Substation equipment infrared defect diagnosis method, device and equipment

Abstract

The application discloses a substation equipment infrared defect diagnosis method, device and equipment, relates to the technical field of substation equipment detection, and includes: taking a single-mode infrared thermal image as input, extracting multi-scale features from a backbone feature extraction network step by step, establishing long-distance spatial position dependence by a coordinate attention enhancement network in a neck network, decoupling a detection network to separate attribute classification and boundary regression, and then reversely mapping a positioning boundary box to an absolute temperature field matrix to extract a temperature difference, combining double threshold value cascade determination to output quantitative grading results. The beneficial effect is to effectively suppress environmental thermal noise interference such as solar radiation and metal high reflection, to strengthen the expression of weak heat sources from the feature level, to avoid positioning box drift and small heat point missed detection, to get rid of the dependence on visible light assistance, to eliminate manual secondary segmentation temperature measurement and table lookup judgment, and to significantly improve the diagnosis accuracy, real-time performance and environmental adaptability of substation live inspection.

Description

Technical Field

[0001] This invention relates to the field of substation equipment testing technology, and in particular to a method, apparatus and equipment for infrared defect diagnosis of substation equipment. Background Technology

[0002] Currently, infrared defect diagnosis of substation equipment mainly adopts a dual-light fusion detection scheme, which involves spatially registering and fusing visible light images with infrared images, and using the high-frequency edge information of the visible light image to assist in the detection and positioning of infrared targets. In the diagnostic decision-making stage, the detected heat-generating areas are manually segmented into secondary images, temperature is measured and calculated point by point, and the temperature difference threshold table in the power industry standard is manually consulted to make a defect classification judgment.

[0003] However, the dual-light fusion scheme is highly dependent on visible light sources and cannot work at night or in low-light environments. Furthermore, there are complex spatial nonlinear registration, focusing, and hardware clock synchronization problems between heterogeneous dual lenses, resulting in poor environmental adaptability. At the same time, existing infrared defect target detection networks, lacking high-frequency edge constraints from visible light, are prone to problems such as frame drift and the submersion of tiny heat points when faced with environmental interference such as strong solar radiation and high reflectivity noise from metals. Moreover, the qualitative classification results output by the detection network are mismatched with the quantitative decision-making requirements of power diagnostic business procedures. In addition, the manual secondary segmentation, temperature measurement calculation, and table lookup decision process is cumbersome, consuming a lot of manpower and time costs, and is easily affected by temperature measurement errors caused by meteorological conditions.

[0004] Therefore, how to accurately locate and quantitatively classify the minor heating defects of voltage-induced heating equipment in substation live-line inspection scenarios where there is a lack of visible light assistance and strong environmental thermal noise interference has become an urgent problem to be solved. Summary of the Invention

[0005] The main purpose of this application is to provide a method, device and equipment for infrared defect diagnosis of substation equipment, which aims to solve the technical problem of how to accurately locate and quantitatively grade the weak heating defects of voltage-heated equipment.

[0006] To achieve the above objectives, this application proposes a method for infrared defect diagnosis of substation equipment, comprising: Acquire single-mode infrared thermal images of substation equipment; The single-modal infrared thermal image is input into a preset defect diagnosis model for feature extraction and defect detection to obtain equipment attribute category labels and defect location bounding box coordinates. The preset defect diagnosis model includes a backbone feature extraction network, a neck network, and a decoupled detection network. The backbone feature extraction network is composed of a series of continuous convolutional layers, stride downsampling layers, and residual bottleneck structures. The neck network is composed of a neck feature pyramid layer and a coordinate attention enhancement network. The decoupled detection network consists of parallel attribute classification branches and boundary regression branches. Based on the coordinates of the defect location bounding box, the highest fault heating temperature and normal operating temperature are extracted from the absolute temperature field matrix; The phase-to-phase temperature difference is calculated based on the highest heat-generating temperature of the fault and the normal operating temperature. The phase-to-phase temperature difference is then subjected to double threshold cascading judgment based on the preset severe anomaly benchmark threshold corresponding to the equipment attribute category label, thereby obtaining the equipment diagnosis result.

[0007] Furthermore, to achieve the above objectives, this application also proposes an infrared defect diagnosis device for substation equipment, the substation equipment infrared defect diagnosis device comprising: The data acquisition module is used to acquire single-mode infrared thermal images of substation equipment; The defect detection module is used to input the single-modal infrared thermal image into a preset defect diagnosis model for feature extraction and defect detection, and to obtain equipment attribute category labels and defect location bounding box coordinates. The preset defect diagnosis model includes a backbone feature extraction network, a neck network, and a decoupled detection network. The backbone feature extraction network is composed of a series of continuous convolutional layers, stride downsampling layers, and residual bottleneck structures. The neck network is composed of a neck feature pyramid layer and a coordinate attention enhancement network. The decoupled detection network consists of parallel attribute classification branches and boundary regression branches. The temperature extraction module is used to extract the highest fault heating temperature and normal operating temperature from the absolute temperature field matrix based on the coordinates of the defect location bounding box. The graded diagnosis module is used to calculate the interphase temperature difference based on the highest heat temperature of the fault and the normal operating temperature, and to perform double threshold cascade judgment on the interphase temperature difference based on the preset severe abnormality benchmark threshold corresponding to the equipment attribute category label, so as to obtain the equipment diagnosis result.

[0008] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the infrared defect diagnosis method for substation equipment as described above.

[0009] In addition, to achieve the above objectives, this application also proposes a substation equipment infrared defect diagnosis device, the device comprising: a memory, a processor, and a substation equipment infrared defect diagnosis program stored in the memory and running on the processor, the substation equipment infrared defect diagnosis program being configured to implement the steps of the substation equipment infrared defect diagnosis method as described above.

[0010] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the substation equipment infrared defect diagnosis method described above.

[0011] This application uses single-modal infrared thermal images as input, extracts multi-scale features step-by-step through a backbone feature extraction network, establishes long-distance spatial position dependence through a coordinate attention enhancement network in the neck network, decouples the detection network to separate attribute classification and boundary regression, and then maps the localized bounding box back to the absolute temperature field matrix to extract temperature difference. Combined with a double-threshold cascade, it outputs quantitative and graded results. The beneficial effects include effectively suppressing environmental thermal noise interference such as solar radiation and high metal reflectivity, strengthening the expression of weak heat sources at the feature level, avoiding localization box drift and missed detection of tiny heat points, eliminating dependence on visible light assistance, eliminating manual secondary segmentation temperature measurement and table lookup decision-making, and significantly improving the diagnostic accuracy, real-time performance, and environmental adaptability of substation live-line inspections. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the first embodiment of the infrared defect diagnosis method for substation equipment according to this application; Figure 2 This is a flowchart illustrating the second embodiment of the infrared defect diagnosis method for substation equipment according to this application; Figure 3 This is a schematic diagram of the module structure of the infrared defect diagnosis device for substation equipment in this application; Figure 4 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the infrared defect diagnosis method for substation equipment in this application embodiment.

[0014] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0015] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0016] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0017] Existing infrared defect diagnosis technologies for substation equipment are limited by the dependence of dual-light fusion on visible light and the registration problem, the insufficient ability of detection networks to capture weak heating features and locate fuzzy thermal boundaries, and the mismatch between qualitative detection and quantitative operational procedure decision-making. These limitations make it difficult to meet the requirements of unmanned intelligent substation inspection for diagnostic accuracy, real-time performance, and all-weather environmental adaptability. Therefore, how to accurately focus and locate weak heating defects in voltage-induced heating equipment based solely on single-modal infrared thermal images under strong environmental thermal noise interference, and how to automatically cascade the detection results with absolute temperature information and power diagnostic procedures to complete end-to-end quantitative and graded defect diagnosis, has become an urgent problem to be solved.

[0018] To address this, this application proposes a method, apparatus, and equipment for infrared defect diagnosis of substation equipment. After acquiring a single-mode infrared thermal image of the substation equipment, it is input into a preset defect diagnosis model for feature extraction and defect detection. This model consists of a backbone feature extraction network, a neck network, and a decoupled detection network. The backbone feature extraction network extracts multi-scale features step-by-step through continuous convolutional layers, step-size downsampling layers, and residual bottleneck structures. The neck network uses feature pyramid recombination combined with a coordinate attention enhancement network employing one-dimensional pooling encoding along horizontal and vertical directions to lock the spatial coordinates of long, columnar, or segmented equipment and suppress environmental thermal noise. The decoupled detection network outputs equipment attribute category labels and high-precision defect location bounding box coordinates through parallel attribute classification and boundary regression branches. Then, using these location bounding box coordinates as an index, the highest fault heating temperature and the normal operating temperature of the same group of normal phases are extracted from the absolute temperature field matrix. The temperature difference between phases of the same type is calculated, and a double threshold cascade judgment is performed according to the preset severe anomaly benchmark threshold corresponding to the equipment attribute category. Finally, a quantitative and graded equipment diagnosis result is output end-to-end.

