CN112183182A - Defect identification method for power transmission line - Google Patents

Abstract

The invention belongs to the technical field of image recognition, and particularly relates to a defect recognition method for a power transmission line. The invention provides an image recognition technology based on a YOLO algorithm. The identification system can not only accurately monitor the damage of the external force and induce the microclimate of the iron tower, but also detect the defects of key components such as the whole body of the iron tower, and greatly improve the accuracy of target identification and defect detection and the accuracy of data collected in different environments. The invention can rapidly identify and detect the images collected in different areas and find abnormality; the intelligent monitoring is realized for a large amount of returned data, and the workload of an operation team is reduced; and an end-to-end image processing mode is adopted, and the second-level processing can be realized for each image.

Description

Defect identification method for power transmission line

Technical Field

The invention belongs to the technical field of computer vision based on deep learning, and particularly relates to a defect identification method for a power transmission line.

Background

At present, the demand of production and living of residents in China for electric energy resources is increasing. With the rapid development of the scientific and technological technology, the intelligent technology is utilized to solve the problems of high inspection difficulty, high workload, heavy data, difficult safety guarantee and the like of the power transmission line, and the intelligent technology becomes an extremely popular solution at present. Therefore, the image recognition of the power transmission line is realized by utilizing the deep learning algorithm, and the method is particularly important for guaranteeing the daily operation of the power grid.

The rapid development of deep learning in recent years realizes more and more artificial intelligence applications. The neural network can simulate the human brain to analyze and learn, and imitates the mechanism of the human brain to interpret data (such as image data), and forms more abstract high-level characteristics or attribute categories by combining bottom-level characteristics to fit various things in daily life of people.

In daily work, a large amount of image information is collected from the front end, which makes image recognition valuable in business. Four major elements of image recognition technology: data, computing power, algorithms, and scenarios. And the recognition model with excellent performance can be fitted by using enough image data and a reasonable algorithm.

In addition, the traditional end-to-end image recognition algorithm has insufficient performance on the factors of complex targets in the environment, different illumination intensities, different target sizes and the like.

Disclosure of Invention

Aiming at the defect of the performance of a general image recognition algorithm, the invention provides an end-to-end image recognition technology. The identification system can not only accurately monitor the damage of the external force and induce the microclimate of the iron tower, but also detect the defects of key components such as the whole body of the iron tower, and greatly improve the accuracy of target identification and defect detection and the accuracy of data collected in different environments.

The technical scheme of the invention is as follows: the defect identification method for the power transmission line is characterized by comprising the following steps of:

s1, constructing a training data set: acquiring an image containing defect characteristics through a camera, wherein the types of the defect characteristics comprise geranium, bird nests, shrubs under a tower, forest fire, ice coating, pedestrians under the tower, mechanical vehicles, houses, microclimate, insulator burst, vibration damper falling, wire stranding, pin falling and license plate damage; marking each type of target features in the images by using a marking frame, marking by using labels, adjusting the sizes of all the images to 224 x 224 by preprocessing, and forming a training data set by all the images and the labels corresponding to the images;

s2, constructing a neural network model: adopting a resnet-50 model which comprises 5 convolution modules, 1 average pooling layer and 1 full connection layer, and finally classifying through a Softmax classifier; the convolution module performs convolution operation through a convolution kernel of 3 × 3, from the first convolution module to the fifth convolution module, an input image is processed into a characteristic image of 7 × 7, and the neuron nodes all adopt a ReLU activation function;

s3, training the neural network model constructed in the step S2 by adopting the training data set constructed in the step S1, and adopting a mean square error function as a cost function of model training:

wherein n is the number of training samples; the observde represents the output of the model at the t input; predicted represents a predicted value under the same input.

Obtaining a trained neural network model;

and S4, inputting the acquired real-time image into the trained neural network model, thereby obtaining a defect identification result.

The invention has the beneficial effects that: (1) the images collected under different areas can be quickly identified and detected, and abnormity can be found; (2) the identification system realizes intelligent monitoring on a large amount of returned data, and reduces the workload of an operation team; (3) and an end-to-end image processing mode is adopted, and the second-level processing can be realized for each image.

