Transmission line equipment icing detection method, device, system and storage medium

Abstract

The present invention relates to the field of image recognition technology, and provides a method, device, system and storage medium for detecting ice coating on transmission line equipment. The method for detecting ice coating on transmission line equipment is executed on the edge side, and includes: collecting images of the transmission line equipment through an embedded device; analyzing the collected images using a deep learning network model deployed in the embedded device to determine the ice coating detection results of the transmission line equipment; wherein the deep learning network model is obtained after training, evaluation and compression on the server side. The present invention can realize real-time detection of ice coating on transmission line equipment, and can accurately identify and locate the ice coating area and position of the transmission line equipment.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a method for detecting icing of transmission line equipment, a device for detecting icing of transmission line equipment, a system for detecting icing of transmission line equipment, and a storage medium. Background Art

[0002] Since icing disasters can cause mechanical and electrical failures in the power grid, seriously affecting the stable operation of the power grid, it is very important to detect icing on transmission line equipment. At present, the main means of icing detection for transmission line equipment is to install sensors and cameras to achieve online monitoring of equipment. This method requires the captured image information or environmental information to be transmitted to the data center for analysis and processing, which has high requirements on data network transmission bandwidth, poor real-time performance, and it is difficult to remove icing in time. Summary of the invention

[0003] The object of the present invention is to provide a method and device for detecting ice coating on transmission line equipment, so as to at least solve the above-mentioned problem of poor real-time performance of ice coating detection on transmission line equipment.

[0004] In order to achieve the above-mentioned objectives, the first aspect of the present invention provides a method for detecting ice coating on transmission line equipment, which is executed on the edge side. The method comprises: collecting images of the transmission line equipment through an embedded device; analyzing the collected images using a deep learning network model deployed in the embedded device to determine the ice coating detection results of the transmission line equipment; wherein the deep learning network model is obtained after training, evaluation and compression on the server side.

[0005] Furthermore, training is performed on the server side, including: training the selected deep learning network model using a training data set in a power line equipment image data set on the server side, wherein the power line equipment image data set is established by the collected image data of the power line equipment.

[0006] Furthermore, the method of using the training data set in the power transmission line equipment image data set to train the selected deep learning network model includes: inputting the training data set into the deep learning network model, and obtaining a score value through forward propagation; inputting the score value into the error function to obtain an error loss value used to judge the degree of recognition loss; determining the gradient vector through back propagation; and adjusting the weight of the error loss value according to the gradient vector so that the error loss value converges or approaches zero.

[0007] Furthermore, an evaluation is performed on the server side, including: evaluating the trained deep learning network model using a test data set in the power line equipment image data set on the server side.

[0008] Furthermore, the use of a test data set in a power line equipment image data set to evaluate the trained deep learning network model includes: using the trained deep learning network model to detect the test data set, calculating an average precision mean of the test data set, and evaluating the trained deep learning network model based on the average precision mean.

[0009] Furthermore, compression is performed on the server side, including: compressing the evaluated deep learning network model on the server side.

[0010] Furthermore, compressing the evaluated deep learning network model includes: performing model pruning, model quantization and coefficient encoding on the evaluated deep learning network model.

[0011] Furthermore, the model pruning of the evaluated deep learning network model includes: judging the contribution value of each convolutional layer of the evaluated deep learning network model, deleting the core of the convolutional layer whose contribution value is less than a set value, and deleting the channel corresponding to the core of the output layer; constructing and regenerating the pruned model; and training the pruned model until convergence.

[0012] Furthermore, the model quantization includes: hybrid quantization, full integer quantization or half-precision float16 quantization.

[0013] Furthermore, a transmission line equipment image data set is established through the collected image data of the transmission line equipment, including: collecting the image data of the transmission line equipment; preprocessing the collected image data; and annotating and verifying the preprocessed image data to establish the transmission line equipment image data set.

[0014] Furthermore, the method also includes: deploying a deep learning network model in an embedded device; wherein the deep learning network model deployed in the embedded device is a deep learning network model selected on the server side that matches the embedded device.