[0019] Based on the above, this application provides a method for diagnosing infrared defects in substation equipment, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the infrared defect diagnosis method for substation equipment according to this application.

[0020] In this embodiment, the infrared defect diagnosis method for substation equipment includes steps S10 to S40: Step S10: Obtain a single-mode infrared thermal image of the substation equipment.

[0021] It should be noted that a single-mode infrared thermal image refers to a thermal radiation distribution image acquired solely by an infrared thermal imager without fusing visible light images. The grayscale value of each pixel in the image corresponds to the radiation temperature of a measurement point on the surface of the device.

[0022] Specifically, a drone or inspection robot equipped with an infrared thermal imager takes live photos of the substation equipment along a preset inspection route. During the photo-taking process, the distance between the thermal imager lens and the target equipment is kept within a preset temperature measurement distance range (5 to 30 meters, which is determined based on the spatial resolution and temperature measurement accuracy requirements of the infrared thermal imager; if the distance is too close, the temperature measurement range will be exceeded, and if the distance is too far, the target pixel ratio will be insufficient). The field of view covers the three-phase equipment in the same group. The temperature data output by the infrared thermal imager is encoded into image frames after emissivity correction and environmental parameter compensation, and then transmitted to the back-end server via a wireless link.

[0023] Step S20: Input the single-modal infrared thermal image into the preset defect diagnosis model for feature extraction and defect detection to obtain the equipment attribute category label and the coordinates of the defect location bounding box.

[0024] It should be noted that the pre-set defect diagnosis model refers to a deep learning detection model pre-trained using historical infrared samples from substation inspections. It can be an improved single-stage target detection network, and its overall structure consists of three parts. The pre-set defect diagnosis model includes a backbone feature extraction network, a neck network, and a decoupled detection network. The backbone feature extraction network is responsible for extracting features from the input image step by step. The backbone feature extraction network is composed of consecutive convolutional layers, stride downsampling layers, and a residual bottleneck structure connected in series. The consecutive convolutional layers can be two or more convolutional layers connected in series, each consisting of convolution operations, batch normalization, and nonlinear activation. The stride downsampling layer can be a convolutional layer with a stride of 2, used to halve the feature map size. The residual bottleneck structure refers to a residual module containing the main convolutional path and skipped direct-connection branches. The main path output and the branch output are added element-wise and then fed into the next stage to alleviate gradient degradation after the network deepens. The neck network consists of a neck feature pyramid layer and a coordinate attention enhancement network connected in series. The neck feature pyramid layer refers to a feature recombination structure that performs top-down fusion and bottom-up enhancement of backbone features at different scales. The coordinate attention enhancement network is deployed after the channel splicing position of the neck feature pyramid layer. The coordinate attention enhancement network consists of one-dimensional pooling coding structures along the horizontal and vertical directions, generating attention weights in two directions and performing weighted modulation on the features. The decoupled detection network consists of parallel attribute classification branches and boundary regression branches. The attribute classification branch and boundary regression branch in the decoupled detection network are two parallel convolutional branches. The former outputs the device attribute category label (i.e., the category identifier of the device component to which the defect area belongs, such as zinc oxide surge arrester, high-voltage bushing, or capacitive voltage transformer), and the latter outputs the coordinates of the defect location bounding box (i.e., the vertex coordinates of the outer rectangle of the defect heating area in the image).

[0025] Further, before step S20, the process includes: first, acquiring historical inspection infrared thermal images of substation equipment; performing small-angle synchronous rotation, grayscale noise superposition, and brightness contrast fine-tuning on defect samples in the historical inspection infrared thermal images to obtain an expanded sample set, and dividing the expanded sample set into a training set and a validation set. It is important to understand that angle synchronization rotation refers to an augmentation method that rotates the image and its bounding box around the image center by the same angle. In this example, the rotation angle is between -10 degrees and 10 degrees. This range is determined based on the actual fluctuation of the shooting angle during on-site inspections. An excessively large angle will cause the bounding box to misalign with the equipment area. Gray-scale noise superposition can be achieved by superimposing Gaussian noise with a mean of 0 and a standard deviation of a preset noise intensity (5 to 15 gray levels, calibrated according to the imaging noise level of the infrared thermal imager) onto the gray-scale values ​​of the image pixels. Brightness and contrast fine-tuning refers to randomly scaling the brightness coefficient and contrast coefficient of the image within a preset adjustment range (0.8 to 1.2, determined based on the statistical fluctuation of infrared images under different lighting and seasonal conditions). The augmented sample set refers to the sample set formed by merging the original sample and the augmented sample. The training set refers to the part of the augmented sample set used for parameter updates, and the validation set refers to the part of the augmented sample set that does not participate in parameter updates and is only used for evaluation. The two can be randomly divided in a 4:1 ratio. Specifically, historical inspection infrared thermal images of equipment such as zinc oxide surge arresters, high-voltage bushings, and capacitive voltage transformers were collected, covering operating conditions under different seasons, loads, and shooting angles. Since there were far more normal samples than defective samples on site, small-angle synchronous rotation, grayscale noise superposition, and brightness contrast fine-tuning were performed on the defective samples to generate an expanded sample set, which was then divided into a training set and a validation set.

[0026] Secondly, based on the single-stage target detection network, a coordinate attention enhancement network is embedded after the channel splicing position of its neck feature pyramid layer, and the network parameters are initialized using random initialization or pre-trained parameters on public datasets to construct an initial defect diagnosis model.

[0027] Next, a joint loss function is constructed using distributed focus loss and perfect intersection-over-union (MIU) loss. The training set is then input into the initial defect diagnosis model in batches according to a preset batch size, yielding the predicted bounding box coordinates and predicted attribute category labels for each batch of samples. It's important to understand that distributed focus loss is the regression loss calculated by modeling the distance from the predicted boundary to the true boundary as a continuous probability distribution; perfect intersection-over-union loss is the regression loss that adds a penalty term for the center point distance and aspect ratio to the intersection-over-union ratio; and the joint loss function is the total loss function obtained by weighted summation of the boundary regression loss and the attribute classification loss. The preset batch size refers to the number of samples input for each forward inference; in this example, 16 samples are used, determined based on a trade-off between memory capacity and training stability.

[0028] Next, based on the difference between the predicted bounding box coordinates and the true bounding box, the boundary regression loss is calculated using the distribution focus loss and the perfect intersection-union ratio loss. Based on the difference between the predicted attribute category labels and the true component attribute labels, the attribute classification loss is calculated using the binary cross-entropy loss. Finally, the boundary regression loss and the attribute classification loss are weighted and summed to obtain the loss value for each batch. The specific formula is as follows: in, This represents the joint loss function value, i.e., the loss value for each batch; This represents the distribution focal point loss value; This represents the loss value for a perfect intersection-union ratio; This represents the binary cross-entropy loss value; These represent the weight coefficients corresponding to the distribution focus loss, complete intersection-union ratio loss, and binary cross-entropy loss, respectively. In this example, they are taken as 1, 1, and 0.5, respectively. This represents the actual distance from the predicted boundary to the corresponding edge of the real space bounding box; , Representing the actual distance The left and right adjacent integer nodes, satisfying ≤ ≤ ; , These represent the boundary regression branches at integer nodes. , The output probability value after being processed by the normalized exponential function; Represents the natural logarithm operation; The node number is an integer. This represents the ratio of the intersection area to the union area of ​​the predicted bounding box and the actual bounding box. , These represent the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. (Superscript) Indicates accurate labeling; This represents the Euclidean distance between two center points; This represents the diagonal length of the smallest bounding rectangle that simultaneously contains both the predicted bounding box and the ground truth bounding box. This indicates a measure of aspect ratio consistency. This represents the tradeoff coefficient for the aspect ratio penalty term; Represents pi; Represents the arctangent function; , These represent the width and height of the predicted bounding box, respectively. , These represent the width and height of the actual bounding box, respectively. This indicates the value of the actual component attribute label. The value is 1 if the sample belongs to the corresponding category, and 0 if it does not. This is the predicted probability that a sample belongs to the corresponding category, output by the attribute classification branch.

