CN113486936A - Icing detection method, device and system for power transmission line equipment and storage medium - Google Patents
Icing detection method, device and system for power transmission line equipment and storage medium Download PDFInfo
- Publication number
- CN113486936A CN113486936A CN202110719482.6A CN202110719482A CN113486936A CN 113486936 A CN113486936 A CN 113486936A CN 202110719482 A CN202110719482 A CN 202110719482A CN 113486936 A CN113486936 A CN 113486936A
- Authority
- CN
- China
- Prior art keywords
- deep learning
- line equipment
- network model
- learning network
- power transmission
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 230000005540 biological transmission Effects 0.000 title claims abstract description 94
- 238000001514 detection method Methods 0.000 title claims abstract description 37
- 238000003860 storage Methods 0.000 title claims abstract description 18
- 238000013135 deep learning Methods 0.000 claims abstract description 118
- 238000000034 method Methods 0.000 claims abstract description 58
- 238000012549 training Methods 0.000 claims abstract description 55
- 238000011156 evaluation Methods 0.000 claims abstract description 16
- 238000007906 compression Methods 0.000 claims abstract description 13
- 230000006835 compression Effects 0.000 claims abstract description 13
- 238000013139 quantization Methods 0.000 claims description 44
- 238000012360 testing method Methods 0.000 claims description 30
- 238000013138 pruning Methods 0.000 claims description 22
- 230000006870 function Effects 0.000 claims description 20
- 238000007781 pre-processing Methods 0.000 claims description 13
- 238000011002 quantification Methods 0.000 claims description 11
- 238000004590 computer program Methods 0.000 claims description 10
- 239000011248 coating agent Substances 0.000 claims description 7
- 238000000576 coating method Methods 0.000 claims description 7
- 230000001172 regenerating effect Effects 0.000 claims description 6
- 238000011897 real-time detection Methods 0.000 abstract description 3
- 230000008569 process Effects 0.000 description 11
- 230000004913 activation Effects 0.000 description 10
- 238000010586 diagram Methods 0.000 description 10
- 238000003062 neural network model Methods 0.000 description 8
- 238000012545 processing Methods 0.000 description 7
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 4
- 210000002569 neuron Anatomy 0.000 description 4
- 238000010361 transduction Methods 0.000 description 4
- 230000001133 acceleration Effects 0.000 description 2
- 238000009825 accumulation Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000009434 installation Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000002265 prevention Effects 0.000 description 2
- 230000026683 transduction Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Business, Economics & Management (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Economics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Public Health (AREA)
- Water Supply & Treatment (AREA)
- Human Resources & Organizations (AREA)
- Marketing (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to the technical field of image recognition, and provides a method, a device and a system for detecting icing of power transmission line equipment and a storage medium. The icing detection method of the power transmission line equipment is executed at the edge side and comprises the following steps: acquiring an image of the power transmission line equipment through embedded equipment; analyzing the acquired image by using a deep learning network model deployed in the embedded equipment, and determining an icing detection result of the power transmission line equipment; the deep learning network model is obtained after training, evaluation and compression are carried out on the server side. The invention can realize the real-time detection of the icing condition of the power transmission line equipment and can accurately identify and position the icing area and position of the power transmission line equipment.
Description
Technical Field
The invention relates to the technical field of image recognition, in particular to a power transmission line equipment icing detection method, a power transmission line equipment icing detection device, a power transmission line equipment icing detection system and a storage medium.
Background
Because icing disasters can cause mechanical faults and electrical faults of a power grid and seriously affect the stable operation of the power grid, the icing detection of the power transmission line equipment is very important. At present, the icing detection means of the power transmission line equipment mainly realizes the online monitoring of the equipment by installing a sensor and a camera, the method needs to transmit the shot image information or environment information to a data center for analysis and processing, the requirement on the transmission bandwidth of a data network is high, the real-time performance is poor, and the icing is difficult to remove in time.
Disclosure of Invention
The invention aims to provide a method and a device for detecting icing of power transmission line equipment, which at least solve the problem of poor real-time performance of the icing detection of the power transmission line equipment.
In order to achieve the above object, a first aspect of the present invention provides a method for detecting icing on a power transmission line device, the method being performed on an edge side, the method including: acquiring an image of the power transmission line equipment through embedded equipment; analyzing the acquired image by using a deep learning network model deployed in the embedded equipment, and determining an icing detection result of the power transmission line equipment; the deep learning network model is obtained after training, evaluation and compression are carried out on the server side.
Further, training is performed on the server side, comprising: and training the selected deep learning network model by utilizing 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 acquired image data of the power line equipment.
Further, the training the selected deep learning network model by using the training data set in the image data set of the power line device includes: inputting the training data set into a 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 judging the recognition loss degree; determining a gradient vector by back propagation; and adjusting the weight of the error loss value according to the gradient vector to make the error loss value converge or tend to zero.