[0015] Furthermore, the deep learning network model matching the embedded device is selected according to the deployment environment and storage space of the embedded device.

[0016] A second aspect of the present invention provides an ice detection device for transmission line equipment, the device comprising: an embedded device, arranged on the edge side, for: collecting images of transmission line equipment; analyzing the collected images using a deployed deep learning network model to determine the ice detection results of the transmission line equipment; wherein the deep learning network model is obtained after training, evaluation and compression on the server side.

[0017] A third aspect of the present invention provides a transmission line equipment icing detection system, the system comprising:

[0018] The above-mentioned transmission line equipment icing detection device; and

[0019] The server is used to: train a selected deep learning network model using a training data set in a transmission line equipment image data set, wherein the transmission line equipment image data set is established by collecting image data of the transmission line equipment; evaluate the trained deep learning network model using a test data set in the transmission line equipment image data set; and compress the evaluated deep learning network model.

[0020] Furthermore, the method of using the training data set in the power transmission line equipment image data set to train the selected deep learning network model includes: inputting the training data set into the deep learning network model, and obtaining a score value through forward propagation; inputting the score value into the error function to obtain an error loss value used to judge the degree of recognition loss; determining the gradient vector through back propagation; and adjusting the weight of the error loss value according to the gradient vector so that the error loss value converges or approaches zero.

[0021] Furthermore, the method of evaluating the trained deep learning network model using a test data set in the power line equipment image data set includes: the server using the trained deep learning network model to detect the test data set, calculating the average precision mean of the test data set, and evaluating the trained deep learning network model according to the average precision mean.

[0022] Furthermore, compressing the evaluated deep learning network model includes: performing model pruning, model quantization and coefficient encoding on the evaluated deep learning network model.

[0023] Furthermore, the model pruning of the evaluated deep learning network model includes: judging the contribution value of each convolutional layer of the evaluated deep learning network model, deleting the core of the convolutional layer whose contribution value is less than a set value, and deleting the channel corresponding to the core of the output layer; constructing and regenerating the pruned model; and training the pruned model until convergence.

[0024] Furthermore, the model quantization includes: hybrid quantization, full integer quantization or half-precision float16 quantization.

[0025] Furthermore, a transmission line equipment image data set is established through the collected image data of the transmission line equipment, including: collecting the image data of the transmission line equipment; preprocessing the collected image data; and annotating and verifying the preprocessed image data to establish the transmission line equipment image data set.

[0026] Furthermore, the server is also used to select a deep learning network model that matches the embedded device according to the deployment environment and storage space of the embedded device.

[0027] The present invention also provides a storage medium on which computer program instructions are stored, and when the computer program instructions are executed, the above-mentioned method for detecting ice coating on transmission line equipment is implemented.

[0028] The transmission line equipment icing detection method provided in the embodiment of the present invention trains, evaluates and compresses the deep learning network model on the server side, deploys the compressed deep learning network model to the embedded device on the edge side, and uses the deep learning network model of the embedded device to analyze and process the transmission line equipment image, thereby realizing real-time detection of the icing condition of the transmission line equipment, and being able to accurately identify and locate the icing area and position of the transmission line equipment. The transmission line equipment icing detection method of the present invention detects icing through the deep learning network model on the edge side, does not need to upload image data to the server side, does not occupy network bandwidth, has good real-time performance, and improves the timeliness of icing problem discovery.

[0029] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the embodiments of the present invention, but do not constitute a limitation on the embodiments of the present invention. In the accompanying drawings:

[0031] Figure 1 It is a flow chart of a method (edge ​​side) for detecting ice coating on transmission line equipment provided by one embodiment of the present invention;

[0032] Figure 2 It is a flow chart of a method (server side) for detecting ice coating on transmission line equipment provided by one embodiment of the present invention;

[0033] Figure 3 It is a block diagram of a transmission line equipment icing detection system provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0034] The specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the present invention, and is not used to limit the present invention.

