Power equipment on-line detection system based on lightweight model

Abstract

The invention discloses a power equipment online detection system based on a lightweight model, and particularly relates to the field of deep learning, and the system comprises a data collection module, a data preprocessing module, a detection model training module, an equipment reasoning module, an abnormity recognition module, a demand cooling analysis module, and a man-machine interaction module. According to the invention, the lightweight model is used to carry out the online detection of the power equipment at the edge computing gateway, the training of the detection model is carried out before the equipment reasoning, and the detection precision under a complex background is improved through the parameter optimization unit, the texture enhancement unit, the multi-scale fusion unit and the residual enhancement unit. When the overheating risk is identified, targeted cooling adjustment is carried out based on the temperature distribution of the abnormal region and the accurate measurement and calculation of the temperature difference of the adjacent regions, so that the cooperative capability of the multi-layer features is improved, the model can still keep higher robustness under noise interference, and the method has a wide application prospect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, more particularly, the present application relates to an online detection system for power equipment based on a lightweight model. BACKGROUND

[0002] The power system is the lifeblood of modern society, and its safe, stable and efficient operation is of great importance. Various types of power equipment in power plants, substations and transmission lines are the core units of this complex system. These devices are inevitably subject to insulation aging, mechanical wear, poor contact and other hidden faults under harsh working conditions such as high voltage, large current and strong electromagnetic field.

[0003] The traditional detection and maintenance method of power equipment mainly uses physical sensors and signal processing for periodic inspection and preventive testing, which can extract feature values related to equipment failure. However, in actual use, it still has some shortcomings. First, the existing device fault acquisition cannot realize continuous monitoring, which may miss sudden failures or rapidly developing defects, and may cause fluctuations during power equipment fault correction, thus requiring correction adjustment. However, the existing correction adjustment usually tends to shut down the equipment, which may affect power supply reliability and cannot reflect the real operating state of the equipment, resulting in errors in power equipment detection and causing greater safety hazards. Second, the physical sensor collects various state parameters of the equipment, so deep domain knowledge is needed to design and select effective features, which has poor generalization ability and hidden dangers that cannot meet the detection requirements. Based on this situation, the power equipment detection model needs to be adjusted to adapt to various equipment, but the existing method cannot effectively integrate data from different sensors and different physical meanings for joint diagnosis. For early and weak fault signs, traditional algorithms are difficult to accurately separate and identify from background noise. SUMMARY

[0004] Therefore, the embodiments of the present application provide an online detection system for power equipment based on a lightweight model, which uses a lightweight model to detect power equipment in real time, locates faults and analyzes demand cooling based on model training, and intelligently warns and regulates the state of the equipment, effectively solving the problems in the background art.

[0005] To achieve the above purpose, the present application provides the following technical scheme: The data collection module is used to retrieve the detection log of the power equipment, and the target power equipment corresponding feature map is collected by the lightweight network to form a power dataset; The data preprocessing module is used to preprocess the collected feature map to expand the power dataset; The detection model training module is configured to train a detection model using the expanded power dataset, and includes a parameter optimization unit, a texture enhancement unit, a multi-scale fusion unit, and a residual enhancement unit. The device inference module is configured to deploy the detection model to an edge computing gateway, and to receive real-time feature maps corresponding to the target power device for preprocessing, so as to transmit the preprocessed real-time feature maps to the detection model for inference. The anomaly identification module is configured to identify anomalies in the feature maps corresponding to the target power device based on the inference results, so as to obtain an abnormal region. The demand cooling analysis module is configured to demarcate a neighboring region based on the abnormal region of the feature maps corresponding to the target power device, to obtain the temperature of the neighboring region, and to analyze the demand cooling temperature based on the temperature. The human-computer interaction module is configured to adjust the power device based on the demand cooling, and to display a warning on a Web management interface in a preset manner after receiving the warning through a cloud platform, and to generate a maintenance work order.

