defect confidence factor in the matrix is separated out. Since the defect confidence factor is strongly associated with the input layer weight of the defect recognition model, and the support function has established a connection with the confidence factor through the correlation coefficient feedback check, the parameter set reflecting only the influence of the defect confidence factor on the input layer weight can be extracted from the mapping relationship by matrix decomposition. Further integrating these parameters into a vector form, the influence weight vector of the defect confidence factor is obtained, each element of the vector corresponds to a weight parameter of the input layer of the model, and the element value directly reflects the degree of influence on the input layer weight when the defect confidence factor changes. For example, the larger the value of an element in the vector, the higher the adjustment amplitude requirement of the defect confidence factor change on the input layer weight, thereby providing a clear parameter adjustment basis for the subsequent adaptive update of the weight.

[0051] The core monitoring index and determination standard are determined. The core index is the correlation coefficient of the defect confidence factor. The correlation coefficient is calculated by the association of the support function and the defect confidence factor, and is used to reflect the matching reliability of the confidence factor and the actual defect state of the equipment. At the same time, two key parameters are preset: one is the correlation coefficient threshold, such as 0.6, which is set according to the equipment operation accuracy requirement and the historical defect identification accuracy; the other is the number of continuous detection periods Q, such as 3 periods, which is determined in combination with the equipment detection frequency and the defect development rate. In the actual detection process, in each detection period, the correlation coefficient of the defect confidence factor and the support function output value is recalculated, and whether the coefficient is lower than the preset threshold is recorded. If the correlation coefficient of a certain detection period is lower than the threshold, the continuous period count is started; if the correlation coefficients of the subsequent detection periods continue to be lower than the threshold, and the number of periods that are continuously lower than the threshold reaches Q, it is determined that the current defect confidence factor cannot accurately adapt to the equipment defect state, and the input layer weight of the defect recognition model associated with the confidence factor may also cause the recognition accuracy to decrease due to the deviation of the confidence factor. At this time, the weight self-adaptive updating mechanism is automatically triggered: based on the determined defect confidence factor influence weight vector, the model input layer weight is differentially adjusted according to the degree of influence of each input layer weight on the confidence factor, in combination with the current photoelectric detection data, infrared thermal imaging, ultraviolet discharge, visible light image data and the recognized defect type mode, for example, if the coefficient of the ultraviolet light spot feature weight in the influence weight vector is high, the weight is preferentially corrected, so that the adjusted input layer weight can accurately match the updated defect state, and the accuracy and reliability of subsequent defect recognition are improved.

[0052] In one possible implementation manner, step S400 further includes: The learning rate is dynamically adjusted according to the change of the operation environment of the power transmission and transformation equipment, and a cosine annealing strategy is adopted: , wherein, is the learning rate of the tthiteration, and are the maximum learning rate and the minimum learning rate respectively, and T is the total number of iterations.

[0053] Specifically, first, the core goal of learning rate adjustment is determined. Since the operation environment of the power transmission and transformation equipment (such as temperature, humidity, illumination, etc.) will dynamically change with seasons, weather and other factors, it may cause fluctuations in photoelectric detection data, and then affect the training accuracy and stability of the defect recognition model. Therefore, the model learning rate needs to be dynamically adjusted according to the environmental change to ensure that the model can continuously adapt to the environmental change and maintain good defect recognition performance.

[0054] The cosine annealing strategy is adopted to realize the dynamic adjustment of the learning rate, and the core formula is: The meanings of the parameters in the formula are clear: represents the learning rate at the t-th iteration, is the actual learning rate value applied to the model training after each adjustment; and respectively represent the preset minimum learning rate and maximum learning rate, which are used to limit the adjustment range of the learning rate, so as to avoid unstable model training caused by excessively high learning rate or slow training convergence caused by excessively low learning rate; T is the total number of iterations, that is, the number of iteration steps required to complete a complete learning rate adjustment cycle. As the number of model training iterations t gradually increases from 0 to T, the learning rate will periodically fluctuate between and : at the beginning of the iteration, the learning rate starts from , and decreases slowly according to the cosine curve with the increase of t, so as to ensure that the model can quickly update the parameters with a relatively high learning rate at the beginning of the training to approach the optimal solution; as the number of iterations approaches T, the learning rate gradually decreases to , at which time the model parameters have tended to be stable, and a smaller learning rate can avoid the oscillation of the parameters around the optimal solution, thereby improving the stability and accuracy of the model training. Through this cosine annealing strategy, the dynamic changes of the operation environment of the power transmission and transformation equipment can be effectively adapted, so that the defect recognition model can maintain high training efficiency and accurate defect recognition ability under different environmental conditions.