[0035] Figure 1 FIG. 1 is a flow chart of a method (edge ​​side) for detecting ice coating on transmission line equipment provided by an embodiment of the present invention. Figure 1As shown, the method for detecting ice coating on transmission line equipment provided by the embodiment of the present invention is performed on the edge side, and the method includes the following steps:

[0036] Step 1: Collect the image of the power transmission line equipment through an embedded device. Specifically, it includes: collecting image data of the power transmission line equipment, preprocessing the image data, and annotating and verifying the preprocessed image data. The data set includes a training data set and a test data set.

[0037] Step 2: deploy a deep learning network model in an embedded device, and use the deep learning network model deployed in the embedded device to analyze the collected images to determine the ice detection results of the transmission line equipment; wherein the deep learning network model deployed in the embedded device is a deep learning network model selected on the server side to match the embedded device, and the deep learning network model is obtained after training, evaluation and compression on the server side.

[0038] The embedded device is, for example, ARM, TPU, etc. The control panel of the embedded device is used to collect the transmission line image, and the collected image is preprocessed, and the image preprocessing includes image enhancement, deblurring, etc. The compressed deep learning network model deployed in the embedded device is used to perform real-time analysis on the preprocessed image, locate and identify the ice-covered position, distinguish the transmission line equipment, and output the ice-covered detection result to the de-icing control system to achieve the purpose of rapid de-icing and reduce the damage of ice-covered to the transmission line equipment.

[0039] Figure 2 FIG. 1 is a flow chart of a method (server side) for detecting ice coating on transmission line equipment provided by an embodiment of the present invention. Figure 2 As shown, the method for detecting ice coating on power line equipment provided by the embodiment of the present invention performs model training, model evaluation and model compression on the server side, specifically comprising: training the selected deep learning network model using the training data set in the power line equipment image data set on the server side, wherein the power line equipment image data set is established by the collected image data of the power line equipment. Evaluating the trained deep learning network model using the test data set in the power line equipment image data set on the server side. Compressing the evaluated deep learning network model on the server side.

[0040] In this implementation, first on the server side, a deep learning network model that matches the embedded device is selected according to the deployment environment and storage space of the embedded device on the edge side, such as SSD, YOLO, Faster-Rcnn, and then the selected deep learning network model is trained, evaluated and compressed.

[0041] In one embodiment, training the selected deep learning network model includes the following steps: inputting a training data set into the deep learning network model, obtaining a score value through forward propagation; inputting the score value into an error function to obtain an error loss value; judging the degree of recognition loss; determining a gradient vector through back propagation; and adjusting weights according to the gradient vector so that the error loss value converges or approaches zero. Specifically, the process of model training using the MobileNet-SSD network and the ADAM optimization algorithm is as follows: 1) Preprocess the data of the training data set; 2) Input the preprocessed data into the deep learning neural network model (each neuron first inputs a value for weighted accumulation, and then inputs the activation function as the output value of the neuron) for forward propagation to obtain a score value; 3) Input the "score value" into the error function loss function (regularization penalty to prevent overfitting), compare it with the expected value to obtain the error, and the sum is obtained if there are multiple values. The degree of loss is identified by error judgment (the smaller the loss value, the better); 4) Determine the gradient vector through back propagation, that is, the error function and each activation function in the neural network are reversely derived to minimize the error; 5) Adjust each weight according to the gradient vector, and use the ADAM algorithm to make the error of the "score value" approach 0, or adjust the trend of convergence to the "score value". Repeat the above process 1)-5) until the set number of times or the average value of the error loss no longer decreases (that is, the error loss reaches the lowest point), and the model training is completed.