[0006] The technical effects and advantages of the present application are as follows: 1. The present application uses a lightweight model to implement online detection of power devices in an edge computing gateway, and performs fault positioning and demand cooling analysis based on the inference results of the feature maps during anomaly identification, so as to intelligently warn and regulate the state of the device, which is not limited to simple alarm. On the one hand, it can meet the demand for precise operation and maintenance to the maximum extent, and on the other hand, it can avoid the consumption of human resources and the impact of device downtime caused by frequent manual inspection, which is beneficial to improving the operation and maintenance efficiency under the premise of ensuring continuous power supply of the power grid. 2. The present application performs demand cooling analysis after performing the anomaly identification operation through the detection model training, and uses a lightweight network to significantly reduce the model inference resource consumption and improve the running speed of the edge platform. It has stronger texture focusing ability and feature fusion ability in the task of detecting fuzzy small targets, and the detection accuracy is improved compared with the basic model, thereby reducing the model parameters, significantly improving the average frame rate in the power device image test set, and being suitable for automatic inspection of power devices in harsh scenes such as complex weather and low definition. BRIEF DESCRIPTION OF DRAWINGS

[0007] Fig. 1 It is a schematic diagram of the overall structure of the present application.

[0008] Fig. 2 It is a flowchart of obtaining a probability vector of the present application.

[0009] Fig. 3 It is a detection model inference flowchart of the present application. DETAILED DESCRIPTION

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] As attached Figs. 1-3 The above describes an online power equipment monitoring system based on a lightweight model, which includes a data collection module, a data preprocessing module, a monitoring model training module, an equipment inference module, an anomaly identification module, a demand cooling analysis module, and a human-computer interaction module.

[0012] The specific embodiments of the present invention include the following: The data collection module is used to retrieve the detection logs of power equipment and collect the corresponding feature maps of the target power equipment in the lightweight network to form a power dataset.

[0013] In this embodiment, it should be specifically noted that the detection log of the power equipment includes the power equipment ID, timestamp, and detection point, and the feature map of the equipment is captured at the same timestamp.

[0014] It should be noted that the lightweight network is the APT-TNet network.

[0015] Furthermore, the collected feature maps are initially automatically labeled, including marking images with relevant log records as abnormal and those without any records as normal. At the same time, a human-computer interaction interface is provided for experts to perform secondary verification and fine labeling, including delineating fault areas, defining fault types, and filtering batch data according to preset rules, thereby constructing an initial training dataset as a power dataset, where the preset rules can be time range, equipment type, and fault type.

[0016] It should be explained that since most power equipment failures are accompanied by abnormal heating during their development, temperature sensing of power equipment can detect the equipment status in real time, identify potential problems in advance, avoid unexpected shutdowns, and greatly improve the reliability and stability of power grid supply.

[0017] It should be noted that when collecting the operating temperature of the power equipment, an infrared thermal sensor is used to perceive and form a thermal map, which can realize non-contact measurement of the internal temperature of the power equipment without directly contacting the equipment or internal elements, avoiding interference or damage caused by contact sensors. On the other hand, by forming a thermal map, the temperature distribution of each region inside the power equipment can be comprehensively understood, which helps to comprehensively evaluate the working state and abnormal conditions of the power equipment. In addition, the infrared thermal sensor has a small size and convenient installation method compared with other sensors, which is easy to deploy inside the power equipment without affecting the layout and structure of the power equipment.

[0018] The data preprocessing module is used to preprocess the collected feature maps, thereby expanding the power dataset.

[0019] It should be noted in this embodiment that the original feature maps obtained by the data collection module are normalized and data enhanced, thereby expanding the power dataset, which specifically includes: Normalization: all feature maps are uniformly scaled to a fixed size required by the model; the absolute temperature of the target power equipment corresponding to the feature map is extracted, wherein the absolute temperature represents the actual temperature value of each region of the power equipment measured directly by the infrared thermal sensor, and the gray value thereof is obtained according to the absolute temperature, and is scaled to the [0, 1] interval by minimum and maximum scaling. Specifically, the temperature value of the current feature map is normalized by comparing the difference between the maximum gray value and the minimum gray value with the maximum gray value. Data enhancement: geometric transformation and pixel transformation are performed on the feature map to simulate feature map collection under different environmental conditions, and Gaussian noise is added to the image.

[0020] It should be explained that normalizing all feature maps can eliminate the influence of different original image sizes, and accelerate model convergence by minimum and maximum scaling to provide a data basis for subsequent lightweight model construction; data enhancement can further improve the generalization performance of the model, and adding noise to the image can improve the anti-interference ability of the model.

[0021] The detection model training module is used to train the detection model using the expanded power dataset.