[0055] In one possible implementation manner, step S400 further includes: setting a defect recognition model under a bidirectional channel attention mechanism.

[0056] Among them, the first one-way channel adopts a convolutional neural network to extract temperature anomaly features of the infrared thermal imaging image, constructs a thermal defect factor in combination with a local extreme value of a temperature field and a temperature gradient threshold, and uses a channel attention mechanism to adaptively adjust the weight of different temperature intervals to highlight the features of abnormal heat zones.

[0057] The second one-way channel converts the ultraviolet discharge image to the frequency domain through wavelet transform, extracts discharge energy features using a frequency domain convolution kernel, and identifies the distribution law of discharge light spots in combination with a spatial attention mechanism.

[0058] Specifically, the defect recognition model under the bidirectional channel attention mechanism is set, which extracts and optimizes the core features of the infrared thermal imaging image and the ultraviolet discharge image through two independent and complementary one-way channels, and simultaneously integrates the attention mechanism to improve the recognition accuracy of the defect features.

[0059] The temperature anomaly feature processing of the first one-way channel focusing on the infrared thermal imaging image includes the following steps. First, a convolutional neural network is used to perform multi-round convolution and pooling operations on the infrared thermal imaging image, so as to gradually extract temperature features at different levels in the image, and to focus on capturing temperature distribution differences of device components such as insulators and wire joints. Then, the temperature field local extreme values obtained in the early stage, such as the specific temperature value of the high-temperature point of the component and the temperature gradient threshold corresponding to the insulator of the device, are fused with the features extracted by convolution, to construct a thermal defect factor that can intuitively reflect the degree of thermal defects. Finally, a channel attention mechanism is introduced, the contribution of features in different temperature intervals to defect recognition is calculated, and the weights of different temperature intervals are adaptively adjusted. For example, higher weights are assigned to abnormal heat zone features that exceed the temperature gradient threshold, and lower weights are assigned to normal temperature interval features, so as to highlight the abnormal heat zone features and improve the pertinence of thermal defect recognition.

[0060] The second one-way channel focuses on the discharge energy feature analysis of the ultraviolet discharge image. First, a wavelet transform algorithm is used to convert the ultraviolet discharge image from the spatial domain to the frequency domain, so that the discharge light spots and background noise can be more clearly separated in the frequency domain, and the influence of environmental interference on feature extraction is reduced. Then, a frequency domain convolution kernel is used to perform convolution operation on the converted frequency domain image, so as to accurately extract the energy distribution features of the discharge region and quantify the size and range of the discharge intensity. At the same time, a spatial attention mechanism is used to analyze the spatial position and distribution density of the discharge light spots in the image, to generate a spatial attention weight map, to give higher attention to the concentrated discharge light spot area, and to reduce the weight of the background area without discharge or with weak discharge, so as to more accurately recognize the distribution rule of the discharge light spots and provide a reliable basis for the judgment of discharge defects. Through the synergistic effect of the two one-way channels, the bidirectional channel attention model can efficiently process two types of photoelectric detection data at the same time, comprehensively capture the key features of thermal defects and discharge defects, and improve the accuracy and efficiency of the overall defect recognition.

[0061] In one possible implementation, the step S500 further includes: Step S510: A multi-scale feature fusion unit is set to perform tensor product operation on the shallow image texture features and the deep defect semantic features.