Drawings

FIG. 1 is a logical schematic block diagram of an identification system;

FIG. 2 is a schematic diagram of a part of test indexes- -weather classification test indexes;

Detailed Description

The specific method of the invention comprises the following steps:

and collecting and labeling data. The data includes the following types: specific birds (geranium), bird nests, shrubs under towers, mountain fires, ice, pedestrians under towers, mechanical vehicles, houses, microclimates (snowy days, sunny days, rainy days and foggy days), insulator burst, vibration damper falling, wire stranding, pin falling and license plate damage. All images are collected from the front-end camera. The target category is labeled with a rectangular label box and accompanied by a name.

And (4) preprocessing data. The algorithm adopts a single CNN model to realize end-to-end target detection. Thus, the input of the model requires a resizing. The input image size is adjusted to 224 x 224 during the pre-processing stage.

And (5) designing a network model. The YOLO algorithm employs a convolutional neural network to extract image features. The fully-connected layer is then used to derive a prediction value. The network used in the invention of this patent is the resnet-50 model. The network contains a total of 5 convolution modules, 1 average pooling layer and 1 full connection layer. In the convolution processing stage, a convolution operation is performed using mainly 3 × 3 convolution kernels. From the first convolution module to the fifth convolution module, the input image is processed into a 7 × 7 feature image. The neuron nodes each employ a ReLU activation function (formula 1 below).

F (x) max (0, x) (equation 1)

The activation function is characterized in that the overall characteristics of the target can be expressed by a sparse matrix. But the last layer uses a nonlinear activation function, Softmax regression, where the fully connected layer corresponds to the overall number of identified classes, and the Softmax function is used for the final classifier. The formula is as follows:

softmax (exp (logits)), dim (equation 2)

Wherein, logits represents the output value of a certain node and corresponds to the output value of Softmax; e represents a natural constant; reduce _ sum represents the sum of the input values of the neuron.

And training the recognition model. And adopting a mean square error function as a cost function of the model training for the real state of the model training. The difference between the predicted value and the actual value is mainly expressed in the training process. The formula for this function is as follows:

the identification model adopts small-batch gradient descent, divides data into a plurality of batches, and updates parameters according to the batches, so that a group of data in one batch jointly determines the direction of the gradient. The method balances the training speed and the training precision, so that the algorithm obtains an excellent model on the premise of not needing overlong training time. The initial value of the learning rate set in the project profile is 0.001, and the learning rate is attenuated as the training progresses. The intention is to accelerate the training by using a larger learning rate at the initial stage of the training and to improve the training precision by reducing the learning rate at the later stage of the training.

Claims (1)

1. The defect identification method for the power transmission line is characterized by comprising the following steps of:

s1, constructing a training data set: acquiring an image containing defect characteristics through a camera, wherein the types of the defect characteristics comprise geranium, bird nests, shrubs under a tower, forest fire, ice coating, pedestrians under the tower, mechanical vehicles, houses, microclimate, insulator burst, vibration damper falling, wire stranding, pin falling and license plate damage; marking each type of target features in the images by using a marking frame, marking by using labels, adjusting the sizes of all the images to 224 x 224 by preprocessing, and forming a training data set by all the images and the labels corresponding to the images;

s2, constructing a neural network model: adopting a resnet-50 model which comprises 5 convolution modules, 1 average pooling layer and 1 full connection layer, and finally classifying through a Softmax classifier; the convolution module performs convolution operation through a convolution kernel of 3 × 3, from the first convolution module to the fifth convolution module, an input image is processed into a characteristic image of 7 × 7, and the neuron nodes all adopt a ReLU activation function;

s3, training the neural network model constructed in the step S2 by adopting the training data set constructed in the step S1, and adopting a mean square error function as a cost function of model training:

wherein n is the number of training samples; the observde represents the output of the model at the t input; predicted represents a predicted value under the same input.

Obtaining a trained neural network model;

and S4, inputting the acquired real-time image into the trained neural network model, thereby obtaining a defect identification result.

Images