[0022] It should be noted in this embodiment that the detection model training includes a parameter optimization unit, a texture enhancement unit, a multi-scale fusion unit, and a residual enhancement unit, which are specifically as follows: Parameter optimization: a reparameterization convolution structure is used to reduce the amount of calculation and the number of parameters while ensuring the feature expression capability, and the specific formula is as follows: , wherein the input feature map , H and W represent the height and width of the input feature map respectively, and the output feature map f is obtained by performing convolution operation on 25% of the channel x1 and splicing with the original input channel x2 out .

[0023] It should be noted that the channel dimension is split into the activation area and the reserved area . onv3×3 3x3 convolution, the output is spliced with x2 to generate , by performing convolution on only 25% of the channel, the parameter amount and the calculation amount are reduced by about 32.8% and 25% respectively, and the feature map after splicing is added to the residual of the identity mapping path to improve the information flow efficiency and reduce the gradient vanishing problem, that is, y=F(x)+x short , wherein x short represents the direct transmission path of the feature map x, that is, the identity mapping, and the structure design can effectively reduce the parameter amount by 32.8% and the calculation burden by 25% compared with the traditional residual network, which is suitable for mobile terminal or unmanned aerial vehicle deployment.

[0024] Texture enhancement: please refer to Fig. 3 , the application introduces texture enhancement in the high-level encoder, through the guided spatial bias mechanism, fuses the position prior, deformable attention and learnable spatial distribution intensity, to strengthen the attention focusing ability of the texture blurred area in the image, and adaptively enhances the texture features through the target driven spatial guidance strategy, thereby flexibly fine-tuning the attention area, and improving the focusing ability of the model on the key texture area, the formula is as follows: , , wherein Q, K, V and B are query (Query), key (Key), value (Value) and bias (Bias) matrices respectively, used to calculate adaptive spatial attention, x sample is the feature after dynamic sampling, x represents the input feature map, and B map () represents a function for generating a spatial bias matrix B.

[0025] It should be noted that through a group of convolution layers (including 3x3 deep convolution, normalization and GELU activation): Δp=Conv offset (Concat(Q, B mapped )), wherein the offset Δp is normalized by the tanh function and limited to the range factor, B mapped indicates that the spatial bias map is matched with the query (Q) matrix dimension after transformation and normalization, then, based on the offset and the reference point p ref , the sampling position p=p ref +Δp is calculated, and the local features are sampled by bilinear interpolation, and the spatial bias map is generated by the learnable target center c=(c y , cx ) and standard deviation σ = (σ y , σ x ) are normalized, and then multiplied by a learnable bias strength b s to obtain the spatial bias map ; finally, multi-head attention is calculated for the sampled features, combined with learnable relative position encoding to enhance spatial perception capabilities.

[0026] Multi-scale fusion: through dynamic scaling to enhance the fusion effect of features of different scales, and by introducing the SDI mechanism, the problem of feature loss is avoided, and the formula is as follows: , where F L , F M and F S represent large-scale, medium-scale and small-scale feature maps respectively, Align represents an alignment operation, and through dynamic calculation of the scaling factor, the feature maps of different scales are aligned, and finally spliced together to obtain the fused feature F fused .

[0027] It should be noted that as the input, the feature map is first subjected to bidirectional adaptive pooling on the large-scale feature map based on the spatial dimensions corresponding to the medium-scale feature map, achieving spatial alignment and being marked as Align FL , and at the same time, a lightweight upsampler DySample (linear interpolation mode) is initialized as needed according to the real-time calculated scaling factor, upsampling the small-scale feature map to the size of the medium-scale feature map, and being marked as Align Fs . Finally, the three aligned features are spliced along the channel dimension to generate the fused feature F dused , which effectively solves the problem of inconsistent resolution and incomplete information coverage between feature maps of different scales, and maintains key texture details.

[0028] It should be noted that the SDI mechanism design process is as follows: unify the features to the same channel dimension; refine the structure through 1x1 convolution and parameter optimization; realize dense interaction through weighted fusion; and output the fused feature for upsampling decoding.

[0029] This structure can effectively suppress the feature redundancy caused by splicing and improve the semantic purity of the fused feature, and is particularly suitable for scenes where electrical components are severely occluded.