[0042] In one embodiment, the trained deep learning network model is evaluated using a test data set, including the following steps: using the trained deep learning network model to detect the test data set, calculating the mean average precision mAP (mean average precision) of the test data set, and evaluating according to the mean average precision mAP. Assume that there are a total of k categories of target objects in the test set. We first calculate the AP of our model for each category of objects, and then add these APs together and divide them by the number of all categories k, and we can get the mAP of the final model. For example, assuming that there are a total of k categories of target objects in the test data set, first calculate the AP (average precision) of the network model for each category of objects, and then add the AP of each category of objects and divide it by the number of all categories k to get the mAP of the final model. For the PR (Precision-Recall Curve) curve, different prediction results can be obtained by adjusting different confidence thresholds, thereby obtaining different precision rates (Precision) and recall rates (Recall). For an object of a certain category in an image, the model predicts the corresponding bounding box, but the intersection over union (IoU) corresponding to the predicted bounding box may be very small or very large. Therefore, assuming that the IoU corresponding to the bounding box is greater than a certain threshold, the predicted bounding box is correct and is classified as TP; if the IoU is less than the threshold, the predicted bounding box is wrong and is classified as FP. For an object in the image, the model does not predict the corresponding bounding box, which is recorded as an FN. The precision (Precision) and recall (Recall) can be calculated through TP, FP and FN. Therefore, by adjusting the IoU threshold, the PR curve of the model for each object category can be obtained. Through the PR curve, the AP value corresponding to the PR curve is obtained, and then the mean average precision (mAP) is obtained. The mean average precision (mAP) is the average value of the mean precision (AP).

[0043] In one embodiment, the evaluated deep learning network model is compressed so that the compressed network model can be accelerated on edge devices with limited computing resources. Network model compression includes operations such as model pruning, model quantization, and coefficient encoding. The weights of most kernels of deep learning network models are very small, all oscillating between -1 and 1. For these parameters with very small absolute values, they can be regarded as contributing very little to the overall model. Model pruning is to delete these parameters with very small absolute values, and then form a new model with the remaining weights to achieve the purpose of model compression, acceleration, and ensuring unchanged accuracy. The basic steps of model pruning include: 1) Obtain the weight of each convolutional layer of the original model, judge the contribution of each convolutional layer to the model, and delete the kernel of the convolutional layer with a small contribution; 2) Since deleting some kernels will cause the number of channels in the output layer to change, it is necessary to delete the channels corresponding to the kernel of the output layer; 3) Build a pruned model, load the pruned weights, and compare the accuracy with the original model; 4) Use a smaller learning rate to regenerate the pruned model; 5) Train the pruned model until convergence and save the weights of each layer.

[0044] Since wider, deeper and larger neural network models require higher computing resources, in order to enable the neural network model to be deployed from the cloud (i.e., the server) to the edge side, and considering the computing resources, storage space, operating power consumption and latency issues of the edge side devices (mobile terminals or embedded devices), the neural network model needs to be quantized. Model quantization is the process of approximating the floating-point model weights with continuous values ​​(or a large number of possible discrete values) or the tensor data flowing through the model with fixed-point approximation (usually int8) to a finite number (or fewer) of discrete values ​​with a lower loss of inference accuracy. It is the process of using a data type with fewer bits to approximate 32-bit limited range floating-point data, while the input and output of the model are still floating point, thereby achieving the goals of reducing the model size, reducing the model memory consumption and speeding up the model inference speed. Usually, after model training, the weights or activation values ​​are often distributed in a limited range, such as the activation value range of [-2.0, 6.0]. Using int8 for model quantization, the fixed-point quantization value range is [-128, 127]. Model quantization methods include tensorflow post-training quantization and tensorflow quantization-aware training. Taking tensorflow post-training quantization as an example, it includes three implementation methods: hybrid quantization, full integer quantization, and half-precision float16 quantization. Hybrid quantization is to quantize floating-point weights to int8 integers. During the inference process, the int8 quantized values ​​need to be dequantized to floating point before calculation. This method can directly reduce the model size by 75% and increase the inference speed by up to 3 times. Full integer quantization is to quantize all weights, activation values, and input values ​​to int8, and execute all model operations under int8 to achieve the best quantization effect. Half-precision float16 quantization is to quantize weights to half-precision float16 form, which can reduce the model size by half and has less precision loss than int8.

