Research on Typical Defect Identification Technology of Composite Insulators for Ultra High Voltage Transmission Lines Based on Spectral Feature Extraction

Support vector machine is a machine learning model with optimization method based on statistical learning with good theoretical foundation [29–30]. Support vector machine is different from back propagation neural network model, not to minimize the learning error, but to minimize the confidence level with the learning error as a constraint. Support vector machine can be understood as a two-layer neural network with weights and number of nodes calculated by statistical learning direction. The core idea of a support vector machine is to construct a hyperplane so that the training set data is as far away from the hyperplane as possible, dividing the training data space into partitions that are uniform on either side. Support vector machines can be used for almost any type of learning task, including classification of genetic data, text categorization, image recognition, and so on.

Support vector machines use hyperplanes to partition data into similar groups. In both dimensions, the task of the support vector machine algorithm is to identify the two groups and search for the maximum edge hyperplane so that the hyperplane produces the maximum separation between the two groups. The goal of the support vector machine is to train a model that assigns each data to a specific class, and this is achieved mainly by creating partitions of the feature space into two subspaces. The advantage of support vector machines is that the partition plane does not actually need to be a linear hyperplane, and by introducing a kernel function, it can be used with various types of nonlinear decision situations. Support vector machines construct linearly separating hyperplanes in a multidimensional vector space, and the best classification is when the distance between the hyperplane and the training data is maximized.

The main motivation of Support Vector Machines is to separate several classes of the training set with a surface that maximizes the boundary between them, i.e., to delimit the dichotomous data by constructing a hyperplane.In addition to this, based on the Structural Risk Minimization Principle (SRM), SVMs emphasize on minimizing the generalization error instead of making the empirical error of the training set minimal.2 Therefore, trained SVMs have some generalization ability when faced with a yet unknown dataset, it has some generalization ability, which is where its strength lies.

The segmentation hyperplane is shown in Fig. 1. Given data sets D, D:={ (xi,yi) }i=1n , x_i ∈ ▯ ^d, y_i ∈ {+1, –1}. Where x_i is the feature vector of a particular input column and y_i is the label/category associated with it, and assuming that the data is two-dimensional and linearly divisible, multiple hyperplanes dividing the sample points can be obtained. Figure 1 shows four hyperplanes that can completely divide the input dataset, and it is clear that there are infinitely many hyperplanes that can divide the dataset. However, it is necessary to find an optimal hyperplane that “splits” the sample points so that the distance between the hyperplane and the closest sample point is maximized to improve the generalization ability of the model. This ideal optimal hyperplane is defined as w^Tx + b. Using geometry, the distance from the sample points to the hyperplane can be called the geometric interval, and to optimize the geometric interval, all that is needed is to find the optimal hyperplane, which is also called the Canonical hyperplane.

Figure 1.

Partition hyperplane

The optimal hyperplane is shown in Fig. 2 and the following two hyperplanes are found separately: 1wTx++b=1wTx−1+b=−1

Figure 2.

Optimal hyperplane

Combining the two equations gives ∀i: 2yi(wTxi+b)≥1

Then, the geometric interval x⁺yx^– is: 3γi=12(〈 w‖w‖·x+ 〉−〈 w‖w‖·x− 〉)=12‖w‖(〈 w·x+ 〉−〈 w·x− 〉)=1‖w‖ where w is the weight vector and b is the bias. Optimising this geometric interval requires minimising the norm of the weight vector, and H is the optimal hyperplane. H₁ and H₂ are the boundary hyperplanes of positive and negative sample points, respectively, and are parallel. There are no sample points between the two hyperplanes, and when the distance between the two hyperplanes is maximised, some sample points will exceed H₁ and H₂, which are called support vectors. Because they are directly involved in the definition of the dividing hyperplane, while other sample points do not affect the generalisation performance of the classifier.

Then, the following optimisation problem is obtained: 4{min12∥w2∥+C∑i=1nξi s.t. yi(〈w·xi〉+b)≥1−ξi,i=1,…,nξi≥0,i=1,…,n

Where, C is the regularisation parameter and ξ_i is the slack variable, there are two ways to solve problem (1) in general, one is to solve it directly and the other is to solve it indirectly by using the Lagrange multiplier method, the former is affected by constraints and the solution process is more complicated. The latter facilitates the problem solving by transforming the original problem into a dyadic problem. Both methods can find the optimal solution. In order to avoid large matrix computation and improve computational efficiency, this paper tries to solve the problem directly. Instead, the chosen approach is to transform it into an unconstrained form using the outer penalty function method.