[0030] Residual enhancement: through information competition and selection mechanism, it is used to solve the problem of redundant information interference in feature layer fusion, and the formula is as follows: F output = F input + F residual , where F input is the input feature map, and F residualis a residual feature map, which represents that information compensation is performed between different scale features through residual connection, so as to optimize the fusion efficiency of cross-scale features.

[0031] It should be noted that the residual enhancement is specifically implemented as follows: 1x1 convolution is performed on the input feature maps C3, C4 and C5, and the channels are unified to c, which is described as: , where Conv 1×1 represents a 1x1 convolution operation, , , C3, C4 and C5 are respectively new feature maps after 1x1 convolution processing of C3, C4 and C5; Then, the feature maps , are respectively batch normalized and upsampled to consistent resolution, which is described as: , where BN is a batch normalization layer, Upsample is an upsample operation, and nearest neighbor interpolation is used to avoid blurring effect, the feature map output by the deep level in the model, the feature map output by the high deep level in the model.

[0032] The aligned feature maps are then element-wise maximum, and an activation function is used, which can be described as: , where LeakyReLU 0.1 is an activation function, is a maximum value after stacking along the channel dimension, which is used to simulate the neuron competition activation mechanism; Finally, the fusion result F fused is added to the original high-resolution feature X3 through residual connection, which retains the spatial details of the bottom layer and compensates for the information loss caused by upsample, which is described as: F residual =F fuesd +X3.

[0033] The device inference module is used to deploy the detection model to the edge computing gateway, and real-time features corresponding to the target power equipment are preprocessed, so that the preprocessed real-time features are transmitted to the detection model for inference.

[0034] It should be noted that in this embodiment, the device inference module deploys the lightweight model file generated by the trained detection model to the inference engine of the edge gateway, and real-time features corresponding to the target power equipment are collected by the infrared thermal sensor; The feature map collected in real time is subjected to the same preprocessing operation as the detection model training module, and the detection model inference including the normal, abnormal and defect corresponding probability vectors is performed.

[0035] It should be further explained that the probability vector is obtained as follows: The lightweight model file is deployed to the edge gateway and output to obtain the score vector, marked as [Zn, Za, Zd].

[0036] It should be explained that the value range of the score vector is arbitrary, reflecting the strength and weakness of the model for each category.

[0037] The values in the score vector are exponentiated by the Softmax function, and the scores in the score vector are extracted and exponentiated, and the exponential sum is calculated, which is specifically represented as: , Where To represents the exponential sum of the scores in the score vector, Zn, Za and Zd represent the scores of normal, abnormal and defect in the detection model, and e represents the natural constant.

[0038] The scores of normal, abnormal and defect in the detection model are respectively calculated by standardization, and the probabilities corresponding to normal, abnormal and defect are respectively obtained, which are specifically represented as: 、 、 Where ZN, ZA and ZD represent the probabilities corresponding to normal, abnormal and defect, respectively.

[0039] The probabilities corresponding to normal, abnormal and defect are integrated into a probability vector, and marked as [ZN, ZA, ZD].

[0040] Further, the detection model inference is specifically operated as follows: according to the probability vector, the class with the highest probability is selected as the preliminary inference result of the model by argmax().

[0041] For example, assuming that the preliminary prediction result is Pre=argmax([ZN, ZA, ZD])=argmax([0.067, 0.111, 0.822])=3, and assuming that the index 3 corresponds to the category "defect", then the target power equipment is preliminarily determined as "defect".

[0042] Set the confidence threshold, extract the maximum probability and minimum probability corresponding to each category, and compare the difference between the two with the confidence threshold. If the difference between the two is greater than or equal to the confidence threshold, the result is reliable, and an abnormal alarm is triggered immediately.

[0043] If the difference is less than the confidence threshold, it is determined that there is a difference in the result, at which point the operation and maintenance personnel are notified for manual intervention to view the original infrared image and the on-site situation, and re-reasoning is performed, and when there are continuous multiple differences in the result, the model needs to be retrained or optimized.

[0044] An abnormality identification module identifies the abnormality of the target power equipment corresponding feature map based on the reasoning result, thereby obtaining an abnormal area.