[0045] On the edge side, the compressed deep learning network model on the server side is deployed to the embedded device; the deep learning network model in the embedded device is used to analyze the image of the transmission line equipment collected by the embedded device to detect the icing condition of the transmission line equipment. The embedded device is, for example, ARM, TPU, etc. The control panel of the embedded device is used to collect the transmission line image, and the collected image is preprocessed. The image preprocessing includes operations such as image enhancement and deblurring. The compressed deep learning network model deployed in the embedded device is used to perform real-time analysis on the preprocessed image, locate and identify the icing position, distinguish the transmission line equipment, and output the icing detection result to the deicing control system to achieve the purpose of rapid deicing and reduce the damage of icing to the transmission line equipment.

[0046] The transmission line equipment icing detection method provided in the embodiment of the present invention trains, evaluates and compresses the deep learning network model on the server side, deploys the compressed deep learning network model to the embedded device on the edge side, and uses the deep learning network model of the embedded device to analyze and process the transmission line equipment image, thereby realizing real-time detection of the icing condition of the transmission line equipment, and being able to accurately identify and locate the icing area and position of the transmission line equipment. The transmission line equipment icing detection method of the present invention detects icing through the deep learning network model on the edge side, does not need to upload image data to the server side, does not occupy network bandwidth, has good real-time performance, and improves the timeliness of icing problem discovery.

[0047] An embodiment of the present invention provides an ice detection device for transmission line equipment, wherein the device includes an embedded device, which is arranged on an edge side and is used to collect images of transmission line equipment, analyze the collected images using a deployed deep learning network model, and determine ice detection results of the transmission line equipment; wherein the deep learning network model is obtained after training, evaluation and compression on the server side.

[0048] Figure 3 FIG. 1 is a block diagram of a transmission line equipment ice detection system provided by an embodiment of the present invention. Figure 3 As shown, an embodiment of the present invention provides an icing detection system for transmission line equipment, the system comprising a data set module, a server and an embedded device. The data set module is used to establish a data set of images of transmission line equipment, specifically comprising: collecting image data of transmission line equipment, preprocessing the image data, and annotating and verifying the preprocessed image data. The data set established by the data set module includes a training data set and a test data set. The server is used to train a selected deep learning network model using the training data set; to evaluate the trained deep learning network model using the test data set; and to compress the evaluated deep learning network model. The embedded device is arranged on the edge side, and is used to collect images of transmission line equipment and deploy a server-side compressed deep learning network model, and to analyze the collected images of transmission line equipment using the deep learning network model to detect the icing condition of the transmission line equipment.

[0049] In this embodiment, the server first selects a deep learning network model that matches the embedded device according to the deployment environment and storage space of the embedded device on the edge side, such as SSD, YOLO, Faster-Rcnn, and then trains, evaluates and compresses the selected deep learning network model.

[0050] In one embodiment, the server trains the selected deep learning network model including the following steps: inputting the training data set into the deep learning network model, obtaining a score value through forward propagation; inputting the score value into the error function to obtain an error loss value; judging the degree of recognition loss; determining the gradient vector through back propagation; and adjusting the weight according to the gradient vector so that the error loss value converges or approaches zero. Specifically, the process of model training using the MobileNet-SSD network and the ADAM optimization algorithm is as follows: 1) Preprocess the data of the training data set; 2) Input the preprocessed data into the deep learning neural network model (each neuron first inputs a value for weighted accumulation, and then inputs the activation function as the output value of the neuron) for forward propagation to obtain a score value; 3) Input the "score value" into the error function loss function (regularization penalty to prevent overfitting), compare it with the expected value to obtain the error, and the sum is obtained if there are multiple values. The degree of loss is identified by error judgment (the smaller the loss value, the better); 4) Determine the gradient vector through back propagation, that is, the error function and each activation function in the neural network are reversely derived to minimize the error; 5) Adjust each weight according to the gradient vector, and use the ADAM algorithm to make the error of the "score value" approach 0, or adjust the trend of convergence to the "score value". Repeat the above process 1)-5) until the set number of times or the average value of the error loss no longer decreases (that is, the error loss reaches the lowest point), and the model training is completed.