[0029] Then, based on the loss value of each batch, the network parameters of the initial defect diagnosis model are updated using the backpropagation algorithm. After each preset decay epoch of training, the learning rate is decayed by a preset decay coefficient to obtain the optimized defect diagnosis model. It's important to understand that the backpropagation algorithm can be a stochastic gradient descent algorithm with momentum; the learning rate refers to the step size coefficient for each parameter update. In this example, the initial value is 0.01, the preset decay epochs are 30, and the preset decay coefficient is 0.1. Both are determined based on experimental statistics of loss convergence speed. Specifically, the training set is input into the initial defect diagnosis model in batches according to the preset batch size. After completing the forward inference of one batch and calculating the loss value for that batch, the accumulated historical gradients on the network parameters of each layer are cleared to zero. Then, the backpropagation algorithm is used to differentiate the joint loss function layer by layer from the output layer to the input layer, starting from the joint loss function, to obtain the gradient of the joint loss function with respect to the network parameters of each layer.

[0030] Then, the network parameters of each layer are updated according to the stochastic gradient descent algorithm with momentum, and the parameter update satisfies: in , The first Second and third The network parameters after the next update; Indicates the first The gradient of the joint loss function with respect to the network parameters at the next update; This represents the current learning rate. The momentum coefficient, , The first Second and third The parameter change amount in each update. After updating the parameters for one batch, the next batch is input and the above process is repeated. One training round is considered complete when all batches in the training set have been processed. Then, the number of completed rounds is counted. For each preset decay round of training, the current learning rate is multiplied by a preset decay coefficient to become the learning rate for subsequent training, thus satisfying the condition... In the formula This represents the decayed learning rate. This represents the learning rate before decay (i.e., the current learning rate). This represents the preset attenuation coefficient. After one round of training, the model with updated network parameters is the optimized defect diagnosis model.

[0031] The optimized defect diagnosis model is then validated using a validation set, and the joint loss value corresponding to the validation set is recorded. Specifically, after each training round, the validation set is input into the currently obtained optimized defect diagnosis model, and forward inference is performed batch by batch. During inference, the network parameters are frozen, i.e., gradient calculation and parameter updates are not performed; only the joint loss function value is calculated for each sample. After all samples have been inferred, the joint loss function values ​​of each sample are summed and divided by the total number of samples in the validation set to obtain the joint loss value corresponding to this training round. Then, the joint loss value of this round and the joint loss values ​​recorded in previous rounds are stored together in a record table. The record table also saves the network parameters corresponding to each round, providing a basis for subsequent stagnation determination and parameter optimization.

[0032] Finally, when the joint loss value corresponding to the validation set no longer decreases within a preset number of consecutive stagnation rounds, or when the number of training iterations reaches a preset maximum number, the optimized defect diagnosis model corresponding to the minimum joint loss value on the validation set is selected as the preset defect diagnosis model. It's important to understand that in this example, the preset stagnation rounds are 20, and the preset maximum number of iterations is 300, determined based on statistical results of the training set size and the timing of loss plateaus. Specifically, after each validation round, the joint loss value of that round is compared with the minimum joint loss value already saved in the record table. If the joint loss value of that round is less than the saved minimum, the minimum is updated with the joint loss value of that round, the corresponding network parameters are saved as the current optimal parameters, and the stagnation count is reset to zero. If the joint loss value of that round is not less than the saved minimum, the stagnation count is incremented by 1. Then, it is determined whether the stopping condition is met. Training ends when either the stagnation count reaches the preset number of stagnation rounds (i.e., the joint loss value no longer decreases within a preset number of consecutive stagnation rounds) or the number of completed training iterations reaches the preset maximum number. After training, the network parameters saved when the joint loss value is minimized are retrieved from the record table. These network parameters are then loaded into the optimized defect diagnosis model. The loaded model serves as the preset defect diagnosis model for subsequent online inspection and diagnosis.

[0033] Step S30: Extract the highest fault temperature and normal operating temperature from the absolute temperature field matrix based on the coordinates of the defect location boundary box.

[0034] It should be noted that the absolute temperature field matrix refers to the two-dimensional floating-point temperature matrix recorded and uploaded to the background memory by the infrared thermal imager. The value of each element in the matrix is ​​the absolute temperature value of the corresponding field of view position.

[0035] Further, step S30 includes: First, using the coordinates of the defect location bounding box as a spatial retrieval index, and then back-mapping it to the absolute temperature field matrix uploaded by the infrared thermal imager through a scale conversion between the relative image coordinates and the physical field of view. It is important to understand that scale conversion refers to a coordinate transformation based on the proportional relationship between image pixel coordinates and matrix element coordinates. Specifically, while acquiring the infrared thermal image, the infrared thermal imager records the absolute temperature of each measurement point within the field of view, generates an absolute temperature field matrix, and uploads it to the backend server's memory. Under normal circumstances, the number of rows and columns of the matrix is ​​consistent with the number of pixels in the infrared thermal image, and the two correspond pixel-by-pixel. In this case, the coordinates of the defect location bounding box can be directly used as the row and column range index of the absolute temperature field matrix. If the imaging resolution and the matrix resolution are inconsistent, then scale conversion is performed according to the following formula: in , Locate the pixel coordinates in the bounding box coordinates of the defect; , These represent the number of pixels in the width and height directions of the infrared thermal image, respectively. , These represent the number of elements in the width and height directions of the absolute temperature field matrix, respectively. , The converted matrix coordinates are rounded to the nearest integer. The converted top-left and bottom-right vertices define a closed region in the absolute temperature field matrix.

[0036] Secondly, within the closed region defined by the defect location bounding box coordinates in the absolute temperature field matrix, the maximum absolute temperature pixel value is extracted using an extreme value optimization algorithm to obtain the highest fault heating temperature. It's important to understand that the extreme value optimization algorithm can be a traversal comparison method, that is, comparing the values ​​of each element within the closed region one by one and retaining the largest. Specifically, within the closed region, the extreme value optimization algorithm traverses all temperature elements, compares the values ​​of each element, and retains the largest; this maximum absolute temperature pixel value is the highest fault heating temperature.

[0037] Then, based on the physical topology of the symmetrical three-phase distribution of substation electrical equipment, the area of ​​normal phase electrical equipment operating in the same group as the faulty phase electrical equipment and under the same load conditions is selected in the absolute temperature field matrix. It is important to understand that a symmetrical three-phase distribution refers to a layout where the three-phase equipment in the same group are arranged at equal intervals in the horizontal direction, and have the same model and structure. Specifically, the position of the normal phase is determined according to the physical topology of the symmetrical three-phase distribution. Using the closed area of ​​the faulty phase as a reference, the number of elements corresponding to the preset phase spacing is shifted horizontally to the left or right. The preset phase spacing is equal to the pixel distance between the centers of two adjacent phase devices in the same group. Based on the phase distance and shooting distance calibration in the equipment installation drawings, the area of ​​normal phase electrical equipment operating in the same group as the faulty phase electrical equipment and under the same load conditions is selected in the absolute temperature field matrix.

[0038] Finally, the highest temperature pixel value at the corresponding measuring point in the normal phase electrical equipment region is extracted from the absolute temperature field matrix to obtain the normal operating temperature. It's important to understand that the corresponding measuring point refers to the location on the normal phase equipment that corresponds structurally to the heat-generating area of ​​the faulty phase, such as the root of the upper flange section. Specifically, the corresponding measuring point in the normal phase electrical equipment region is located, and the highest temperature pixel value is extracted within that location using an extreme value optimization algorithm to obtain the normal operating temperature.

[0039] Step S40: Calculate the phase-to-phase temperature difference based on the highest fault temperature and the normal operating temperature, and perform double threshold cascade judgment on the phase-to-phase temperature difference based on the preset severe abnormality benchmark threshold corresponding to the equipment attribute category label to obtain the equipment diagnosis result.