[0062] Step S520: At the same time, the classification results of the defect severity level and the defect area regression prediction value are output through the full connection layer respectively. When the power transmission equipment detection scene triggers the environmental light replacement condition or the humidity replacement condition, the incremental learning mechanism is automatically triggered to perform local parameter fine-tuning in combination with the detection environment change response mechanism.

[0063] Specifically, a multi-scale feature fusion unit is set up, which needs to integrate the features extracted at different levels in the early defect recognition model, and make up for the shortcomings of single-scale features in detail capture or semantic understanding. Among them, the shallow image texture features mainly come from the bottom of the image pyramid or the shallow network of the model, such as the initial convolutional layer of the bidirectional channel attention model. This kind of feature retains the microscopic detail information of the device components, such as the surface texture of the wire joint, the trace direction of the insulator umbrella skirt, and the fine scratches on the sleeve shell, which can provide accurate detail support for defect positioning. The deep defect semantic features come from the deep network of the model, such as the late feature extraction layer of the bidirectional channel attention model, which is a high-order feature optimized by multiple convolutions and attention mechanisms, and has been separated from the original image pixel level, which can abstractly reflect the essential attributes of defects, such as the temperature anomaly area semantics corresponding to thermal defects, the light spot distribution semantics corresponding to discharge defects, and the contour breaking semantics corresponding to mechanical damage defects, which can realize accurate classification of defect types. Subsequently, the multi-scale feature fusion unit is started to perform tensor product operation on the two types of features: before the operation, the dimension matching processing of the shallow image texture features and the deep defect semantic features is needed to ensure that the two are consistent in feature dimension and spatial resolution, avoiding fusion bias caused by dimension difference; in the tensor product operation process, each detail dimension of the shallow feature and each semantic dimension of the deep feature are associated and calculated to generate a fusion feature tensor that combines detail information and semantic information. This tensor not only contains the microscopic texture details of the device components, but also integrates the semantic attributes of the defects, which can not only accurately locate the defect position through texture features, but also clearly identify the defect type through semantic features, realizing the dual support of "positioning-qualification". Through this fusion process, the problem of insufficient semantic information of shallow features and loss of details of deep features is effectively solved.

[0064] The shallow image texture features and deep defect semantic features fused by tensor product operation are input into the fully connected layer of the defect recognition model. The fully connected layer contains multiple neuron structures, which realizes high-order mapping and information integration of the fusion features through nonlinear transformation and weight matrix operation. Among them, one branch of the fully connected layer is responsible for defect severity classification: referring to the pre-configured defect confidence factor and the recognized defect type pattern, combining the preset grade determination standard, such as dividing thermal defects into slight, moderate, and severe according to the temperature exceeding the threshold range, outputting the classification result of defect severity, and clearly defining the impact of the current defect on the equipment operation; the other branch is responsible for defect area regression prediction: based on the pixel distribution, geometric contour, and other information of the defect area in the fusion features, the regression algorithm such as neural network regression is used to calculate and output the actual area regression prediction value of the defect, quantifying the physical size of the defect.

[0065] Meanwhile, a detection environment change response mechanism is started, and environmental parameters of a detection scene of power transmission and transformation equipment are collected and monitored in real time. Two key indicators, i.e., environmental light intensity and air humidity, are focused on. A preset stable light intensity range, such as 500-10000 lux, and a normal humidity threshold, such as 30%-70%, are set. When it is monitored that the environmental light intensity exceeds the stable range, such as strong light direct radiation causes light to rise sharply, rainy weather causes light to drop sharply, or humidity breaks the normal threshold, such as high humidity fog causes humidity to exceed the threshold, it is determined that the environmental replacement condition is triggered. At this time, an incremental learning mechanism is automatically started: without retraining the entire defect recognition model, only a small amount of photoelectric detection data collected after the environment changes, such as 10-50 groups of samples, are called to fine-tune the local parameters of the model, such as part of the weight parameters of the full connection layer and the correlation coefficient of the feature fusion unit. Through gradient descent update of the small batch data, the model quickly adapts to the new detection environment, avoids feature extraction deviation or recognition accuracy decline caused by changes in environmental light and humidity, and ensures the continuous accuracy of the defect severity classification result and the defect area regression prediction value, thereby providing reliable data support for defect operation and maintenance of power transmission and transformation equipment.