[0045] It needs to be specifically explained in this embodiment that the abnormal area is obtained through abnormality identification of the feature map, and the specific operation is as follows: According to the reasoning result, the target category is obtained, and an abnormality judgment map corresponding to the original heat map is generated through Grad-CAM; Feature extraction is performed on the abnormality judgment map, specifically, a temperature threshold is set, the heat map is binarized to obtain a binary image, and the temperature threshold can be 20%, indicating that only the area with a temperature value in the top 20% is retained; The brightest connected region in the binary image is obtained through connected region analysis, and the minimum bounding rectangle of the connected region is calculated, thereby obtaining the bounding box of the abnormal area; The temperature data in the original heat map is called, and the temperature of all pixels in the abnormal area is extracted from the temperature data in the original heat map according to the bounding box of the abnormal area.

[0046] It needs to be explained that Grad-CAM is a visualization technology for generating a heat map to show the importance of image regions to the decision of a deep learning model. Global average pooling is performed on the gradient of the feature map to obtain a set of weights. These weights are multiplied by the feature map of the corresponding last convolutional layer and summed to synthesize a heat map as an abnormality judgment map. Finally, the abnormality judgment map is superimposed on the original image in a color mapping manner to obtain the final visualization result. Through Grad-CAM, the fault point of the power equipment can be identified and visualized, greatly improving the abnormality identification efficiency and providing data support and abnormal areas for the subsequent demand cooling analysis module. Connected region analysis can identify connected regions in an image, and then Grad-CAM combined with connected regions can accurately obtain the fault location in the power equipment, thereby laying a foundation for subsequent measurement and analysis.

[0047] A demand cooling analysis module is used to delimit adjacent areas according to the abnormal area of the target power equipment corresponding feature map, and then the temperature of the adjacent area is obtained, and the demand cooling temperature is analyzed accordingly.

[0048] It needs to be explained that the adjacent area refers to an area that shares a boundary profile with the abnormal area.

[0049] It needs to be specifically pointed out in this embodiment that the adjacent area is determined as follows: the area sharing the boundary profile with the abnormal area is extracted as the adjacent area.

[0050] Further, the analysis of the temperature reduction temperature is implemented as follows: The number of adjacent areas is counted, and the size of the adjacent area is extracted from the internal thermal map of the power equipment. The temperature of the adjacent area is extracted from the internal thermal map of the power equipment, and the highest temperature and the lowest temperature are extracted by comparing the temperatures of the adjacent areas. The temperature fitting degree corresponding to the adjacent area is calculated according to the highest temperature and the lowest temperature, which is specifically represented as: , Where Sm represents the temperature fitting degree corresponding to the adjacent area, T max and T min represent the highest temperature and the lowest temperature respectively, and e represents the natural constant, wherein the smaller the difference between the highest temperature and the lowest temperature, the greater the temperature fitting degree. The temperature fitting degree corresponding to the adjacent area is compared with the effective temperature fitting degree, and the effective temperature fitting degree is initially configured by the system, which aims to assist the analysis of the temperature reduction target. For example, the effective temperature fitting degree is 0.8. If the temperature fitting degree corresponding to the adjacent area is greater than or equal to the effective temperature fitting degree, the target temperature is obtained by calculating the average temperature of each adjacent area. Otherwise, the size of each adjacent area is compared, and the temperature of the largest size adjacent area is extracted as the target temperature, wherein the target temperature is the temperature that needs to be reached.

[0051] The man-machine interaction module adjusts the power equipment according to the demand for temperature reduction, and displays in a preset manner based on the lightweight network after receiving the early warning through the cloud platform, and generates a maintenance work order.

[0052] It needs to be specifically pointed out in this embodiment that the man-machine interaction module includes a warning unit and a control unit, wherein the warning unit is connected with the mobile terminal based on the lightweight network, and when it is detected that the power equipment has an abnormality, the mobile terminal will give an early warning. The mobile terminal pushes the key alarm information to the mobile device of the operation and maintenance personnel in real time. The push information contains a brief explanation and a deep link to the direct detail page, and the power equipment ID is extracted according to the power equipment detection log to ensure that the relevant personnel can know the situation at the first time and handle it in time. The regulation unit automatically generates an instruction and pops up a window to request final confirmation for execution based on the demand cooling temperature preset automatic strategy rule transmitted by the demand cooling analysis module when the alarm information hits the rule condition, and remotely intervenes in the operation state of the power equipment to ensure safety, at this time, the preset mode is displayed on the Web management interface, the preset mode is one of picture display, voice display and text display, when selecting the preset mode display, it needs to be selected according to the situation, if the picture display can make the operation and maintenance personnel understand better, then the preset mode is picture display to generate a maintenance work order, if the text display can make the operation and maintenance personnel obtain the detection result of the power equipment better, then the preset mode is text display to generate a maintenance work order.