Further, the evaluation is performed at the server side, comprising: and evaluating the trained deep learning network model by utilizing a test data set in the image data set of the power transmission line equipment at the server side.
Further, the evaluating the trained deep learning network model by using the test data set in the image data set of the power line device includes: and detecting the test data set by using the trained deep learning network model, calculating an average precision mean value of the test data set, and evaluating the trained deep learning network model according to the average precision mean value.
Further, the compression is performed at the server side, comprising: and compressing the evaluated deep learning network model on the server side.
Further, the compressing the evaluated deep learning network model includes: and performing model pruning, model quantization and coefficient coding on the evaluated deep learning network model.
Further, the performing model pruning on the evaluated deep learning network model includes: judging the contribution value of each convolution layer of the deep learning network model after evaluation, deleting the core of the convolution layer with the contribution value smaller than a set value, and deleting the channel corresponding to the core of the output layer; constructing and regenerating a model after pruning; and training the pruned model until convergence.
Further, the model quantification includes: mixed quantization, full integer quantization or half precision float16 quantization.
Further, establishing a powerline device image dataset from the acquired image data of the powerline device, comprising: acquiring image data of the power transmission line equipment; preprocessing the acquired image data; and marking and verifying the preprocessed image data to establish an image data set of the power line equipment.
Further, the method further comprises: deploying a deep learning network model in the embedded equipment; the deep learning network model deployed in the embedded device is the deep learning network model which is selected at the server side and matched with the embedded device.
Further, the deep learning network model matched with the embedded device is selected according to the deployment environment and the storage space of the embedded device.
A second aspect of the present invention provides an icing detection device for a power transmission line equipment, the device comprising: the embedded equipment is arranged on the edge side and used for: collecting an image of the power transmission line equipment; analyzing the acquired image by using the deployed deep learning network model, and determining the icing detection result of the power transmission line equipment; the deep learning network model is obtained after training, evaluation and compression are carried out on the server side.
A third aspect of the present invention provides a power transmission line equipment icing detection system, the system comprising:
the icing detection device of the power transmission line equipment; and
a server to: training the selected deep learning network model by utilizing a training data set in an image data set of the power line equipment, wherein the image data set of the power line equipment is established by the acquired image data of the power line equipment; evaluating the trained deep learning network model by using a test data set in the image data set of the power transmission line equipment; and compressing the evaluated deep learning network model.
Further, the training the selected deep learning network model by using the training data set in the image data set of the power line device includes: inputting the training data set into a 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 judging the recognition loss degree; determining a gradient vector by back propagation; and adjusting the weight of the error loss value according to the gradient vector to make the error loss value converge or tend to zero.
Further, the evaluating the trained deep learning network model by using the test data set in the image data set of the power line device includes: and the server detects the test data set by using the trained deep learning network model, calculates the average precision mean value of the test data set, and evaluates the trained deep learning network model according to the average precision mean value.
Further, the compressing the evaluated deep learning network model includes: and performing model pruning, model quantization and coefficient coding on the evaluated deep learning network model.
Further, the performing model pruning on the evaluated deep learning network model includes: judging the contribution value of each convolution layer of the deep learning network model after evaluation, deleting the core of the convolution layer with the contribution value smaller than a set value, and deleting the channel corresponding to the core of the output layer; constructing and regenerating a model after pruning; and training the pruned model until convergence.
Further, the model quantification includes: mixed quantization, full integer quantization or half precision float16 quantization.
Further, establishing a powerline device image dataset from the acquired image data of the powerline device, comprising: acquiring image data of the power transmission line equipment; preprocessing the acquired image data; and marking and verifying the preprocessed image data to establish an image data set of the power line equipment.
Further, the server is also used for selecting a deep learning network model matched with the embedded device according to the deployment environment and the storage space of the embedded device.
The present invention also provides a storage medium having stored thereon computer program instructions which, when executed, implement the above-described method of detecting icing on a power transmission line device.
According to the method for detecting the icing of the power transmission line equipment, the deep learning network model is trained, evaluated and compressed at the server side, the compressed deep learning network model is deployed to the embedded equipment at the edge side, the deep learning network model of the embedded equipment is used for analyzing and processing the image of the power transmission line equipment, the real-time detection of the icing condition of the power transmission line equipment is realized, and the icing area and position of the power transmission line equipment can be accurately identified and positioned. According to the method for detecting the icing of the power transmission line equipment, the icing is detected through the deep learning network model on the edge side, image data do not need to be uploaded to a server side, network bandwidth is not occupied, the real-time performance is good, and the timeliness of finding the icing problem is improved.
Additional features and advantages of embodiments of the invention will be set forth in the detailed description which follows.