[0040] It should be noted that the phase-to-phase temperature difference refers to the difference between the highest fault temperature and the normal operating temperature, which can be obtained by direct subtraction. The preset severe anomaly benchmark threshold can be the severe anomaly temperature difference limit value specified in the power industry standard for the corresponding type of voltage-induced heating equipment. For example, 4 Kelvin is used for a certain type of equipment. This value is determined according to industry standards and combined with the equipment insulation characteristics. The double threshold cascade judgment refers to a three-level judgment logic that compares the phase-to-phase temperature difference with 1 times and 2 times the preset severe anomaly benchmark threshold in turn: if the phase-to-phase temperature difference is less than the threshold, it is recorded as a general anomaly; if it is greater than or equal to the threshold but less than twice the threshold, it is recorded as a severe anomaly; if it is greater than or equal to twice the threshold, it is recorded as a dangerous anomaly and an alarm is triggered.

[0041] Further, step S40 includes: first, subtracting the normal operating temperature from the highest fault temperature to obtain the phase-to-phase temperature difference, denoted as... ,satisfy In the formula The highest temperature at which the fault occurs. This is the normal operating temperature.

[0042] Secondly, based on the device attribute category label, the preset severe anomaly baseline threshold corresponding to the device attribute category label in the locally stored industry standard is retrieved. Specifically, the device type is determined based on the device attribute category label, and the entry corresponding to the voltage-induced heating device of that type is found in the locally stored industry standard threshold table. The value in the entry is read as the preset severe anomaly baseline threshold. The threshold table can be pre-entered according to the temperature difference limits specified for various types of equipment in the current power industry standards.

[0043] Then, the ambient temperature of the current inspection environment is obtained. Based on the deviation between the ambient temperature and the preset reference ambient temperature, the preset severe anomaly reference threshold is corrected to obtain the corrected reference threshold. Specifically, the ambient temperature of the current inspection environment measured by the temperature sensor is obtained, the deviation between the ambient temperature and the preset reference ambient temperature is calculated, and the preset severe anomaly reference threshold is corrected according to the following formula: in This is the corrected baseline threshold; A preset threshold for severe anomalies; Ambient temperature; The preset reference ambient temperature; The temperature correction factor is determined based on the statistical regression results of the measured temperature difference of the same type of equipment under different ambient temperatures. The threshold increases as the ambient temperature rises and decreases as the ambient temperature falls.

[0044] When the temperature difference between phases of the same type is less than the corrected baseline threshold, the abnormal state of the substation equipment is determined to be a general abnormality. When the temperature difference between phases of the same type is greater than or equal to the corrected baseline threshold but less than twice the corrected baseline threshold, the abnormal state of the substation equipment is determined to be a serious abnormality. When the temperature difference between phases of the same type is greater than or equal to twice the corrected baseline threshold, the abnormal state of the substation equipment is determined to be a dangerous abnormality and an alarm is triggered. Finally, the heated areas corresponding to the defect location bounding box coordinates are highlighted, and the equipment attribute category label, the highest fault temperature, the temperature difference between phases of the same type, and the abnormal status are summarized to obtain the equipment diagnosis results. Specifically, the current infrared thermal image is loaded on the front-end inspection interface, and highlighted labels are superimposed on the corresponding positions of the image according to the defect location bounding box coordinates. The frame color is selected according to the abnormal status, so that defects of different levels can be visually distinguished. When a dangerous abnormality is determined, an alarm prompt will pop up simultaneously. Then, the equipment attribute category label, the highest fault temperature, the temperature difference between phases of the same type, and the abnormal status are filled into the corresponding fields of the structured form. The form fields also include inspection time, equipment number, and shooting location. The data in each field comes from the intermediate results of this diagnosis and the registration information of the inspection task. Finally, the completed structured form is stored as the equipment diagnosis result in the background inspection record database and pushed to the operation and maintenance platform for operation and maintenance personnel to view and arrange subsequent defect elimination.

[0045] This embodiment uses a single-modal infrared thermogram as input, extracting multi-scale features step-by-step through a backbone feature extraction network. A coordinate attention enhancement network in the neck network establishes long-distance spatial position dependence, and a decoupled detection network separates attribute classification and boundary regression. The localized bounding box is then back-mapped to an absolute temperature field matrix to extract temperature difference. Combined with a double-threshold cascade, quantitative grading results are output. Beneficial effects include effectively suppressing environmental thermal noise interference from solar radiation and high metal reflectivity, enhancing the representation of weak heat sources at the feature level, avoiding bounding box drift and missed detection of tiny heat points, eliminating reliance on visible light assistance, and eliminating the need for manual secondary segmentation temperature measurement and table lookup decisions. This significantly improves the diagnostic accuracy, real-time performance, and environmental adaptability of substation live-line inspections.

[0046] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 The infrared defect diagnosis method for substation equipment, step S20, further includes steps S201 to S204: Step S201: Multi-scale feature extraction is performed on the single-modality infrared thermal image through a backbone feature extraction network to obtain a multi-scale feature map.

[0047] It is important to understand that multi-scale feature extraction refers to the operation of extracting features at different depths of the network. Shallow features retain texture and edge information, while deep features retain semantic information. Multi-scale feature maps refer to the set of feature maps output by several network layers with different downsampling ratios in the backbone feature extraction network.

[0048] Further, step S201 includes: first, performing size normalization processing on the single-modality infrared thermal image to obtain processed data, and then reading the processed data row by row and column by column according to pixel grayscale values ​​to convert it into a pixel grayscale feature matrix. It should be understood that size normalization processing refers to scaling the input image to a fixed size required by the network; in this example, it is scaled to 640×640 pixels, which is consistent with the input specifications of the single-stage object detection network. The pixel grayscale feature matrix is ​​a two-dimensional matrix formed by arranging the grayscale values ​​of each pixel in the image in rows and columns.

[0049] Secondly, the pixel grayscale feature matrix is ​​convolved layer by layer through consecutive convolutional layers. After each convolutional layer, batch normalization and nonlinear activation are performed to extract texture and edge features from the infrared thermal image, thus obtaining shallow spatial geometric features. It is important to understand that shallow spatial geometric features refer to the features extracted by the shallow layers of the network that contain texture and edge information.

[0050] Next, the shallow spatial geometric features are downsampled stepwise through a stride downsampling layer to obtain the downsampled features. It's important to understand that the downsampling ratio of the stride downsampling layer includes a first downsampling ratio, a second downsampling ratio, and a third downsampling ratio. The first downsampling ratio is less than the second downsampling ratio, and the second downsampling ratio is less than the third downsampling ratio. In this example, the preset first downsampling ratio is 8x, the preset second downsampling ratio is 16x, and the preset third downsampling ratio is 32x. These three ratios are determined based on the size distribution of the detected target in the image; the lower ratio layer handles small parts, and the higher ratio layer handles heavy equipment. It's also important to understand that the shallow spatial geometric features are input into the stride downsampling layer. The first downsampling is performed by a convolutional layer with a stride of 2. The convolutional kernel slides across the feature map with a stride of 2, resulting in a feature map whose width and height are each halved, and the number of channels is doubled, yielding the features after the first downsampling. Then, using the features after the first downsampling as input, the second and third downsampling are repeated. The downsampling operation is the same as the first time, and the features after the second and third downsampling are obtained respectively. The features after the three downsamplings are each sent to the corresponding residual bottleneck structure for subsequent calculation.

[0051] Following each downsampling stage, the downsampled features are input into a residual bottleneck structure. This structure divides the downsampled features into two paths. The first path undergoes convolution operations through two cascaded convolutional branches within the bottleneck structure, yielding convolutional branch features. The second path is directly passed through a skip-connection branch within the bottleneck structure, resulting in directly connected features. Finally, the convolutional branch features and the directly connected features are added element-wise to obtain deep, high-order semantic features. Specifically, after each downsampling stage, the downsampled features are input into the residual bottleneck structure, which is then replicated into two identical paths. The first path first compresses the number of channels using a 1×1 convolutional layer, then restores the number of channels using a 3×3 convolutional layer. Each convolutional layer is followed by batch normalization and non-linear activation to obtain convolutional branch features. The second path does not undergo any convolution operations and is directly passed through a skip-connection branch to obtain directly connected features. Then, the convolutional branch features and the direct-connected features are added one by one according to the corresponding positions. Since the size and number of channels of the two features are the same, the positions are one-to-one when added. The addition result is output after non-linear activation to obtain deep high-order semantic features.