The NIR spectra were collected in the laboratory, where the indoor temperature and humidity were basically constant, the ambient temperature was controlled at (20±1)°C and the average relative humidity was 50%. As the 1000nm∼1600nm spectrum carries important information of wood, it can better predict the properties of wood surface defects. Therefore, the One-chip micro-integrated fibre optic spectrometer produced by INSION (Germany) was selected for the test to perfectly achieve the collection of near-infrared spectral data of surface defects on composite insulators of ultra-high-voltage transmission lines. The spectrometer uses a bifurcated fibre optic probe to collect the NIR spectra of the sample surface [31]. The fibre optic probes were placed into the grooves of the mobile unit, and the probes performed vertical, non-contact measurements on 400 defective samples as well as 100 defect-free samples, and the spectral data were acquired as shown in Figure 3. The scanning was calibrated with a commercial PTFE white board, and then the samples were subjected to NIR spectral acquisition in the range of 900 nm to 1800 nm, and multiple measurements were performed for each sample and each defect, and the spectra obtained from multiple measurements were averaged into 1 spectrum, which represents the defective NIR absorption spectra of the specimens.

Figure 3.

Spectral data acquisition

Data preprocessing is a key step in spectral data analysis, aiming to improve the data quality and the reliability of the analysis results. Firstly, noise in the spectral signal is eliminated by denoising algorithms, such as wavelet transform and Savitzky-Golay filter. Secondly, normalisation and standardisation methods, such as max-min normalisation and Z-score normalisation, are applied to ensure the comparability of different spectral data. In addition, the effects of sample particle size and surface properties are eliminated based on scattering correction techniques (e.g., multiple scattering correction MSC and standard normal transform SNV). Finally, key features are extracted by feature selection and dimensionality reduction methods (e.g. Principal Component Analysis PCA and Independent Component Analysis ICA) to reduce data dimensionality and improve computational efficiency. These steps provide high-quality input data for subsequent training of composite insulator identification for UHV transmission lines.

Now the HOG features are introduced into the UHV line insulator defect recognition, mainly for insulator defect recognition.The main process of HOG features is: 1)

Image normalisation. The input image is normalised using the Gamma correction method, which can effectively reduce the effects of local shadows and illumination, Gamma as shown in the compression equation (5): 5I′(x,y)=I(x,y)gamma

In complex scenes where the contour features are not obvious, using a single HOG feature may affect the recognition accuracy, so we introduce LBP local texture features to compensate for the possible effects.

The original LBP algorithm only focuses on the correlation between the centre pixel and the neighbouring pixels, and only compares the size of the centre pixel grey value with the neighbouring pixel grey value in the window, and does not take into account the relationship between the neighbouring pixels, which inevitably leads to the loss of some local feature information, affecting the final classification and recognition. This problem may lead to the existence of the same LBP coding value for different pixels, the LBP coding is shown in Fig. 4, (Fig. a centre pixel value is 210, Fig. b centre pixel value is 17), in comparison, the pixels in (Fig. a) are a little brighter than those in (Fig. b), however, the appearance of the same LBP coding value, which will classify pixels that are relatively brighter into one category, and it is easy to make mistakes in the identification and classification of the pixels.

Figure 4.

LBP coding

To solve this problem we improve the LBP algorithm, firstly the central window pixel g_c value and the neighbourhood pixel value g_i are differed and the result is taken as an absolute value, and then the average of all the absolute values is taken as the threshold value M, as shown in the formula (11) below: 11M=1p∑i=0p−1| gi−gc |

Then the window neighbourhood pixel value g_p and the centre pixel point g_c are subtracted clockwise and the absolute value is taken, and then the result is compared with the threshold value M, and if the result is greater than 0, it is coded as 1, and if the result is less than 0, it is coded as 1. The process of improving the LBP coding is shown in Fig. 5, where the threshold value M is 16, and the binary coding is 0011000. the equation is shown in (12): 12S(gp−gc)={ 1| gp−gc |≥M0| gp−gc |<M

Figure 5.

Improved LBP coding process

The feature serial fusion process is shown in Fig. 6, the HOG features can describe the contour information of the object very well, but the object also has rich texture information, in this regard, we give a way to fuse the HOG features and the improved LBP features, which can obtain the gradient information of the object as well as the texture information of the object, and is used to improve the correct rate of fault detection. The methods of feature fusion are weighted fusion, parallel fusion and serial fusion. In this paper, HOG features and improved LBP features are fused in series mainly by serial fusion method, which has the advantage of protecting all the feature information and can improve the accuracy of detection.

Figure 6.

Feature series fusion process

In feature fusion. Firstly HOG and improved LBP features are extracted separately and then the two feature histograms are concatenated. In the procedure feature histograms are represented by vectors, so fusing two features is actually summing their feature vectors.