Drawings
The accompanying drawings, which are included to provide a further understanding of the embodiments of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the embodiments of the invention without limiting the embodiments of the invention. In the drawings:
fig. 1 is a flow chart of a method for detecting icing on a power line device (edge side) according to an embodiment of the present invention;
FIG. 2 is a flow chart of a method (server side) for detecting icing on a powerline device according to one embodiment of the present invention;
fig. 3 is a block diagram of an icing detection system for a power transmission line installation according to an embodiment of the present invention.
Detailed Description
The following detailed description of embodiments of the invention refers to the accompanying drawings. It should be understood that the detailed description and specific examples, while indicating the present invention, are given by way of illustration and explanation only, not limitation.
Fig. 1 is a flowchart of an icing detection method (edge side) for a power line device according to an embodiment of the present invention. As shown in fig. 1, the method for detecting icing on a power transmission line device according to an embodiment of the present invention is performed on an edge side, and includes the following steps:
step one, collecting an image of the power transmission line equipment through embedded equipment. The method specifically comprises the following steps: the method comprises the steps of collecting image data of the power transmission line equipment, preprocessing the image data, and marking and verifying the preprocessed image data. The data sets include a training data set and a testing data set.
Deploying a deep learning network model in the embedded equipment, analyzing the acquired image by using the deep learning network model deployed in the embedded equipment, and determining the icing detection result of the power transmission line equipment; the deep learning network model deployed in the embedded equipment is a deep learning network model which is selected at a server side and matched with the embedded equipment, and the deep learning network model is obtained after training, evaluation and compression are carried out on the server side.
Such as an ARM, TPU, etc. The method comprises the steps of collecting power transmission line images by using a control panel of embedded equipment, and preprocessing the collected images, wherein the image preprocessing comprises operations of image enhancement, deblurring and the like. The method comprises the steps of utilizing a compressed deep learning network model deployed in embedded equipment to analyze a preprocessed image in real time, locating and identifying an icing position, distinguishing power transmission line equipment, and outputting an icing detection result to a deicing control system, so that the aim of quickly deicing is fulfilled, and damage of icing to the power transmission line equipment is reduced.
Fig. 2 is a flowchart of a method (server side) for detecting icing on a power line device according to an embodiment of the present invention. As shown in fig. 2, the method for detecting icing of power transmission line equipment according to the embodiment of the present invention performs model training, model evaluation, and model compression on a server side, and specifically includes: and training the selected deep learning network model by utilizing 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 acquired image data of the power line equipment. And evaluating the trained deep learning network model by utilizing a test data set in the image data set of the power transmission line equipment at the server side. And compressing the evaluated deep learning network model on the server side.
In the embodiment, at the server side, a deep learning network model matched with the embedded device, such as SSD, YOLO, fast-Rcnn, is selected according to the deployment environment and the storage space of the embedded device at the edge side, and then the selected deep learning network model is trained, evaluated and compressed.
In one embodiment, the selected deep learning network model is trained, comprising the steps of: inputting the training data set into a 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; judging the degree of identification loss; determining a gradient vector by back propagation; and adjusting the weight according to the gradient vector to make the error loss value converge or tend to zero. Specifically, the process of model training by using the MobileNet-SSD network and the ADAM optimization algorithm is as follows: 1) preprocessing the data of the training data set; 2) inputting the preprocessed data into a deep learning neural network model (each neuron inputs a value for weighted accumulation and then inputs an activation function as an output value of the neuron) for forward propagation to obtain a score value; 3) inputting the 'score value' into an error function loss function (regularization punishment, over-fitting prevention), comparing the 'score value' with an expected value to obtain errors, and judging and identifying the loss degree (the smaller the loss value is, the better the loss degree) through the errors if a plurality of the 'score values' are sums; 4) determining a gradient vector through back propagation, namely reversely deriving an error function and each activation function in the neural network to minimize the error; 5) and adjusting each weight according to the gradient vector, and adopting an ADAM algorithm to enable the error of the score value to tend to 0, or adjusting the error toward the convergence trend of the score value. And repeating the processes 1) to 5) until the set times or the average value of the error loss does not fall (namely the error loss reaches the lowest point), and finishing the model training.
In one embodiment, the evaluation of the trained deep learning network model by using the test data set comprises the following steps: detecting the test data set by using the trained deep learning network model, calculating an average precision mean value mAP (mean average precision) of the test data set, and evaluating according to the average precision mean value mAP. Assume a total of k classes of target objects in the test set. We first calculate the AP of our model for each class of objects, then add these APs together and divide by the number k of all classes to get the mAP of the final model. For example, assuming that there are k types of target objects in the test dataset, the AP (average precision) of the network model for each type of object is calculated, and then the APs of each type of object are added and divided by the number k of all types to obtain the maps of the final model. For a PR (Precision-Recall Curve), different prediction results can be obtained by adjusting different confidence thresholds, so that different Precision rates (Precision) and Recall rates (Recall) are obtained. For a class of objects in the image, the model predicts the corresponding bounding box, but the intersection ratio (IoU) for this predicted bounding box may be small or large. Therefore, assuming that IoU corresponding to the bounding box is greater than a certain threshold, this predicted bounding box is paired and divided into TPs; if IoU is less than the threshold, the predicted bounding box is false and is classified as FP. For an object in the image, the model does not predict the corresponding bounding box, which is denoted as primary FN. Precision (Precision) and Recall (Recall) can be calculated from TP, FP and FN. By adjusting IoU the threshold value, the model's PR curve can be derived for each object class. And obtaining an AP value corresponding to the PR curve through the PR curve, thereby obtaining an average precision average (mAP). The average mean of precision (mAP) is the average of the mean of precision (AP).