[0052] Then, deep high-order semantic features corresponding to the first, second, and third downsampling ratios are selected respectively to obtain hierarchical features at three scales. It's important to understand that deep high-order semantic features refer to features containing category semantic information extracted after multi-level downsampling and residual calculation; hierarchical features refer to the features output by the network layer corresponding to a certain downsampling ratio. Specifically, after the three downsampling operations and the corresponding residual bottleneck structure calculations are completed, the outputs of three network layers are derived: deep high-order semantic features output by the residual bottleneck structure corresponding to 8x downsampling, 16x downsampling, and 32x downsampling. These three outputs represent the hierarchical features at three scales. Among them, the hierarchical feature size corresponding to 8x downsampling is the largest, with each feature point corresponding to an 8×8 pixel region in the original image; the hierarchical feature size corresponding to 32x downsampling is the smallest, with each feature point corresponding to a 32×32 pixel region in the original image; and the hierarchical feature size corresponding to 16x downsampling falls between the two.

[0053] Finally, the hierarchical features of the three scales are arranged in ascending order of the preset downsampling factor, namely the hierarchical features corresponding to 8x downsampling, 16x downsampling, and 32x downsampling. The arranged sequence is the multi-scale feature map. The order of the hierarchical features in the sequence corresponds to the processing order when the neck feature pyramid layer is fused from top to bottom.

[0054] Step S202: Input the multi-scale feature map into the neck feature pyramid layer for feature recombination to obtain the channel splicing feature map.

[0055] It should be noted that feature recombination refers to the operation of cross-layer fusion and splicing of feature maps at different scales; channel splicing feature map refers to the feature map formed by splicing the recombinated features at each scale along the channel dimension.

[0056] Further, step S202 includes: First, upsampling the hierarchical feature with the largest downsampling factor in the multi-scale feature map to obtain the upsampled feature. It should be understood that upsampling can be bilinear interpolation upsampling, which doubles both the width and height of the feature map to make the feature map sizes consistent across the two scales. Specifically, the hierarchical feature with the largest downsampling factor in the multi-scale feature map, i.e., the hierarchical feature corresponding to 32x downsampling, is taken and upsampled using bilinear interpolation, doubling both its width and height to make its size consistent with the hierarchical feature corresponding to 16x downsampling, thus obtaining the upsampled feature. Then, the upsampled feature is added element-wise to the hierarchical feature corresponding to 16x downsampling to obtain the fused feature corresponding to the 16x scale.

[0057] Secondly, the upsampled features are added element-wise with the features of the adjacent next-level downsampled features to obtain the fused features. It is important to understand that the fused features refer to the features obtained by adding deep features and shallow features element-wise in the top-down path.

[0058] The upsampling and element-wise addition fusion process is repeated until all hierarchical features at each scale in the multi-scale feature map are fused, resulting in the fused features corresponding to each scale. Specifically, the upsampling and element-wise addition fusion process is repeated, that is, the fused features corresponding to the 16x scale are upsampled again and added element-wise with the hierarchical features corresponding to the 8x downsampling to obtain the fused features corresponding to the 8x scale. At this point, all hierarchical features at each scale are fused, and the hierarchical features corresponding to the 32x scale are directly used as the fused features corresponding to the 32x scale.

[0059] Next, the fusion feature with the smallest downsampling factor among the fusion features corresponding to each scale is downsampled to obtain the downsampled fusion feature.

[0060] Next, the downsampled fusion feature is added element-wise to the fusion feature at the adjacent previous downsampling factor to obtain the enhanced feature. It's important to understand that the enhanced feature refers to the feature obtained by downsampling the fusion feature and adding it element-wise to the fusion feature at the adjacent scale in the bottom-up path. Specifically, moving to the bottom-up path, the fusion feature corresponding to the 8x scale is downsampled by a convolution with a stride of 2, reducing both width and height by half to obtain the downsampled fusion feature. This downsampled fusion feature is then added element-wise to the fusion feature corresponding to the 16x scale to obtain the enhanced feature corresponding to the 16x scale. This process is repeated to obtain the enhanced feature corresponding to the 32x scale. The fusion feature corresponding to the 8x scale is directly used as the enhanced feature corresponding to the 8x scale.

[0061] Then, the information richness of each scale is calculated based on the variance of pixel grayscale values ​​in the enhanced features corresponding to each scale, and the fusion weight of each scale is determined based on the information richness. It's important to understand that information richness refers to the amount of effective information in a feature, measured by the variance of feature grayscale values; a larger variance indicates a more significant difference between strong and weak responses in the feature. The fusion weight refers to the coefficient used when weighting each scale before stitching; in this example, it is determined according to the proportion of the information richness of each scale to the sum of the information richness of all scales. Specifically, for the enhanced feature corresponding to each scale, the mean of all its pixel grayscale values ​​is first calculated, and then the variance of the pixel grayscale values ​​is calculated. The specific formula is as follows: in For the first Information richness at each scale Choose 1, 2, or 3; For the first The total number of pixels in the enhanced features corresponding to each scale; For the first Among the enhanced features corresponding to the i-th scale, the i-th The grayscale value of each pixel; For the first The mean of all pixel grayscale values ​​in the enhancement features corresponding to each scale.

[0062] Then, the fusion weights for each scale are calculated using the following formula: in For the first Fusion weights at each scale; , , The information richness is represented by three scales, with the scale having a higher information richness and thus a greater fusion weight.

[0063] Finally, the enhanced features at the corresponding scales are multiplied element-wise by fusion weights and weighted. The weighted enhanced features at each scale are then interpolated and scaled to the same size and stitched together along the channel dimension to obtain the channel stitched feature map.

[0064] Step S203: The channel splicing feature map is subjected to directional aggregation encoding and weight modulation through a coordinate attention enhancement network to obtain an enhanced feature map.

[0065] It should be noted that directional aggregation encoding refers to the operation of performing one-dimensional pooling on features along the horizontal and vertical directions respectively, compressing two-dimensional features into one-dimensional encodings in two directions; weight modulation refers to the operation of multiplying the generated attention weights with the original features point by point, adjusting the feature response intensity at each position; and enhanced feature map refers to the feature map output after weight modulation.

[0066] Further, step S203 includes: First, using a horizontal one-dimensional pooling kernel with a preset first pooling size in the one-dimensional pooling encoding structure, the channel-to-channel concatenated feature map is average-pooled along the horizontal direction to obtain the horizontal component. Then, using a vertical one-dimensional pooling kernel with a preset second pooling size in the one-dimensional pooling encoding structure, the channel-to-channel concatenated feature map is average-pooled along the vertical direction to obtain the vertical component. It should be understood that the preset first pooling size refers to the size of the horizontal one-dimensional pooling kernel, which is 1×W, where W is the width of the channel-to-channel concatenated feature map; the preset second pooling size refers to the size of the vertical one-dimensional pooling kernel, which is H×1, where H is the height of the channel-to-channel concatenated feature map; the horizontal component refers to the one-dimensional feature obtained after pooling along the horizontal direction, retaining the vertical position information; the vertical component refers to the one-dimensional feature obtained after pooling along the vertical direction, retaining the horizontal position information. Specifically, the channel splicing feature map is averaged channel by channel along the horizontal direction using a 1×W horizontal one-dimensional pooling kernel. The W elements of each channel in each row are averaged to obtain the horizontal component that retains the position information of each row. The channel splicing feature map is averaged channel by channel along the vertical direction using an H×1 vertical one-dimensional pooling kernel. The H elements of each channel in each column are averaged to obtain the vertical component that retains the position information of each column.

[0067] Secondly, the horizontal and vertical components are spliced ​​along the spatial dimensions to obtain the spliced ​​features.

[0068] Next, the concatenated features are compressed using a preset dimensionality-reducing convolutional kernel at a preset channel compression ratio. Then, a non-linear activation function is used to jointly encode the compressed features, resulting in a comprehensive encoded intermediate feature. It's important to understand that the preset dimensionality-reducing convolutional kernel can be a 1×1 kernel; the preset channel compression ratio refers to the ratio of the number of channels before and after compression (32 in this example, determined by a trade-off between the computational cost of the attention module and the feature representation capability); and the comprehensive encoded intermediate feature refers to the intermediate feature obtained by concatenating and compressing the components from both directions. Specifically, the concatenated features are input into the preset dimensionality-reducing convolutional kernel for a 1×1 convolution operation. The number of convolutional kernels equals the original number of channels divided by the preset channel compression ratio (32 in this example). Each convolutional kernel weights and combines the information from all channels along the channel dimension, outputting the channel-compressed feature. Then, the ReLU activation function is applied to the channel-compressed features, setting negative values ​​to zero and retaining positive values ​​to complete the joint encoding of information in both horizontal and vertical directions, resulting in comprehensive encoded intermediate features. The spatial dimension of the comprehensive encoded intermediate features contains both row and column position encoding information for subsequent segmentation.