[0051] In one embodiment, the server evaluates the trained deep learning network model using a test data set, including the following steps: using the trained deep learning network model to detect the test data set, calculating the mean average precision mAP (mean average precision) of the test data set, and evaluating according to the mean average precision mAP. Assume that there are a total of k categories of target objects in the test set. We first calculate the AP of our model for each category of objects, and then add these APs together and divide them by the number of all categories k, and we can get the mAP of the final model. For example, assuming that there are a total of k categories of target objects in the test data set, first calculate the AP (average precision) of the network model for each category of objects, and then add the AP of each category of objects and divide it by the number of all categories k to get the mAP of the final model. For the PR (Precision-Recall Curve) curve, different prediction results can be obtained by adjusting different confidence thresholds, thereby obtaining different precision rates (Precision) and recall rates (Recall). For an object of a certain category in an image, the model will predict the corresponding bounding box, but the intersection over union (IoU) corresponding to this predicted bounding box may be very small or very large. Therefore, assuming that the IoU corresponding to the bounding box is greater than a certain threshold, the predicted bounding box is correct and is classified as TP; if the IoU is less than the threshold, the predicted bounding box is wrong and is classified as FP. For an object in the image, the model does not predict the corresponding bounding box, which is recorded as an FN. The precision (Precision) and recall (Recall) can be calculated through TP, FP and FN. Therefore, by adjusting the IoU threshold, the PR curve of the model for each object category can be obtained. Through the PR curve, the AP value corresponding to the PR curve is obtained, and then the mean average precision (mAP) is obtained. The mean average precision (mAP) is the average value of the mean precision (AP).

[0052] In one embodiment, the server compresses the evaluated deep learning network model so that the compressed network model can be accelerated on edge devices with limited computing resources. Network model compression includes operations such as model pruning, model quantization, and coefficient encoding. Most kernel weights of deep learning network models are very small, all oscillating between -1 and 1. For these parameters with very small absolute values, they can be regarded as contributing very little to the overall model. Model pruning is to delete these parameters with very small absolute values, and then form a new model with the remaining weights to achieve the purpose of model compression, acceleration, and ensuring unchanged accuracy. The basic steps of model pruning include: 1) Obtain the weight of each convolutional layer of the original model, judge the contribution of each convolutional layer to the model, and delete the kernel of the convolutional layer with a small contribution; 2) Since deleting some kernels will cause the number of channels in the output layer to change, it is necessary to delete the channels corresponding to the kernel of the output layer; 3) Build a pruned model, load the pruned weights, and compare the accuracy with the original model; 4) Use a smaller learning rate to regenerate the pruned model; 5) Train the pruned model until convergence and save the weights of each layer.

[0053] Since wider, deeper and larger neural network models require higher computing resources, in order to enable the neural network model to be deployed from the cloud (i.e., the server) to the edge side, and considering the computing resources, storage space, operating power consumption and latency issues of the edge side devices (mobile terminals or embedded devices), the neural network model needs to be quantized. Model quantization is the process of approximating the floating-point model weights with continuous values ​​(or a large number of possible discrete values) or the tensor data flowing through the model with fixed-point approximation (usually int8) to a finite number (or fewer) of discrete values ​​with a lower loss of inference accuracy. It is the process of using a data type with fewer bits to approximate 32-bit limited range floating-point data, while the input and output of the model are still floating point, thereby achieving the goals of reducing the model size, reducing the model memory consumption and speeding up the model inference speed. Usually, after model training, the weights or activation values ​​are often distributed in a limited range, such as the activation value range of [-2.0, 6.0]. Using int8 for model quantization, the fixed-point quantization value range is [-128, 127]. Model quantization methods include tensorflow post-training quantization and tensorflow quantization-aware training. Taking tensorflow post-training quantization as an example, it includes three implementation methods: hybrid quantization, full integer quantization, and half-precision float16 quantization. Hybrid quantization is to quantize floating-point weights to int8 integers. During the inference process, the int8 quantized values ​​need to be dequantized to floating point before calculation. This method can directly reduce the model size by 75% and increase the inference speed by up to 3 times. Full integer quantization is to quantize all weights, activation values, and input values ​​to int8, and execute all model operations under int8 to achieve the best quantization effect. Half-precision float16 quantization is to quantize weights to half-precision float16 form, which can reduce the model size by half and has less precision loss than int8.