The experimental software environment for this paper: Windows10 (64-bit) system, opencv2.4.3, Matlab2014, visualstudio2008. In order to simulate a more realistic line environment, the Chongqing Electric Power Research Institute (CQEPRI) in Chongqing Panxi substation to build a ground-based simulation of the transmission environment. The ground simulation environment is arranged according to the structure of the actual high-voltage transmission line, the simulated transmission line is the same as the real transmission conductor, the total length is about 20 m, the diameter of the transmission line is about 27 mm, and the fixtures on the line are the same as those on the real line, which are mainly: spacer, tensioning clamps and insulators. In this ground simulation environment, the inspection robot can simulate the movement on the high-voltage transmission line.

In this paper, the experiments were carried out to collect the research dataset using spectroscopic technique, and the dataset was divided into training sample set, and test sample set according to the ratio of 2:8. The samples contain four types of fixtures: spacers, vibration-proof hammers, overhanging wire clamps and insulators. Each type of sample contains 1000 training samples, 130 test samples, and the sample image resolution is 256 × 256.

The HOG features of vibration-proof hammers, spacers, insulators and tension clamps are extracted respectively, and the LBP local texture features are introduced on the basis of the HOG features to obtain the LBP-HOG features, and the HOG features and the dimensionally reduced LBP-HOG features are inputted into the six classifiers corresponding to the six combinations of the gold tools respectively, and the recognition accuracy of each classifier corresponding to each feature is calculated respectively, and a total of 12 A total of 12 sets of experiments were conducted. Classification accuracy and algorithm operation speed are the focus of this paper, therefore, the classification accuracy S, the number of support vectors N, the classifier training time t-train, and the classification test time t-test of the six classifiers with HOG features and LBP-HOG features as inputs respectively are analysed, and the classification results of the 12 groups of experiments are given in Table 1, where A, B, C, and D are represented respectively by insulators, overhanging wire clips, vibration-proof hammers and spacers, for SVM classifier training, there are six combination forms of the four types of samples: AB, AC, AD, BC, BD, CD. The larger the difference between the positive and negative sample features of the classifier, the more obvious the classification effect. The detection accuracy using LBP-HOG features is similar to that using HOG features, but the number of support vectors for the classifier is less when LBP-HOG features are used as inputs, and the classifier training speed and testing speed are faster. Theoretically the training complexity of SVM is only affected by the number of support vectors in the sample, independent of the dimension of the sample features, but the increase in the type of input samples and the number of features may lead to an increase in the number of support vectors. The resolution of the experimental samples is 256*256, so the extracted HOG features are 34596 dimensions, there are 1000 training samples of one class, and there are a total of 2000 positive and negative samples when SVM training is performed, and the total dimension of the input sample features is: 34596*2000=69,192,000. The LBP-HOG features are 12475 dimensions, and the total dimension of the experimental LBP-HOG as the the total dimension of the input feature is 24498000, and the dimension of the feature is greatly reduced by using LBP-HOG.

While the expressive ability of LBP-HOG features is affected by external factors such as light, its own block size as well as the size of the unit also affects the final recognition accuracy, so experiments have been conducted to address this issue. The image size in the image dataset is uniformly 32*32, so the detection window has been fixed to 32*32. The number of gradient direction divisions n is 9, and the recognition accuracy under different block and cell sizes is shown in Table 2. The block area will move on the detection window with a step distance, when the step size is smaller than the block size, there will be an area overlap before and after the movement, and vice versa, there will not, so there will be feature information redundancy when overlap occurs, resulting in an increase in the vector dimensions, which also affects the recognition effect. As can be seen from the table, the recognition accuracy is highest when the block size is 32*32, the cell size is 16*16 and the step size is 16. This is because the size of the detection window is 32*32, the block size is one-half of its position, and it moves with a step size of 8, which is more conducive to extracting the features about the insulator contour, and the LBP-HOG features obtained are more expressive, with a higher recognition accuracy.

In order to better quantify the performance of the model in this paper, it was decided to use the recall and precision expressed in equation (13), average precision (AP) expressed in equation (14), and average precision mean expressed in equation (15), respectively, to reflect the effectiveness of the model on the composite insulators of the ultra-high-voltage transmission lines, and the higher the value of AP expresses the higher the model recognition accuracy. Where, TP represents the number of defective insulators correctly identified, FP represents the number of defective insulators mistakenly detected, FN represents the number of defective insulators missed, B represents the type of detection, O represents the number of types AP value also represents the area under the curve of RP, and F1 represents the reconciled average of P and R in Equation (16), and MacroF1 represents the average of all F1 values. Namely: 13R=TPTP+FN,P=TPTP+FP 14AP=∑ (Ri−Ri−1)Pi 15mAP=1O∑BAPB 16F1=2P·RP+R, MancroF1=∑ F1O