[0069] Next, the synthesized intermediate features are spatially segmented into horizontal and vertical one-dimensional encoding vectors. The number of channels in both vectors is then restored, and they are processed using a Sigmoid activation function to obtain horizontal and vertical attention weights. It's important to understand that the Sigmoid activation function maps the input to the range of 0 to 1. Specifically, the synthesized intermediate features are first segmented spatially, with the segmentation position corresponding to the position when the horizontal and vertical components were concatenated. This results in horizontal and vertical one-dimensional encoding vectors, with lengths equal to the height and width of the channel concatenation feature map, respectively. A 1×1 convolution is then performed on the horizontal one-dimensional encoding vector to restore the number of channels to their original value. This is followed by Sigmoid activation, mapping the values ​​of each element to the range of 0 to 1, resulting in the horizontal attention weights. Each element in the horizontal attention weight corresponds to the response intensity of a column. Then, the same 1×1 convolution and sigmoid activation function are applied to the one-dimensional encoding vector in the vertical direction to obtain the vertical attention weights. Each element in the vertical attention weights corresponds to the response intensity of a row position.

[0070] Next, the ratio of the length to the width of each connected high-response region in the channel stitching feature map is statistically analyzed to obtain the shape prior coefficient. Based on the shape prior coefficient, the first weighting coefficient for the vertical attention weight and the second weighting coefficient for the horizontal attention weight are determined. It is important to understand that a connected high-response region refers to a pixel region in the channel stitching feature map whose grayscale value exceeds a preset response threshold and is interconnected. The preset response threshold is 1.5 times the average grayscale value of the feature, determined statistically based on the grayscale difference between the heated area and the background. The shape prior coefficient is the ratio of the length to the width of the connected high-response region, used to reflect the slenderness of the device. The larger the shape prior coefficient, the more slender the device, and the larger the first weighting coefficient is relative to the second weighting coefficient. The first weighting coefficient and the second weighting coefficient are the coefficients used when weighting the vertical attention weight and the horizontal attention weight, respectively. Specifically, firstly, the grayscale average value of the channel stitching feature map is statistically analyzed. Pixels with grayscale values ​​exceeding the preset response threshold are set to 1, and the rest are set to 0, resulting in a binary image. The binary image is then scanned twice to mark the connected components. The interconnected pixels set to 1 are grouped into the same region, resulting in each connected high-response region. Next, for each connected high-response region, calculate its minimum bounding rectangle. Read the number of pixels along the longer side of the minimum bounding rectangle as the length and the number of pixels along the shorter side as the width. Calculate the ratio of length to width, and take the maximum value of this ratio across all connected high-response regions as the shape prior coefficient r. Then, according to the specific formula: in Indicates the first weighting coefficient; This represents the second weighting coefficient. When there are no connected high-response regions in the graph, the shape prior coefficient is set to 1. In this case, both the first and second weighting coefficients are 0.5, and the attention weights in the two directions are treated equally.

[0071] Then, the vertical attention weights are weighted using a first weighting coefficient to obtain the corrected vertical attention weights, and the horizontal attention weights are weighted using a second weighting coefficient to obtain the corrected horizontal attention weights. Specifically, each element of the vertical attention weights is multiplied by the first weighting coefficient to obtain the corrected vertical attention weights, satisfying the following conditions: In the formula To correct the vertical attention weights, The first weighting coefficient, These are the vertical attention weights. Similarly, multiplying each element of the horizontal attention weights by the second weighting coefficient yields the corrected horizontal attention weights, satisfying... In the formula To correct the horizontal attention weights, This is the second weighting coefficient. For horizontal attention weights.

[0072] Finally, the corrected vertical attention weights and corrected horizontal attention weights are used as bidirectional attention weights. These bidirectional attention weights are then modulated point-by-point by multiplying with the channel-stitched feature map on the spatial grid to obtain the enhanced feature map. Specifically, the corrected vertical attention weights and corrected horizontal attention weights are used as bidirectional attention weights. For each pixel position on the spatial grid of the channel-stitched feature map, the corrected vertical attention weight element corresponding to its row and the corrected horizontal attention weight element corresponding to its column are taken, and these three are multiplied point-by-point. The specific formula is as follows: in, Indicates the first in the enhanced feature map Line 1 Feature values ​​of column positions; The first channel in the splicing feature map Line 1 Feature values ​​of column positions; Indicates the first The corresponding adjusted vertical attention weights for each row; Indicates the first The corresponding adjusted attention weights for each column; , These are the row and column numbers of the spatial grid, respectively. After multiplying all pixel positions on the spatial grid point by point, the feature values ​​at each position are rescaled according to weights in both directions. The output is the enhanced feature map, whose size and number of channels are the same as those of the channel concatenation feature map. Figure 1 This is intended for use by the subsequent decoupling detection network.

[0073] Step S204: Classify and regress the enhanced feature map using a decoupled detection network to obtain the device attribute category label and the coordinates of the defect location bounding box.

[0074] It should be noted that step S204 includes: First, performing classification calculations on the enhanced feature map through the attribute classification branch to obtain the device attribute category label. It is important to understand that classification calculation refers to calculating the confidence score of each category at each location of the feature map and taking the category corresponding to the highest confidence score. Specifically, the enhanced feature map is input into the attribute classification branch, and the convolutional layer at the end of the branch outputs the confidence score of each category for each feature location. Categories include device types such as zinc oxide surge arresters, high-voltage bushings, and capacitive voltage transformers. The confidence scores of each category at each feature location are compared, and the category with the highest confidence score is taken as the device attribute category label for that location.

[0075] Next, the boundary regression branch models the physical distance from the predicted boundary to the ground truth boundary as a continuous probability distribution, resulting in the boundary distance probability distribution. It's important to understand that physical distance refers to the distance from the predicted boundary to the ground truth boundary, expressed in pixels; the boundary distance probability distribution refers to the distribution of probabilities corresponding to each integer node after discretizing the possible values ​​of the physical distance into integer nodes. Specifically, the boundary regression branch regresses the positions of the four boundaries of the heated region on the enhanced feature map position by position. For each boundary, the range of values ​​for the physical distance from the predicted boundary to the ground truth boundary is discretized into a preset number of integer nodes (16, determined by a trade-off between the accuracy requirements and computational cost of boundary regression). The unnormalized values ​​corresponding to each integer node are output, forming the original output of the boundary distance probability distribution.

[0076] Then, the conditional probability density at two adjacent integer nodes in the boundary distance probability distribution is optimized using a normalized exponential function, driving the probability centroid to converge towards the ground truth labeled boundary, thus obtaining the boundary regression result. It's important to understand that the normalized exponential function converts a set of values ​​into probability values ​​that sum to 1; the conditional probability density refers to the probability value of the boundary distance probability distribution at a specified integer node; the probability centroid is the position obtained by weighting the integer nodes according to their probabilities; and the boundary regression result refers to the predicted positions of the four boundaries of the heated region output by the boundary regression branch. Specifically, a normalized exponential function is applied to the original output boundary distance probability distribution, converting the values ​​of each integer node into a conditional probability density that sums to 1. During training, the conditional probability density at two adjacent integer nodes in the ground truth location is optimized using a distribution focus loss, giving nodes near the ground truth location a higher probability, driving the probability centroid to converge towards the ground truth labeled boundary. During the inference phase, the integer nodes are weighted and summed according to their conditional probability densities to obtain the predicted position of each boundary. The predicted positions of the four boundaries constitute the boundary regression result.

[0077] Finally, the coordinates of the defect location bounding box are determined based on the boundary regression results. Specifically, the predicted positions of the four boundaries in the boundary regression results are converted into the coordinates of the top left and bottom right vertices of the circumscribed rectangle. Overlapping redundant predicted boxes are removed by non-maximum suppression, and the predicted box with the highest confidence is retained. The vertex coordinates of the predicted box are the coordinates of the defect location bounding box.

[0078] This embodiment extracts multi-scale features from single-modality infrared thermal images using a backbone feature extraction network. Information from each scale is then fused through a neck feature pyramid layer reconstruction. A coordinate attention enhancement network aggregates and encodes these features along the horizontal and vertical directions, modulating the weights. Finally, a decoupled detection network separates and completes attribute classification and boundary regression. Multi-scale extraction and pyramid reconstruction take into account the feature representation of components of different sizes. The directional attention enhances the spatial position dependence of elongated devices and reduces environmental thermal noise. Decoupled detection avoids mutual interference between classification and regression, improving the detection capability and positioning accuracy of subtle heating defects without relying on visible light assistance.