[0066] In the second embodiment, based on the same inventive concept as the defect recognition method of the photoelectric detection image of the power transmission and transformation equipment in the foregoing embodiments, as shown in the second embodiment, the present application provides a defect recognition system of a photoelectric detection image of power transmission and transformation equipment. The system and method embodiments in the present application are based on the same inventive concept. The system comprises: Figure 2 A photoelectric detection data acquisition module 10 is configured to acquire photoelectric detection data of a target power transmission and transformation equipment, including an infrared thermal imaging image, an ultraviolet discharge image, and a visible light appearance image.

[0067] A feature matrix preparation module 20 is configured to prepare a feature matrix by coupling analysis based on a temperature gradient threshold value corresponding to an insulator of the equipment, a discharge light spot critical area corresponding to a sleeve surface, a gray scale deviation coefficient corresponding to a conductor joint, and a device topology connection relationship.

[0068] A defect type pattern recognition module 30 is configured to simultaneously recognize a defect type pattern based on the device topology connection relationship and a change trend of the gray scale deviation coefficient in multiple detection periods.

[0069] A defect confidence factor configuration module 40 is configured to configure a defect confidence factor based on the feature matrix, the defect type pattern, and the photoelectric detection data. The defect confidence factor is strongly associated with an input layer weight of a defect recognition model.

[0070] A hierarchical reminding module 50 is configured to determine a defect severity level and perform hierarchical reminding by using the defect confidence factor.

[0071] ​Further, the system is also used to realize the following functions: According to the thermal field distribution map of the key components of the equipment, the temperature gradient threshold corresponding to the insulator of the equipment is obtained; the pixel proportion of the discharge light spot on the surface of the bushing is extracted, and the correlation function of the critical area of the discharge light spot and the operating voltage of the equipment is set; the distribution function of the gray deviation coefficient is defined through the gray mean deviation of the wire joint.

[0072] Further, the system is also used to realize the following functions: The classification of the defect type mode is thermal defect, discharge defect, mechanical damage defect and material aging defect; in the infrared thermal imaging image, the thermal defect and the material aging defect in the defect type mode are distinguished through the entropy change rate; at the same time, the discharge defect in the defect type mode is identified according to the spot shape factor and the gray level co-occurrence matrix of the ultraviolet discharge image.

[0073] Further, the system is also used to realize the following functions: Based on the device topology connection relationship, the visible light appearance image is decomposed into multiple image pyramid layers, and the edge feature component related to the profile of the equipment component is extracted; based on the mechanical damage defect in the defect type mode, the edge feature component is compared and analyzed; if the continuous fracture length of the edge feature component in the preset detection area exceeds the fracture threshold and the profile fitting error increases sharply, the dynamic correction mechanism of the gray deviation coefficient corresponding to the wire joint is triggered.

[0074] Further, the system is also used to realize the following functions: The infrared temperature peak value, the ultraviolet light spot energy and the visible light edge gradient value are taken as input variables; based on the input variables, the feature matrix is fused, and the support function of defect recognition is updated; the correlation coefficient feedback check of the defect confidence factor is carried out by using the support function.

[0075] Further, the system is also used to realize the following functions: The mapping relationship matrix of the support function and the input layer weight of the defect recognition model is set, and the influence weight vector of the defect confidence factor is determined by decomposition; when the correlation coefficient of the defect confidence factor is lower than the correlation coefficient threshold limit for Q consecutive detection periods, the weight self-adaptive updating mechanism is triggered.

[0076] Further, the system is also used to realize the following functions: According to the change of the operation environment of the power transmission and transformation equipment, the learning rate is dynamically adjusted, and the cosine annealing strategy is adopted: , wherein, is the learning rate of the tth iteration, and are the maximum learning rate and the minimum learning rate respectively, and T is the total number of iterations.