[0053] It should be noted that the automatic strategy rule can be set by a lightweight model based on the power equipment detection log.

[0054] It should be noted that the lightweight network can be integrated into the power inspection unmanned aerial vehicle platform, equipped with a high-definition visible light camera, an embedded inference, a communication module and a ground station system, and multi-source data fusion analysis is performed through the APT-TNet model, the system can quickly process input images and output detailed analysis results, which are displayed on the user interface, the system has the advantages of model lightweight, fast response speed and high recognition accuracy, significantly improves the efficiency and accuracy of power equipment online monitoring, and provides a strong guarantee for the stable operation of the power grid.

[0055] Secondly, in the drawings of the disclosed embodiments, only the structures related to the disclosed embodiments are involved, other structures can be referred to the general design, and the same embodiments and different embodiments of the present application can be combined with each other under the condition of no conflict. Finally, the above only describes the preferred embodiments of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A lightweight model-based online detection system for power equipment, characterized in that, The system comprises a data collection module, a data preprocessing module, a detection model training module, a device inference module, an anomaly identification module, a demand cooling analysis module, and a human-computer interaction module. The data collection module is configured to call detection logs of the power equipment and collect feature maps corresponding to the target power equipment from the lightweight network to form a power dataset. The data preprocessing module is configured to preprocess the collected feature maps to expand the power dataset. The detection model training module is configured to train the detection model using the expanded power dataset, including a parameter optimization unit, a texture enhancement unit, a multi-scale fusion unit, and a residual enhancement unit. The device inference module is configured to deploy the detection model to the edge computing gateway, receive the feature maps corresponding to the target power equipment in real time, preprocess the real-time feature maps, and then transmit the preprocessed real-time feature maps to the detection model for inference. The anomaly identification module is configured to identify the feature maps corresponding to the target power equipment based on the inference results to obtain abnormal areas. The demand cooling analysis module is configured to delimit adjacent areas according to the abnormal areas of the feature maps corresponding to the target power equipment, obtain the temperature of the adjacent areas, and analyze the demand cooling temperature based on the temperature. The human-computer interaction module is configured to adjust the power equipment according to the demand cooling, display the pre-set mode on the Web management interface after receiving the warning through the cloud platform, and generate a maintenance work order.

2. The online detection system of power equipment based on lightweight model according to claim 1, characterized in that: The expanded power dataset specifically includes: Normalization processing: all feature maps are uniformly scaled to a fixed size required by the model; the absolute temperature of the feature maps corresponding to the target power equipment is extracted, and the gray value is obtained based on the absolute temperature; the minimum and maximum scaling is performed to normalize the temperature value of the current feature map to the [0, 1] interval, which is specifically represented as the difference between the maximum gray value and the minimum gray value compared with the maximum gray value to obtain the temperature value of the current feature map after normalization; Data enhancement: geometric transformation and pixel transformation are performed on the feature maps to simulate feature map collection under different environmental conditions, and Gaussian noise is added to the image.

3. The online detection system of power equipment based on lightweight model according to claim 1, characterized in that: The detection model training includes a parameter optimization unit, a texture enhancement unit, a multi-scale fusion unit, and a residual enhancement unit, specifically as follows: Parameter optimization unit: adopt re-parameterization convolution structure, specific formula as follows: , wherein the input feature map , H and W represent the height and width of the input feature map respectively, the output feature map f out is obtained by performing convolution operation on 25% of the channels x1 and splicing with the original input channel x2. Texture enhancement unit: adaptively enhance the texture features by target-driven spatial guidance strategy, formula as follows: , where Q, K, V, B are Query, Key, Value and Bias matrix respectively, used to calculate adaptive spatial attention, where x sample is the feature after dynamic sampling, x represents the input feature map, B map () represents a function for generating a spatial bias matrix B; Multi-scale fusion unit: enhance the fusion effect of different scale features through dynamic scaling, and introduce SDI mechanism, the formula is as follows: , where F L , F M and F S represent large-scale, medium-scale and small-scale feature maps respectively, Align represents an alignment operation, different scale feature maps are aligned by dynamically calculating a scaling factor, and finally spliced together to obtain a fused feature F fused ; Residual enhancement unit: feature optimization between different scales through information competition and selection mechanism, formula as follows: F output =F input +F residual , where F input is the input feature map, F residual is the residual feature map, indicating information compensation between different scale features through residual connection.