[0079] Based on the first embodiment of this application, this application also provides an infrared defect diagnosis device for substation equipment. Please refer to... Figure 3 The device includes: The data acquisition module 10 is used to acquire single-mode infrared thermal images of substation equipment.

[0080] The defect detection module 20 is used to input the single-modal infrared thermal image into the preset defect diagnosis model for feature extraction and defect detection, and obtain the equipment attribute category label and the coordinates of the defect location bounding box. The preset defect diagnosis model includes a backbone feature extraction network, a neck network and a decoupled detection network. The backbone feature extraction network is composed of a continuous convolutional layer, a stride downsampling layer and a residual bottleneck structure connected in series. The neck network is composed of a neck feature pyramid layer and a coordinate attention enhancement network connected in series. The decoupled detection network consists of a parallel attribute classification branch and a boundary regression branch.

[0081] The temperature extraction module 30 is used to extract the highest fault temperature and normal operating temperature from the absolute temperature field matrix based on the coordinates of the defect location bounding box.

[0082] The graded diagnosis module 40 is used to calculate the temperature difference between the same type of phases based on the highest heat temperature of the fault and the normal operating temperature, and to perform double threshold cascade judgment on the temperature difference between the same type of phases according to the preset severe abnormality benchmark threshold corresponding to the equipment attribute category label, so as to obtain the equipment diagnosis result.

[0083] The infrared defect diagnosis device for substation equipment provided in this application, employing the infrared defect diagnosis method for substation equipment in the above embodiments, can solve the technical problem of how to accurately locate and quantitatively grade the weak heating defects of voltage-induced heating equipment. Compared with the prior art, the beneficial effects of the infrared defect diagnosis device for substation equipment provided in this application are the same as those of the infrared defect diagnosis method for substation equipment provided in the above embodiments, and other technical features in the infrared defect diagnosis device for substation equipment are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0084] This application provides an infrared defect diagnosis device for substation equipment. The infrared defect diagnosis device for substation equipment includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the infrared defect diagnosis method for substation equipment in the above embodiment 1.

[0085] The following is for reference. Figure 4 The diagram illustrates a structural schematic suitable for implementing the infrared defect diagnosis device for substation equipment in the embodiments of this application. The infrared defect diagnosis device for substation equipment in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), vehicle terminals (e.g., vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 4 The infrared defect diagnosis device for substation equipment shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0086] like Figure 4As shown, the substation equipment infrared defect diagnosis device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the substation equipment infrared defect diagnosis device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the substation equipment infrared defect diagnostic equipment to exchange data wirelessly or via wired communication with other devices. Although various types of substation equipment infrared defect diagnostic equipment are shown in the figures, it should be understood that implementation or possession of all of them is not required. More or fewer may be implemented alternatively.

[0087] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0088] The infrared defect diagnosis device for substation equipment provided in this application, employing the infrared defect diagnosis method for substation equipment described in the above embodiments, can solve the technical problem of how to accurately locate and quantitatively grade the weak heating defects in voltage-induced heating equipment. Compared with the prior art, the beneficial effects of the infrared defect diagnosis device for substation equipment provided in this application are the same as those of the infrared defect diagnosis method for substation equipment provided in the above embodiments, and other technical features of this infrared defect diagnosis device for substation equipment are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0089] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0090] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0091] This application provides a computer-readable medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the substation equipment infrared defect diagnosis method in the above embodiments.

[0092] The computer-readable medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable medium may be any tangible medium containing or storing a program that can be executed by instructions, used by a device, or used in conjunction with it. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0093] The aforementioned computer-readable medium may be included in the infrared defect diagnosis equipment for substation equipment; or it may exist independently and not be assembled into the infrared defect diagnosis equipment for substation equipment.

[0094] The aforementioned computer-readable medium carries one or more programs that, when executed by the substation equipment infrared defect diagnosis device, enable the substation equipment infrared defect diagnosis device to write computer program code for performing the operations of this application in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0095] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, all blocks in the flowcharts or block diagrams may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that all blocks in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using dedicated hardware-based implementations that perform the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0096] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0097] The readable medium provided in this application is a computer-readable medium, which stores computer-readable program instructions (i.e., a computer program) for executing the above-described infrared defect diagnosis method for substation equipment. This solves the technical problem of accurately locating and quantitatively grading the diagnosis of subtle heating defects in voltage-heated equipment. Compared with the prior art, the beneficial effects of the computer-readable medium provided in this application are the same as those of the infrared defect diagnosis method for substation equipment provided in the above embodiments, and will not be elaborated upon here.

[0098] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the infrared defect diagnosis method for substation equipment as described above.

[0099] The computer program product provided in this application can solve the technical problem of how to accurately locate and quantitatively grade the subtle heating defects in voltage-heated equipment. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the infrared defect diagnosis method for substation equipment provided in the above embodiments, and will not be repeated here.

[0100] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for infrared defect diagnosis of substation equipment, characterized in that, The method includes: Acquire single-mode infrared thermal images of substation equipment; The single-modal infrared thermal image is input into a preset defect diagnosis model for feature extraction and defect detection to obtain equipment attribute category labels and defect location bounding box coordinates. The preset defect diagnosis model includes a backbone feature extraction network, a neck network, and a decoupled detection network. The backbone feature extraction network is composed of a series of continuous convolutional layers, stride downsampling layers, and residual bottleneck structures. The neck network is composed of a neck feature pyramid layer and a coordinate attention enhancement network. The decoupled detection network consists of parallel attribute classification branches and boundary regression branches. Based on the coordinates of the defect location bounding box, the highest fault heating temperature and normal operating temperature are extracted from the absolute temperature field matrix; The phase-to-phase temperature difference is calculated based on the highest heat-generating temperature of the fault and the normal operating temperature. The phase-to-phase temperature difference is then subjected to double threshold cascading judgment based on the preset severe anomaly benchmark threshold corresponding to the equipment attribute category label, thereby obtaining the equipment diagnosis result.

2. The method as described in claim 1, characterized in that, The steps of inputting the single-modal infrared thermal image into a preset defect diagnosis model for feature extraction and defect detection to obtain equipment attribute category labels and defect location bounding box coordinates include: Multi-scale feature maps are obtained by performing multi-scale feature extraction on the single-modality infrared thermal image through a backbone feature extraction network. The multi-scale feature map is input into the neck feature pyramid layer for feature reconstruction to obtain the channel stitched feature map. The channel splicing feature map is subjected to directional aggregation encoding and weight modulation by a coordinate attention enhancement network to obtain an enhanced feature map. The coordinate attention enhancement network consists of a one-dimensional pooling encoding structure along the horizontal and vertical directions. The enhanced feature map is classified and regressed by a decoupled detection network to obtain device attribute category labels and defect location bounding box coordinates.

3. The method as described in claim 2, characterized in that, The step of extracting multi-scale features from the single-modality infrared thermal image using a backbone feature extraction network to obtain a multi-scale feature map includes: The single-mode infrared thermal image is subjected to size normalization processing to obtain processed data, and the processed data is converted into a pixel grayscale feature matrix; By performing layer-by-layer convolution operations on the pixel grayscale feature matrix through continuous convolutional layers, texture features and edge features in the infrared thermal image are extracted to obtain shallow spatial geometric features. The shallow spatial geometric features are downsampled stepwise by a step-size downsampling layer to obtain downsampled features. The downsampling factor of the step-size downsampling layer includes a first downsampling factor, a second downsampling factor, and a third downsampling factor. The first downsampling factor is less than the second downsampling factor, and the second downsampling factor is less than the third downsampling factor. After each downsampling stage, the downsampled features are input into the residual bottleneck structure. The residual bottleneck structure divides the downsampled features into two paths. The first path performs convolution operations through two cascaded convolution branches in the residual bottleneck structure to obtain convolution branch features. The second path directly passes through the skip direct connection branch of the residual bottleneck structure to obtain direct connection features. The convolutional branch features are added element by element to the directly connected features to obtain deep high-order semantic features; The deep high-order semantic features corresponding to the first downsampling ratio, the second downsampling ratio, and the third downsampling ratio are selected respectively to obtain hierarchical features at three scales; The hierarchical features of the three scales are arranged in ascending order according to the preset downsampling factor to obtain a multi-scale feature map.