[0054] The embedded devices of this system are ARM and TPU. The embedded devices collect images of transmission lines through the control panel, and then preprocess the collected images. Image preprocessing includes image enhancement, deblurring and other operations. The embedded devices use their own deployed deep learning network models to analyze the preprocessed images in real time, locate and identify the ice-covered positions, distinguish the transmission line equipment, and output the ice-covered detection results to the de-icing control system to achieve the purpose of rapid de-icing and reduce the damage of ice-covered transmission line equipment.

[0055] The transmission line equipment icing detection system provided in the embodiment of the present invention trains, evaluates and compresses the deep learning network model through the server, deploys the deep learning network model compressed by the server in the embedded device, and analyzes and processes the transmission line equipment image through the deep learning network model in the embedded device, thereby realizing real-time detection of the icing condition of the transmission line equipment, and being able to accurately identify and locate the icing area and position of the transmission line equipment. The transmission line equipment icing detection system of the present invention detects icing through the deep learning network model on the edge side, does not need to upload the image data to the server side, does not occupy the network bandwidth, has good real-time performance, and improves the timeliness of icing problem discovery.

[0056] An embodiment of the present invention further provides a machine-readable storage medium having computer program instructions stored thereon, and the computer program instructions, when executed, implement the above-mentioned method for detecting ice coating on transmission line equipment.

[0057] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0058] The present application is described with reference to the flowcharts and / or block diagrams of the methods, systems and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0059] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0060] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting ice coating on transmission line equipment, characterized in that: The method comprises: On the server side, a selected deep learning network model is trained using a training data set in a pre-established power line equipment image data set, the trained deep learning network model is evaluated using a test data set in the power line equipment image data set, and the evaluated deep learning network model is compressed; Deploy the compressed deep learning network model to embedded devices at the edge; Collect images of power transmission line equipment in real time through embedded devices on the edge side; Use the deep learning network model deployed in the embedded device on the edge side to analyze the real-time collected images, locate and identify the ice-covered position, distinguish the transmission line equipment, identify and locate the ice-covered area and position of the transmission line equipment, and determine the ice-covered detection result of the transmission line equipment; Outputting the ice detection result to the de-icing control system, in: The method of evaluating the trained deep learning network model by using the test data set in the transmission line equipment image data set includes: detecting the test data set by using the trained deep learning network model, calculating the average precision mean of the test data set, and evaluating the trained deep learning network model according to the average precision mean; The calculating the mean average precision of the test data set includes: calculating the mean average precision value AP of the trained deep learning network model for each category of objects, then adding the mean average precision values ​​of the objects in each category and dividing the sum by the number k of all categories to obtain the mean average precision mAP of the trained deep learning network model; The compression of the evaluated deep learning network model includes model pruning, model quantization and coefficient encoding of the evaluated deep learning network model, and the model pruning of the evaluated deep learning network model includes: obtaining the weight of each convolutional layer of the evaluated deep learning network model, deleting the core of the convolutional layer with a weight between -1 and 1, and deleting the channel corresponding to the core of the output layer; constructing and regenerating the pruned model; and training the pruned model until convergence.

2. The method for detecting ice coating on transmission line equipment according to claim 1, characterized in that: The method of training the selected deep learning network model using a training data set in a pre-established power transmission line equipment image data set includes: Inputting the training data set into the deep learning network model, and obtaining a score value through forward propagation; Inputting the score value into an error function to obtain an error loss value for determining the degree of recognition loss; Determine the gradient vector by back-propagation; The weight of the error loss value is adjusted according to the gradient vector so that the error loss value converges or approaches zero.