4. The online detection system of power equipment based on lightweight model according to claim 1, characterized in that: The device inference module deploys the lightweight model file generated by the trained detection model to the inference engine of the edge gateway, and collects the feature maps corresponding to the target power equipment in real time through the infrared thermal sensor; The same preprocessing operation as the detection model training module is performed on the real-time collected feature maps, and the detection model inference including the normal, abnormal, and defect corresponding probability vectors is performed.

5. The online detection system of power equipment based on lightweight model according to claim 4, characterized in that: The probability vector is obtained as follows: After deploying the lightweight model file to the edge gateway, the score vector is output, which is marked as [Zn, Za, Zd]; The values in the score vector are exponentiated by the Softmax function, and the scores in the score vector are extracted and exponentiated, thereby calculating the exponential sum, which is specifically represented as: , Where To represents the sum of the exponents of each score in the score vector, Zn, Za, and Zd represent the scores of normal, abnormal, and defect in the detection model, and e represents the natural constant; The scores of normal, abnormal and defect in the detection model are respectively normalized to obtain the probabilities corresponding to normal, abnormal and defect, and are specifically expressed as: , , wherein ZN, ZA and ZD represent the probabilities corresponding to normal, abnormal and defect, respectively. The probabilities corresponding to normal, abnormal, and defect are integrated into a probability vector, which is marked as [ZN, ZA, ZD].

6. The online detection system of power equipment based on lightweight model according to claim 4, characterized in that: The detection model inference is specifically operated as follows: according to the probability vector, the class with the highest probability is selected as the preliminary inference result of the model by argmax(). Set a confidence threshold, extract the maximum and minimum probabilities from the corresponding probabilities of each category, and compare the difference between the two with the confidence threshold. If the difference is greater than or equal to the confidence threshold, the result is considered reliable and an anomaly warning is triggered immediately. If the difference between the two is less than the confidence threshold, the judgment result is different. At this time, the operation and maintenance personnel are notified to manually intervene to check the original infrared image and the on-site situation, and to re-infer. If there are multiple consecutive differences in the results, the model needs to be retrained or optimized.

7. The online detection system of power equipment based on lightweight model according to claim 1, characterized in that: The abnormal region is obtained by identifying anomalies in the feature map, specifically through the following operations: The target category is obtained based on the reasoning results, and an anomaly judgment map corresponding to the original heatmap is generated using Grad-CAM. Feature extraction is performed on the anomaly detection map, specifically by setting a temperature threshold and binarizing the heat map to obtain a binary image. The brightest connected region in the binary image is obtained by connecting region analysis, and the minimum bounding rectangle of the connected region is calculated, thereby obtaining the bounding box of the abnormal region. Retrieve the temperature data from the original heatmap, and based on the bounding box of the abnormal region, extract the temperature of all pixels within that abnormal region from the temperature data in the original heatmap.

8. The online detection system of power equipment based on lightweight model of claim 1, wherein: The process for implementing the temperature analysis of the required cooling is as follows: Count the number of adjacent regions and extract the size of adjacent regions from the internal heat map of the power equipment; Extract the temperature of adjacent areas from the internal thermal map of the power equipment, compare the temperatures of adjacent areas, and extract the highest and lowest temperatures from them; Calculate the temperature compatibility between adjacent areas based on the highest and lowest temperatures; The temperature fit of adjacent regions is compared with the effective temperature fit. The effective temperature fit is initially configured in the system and its purpose is to assist in the analysis of the cooling target. If the temperature fit of adjacent regions is greater than or equal to the effective temperature fit, the average temperature of each adjacent region is calculated to obtain the target temperature. Otherwise, the size of each adjacent region is compared, and the temperature of the adjacent region with the largest size is extracted as the target temperature, where the target temperature is the cooling temperature that needs to be achieved.

Images