In one embodiment, the evaluated deep learning network model is compressed, so that the compressed network model can be accelerated to run on the edge side device with limited computing resources. The network model compression comprises operations of model pruning, model quantization, coefficient coding and the like. The weights of most kernel of the deep learning network model are small and oscillate between-1 and 1, and for the parameters with small absolute values, the contribution of the parameters to the whole model is small. The model pruning means that the parameters with small absolute values are deleted, and then the rest weights form a new model so as to achieve the purposes of model compression, acceleration and unchanged accuracy. The basic steps of model pruning include: 1) obtaining the weight of each convolution layer of the original model, judging the contribution of each convolution layer to the model, and deleting the kernel (core) of the convolution layer with smaller contribution; 2) after part of kernel is deleted, the number of channels of the output layer is changed, so that the channels corresponding to the kernel of the output layer need to be deleted; 3) constructing a model after pruning, loading the weight after pruning, and comparing the precision with the original model; 4) regenerating the pruned model with a smaller learning rate; 5) and training the pruned model until convergence, and storing the weight of each layer.
Because the wider, deeper and larger neural network models have higher requirements on computing resources, in order to enable the neural network models to be deployed from a cloud (i.e., a server) to an edge side, the problems of computing resources, storage space, operating power consumption and time delay of edge side devices (mobile terminals or embedded devices) are considered, and therefore the neural network models need to be quantized. The model quantization is a process of approximating the floating-point model weight of continuous values (or a large number of possible discrete values) or tensor data flowing through the model (usually int8) to a finite number of (or fewer) discrete values with low inference precision loss, and it is a process of approximating 32-bit finite-range floating-point data with a data type of less number of bits, and the input and output of the model are still floating-point, so as to achieve the objectives of reducing the size of the model, reducing the memory consumption of the model, and increasing the inference speed of the model. In general, the weights or activation values after model training are often distributed in a limited range, such as the activation value range of [ -2.0,6.0], and model quantization is performed by using int8, and the fixed point quantization value range of [ -128,127 ]. Model quantification methods include Post-transduction quantification (postflow quantification) and transduction-aware quantification (quantification-aware transduction). Taking quantization after tenserflow training as an example, the method comprises three implementation modes of mixed quantization, full integer quantization and half precision float16 quantization. The mixed quantization is to quantize the weight of the floating point type into int8 integer, and in the inference process, the int8 quantization value needs to be inversely quantized into the floating point type and then calculated, so that the size of the model can be directly reduced by 75%, and the inference speed is increased by 3 times at most. The full integer quantization is to quantize the weights, activation values and input values to int8, and perform all model operations under int8 to achieve the best quantization effect. The quantization of half-precision float16 is a quantization of the weights into the form of half-precision float16, which can reduce the model size by half with less loss of precision compared to int 8.
Deploying a deep learning network model compressed by a server side into embedded equipment on an edge side; and analyzing the image of the power transmission line equipment acquired by the embedded equipment by using the deep learning network model in the embedded equipment, and detecting the icing condition of the power transmission line equipment. Such as an ARM, TPU, etc. The method comprises the steps of collecting power transmission line images by using a control panel of embedded equipment, and preprocessing the collected images, wherein the image preprocessing comprises operations of image enhancement, deblurring and the like. The method comprises the steps of utilizing a compressed deep learning network model deployed in embedded equipment to analyze a preprocessed image in real time, locating and identifying an icing position, distinguishing power transmission line equipment, and outputting an icing detection result to a deicing control system, so that the aim of quickly deicing is fulfilled, and damage of icing to the power transmission line equipment is reduced.
According to the method for detecting the icing of the power transmission line equipment, the deep learning network model is trained, evaluated and compressed at the server side, the compressed deep learning network model is deployed to the embedded equipment at the edge side, the deep learning network model of the embedded equipment is used for analyzing and processing the image of the power transmission line equipment, the real-time detection of the icing condition of the power transmission line equipment is realized, and the icing area and position of the power transmission line equipment can be accurately identified and positioned. According to the method for detecting the icing of the power transmission line equipment, the icing is detected through the deep learning network model on the edge side, image data do not need to be uploaded to a server side, network bandwidth is not occupied, the real-time performance is good, and the timeliness of finding the icing problem is improved.