3. The method for detecting ice coating on transmission line equipment according to claim 1, characterized in that: The model quantization includes: hybrid quantization, full integer quantization or half-precision float16 quantization.

4. The method for detecting ice coating on transmission line equipment according to claim 1, characterized in that: The power transmission line equipment image dataset is pre-established in the following manner: Collect image data of power transmission line equipment; Preprocessing the collected image data; The preprocessed image data are labeled and verified to establish a power line equipment image dataset.

5. The method for detecting ice coating on transmission line equipment according to claim 1, characterized in that: The deep learning network model deployed in the embedded device is a deep learning network model selected on the server side to match the embedded device.

6. The method for detecting ice coating on transmission line equipment according to claim 5, characterized in that: The deep learning network model that matches the embedded device is selected based on the deployment environment and storage space of the embedded device.

7. A transmission line equipment ice detection system, characterized in that: The system comprises: A transmission line equipment ice detection device, including an embedded device arranged at an edge side, for collecting images of the transmission line equipment in real time; A data set module, used to pre-establish a training data set and a test data set based on the collected image data of the power transmission line equipment; and The server is used to train the selected deep learning network model using the training data set; evaluate the trained deep learning network model using the test data set; and compress the evaluated deep learning network model; The embedded device on the edge side is also used to use the deployed compressed deep learning network model to analyze the images of the power transmission line equipment collected in real time, locate and identify the ice-covered position, distinguish the power transmission line equipment, thereby identifying and locating the ice-covered area and position of the power transmission line equipment, determining the ice-covered detection result of the power transmission line equipment, and outputting the ice-covered detection result to the de-icing control system. in: The evaluating the trained deep learning network model by using the test data set includes: detecting the test data set by using the trained deep learning network model, calculating the average precision mean of the test data set, and evaluating the trained deep learning network model according to the average precision mean; The calculating the mean average precision of the test data set includes: calculating the mean average precision value AP of the trained deep learning network model for each category of objects, then adding the mean average precision values ​​of the objects in each category and dividing the sum by the number k of all categories to obtain the mean average precision mAP of the trained deep learning network model; The compression of the evaluated deep learning network model includes model pruning, model quantization and coefficient encoding of the evaluated deep learning network model, and the model pruning of the evaluated deep learning network model includes: obtaining the weight of each convolutional layer of the evaluated deep learning network model, deleting the core of the convolutional layer with a weight between -1 and 1, and deleting the channel corresponding to the core of the output layer; constructing and regenerating the pruned model; and training the pruned model until convergence.

8. The transmission line equipment icing detection system according to claim 7, characterized in that: The method of training the selected deep learning network model using the training data set includes: Inputting the training data set into the deep learning network model, and obtaining a score value through forward propagation; Inputting the score value into an error function to obtain an error loss value for determining the degree of recognition loss; Determine the gradient vector by back-propagation; The weight of the error loss value is adjusted according to the gradient vector so that the error loss value converges or approaches zero.

9. The transmission line equipment icing detection system according to claim 7, characterized in that: The model quantization includes: hybrid quantization, full integer quantization or half-precision float16 quantization.

10. The transmission line equipment icing detection system according to claim 7, characterized in that: A training data set and a test data set are pre-established based on the collected image data of the power transmission line equipment, including: The data set module collects image data of power transmission line equipment; Preprocessing the collected image data; The preprocessed image data are labeled and verified to establish the training dataset and test dataset in the power line equipment image dataset.

11. The transmission line equipment icing detection system according to claim 7, characterized in that: The server is also used to select a deep learning network model that matches the embedded device according to the deployment environment and storage space of the embedded device.

12. A storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed, the method for detecting ice coating on power line equipment according to any one of claims 1 to 6 is implemented.

Images