[0077] Further, the system is also used to realize the following functions: The defect identification model under the setting of the bidirectional channel attention mechanism is set; wherein the first one-way channel adopts the convolutional neural network to extract the temperature anomaly features of the infrared thermal imaging image, combines the temperature field local extreme value and the temperature gradient threshold to construct the thermal defect factor, and uses the channel attention mechanism to adaptively adjust the weight of different temperature intervals to highlight the abnormal heat area features; the second one-way channel converts the ultraviolet discharge image to the frequency domain through wavelet transform, adopts the frequency domain convolution kernel to extract the discharge energy features, and combines the spatial attention mechanism to identify the distribution law of the discharge light spot.

[0078] Further, the system is also used to realize the following functions: The multi-scale feature fusion unit is set to perform tensor product operation on the shallow image texture features and the deep defect semantic features; at the same time, the full connection layer is used to respectively output the classification results of the defect severity level and the defect area regression prediction value, and the detection environment change response mechanism is combined, when the power transmission equipment detection scene triggers the environmental light replacement condition or the humidity replacement condition, the incremental learning mechanism is automatically triggered to perform local parameter fine-tuning.

[0079] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. And the above describes the specific embodiments of the present application. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0080] The above only describes the preferred embodiments of the present application, and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0081] The present application and the drawings are only exemplary descriptions of the present application, and are considered to cover any and all modifications, changes, combinations or equivalents within the scope of the present application. Obviously, those skilled in the art can make various modifications and changes to the present application without departing from the scope of the present application. Thus, if these modifications and changes of the present application belong to the scope of the present application and its equivalent technology, the present application intends to include these modifications and changes.

Claims

1. A method for defect identification from photoelectric detection images of power transmission and transformation equipment, characterized in that, The method includes: The acquisition of photoelectric detection data for the target power transmission and transformation equipment includes infrared thermal imaging images, ultraviolet discharge images, and visible light appearance images; Based on the temperature gradient threshold corresponding to the equipment insulator, the critical area of ​​the discharge spot corresponding to the bushing surface, and the grayscale deviation coefficient corresponding to the conductor joint, coupled analysis is performed in conjunction with the equipment topology connection relationship to derive the feature matrix. Simultaneously, based on the device topology connection relationship and the changing trend of the grayscale deviation coefficient over multiple detection cycles, defect type patterns are identified; Based on the feature matrix, combined with the defect type pattern and the photoelectric detection data, a defect confidence factor is configured, and the defect confidence factor is strongly correlated with the input layer weights of the defect identification model. The defect confidence factor is used to determine the severity level of the defect and to issue a graded alert.

2. The defect identification method for photoelectric detection images of power transmission and transformation equipment as described in claim 1, characterized in that, Based on the temperature gradient threshold corresponding to the equipment insulator, the critical area of ​​the discharge spot corresponding to the bushing surface, and the grayscale deviation coefficient corresponding to the conductor joint, the method further includes: Based on the thermal field distribution map of the key components of the equipment, obtain the temperature gradient threshold corresponding to the equipment insulator; Extract the percentage of discharge spot pixels on the surface of the sleeve, and set a correlation function between the critical area of ​​the discharge spot and the operating voltage of the equipment; The distribution function of the grayscale deviation coefficient is defined by the grayscale mean deviation of the wire connector.

3. The defect identification method for photoelectric detection images of power transmission and transformation equipment as described in claim 2, characterized in that, The defect type patterns are classified as thermal defects, discharge defects, mechanical damage defects, and material aging defects. In the infrared thermal imaging image, thermal defects and material aging defects in the defect type pattern are distinguished by the rate of change of entropy value. Simultaneously, discharge defects in the defect type pattern are identified based on the spot shape factor and gray-level co-occurrence matrix of the ultraviolet discharge image.

4. The defect identification method for photoelectric detection images of power transmission and transformation equipment as described in claim 3, characterized in that, Based on the device topology connection relationship, the visible light appearance image is decomposed into multiple image pyramid layers, and edge feature components related to the device component contours are extracted. Based on the mechanical damage defects in the defect type pattern, a comparative analysis is performed with the edge feature components; If the edge feature component experiences a continuous break in the preset detection area with a break length exceeding the break threshold and the contour fitting error increases sharply, then the dynamic correction mechanism for the grayscale deviation coefficient corresponding to the wire connector is triggered.