The embodiment of the invention provides an icing detection device for power transmission line equipment, which comprises embedded equipment, wherein the embedded equipment is arranged on the edge side and is used for acquiring images of the power transmission line equipment, analyzing the acquired images by using a deployed deep learning network model and determining an icing detection result of the power transmission line equipment; the deep learning network model is obtained after training, evaluation and compression are carried out on the server side.
Fig. 3 is a block diagram of an icing detection system for a power transmission line installation according to an embodiment of the present invention. As shown in fig. 3, the embodiment of the present invention provides an icing detection system for a power line device, which includes a data set module, a server and an embedded device. The data set module is used for establishing a data set of the power line equipment image, and specifically comprises the following steps: the method comprises the steps of collecting image data of the power transmission line equipment, preprocessing the image data, and marking and verifying the preprocessed image data. The data set established by the data set module comprises a training data set and a testing data set. The server is used for training the selected deep learning network model by utilizing the training data set; evaluating the trained deep learning network model by using the test data set; and compressing the evaluated deep learning network model. The embedded equipment is arranged on the edge side and used for acquiring images of the power transmission line equipment and deploying a server-side compressed deep learning network model, analyzing the acquired images of the power transmission line equipment by using the deep learning network model and detecting the icing condition of the power transmission line equipment.
In this embodiment, the server first selects a deep learning network model, such as SSD, YOLO, fast-Rcnn, matching the edge-side embedded device according to the deployment environment and the storage space of the embedded device, and then trains, evaluates, and compresses the selected deep learning network model.
In one embodiment, the server training the selected deep learning network model comprises the following steps: inputting the training data set into a 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; judging the degree of identification loss; determining a gradient vector by back propagation; and adjusting the weight according to the gradient vector to make the error loss value converge or tend to zero. Specifically, the process of model training by using the MobileNet-SSD network and the ADAM optimization algorithm is as follows: 1) preprocessing the data of the training data set; 2) inputting the preprocessed data into a deep learning neural network model (each neuron inputs a value for weighted accumulation and then inputs an activation function as an output value of the neuron) for forward propagation to obtain a score value; 3) inputting the 'score value' into an error function loss function (regularization punishment, over-fitting prevention), comparing the 'score value' with an expected value to obtain errors, and judging and identifying the loss degree (the smaller the loss value is, the better the loss degree) through the errors if a plurality of the 'score values' are sums; 4) determining a gradient vector through back propagation, namely reversely deriving an error function and each activation function in the neural network to minimize the error; 5) and adjusting each weight according to the gradient vector, and adopting an ADAM algorithm to enable the error of the score value to tend to 0, or adjusting the error toward the convergence trend of the score value. And repeating the processes 1) to 5) until the set times or the average value of the error loss does not fall (namely the error loss reaches the lowest point), and finishing the model training.
In one embodiment, the server evaluates the trained deep learning network model by using the test data set, and the method comprises the following steps: detecting the test data set by using the trained deep learning network model, calculating an average precision mean value mAP (mean average precision) of the test data set, and evaluating according to the average precision mean value mAP. Assume a total of k classes of target objects in the test set. We first calculate the AP of our model for each class of objects, then add these APs together and divide by the number k of all classes to get the mAP of the final model. For example, assuming that there are k types of target objects in the test dataset, the AP (average precision) of the network model for each type of object is calculated, and then the APs of each type of object are added and divided by the number k of all types to obtain the maps of the final model. For a PR (Precision-Recall Curve), different prediction results can be obtained by adjusting different confidence thresholds, so that different Precision rates (Precision) and Recall rates (Recall) are obtained. For a class of objects in the image, the model predicts the corresponding bounding box, but the intersection ratio (IoU) for this predicted bounding box may be small or large. Therefore, assuming that IoU corresponding to the bounding box is greater than a certain threshold, this predicted bounding box is paired and divided into TPs; if IoU is less than the threshold, the predicted bounding box is false and is classified as FP. For an object in the image, the model does not predict the corresponding bounding box, which is denoted as primary FN. Precision (Precision) and Recall (Recall) can be calculated from TP, FP and FN. By adjusting IoU the threshold value, the model's PR curve can be derived for each object class. And obtaining an AP value corresponding to the PR curve through the PR curve, thereby obtaining an average precision average (mAP). The average mean of precision (mAP) is the average of the mean of precision (AP).