In order to verify the feasibility of the insulator defect recognition model, 100 insulator pictures are randomly selected for testing, with a total of 271 insulator strings, of which 123 strings are glass-type and 148 strings are composite-type, and a comparison of the recognition effect of the two models in the same scene is shown in Table 3, in which the recognition model in this paper is LBP-HOG-SVM, and the control recognition model is Faster R-CNN. In this paper's model, the recall and accuracy of defects of glass-type and composite insulators can reach more than 90%, which is significantly improved compared with the original Faster R-CNN model. The recall and precision of this model are 9.76% and 9.2% higher than the original model for glass insulators, and 11.49% and 13.06% higher than the original model for composite insulators. It can be seen that the effect of this paper's model under the glass-type and composite insulator datasets is better than that of Faster R-CNN, which can effectively identify the defects of the insulators of UHV lines.

In order to test the strengths and weaknesses of the recognition performance of this paper's model, a comparison test is carried out with a variety of mainstream recognition algorithms, including YOLOV2, SSD300, R-FCN, Faster R-CNN.Table 4 gives the results of this paper's self-constructed sample data for insulator defect recognition. From the data comparison in the table, it can be seen that the mAP value and MacroF1 value of this paper are better than the above four algorithms, in which the YOLOV2 model is close to the algorithm of this paper in all the values, the AP value of glass-type insulators exceeds that of this paper's model by 0.69%, but the composite insulators are lower than that of this paper by 1.62%.The YOLOV2 model, although it performs well, has the ability to detect defects of some of the occluded and scale-gap insulators is not as good as the algorithm in this paper. Comprehensive data in Table 4 shows that the overall performance of this model in insulator defect recognition is optimal.

In order to compare this paper with the current mainstream model more intuitively, PR curves are depicted to compare the performance of different algorithms, and when the curve with the largest coverage area of the coordinate axis, it can indicate the superior performance of the algorithm, PR curves are shown in Fig. 7, in which (a) ∼ (b) are composite insulators, glass insulators, respectively. From Fig. 7, it can be seen that all the models in this paper are superior to other models, indicating that the improvement of this paper enhances the algorithm's ability to recognise insulators of different scales and partial shading.

Figure 7.

Precision-Recall curve

Hyperspectral data of insulator samples were extracted as sample data for model application analysis, and during the process of collecting sample data for the training set, the hyperspectral data were collected uniformly on the insulator discs as much as possible in order to reduce the data collection error because the data in the same region have high similarity. Fifteen spectral data were extracted from each insulator hyperspectral image as hyperspectral samples for that insulator, totalling 180 sets of hyperspectral samples. Twelve spectral lines are randomly extracted under each defect level as test set spectral data, and the remaining spectral data are used as training set data for detection model building, with 48 sets of test set samples and containing four insulator defects.

Since the model transfer method based on direct standardisation is a model transfer method with standardised samples, the number of samples in the standard set as well as the type of samples is too small will affect the effect of the model transfer, which means that it is difficult to improve the detection accuracy of the model transfer on the basis of the original with the current limited standard set of samples. Therefore, in this study, the degree of insulator defects is detected by combining different types of feature information through information fusion on the basis of model transfer. The insulator defect degree recognition results are shown in Fig. 8, where Fig. 8(a) shows the traditional recognition model and Fig. 8(b) shows the recognition model of this paper, and the values of 1∼4 on the vertical axis indicate four types of insulator defects, while the horizontal axis indicates 48 test samples (as explained above). It can be clearly seen that the traditional recognition model incorrectly recognises on 2, 4, 5, 6, 14, 16, 19, 23, 28, 31, 35, 39, 42, 43, 46, 47, 31 samples, while on the remaining 1, 3, 7, 8, 9, 10, 11, 12, 13, 15, 17, 18, 20, 21, 22, 24, 25, 26, 27, 29, 30, 32, 33, 34, 36, 37, 38, 40, 41, 44, 44 samples are successfully identified, then the identification rate of insulator defects by the conventional identification model is 64.58% (31 ÷ 48 = 64.58%). However, the model in this paper identifies errors in samples 9, 21, 32 and 46, and identifies correctly in the remaining 44 samples, and its insulator defect recognition rate of this model is 91.67% (44 ÷ 48 = 91.67%), which results in the recognition rate of the model in this paper does not reach more than 95.00% due to the sample data is too small, and also indicates that the fusion LBP-HOG features will be put into the SVM classifiers to be training, which effectively improves the recognition effect of insulator defect degree and plays an important role in power equipment maintenance and monitoring.

Figure 8.

Identification results of the defect degree