4. The method as described in claim 2, characterized in that, The step of performing directional aggregation encoding and weight modulation on the channel spliced ​​feature map using a coordinate attention enhancement network to obtain an enhanced feature map includes: The channel splicing feature map is averaged channel by channel along the horizontal direction using a horizontal one-dimensional pooling kernel with a preset first pooling size in the one-dimensional pooling coding structure to obtain the horizontal component. The channel splicing feature map is then averaged channel by channel along the vertical direction using a vertical one-dimensional pooling kernel with a preset second pooling size in the one-dimensional pooling coding structure to obtain the vertical component. The horizontal component and the vertical component are spliced ​​together along the spatial dimension to obtain the spliced ​​feature; The concatenated features are compressed by a preset dimensionality reduction convolution kernel according to a preset channel compression ratio, and the compressed features are jointly encoded by a non-linear activation function to obtain the comprehensive encoded intermediate features. The integrated encoding intermediate features are divided into horizontal one-dimensional encoding vectors and vertical one-dimensional encoding vectors along the spatial dimension. The number of channels is restored for the horizontal one-dimensional encoding vectors and the vertical one-dimensional encoding vectors respectively, and then processed by the Sigmoid activation function to obtain the horizontal attention weights and vertical attention weights. The ratio of the length to the width of each connected high-response region in the channel splicing feature map is statistically analyzed to obtain the shape prior coefficient. Based on the shape prior coefficient, the first weighting coefficient of the vertical attention weight and the second weighting coefficient of the horizontal attention weight are determined. The larger the shape prior coefficient, the larger the first weighting coefficient is relative to the second weighting coefficient. The vertical attention weights are weighted by the first weighting coefficient to obtain the corrected vertical attention weights, and the horizontal attention weights are weighted by the second weighting coefficient to obtain the corrected horizontal attention weights. The modified vertical attention weights and the modified horizontal attention weights are used as bidirectional attention weights. The bidirectional attention weights are then multiplied and modulated point-by-point on the spatial grid by the channel splicing feature map to obtain the enhanced feature map.

5. The method as described in claim 2, characterized in that, The step of classifying and regressing the enhanced feature map through a decoupled detection network to obtain device attribute category labels and defect location bounding box coordinates includes: The enhanced feature map is classified and calculated using the attribute classification branch to obtain device attribute category labels; The boundary regression branch models the physical distance from the predicted boundary to the actual labeled boundary as a continuous probability distribution, thus obtaining the boundary distance probability distribution. The conditional probability density at two integer nodes adjacent to the true location in the boundary distance probability distribution is optimized by a normalized exponential function, driving the probability centroid to converge toward the true labeled boundary, thus obtaining the boundary regression result. The coordinates of the defect location bounding box are determined based on the boundary regression results.

6. The method as described in claim 1, characterized in that, Before the step of inputting the single-modal infrared thermal image into a preset defect diagnosis model for feature extraction and defect detection to obtain equipment attribute category labels and defect location bounding box coordinates, the method further includes: Collect historical inspection infrared thermal images of substation equipment; The defect samples in the historical inspection infrared thermal images are subjected to small-angle synchronous rotation, grayscale noise superposition, and brightness contrast fine-tuning to obtain an expanded sample set, which is then divided into a training set and a validation set. Construct an initial defect diagnosis model; A joint loss function is constructed by using distributed focus loss and complete intersection-union loss. The training set is then input into the initial defect diagnosis model in batches according to a preset batch size to obtain the predicted bounding box coordinates and predicted attribute category labels for each batch of samples. Based on the difference between the predicted bounding box coordinates and the real bounding box, the boundary regression loss value is calculated using the distribution focus loss and the perfect intersection-union ratio loss. Based on the difference between the predicted attribute category label and the real component attribute label, the attribute classification loss value is calculated using the binary cross-entropy loss. The boundary regression loss value and the attribute classification loss value are then weighted and summed to obtain the loss value for each batch. Based on the loss value of each batch, the network parameters of the initial defect diagnosis model are updated through the backpropagation algorithm, and the learning rate is decayed according to the preset decay coefficient after each preset decay round of training to obtain the optimized defect diagnosis model. The optimized defect diagnosis model is validated using the validation set, and the joint loss value corresponding to the validation set is recorded. When the joint loss value corresponding to the validation set no longer decreases within a preset number of consecutive stagnation rounds, or when the number of training iterations reaches a preset maximum number of iterations, the optimized defect diagnosis model corresponding to the minimum joint loss value on the validation set will be used as the preset defect diagnosis model.

7. The method as described in claim 1, characterized in that, The step of extracting the highest fault temperature and normal operating temperature from the absolute temperature field matrix based on the coordinates of the defect location bounding box includes: Using the coordinates of the defect location bounding box as a spatial retrieval index, the relative image coordinates and the physical field of view are converted and mapped back to the absolute temperature field matrix uploaded by the infrared thermal imager. Within the closed region defined by the defect location bounding box coordinates in the absolute temperature field matrix, the maximum absolute temperature pixel value is extracted by an extreme value optimization algorithm to obtain the highest heating temperature of the fault. Based on the physical topology of the three-phase parallel symmetrical distribution of electrical equipment in the substation, the area of ​​normal phase electrical equipment that is in the same group as the faulty phase electrical equipment and is under the same load condition is selected in the absolute temperature field matrix. The highest temperature pixel value at the same measuring point in the normal phase electrical equipment region is extracted from the absolute temperature field matrix to obtain the normal operating temperature.

8. The method as described in claim 1, characterized in that, The step of calculating the interphase temperature difference based on the highest fault temperature and the normal operating temperature, and performing a double threshold cascade judgment on the interphase temperature difference based on the preset severe anomaly benchmark threshold corresponding to the equipment attribute category label to obtain the equipment diagnosis result includes: Subtracting the normal operating temperature from the highest temperature of the fault, we obtain the temperature difference between phases of the same type. Based on the device attribute category label, retrieve the preset severe anomaly baseline threshold corresponding to the device attribute category label from the industry standard stored locally; The ambient temperature of the current inspection environment is obtained, and the preset severe anomaly benchmark threshold is corrected based on the deviation between the ambient temperature and the preset benchmark ambient temperature to obtain the corrected benchmark threshold. When the temperature difference between the same phases is less than the corrected reference threshold, the abnormal state of the substation equipment is determined to be a general abnormality. When the phase-to-phase temperature difference is greater than or equal to the corrected reference threshold and less than twice the corrected reference threshold, the abnormal state of the substation equipment is determined to be a serious abnormality. When the temperature difference between phases is greater than or equal to twice the corrected baseline threshold, the abnormal state of the substation equipment is determined to be a dangerous abnormality and an alarm is triggered. The heated areas corresponding to the coordinates of the defect location boundary box are highlighted, and the equipment attribute category label, the highest heating temperature of the fault, the temperature difference between phases of the same type, and the abnormal state are summarized to obtain the equipment diagnosis results.

9. An infrared defect diagnosis device for substation equipment, characterized in that, The device is applied to the infrared defect diagnosis method for substation equipment as described in any one of claims 1 to 8, and the device comprises: The data acquisition module is used to acquire single-mode infrared thermal images of substation equipment; The defect detection module is used to input the single-modal infrared thermal image into a preset defect diagnosis model for feature extraction and defect detection, and to obtain equipment attribute category labels and defect location bounding box coordinates. The preset defect diagnosis model includes a backbone feature extraction network, a neck network, and a decoupled detection network. The backbone feature extraction network is composed of a series of continuous convolutional layers, stride downsampling layers, and residual bottleneck structures. The neck network is composed of a neck feature pyramid layer and a coordinate attention enhancement network. The decoupled detection network consists of parallel attribute classification branches and boundary regression branches. The temperature extraction module is used to extract the highest fault heating temperature and normal operating temperature from the absolute temperature field matrix based on the coordinates of the defect location bounding box. The graded diagnosis module is used to calculate the interphase temperature difference based on the highest heat temperature of the fault and the normal operating temperature, and to perform double threshold cascade judgment on the interphase temperature difference based on the preset severe abnormality benchmark threshold corresponding to the equipment attribute category label, so as to obtain the equipment diagnosis result.

10. An infrared defect diagnosis device for substation equipment, characterized in that, The device includes: a memory, a processor, and a substation equipment infrared defect diagnosis program stored in the memory and running on the processor, the substation equipment infrared defect diagnosis program being configured to implement the steps of the substation equipment infrared defect diagnosis method as described in any one of claims 1 to 8.

Images