In one embodiment, the server compresses the evaluated deep learning network model so that the compressed network model can be accelerated to run on the edge side device with limited computing resources. The network model compression comprises operations of model pruning, model quantization, coefficient coding and the like. The weights of most kernel of the deep learning network model are small and oscillate between-1 and 1, and for the parameters with small absolute values, the contribution of the parameters to the whole model is small. The model pruning means that the parameters with small absolute values are deleted, and then the rest weights form a new model so as to achieve the purposes of model compression, acceleration and unchanged accuracy. The basic steps of model pruning include: 1) obtaining the weight of each convolution layer of the original model, judging the contribution of each convolution layer to the model, and deleting the kernel (core) of the convolution layer with smaller contribution; 2) after part of kernel is deleted, the number of channels of the output layer is changed, so that the channels corresponding to the kernel of the output layer need to be deleted; 3) constructing a model after pruning, loading the weight after pruning, and comparing the precision with the original model; 4) regenerating the pruned model with a smaller learning rate; 5) and training the pruned model until convergence, and storing the weight of each layer.
Because the wider, deeper and larger neural network models have higher requirements on computing resources, in order to enable the neural network models to be deployed from a cloud (i.e., a server) to an edge side, the problems of computing resources, storage space, operating power consumption and time delay of edge side devices (mobile terminals or embedded devices) are considered, and therefore the neural network models need to be quantized. The model quantization is a process of approximating the floating-point model weight of continuous values (or a large number of possible discrete values) or tensor data flowing through the model (usually int8) to a finite number of (or fewer) discrete values with low inference precision loss, and it is a process of approximating 32-bit finite-range floating-point data with a data type of less number of bits, and the input and output of the model are still floating-point, so as to achieve the objectives of reducing the size of the model, reducing the memory consumption of the model, and increasing the inference speed of the model. In general, the weights or activation values after model training are often distributed in a limited range, such as the activation value range of [ -2.0,6.0], and model quantization is performed by using int8, and the fixed point quantization value range of [ -128,127 ]. Model quantification methods include Post-transduction quantification (postflow quantification) and transduction-aware quantification (quantification-aware transduction). Taking quantization after tenserflow training as an example, the method comprises three implementation modes of mixed quantization, full integer quantization and half precision float16 quantization. The mixed quantization is to quantize the weight of the floating point type into int8 integer, and in the inference process, the int8 quantization value needs to be inversely quantized into the floating point type and then calculated, so that the size of the model can be directly reduced by 75%, and the inference speed is increased by 3 times at most. The full integer quantization is to quantize the weights, activation values and input values to int8, and perform all model operations under int8 to achieve the best quantization effect. The quantization of half-precision float16 is a quantization of the weights into the form of half-precision float16, which can reduce the model size by half with less loss of precision compared to int 8.
The embedded equipment of the system is ARM and TPU. The embedded equipment collects the power transmission line image through the control panel and then preprocesses the collected image, wherein the image preprocessing comprises operations of image enhancement, deblurring and the like. The embedded equipment utilizes a self-deployed deep learning network model to analyze the preprocessed image in real time, positions and identifies the icing position, distinguishes the power transmission line equipment, and outputs the icing detection result to the deicing control system, so that the aim of quickly deicing is fulfilled, and damage of icing to the power transmission line equipment is reduced.
According to the ice coating detection system for the power transmission line equipment, the deep learning network model is trained, evaluated and compressed through the server, the deep learning network model compressed through the server is deployed in the embedded equipment, the images of the power transmission line equipment are analyzed and processed through the deep learning network model in the embedded equipment, the ice coating condition of the power transmission line equipment is detected in real time, and the ice coating area and the position of the power transmission line equipment can be accurately identified and positioned. According to the ice coating detection system for the power transmission line equipment, disclosed by the invention, ice coating is detected through the deep learning network model at the edge side, image data does not need to be uploaded to a server side, network bandwidth is not occupied, the real-time performance is good, and the timeliness of ice coating problem discovery is improved.
Embodiments of the present invention also provide a machine-readable storage medium having stored thereon computer program instructions that, when executed, implement the above-described method for detecting icing on a power transmission line device.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, systems and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solutions of the present invention and not for limiting the same, and although the present invention is described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications and equivalents may be made to the embodiments of the invention without departing from the spirit and scope of the invention, which is to be covered by the claims.
Claims (22)
1. A method for detecting icing on a power transmission line device, the method being performed on an edge side, the method comprising:
acquiring an image of the power transmission line equipment through embedded equipment;
analyzing the acquired image by using a deep learning network model deployed in the embedded equipment, and determining an icing detection result of the power transmission line equipment; the deep learning network model is obtained after training, evaluation and compression are carried out on the server side.
2. The method for detecting icing on a power transmission line equipment according to claim 1, wherein the training is performed on the server side, and comprises:
and training the selected deep learning network model by utilizing 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 acquired image data of the power line equipment.
3. The method for detecting icing on power line equipment according to claim 2, wherein the training of the selected deep learning network model using a training data set in an image data set of the power line equipment comprises:
inputting the training data set into a 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 judging the recognition loss degree;
determining a gradient vector by back propagation;
and adjusting the weight of the error loss value according to the gradient vector to make the error loss value converge or tend to zero.