5. The defect identification method for photoelectric detection images of power transmission and transformation equipment as described in claim 1, characterized in that, Based on the feature matrix, combined with the defect type pattern and the photoelectric detection data, a defect confidence factor is configured, and the method further includes: Infrared temperature peak, ultraviolet spot energy, and visible light edge gradient value are used as input variables; Based on the input variables, the feature matrix is ​​fused, and the support function for defect identification is updated; The defect confidence factor is validated using the support function with correlation coefficient feedback.

6. The defect identification method for photoelectric detection images of power transmission and transformation equipment as described in claim 5, characterized in that, The defect confidence factor is strongly correlated with the input layer weights of the defect identification model, and the method includes: Set the mapping relationship matrix between the support function and the input layer weights of the defect identification model, and decompose and determine the influence weight vector of the defect confidence factor; When the correlation coefficient of the defect confidence factor is lower than the correlation coefficient threshold for Q consecutive detection cycles, the weight adaptive update mechanism is triggered.

7. The defect identification method for photoelectric detection images of power transmission and transformation equipment as described in claim 6, characterized in that, The learning rate is dynamically adjusted based on changes in the operating environment of power transmission and transformation equipment, and a cosine annealing strategy is adopted. ,in, Let be the learning rate for the t-th iteration. and These represent the maximum and minimum learning rates, respectively, and T is the total number of iterations.

8. The defect identification method for photoelectric detection images of power transmission and transformation equipment as described in claim 7, characterized in that, Set up a defect identification model under a bidirectional channel attention mechanism; Among them, the first unidirectional channel uses a convolutional neural network to extract the temperature anomaly features of infrared thermal imaging images, combines local extreme values ​​of the temperature field and temperature gradient thresholds to construct a thermal defect factor, and uses a channel attention mechanism to adaptively adjust the weights of different temperature ranges to highlight the features of abnormal hot areas. The second unidirectional channel converts the ultraviolet discharge image to the frequency domain through wavelet transform, extracts the discharge energy features using frequency domain convolution kernels, and identifies the distribution pattern of the discharge spot by combining spatial attention mechanism.

9. The defect identification method for photoelectric detection images of power transmission and transformation equipment as described in claim 8, characterized in that, The method further includes: A multi-scale feature fusion unit is set up to perform tensor product operation on shallow image texture features and deep defect semantic features; Meanwhile, the classification results of the severity level of defects and the regression prediction value of defect area are output through the fully connected layer. Combined with the response mechanism for changes in the detection environment, when the detection scenario of power transmission and transformation equipment triggers changes in ambient light or humidity, the incremental learning mechanism is automatically triggered to fine-tune local parameters.

10. A defect recognition system for photoelectric detection images of power transmission and transformation equipment, characterized in that, The system is used to implement the defect identification method for photoelectric detection images of power equipment according to any one of claims 1-9, and the system comprises: The photoelectric detection data acquisition module is used to acquire photoelectric detection data of the target power transmission and transformation equipment, including infrared thermal imaging images, ultraviolet discharge images, and visible light appearance images. The feature matrix formulation module is used to formulate the feature matrix by performing coupled analysis based on the temperature gradient threshold corresponding to the equipment insulator, the critical area of ​​the discharge spot corresponding to the bushing surface, the gray scale deviation coefficient corresponding to the conductor joint, and the equipment topology connection relationship. The defect type pattern recognition module is used to simultaneously identify defect type patterns based on the device topology connection relationship and the changing trend of the grayscale deviation coefficient over multiple detection cycles. The defect confidence factor configuration module is used to configure a defect confidence factor based on the feature matrix, combined with the defect type pattern and the photoelectric detection data. The defect confidence factor is strongly correlated with the input layer weights of the defect recognition model. The graded alert module is used to determine the severity level of a defect and provide graded alerts using the defect confidence factor.

Images