4. The method for detecting icing on a power transmission line equipment according to claim 2, wherein the evaluation is performed on the server side and comprises:
and evaluating the trained deep learning network model by utilizing a test data set in the image data set of the power transmission line equipment at the server side.
5. The method for detecting icing on power line equipment according to claim 4, wherein the evaluating the trained deep learning network model using the test data set in the power line equipment image data set comprises:
and detecting the test data set by using the trained deep learning network model, calculating an average precision mean value of the test data set, and evaluating the trained deep learning network model according to the average precision mean value.
6. The method for detecting icing on a power transmission line equipment according to claim 2, wherein the compressing is performed on the server side, and comprises:
and compressing the evaluated deep learning network model on the server side.
7. The method for detecting icing on power transmission line equipment according to claim 6, wherein the compressing the evaluated deep learning network model comprises: and performing model pruning, model quantization and coefficient coding on the evaluated deep learning network model.
8. The method for detecting icing on power transmission line equipment according to claim 7, wherein the performing model pruning on the evaluated deep learning network model comprises:
judging the contribution value of each convolution layer of the deep learning network model after evaluation, deleting the core of the convolution layer with the contribution value smaller than a set value, and deleting the channel corresponding to the core of the output layer;
constructing and regenerating a model after pruning;
and training the pruned model until convergence.
9. The method for detecting icing on a power transmission line device according to claim 7, wherein said model quantifying comprises: mixed quantization, full integer quantization or half precision float16 quantization.
10. The method for detecting icing on power line equipment according to claim 2, wherein the step of establishing a power line equipment image data set from the acquired power line equipment image data comprises the steps of:
acquiring image data of the power transmission line equipment;
preprocessing the acquired image data;
and marking and verifying the preprocessed image data to establish an image data set of the power line equipment.
11. The method for detecting icing on a power transmission line device according to claim 1, further comprising:
deploying a deep learning network model in the embedded equipment; the deep learning network model deployed in the embedded device is the deep learning network model which is selected at the server side and matched with the embedded device.
12. The method for detecting icing on power line equipment according to claim 11, wherein the deep learning network model matched with the embedded equipment is selected according to the deployment environment and the storage space of the embedded equipment.
13. An icing detection device for a power transmission line equipment, the device comprising:
the embedded equipment is arranged on the edge side and used for:
collecting an image of the power transmission line equipment;
analyzing the acquired image by using the deployed deep learning network model, and determining the icing detection result of the power transmission line equipment; the deep learning network model is obtained after training, evaluation and compression are carried out on the server side.
14. An ice coating detection system for a power transmission line device, the system comprising:
the power transmission line equipment icing detection apparatus of claim 13; and
a server to:
training the selected deep learning network model by utilizing a training data set in an image data set of the power line equipment, wherein the image data set of the power line equipment is established by the acquired image data of the power line equipment;
evaluating the trained deep learning network model by using a test data set in the image data set of the power transmission line equipment;
and compressing the evaluated deep learning network model.
15. The power line equipment icing detection system of claim 14, wherein training the selected deep learning network model with a training data set in a power line equipment image data set comprises:
inputting the training data set into a 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 judging the recognition loss degree;
determining a gradient vector by back propagation;
and adjusting the weight of the error loss value according to the gradient vector to make the error loss value converge or tend to zero.
16. The power transmission line equipment icing detection system of claim 15, wherein the evaluation of the trained deep learning network model using the test data set in the power transmission line equipment image data set comprises:
and the server detects the test data set by using the trained deep learning network model, calculates the average precision mean value of the test data set, and evaluates the trained deep learning network model according to the average precision mean value.
17. The power transmission line equipment icing detection system of claim 16, wherein the compressing the evaluated deep learning network model comprises: and performing model pruning, model quantization and coefficient coding on the evaluated deep learning network model.
18. The system according to claim 17, wherein said model pruning of the evaluated deep learning network model comprises:
judging the contribution value of each convolution layer of the deep learning network model after evaluation, deleting the core of the convolution layer with the contribution value smaller than a set value, and deleting the channel corresponding to the core of the output layer;
constructing and regenerating a model after pruning;
and training the pruned model until convergence.
19. The power transmission line equipment icing detection system of claim 17, wherein the model quantification comprises: mixed quantization, full integer quantization or half precision float16 quantization.
20. The powerline device icing detection system of claim 14, wherein establishing the powerline device image dataset from the collected powerline device image data comprises:
acquiring image data of the power transmission line equipment;
preprocessing the acquired image data;
and marking and verifying the preprocessed image data to establish an image data set of the power line equipment.
21. The system according to claim 14, wherein the server is further configured to select a deep learning network model matching the embedded device according to a deployment environment and a storage space of the embedded device.
22. A storage medium having computer program instructions stored thereon that, when executed, implement the method for detecting icing on a power line device of any of claims 1-12.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110719482.6A CN113486936A (en) | 2021-06-28 | 2021-06-28 | Icing detection method, device and system for power transmission line equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110719482.6A CN113486936A (en) | 2021-06-28 | 2021-06-28 | Icing detection method, device and system for power transmission line equipment and storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113486936A true CN113486936A (en) | 2021-10-08 |
Family
ID=77936412
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110719482.6A Pending CN113486936A (en) | 2021-06-28 | 2021-06-28 | Icing detection method, device and system for power transmission line equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113486936A (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106595551A (en) * | 2016-12-13 | 2017-04-26 | 武汉大学 | Icing thickness detection method for power transmission line icing image based on deep learning |
| CN110070530A (en) * | 2019-04-19 | 2019-07-30 | 山东大学 | A kind of powerline ice-covering detection method based on deep neural network |
| CN110175636A (en) * | 2019-05-08 | 2019-08-27 | 深圳欧翼思特科技有限公司 | A kind of Internet of Things deep neural network distribution differentiation inference system and method |
| CN111709522A (en) * | 2020-05-21 | 2020-09-25 | 哈尔滨工业大学 | Deep learning target detection system based on server-embedded cooperation |
| CN112232509A (en) * | 2020-10-10 | 2021-01-15 | 苏州浪潮智能科技有限公司 | Edge calculation model compression optimization method, device, equipment and readable medium |
| CN112528971A (en) * | 2021-02-07 | 2021-03-19 | 北京智芯微电子科技有限公司 | Power transmission line abnormal target detection method and system based on deep learning |
-
2021
- 2021-06-28 CN CN202110719482.6A patent/CN113486936A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106595551A (en) * | 2016-12-13 | 2017-04-26 | 武汉大学 | Icing thickness detection method for power transmission line icing image based on deep learning |
| CN110070530A (en) * | 2019-04-19 | 2019-07-30 | 山东大学 | A kind of powerline ice-covering detection method based on deep neural network |
| CN110175636A (en) * | 2019-05-08 | 2019-08-27 | 深圳欧翼思特科技有限公司 | A kind of Internet of Things deep neural network distribution differentiation inference system and method |
| CN111709522A (en) * | 2020-05-21 | 2020-09-25 | 哈尔滨工业大学 | Deep learning target detection system based on server-embedded cooperation |
| CN112232509A (en) * | 2020-10-10 | 2021-01-15 | 苏州浪潮智能科技有限公司 | Edge calculation model compression optimization method, device, equipment and readable medium |
| CN112528971A (en) * | 2021-02-07 | 2021-03-19 | 北京智芯微电子科技有限公司 | Power transmission line abnormal target detection method and system based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| 焦李成: "《深度神经网络FPGA设计与实现》", 31 October 2020, 《西安电子科技大学出版社》, pages: 219 - 222 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108090508B (en) | classification training method, device and storage medium | |
| CN108090470B (en) | Face alignment method and device | |
| CN107636690B (en) | Full reference image quality assessment based on convolutional neural network | |
| US6397200B1 (en) | Data reduction system for improving classifier performance | |
| CN111898578B (en) | Crowd density acquisition method and device and electronic equipment | |
| CN110135505B (en) | Image classification method and device, computer equipment and computer readable storage medium | |
| CN110956255A (en) | Difficult sample mining method and device, electronic equipment and computer readable storage medium | |
| CN113328908B (en) | Abnormal data detection method and device, computer equipment and storage medium | |
| CN115758130A (en) | Rolling bearing fault diagnosis method and system based on unsupervised transfer learning | |
| CN116258941A (en) | Yolox target detection lightweight improvement method based on Android platform | |
| CN115484112A (en) | Payment big data security protection method and system and cloud platform | |
| CN112801231B (en) | Decision model training method and device for business object classification | |
| CN112560881A (en) | Object identification method and device and data processing method | |
| CN113780287A (en) | Optimal selection method and system for multi-depth learning model | |
| CN117319223A (en) | Digital twinning technology-based pilot cockpit visualization method and system | |
| CN114565092A (en) | Neural network structure determining method and device | |
| CN117521512A (en) | Bearing residual service life prediction method based on multi-scale Bayesian convolution transducer model | |
| CN113486936A (en) | Icing detection method, device and system for power transmission line equipment and storage medium | |
| US20220269988A1 (en) | Abnormality degree calculation system and abnormality degree calculation method | |
| KR20200072717A (en) | Artificial inteligence learning methods and apparatus using analysis of data frequency value | |
| CN111368837A (en) | Image quality evaluation method and device, electronic equipment and storage medium | |
| CN114973173A (en) | Method and device for classifying driving scene data, electronic equipment and storage medium | |
| US11100321B2 (en) | Information processing method and information processing system | |
| KR20210158740A (en) | Apparatus and method for clustering validation based on machine learning performance | |
| CN112733729A (en) | Model training and regression analysis method, device, storage medium and equipment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |