Open Access
Add to BibSonomy
Abstract
The acidic environment is one of the main factors leading to the aging of silicone rubber (SiR) insulators. Aging can reduce the surface hydrophobicity and pollution flashover resistance of insulators, threatening the safe and stable operation of the power grid. Therefore, evaluating the aging state of insulators is essential to prevent flashover accidents on the transmission line. This paper is based on an optical hyperspectral imaging (HSI) technology for pixel-level assessment of insulator aging status. Firstly, the SiR samples were artificially aged in three typical acidic solutions with different concentrations of HNO3, H2SO4, and HCl, and six aging grades of SiR samples were prepared. The HSI of SiR at each aging grade was extracted using a hyperspectral imager. To reduce the calculation complexity and eliminate the interference of useless information in the band, this paper proposes a joint random forest- principal component analysis (RF-PCA) dimensionality reduction method to reduce the original 256-dimensional hyperspectral data to 7 dimensions. Finally, to capture local features in hyperspectral images more effectively and retain the most significant information of the spectral lines, a convolutional neural network (CNN) was used to build a classification model for pixel-level assessment of the SiR's aging state of and visual prediction of insulators’ defects. The research method in this paper provides an important guarantee for the timely detection of safety hazards in the power grid.
© 2024 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Silicone rubber (SiR) insulators play an important role in electrical insulation and mechanical support in power transmission lines due to their lightweight, high strength, and resistance to pollution flashover [1]. Currently, the count of operational SiR insulators in China has surpassed ten million [2]. However, with the increase in severe weather such as haze, the corona discharge on the humid polymer surfaces can generate HNO3. Insulators exposed to high-humidity outdoor environments are susceptible to corrosion from acidic pollutants in the atmosphere. Consequently, SiR insulators undergo gradual aging, with pores and micro-cracks appearing on the surface [3]. As the insulators age, the infiltration of moisture and impurities from the air becomes more prevalent, leading to an increase in leakage current and a decrease in insulation performance. This condition poses a significant risk of flashover accidents along the power lines, which can have a substantial impact on the stable operation of the power grid and result in substantial economic losses [4]. Hence, conducting assessments of the aging condition of insulators in acidic environments is crucial to ensure the safe and reliable operation of the power grid.
Several detection methods have been used to assess the aging state of insulators, including the hydrophobicity method [5], the leakage current method [6–7], the infrared imaging method [8–9], the ultraviolet imaging method [10] and electron scanning microscope method [11]. The hydrophobicity method can assess the insulator surfaces’ hydrophobic properties, but it cannot capture potential defects in insulators. The leakage current method can reflect the overall insulation performance of the insulator by the current value, but cannot determine the specific location of the defect. Infrared imaging and ultraviolet imaging techniques are primarily utilized for detecting already damaged insulators and may not accurately assess potential defects beforehand. Furthermore, the electron scanning microscope method can reflect the microscopic condition of the material surface, However, it is expensive and not suitable for field testing. In contrast, HSI technology is a more accurate and convenient means of aging assessment, which can transform insulators’ mapping features into spectral data of multiple wavelengths. By combining HSI with deep learning algorithms, pixel-level assessment of composite insulators’ aging status in acidic environments can be achieved.
HSI technology was first widely used in the field of remote sensing and subsequently expanded into various domains, including agriculture, food safety, and military applications [12]. Currently, Yin et al. used the HSI technique combined with the partial least squares regression (PLSR) model to achieve the detection of insulator surface pollution degree and pollution distribution, the error rate between the predicted and actual values of equivalent salt deposition density (ESDD) is less than 10% [13]. Additionally, Hao et al. employed convolutional neural networks (CNN) for spectral-spatial HSI feature extraction and classification of in vivo human brain hyperspectral images, aiming to achieve accurate identification of glioblastoma multiforme (GBM) tumors. The overall accuracy of this method reached 96.69% [14]. Furthermore, to accurately identify corn varieties, Zhang et al. utilized HSI technology to extract spectral features from the region of interest on the embryo surface of corn seeds. Combining this with a five-layer CNN for feature recognition, they established nine corn variety identification models, achieving an accuracy exceeding 96% [15].
Previous studies have indicated that HSI techniques have been applied to external insulation, However, the majority of these applications focus on detecting the degree and components of pollution present on insulator surfaces. Further studies are lacking in the assessment of the insulators’ aging state, as the process is more complex than insulator pollution detection. This paper utilizes the static contact angle method to establish a correlation between hyperspectral spectral lines and the aging levels of insulators, enabling the study of the aging state of SiR insulators. Additionally, there is also ample room for improving and optimizing existing research on HSI technology. Usually, it is necessary to reduce the computational complexity by dimensionality reduction of hyperspectral data, while most of the researchers use PCA and ICA dimensionality reduction methods [16,17]. While these methods are effective in reducing the dimensionality of high-dimensional data [18], it is worth noting that they may still retain irrelevant information from the original data in the dimensionality reduction results. To address this limitation, this paper proposes a novel joint RF-PCA dimensionality reduction method, which integrates the importance and correlation of each feature to achieve the maximum utilization of hyperspectral data while reducing the risk of overfitting.
To sum up, this paper uses an optimized HSI technique combined with a deep learning algorithm to evaluate the aging state of insulators in acidic environments. Specifically, six aging grades of SiR samples were prepared by using three typical acidic solutions of HNO3, H2SO4, and HCl. The hyperspectral spectra of these samples with different aging grades were extracted. To reduce the dimensionality of the hyperspectral data, a joint RF-PCA dimensionality reduction method is proposed in this paper to obtain the effective features of the spectral lines. Finally, the dimensionality-reduced hyperspectral data are divided into training and testing sets according to different ratios, a CNN classification model is established to evaluate the aging state of SiR samples, and the method is applied to the visualization and prediction of the aging state of composite insulators.
2. Sample preparation and aging degree classification
2.1 Aging experimental platform and sample preparation
By IEC 61109-2008 standards [19], an Artificial accelerated aging experimental platform is built in this paper, as shown in Fig. 1. In this experiment, three typical acidic solutions, HNO3, H2SO4, and HCl, were utilized to create accelerated aging environments for the samples. The concentrations of the acidic solutions were set at 1 mol/L and 0.1 mol/L, respectively, while ionized water was employed as the control group. Deionized water and solutions of various concentrations of HNO3, H2SO4, and HCl were placed in a jar with a diameter of 100 mm at the bottom and a height of 130 mm, respectively. The SiR samples were cut into 30 mm × 30 mm, washed through anhydrous ethanol air dried, and then clamped by forceps into different solutions for immersion. The soaking time for the SiR samples was designed as 24, 48, 72, 96, 120, and 168 hours. A total of seven sets of SiR samples were prepared for each type of solution. Among them, Fig. 2 illustrates the surface conditions of the SiR samples after different immersion times in the HNO3 solution.
In addition, this paper followed the IEC 60507-2013 [20] standard to brush the composite insulator sheds with an acidic solution, as shown in Fig. 3(a). Different types of acidic solutions were manually brushed onto the insulator sheds multiple times during the experimental process (three times per day, for one week) This was done to simulate the aging environment that insulators experience during actual operation. The acidic liquid flows and drips on the surface of the insulator sheds, resulting in different surface conditions. The outcomes of these interactions are depicted in Fig. 3(b).
2.2 Aging degree classification criteria
In this study, the static contact angle was selected as the criterion for classifying the aging degree of SiR samples due to its simplicity and intuitive nature. During the preparation stage, a five-point sampling method was utilized to measure the static contact angles of the samples at different aging times, meanwhile according to IEC/TS 62073-2016 [21], the static contact angle of SiR samples in this experiment was classified into six aging classes. The results of this classification are presented in Fig. 4.
Fig. 4. Static contact angle classification results of SiR sheets. (a) Immersion in 1 mol/L acidic solution. (b) Fluctuation range. (c) Immersion in 0.1 mol/L acidic solution. (d) Fluctuation range.
The results indicate that when subjected to the same acidic solution, SiR samples experience relatively small variations in surface static contact angle errors across different aging times. The average values obtained through the five-point measurement method can be effectively utilized as the basis for the classification of aging levels in SiR samples. The deionized water has minimal aging effects on the SiR samples. Even after immersion for 168 hours, the static contact angle of SiR shows almost no change, maintaining a high hydrophobicity. In contrast, the outcomes of the acidic solution corrosion reveal that the aging impact of a 1 mol/L HNO3 solution on SiR is the most severe, causing the static contact angle to drop below 90° after 168 hours of aging. Conversely, the static contact angle of the sample aged for 168 hours under the 0.1 mol/L acidic environment remains above 100°, comparable to the effect of aging for 96 hours in the 1 mol/L acidic solution. This implies that the aging degree of the samples exposed to the 0.1 mol/L acidic solution is much lower than that of the samples in the 1 mol/L acidic solution.
To further observe the surface condition alteration of SiR samples under an acidic environment, scanning electron microscope analysis and 3D profile scanning were conducted on the samples after soaking in 1 mol/L HNO3 solution respectively. The SEM results show that after 24 hours of immersion in HNO3, the SiR samples exhibited slight surface distortion. However, the overall surface appeared relatively flat with no prominent pores, as shown in Fig. 5 (a). However, after 168 hours of immersion, the SiR samples exhibited severe surface pores, along with a few cracks, a noticeable increase in the number of pits, and significant distortion in the surface morphology, as shown in Fig. 5 (b).
The 3D profilometer employs automatic focusing on the surface of SiR and the identification of interference fringes. Subsequently, upon the adjustment of interference fringe width, scanning measurements are initiated. Following the completion of the scanning process, a 3D image of the sample is generated, as shown in Fig. 6. The results of the 3D profile scanning show that the cratered parts below the sample plane appeared as dark blue to green areas, while the raised parts were represented by yellow and red colors. The original samples exhibited almost indistinguishable craters and raised parts on their surfaces. However, the samples that were soaked in HNO3 solution for 168 hours exhibited distinct graded features in their surface contours. In addition to visualizing the surface morphology of the sample using three-dimensional profiles, the aging status of the sample can also be quantitatively analyzed through surface roughness measurements. Based on the height data acquired from the surface profiler, the arithmetic mean roughness (Sa) and root mean square profile (Sq) of the samples were calculated following the ISO 25178-2:2012 standard [22]. These parameters were used to quantify the roughness of the measured surface area. The calculation formulas are respectively as follows:
where Z(x,y) is the height data of the sample and A is the area of the sampling area.In Fig. 6, the Sa and Sq of the original sample were measured as 39.51 nm and 56.28 nm, respectively. However, after immersing the samples in an HNO3 solution for 168 hours, the Sa and Sq values increased to 376.79 nm and 511.82 nm, respectively. This indicates a significant increase in surface roughness, accompanied by the formation of noticeable pits on the sample surface.
3. Hyperspectral data extraction and pre-processing
3.1 Hyperspectral data extraction
HSI technology relies on capturing and analyzing the variations in reflection and scattering properties of different objects when illuminated with light, enabling precise measurements of the objects’ characteristics [23,24]. Take the SiR sheet tested in this paper as an example, as shown in Fig. 7. Hyperspectral data of the sample were extracted using a hyperspectral imager, capturing information across 256 bands ranging from visible to near-infrared (400 nm to 1040 nm). Due to the different wavelength-reflectance curves of samples with different aging states, the subsequent pre-processing of data and the establishment of classification models can be achieved to evaluate the aging state of SiR sheets.
Based on the principle of HSI technology, SiR samples with different aging times were placed on the hyperspectral experimental platform for experimentation, as illustrated in Fig. 8. The experimental platform consists of a hyperspectral imager, two 800W halogen lamp that simulates sunlight sources, a correction whiteboard, and a computer. The hyperspectral imager is of the Dualix Spectral Imaging brand, with the model GaiaField-F-V10, featuring a spectral resolution of 3.8 nm and a spectral range from 400 to 1040 nm (256 bands). During the extraction of hyperspectral spectral data, the distance between the lens of the hyperspectral imager and the sample was maintained at 40 cm, and the distance between the lamp and the samples was 50 cm. The halogen lamp was positioned obliquely downward at a 45° angle of symmetry.
The image capture device is susceptible to interference from environmental factors such as dark currents and noise. The flat field correction can discount the effects of spatial non-homogeneities in the illumination, eliminating issues of image non-uniformity caused by variations in sensor pixel sensitivity and enhancing the image quality. The original hyperspectral spectra of the samples are illustrated in Fig. 9. In this study, spectral data in the range of 400∼900 nm, characterized by a high signal-to-noise ratio in the hyperspectral results, were selected for subsequent preprocessing steps.
3.2 Black and white correction
Black and white correction is a method that utilizes the reflectance or transmittance of a standard reference material. Its purpose is to eliminate regional differences and enhance the accuracy of data, enabling meaningful comparison and analysis of data from different regions [25]. The formula for black and white correction is:
where Xc represents the reflectance of the SiR sheet after the hyperspectral imager completes black and white correction, X0 represents the original reflectance of the SiR sheet, XW represents the reflectance intensity of the whiteboard, where the reflectance is close to 1, and XB represents the reflectance intensity of the standard black. In this study, the black and white correction processing of the hyperspectral spectral data was performed using ENVI software. The results of this correction are shown in Fig. 10, where each color represents each aging level of the SiR samples.3.3 Standard normal variate transformation and Savitzky-Golay convolution
The standard normal variate transformation (SNV) method is used in this study to manipulate the spectral matrix. This method involves processing each spectral line to mitigate the impact of surface scattering and optical path differences on diffuse reflectance spectra. By eliminating spectral interference and instrument errors in the spectral data, SNV ensures consistent spectral intensity scales across different spectra [26]. The calculation process for variable standardization is as follows:
- (1) Calculate the mean of spectral data $\bar{x}$ for all samples:
- (2) Obtain the standard deviation s of the spectral data vector of the sample:
- (3) Standard variation of the spectral data of the samples:
The Savitzky-Golay (SG) convolution algorithm is a widely used digital filtering technique for signal smoothing. It operates by fitting a polynomial function to the signal, enabling the removal of high-frequency noise components while preserving the overall trend of the signal. One of the key advantages of the SG algorithm is its ability to strike a balance between noise removal and signal fidelity by adjusting the degree of smoothing applied. This adaptability makes it suitable for various applications where signal smoothing is required while maintaining the data's essential features [27].
The results of applying the SNV and SG algorithms to the hyperspectral spectra of six aging levels of SiR, along with the results of averaging the hyperspectral spectral lines, are shown in Fig. 11. The results suggest that the hyperspectral spectra of the SiR samples, following corrosion by acidic solutions, exhibit a consistent overall trend. These spectra effectively capture the reflection and scattering characteristics of the samples at different aging stages. Samples with higher degrees of aging exhibit a reduced fluctuation in surface reflectance, approximately 35.7% less compared to samples with lower aging degrees. This reduction in reflectance fluctuation is primarily attributed to surface defects such as pits and pores that become more pronounced in severely aged samples. These defect features significantly influence the reflectance captured by HSI. Compared to the original spectra, the spectra subjected to SNV and SG preprocessing show smoothing effects and noise reduction, thereby diminishing the presence of spectral artifacts commonly referred to as “ burr.” However, due to the inherently noisy nature of the initial data, the results of the correction are not optimal.
3.4 Dimension reduction
In this study, random forest (RF) is employed as a nonlinear method for the extraction of feature wavelength. Decision trees are constructed by randomly sampling the dataset and selecting features. Subsequently, these trees are utilized to create a Bagging ensemble. The importance of features is computed based on the average change in out-of-bag data errors of the random forest. This process generates sorted output, reflecting the importance of features [28]. On the other hand, principal component analysis (PCA) is a linear technique for reducing dimensionality. It transforms data from the original feature space to a new coordinate system known as the principal component space, which maximizes the variance of the data in this new system [29]. By ranking the principal components and retaining those with higher variances, PCA effectively reduces the dimensionality of the data. To enhance the dimensionality reduction effect of hyperspectral data and mitigate the interference caused by irrelevant and redundant information during the feature extraction process, an RF-PCA method is proposed in this paper. By integrating RF and PCA, the limitations of each technique are addressed, leading to an improved analysis of hyperspectral data.
The feature wavelength of HSI was extracted using the RF algorithm. Based on the importance of each wavelength, the top 30 were selected as feature wavelengths, as presented in Table 1. The results indicate that the feature wavelength primarily lies within the wavelength range of 650 nm to 750 nm. The importance of the top 30 feature wavelengths is illustrated in Fig. 12. Among them, the most critical wavelengths are situated at 709.2 nm and 706.8 nm, with importance values surpassing 0.5. The feature wavelengths are extracted from the redundant data of HSI using the RF algorithm to minimize the interference caused by noise.
Table 1. Characteristic wavelengths
The PCA was used to downscale the principal components of the feature bands. The contributions associated with the principal components were linearly weighted using the corresponding weights, resulting in the accumulation of principal component contributions, as shown in Fig. 13. The findings indicate that the contributions of the first 7 principal components amount to over 90%. Therefore, by selecting these 7 principal components, the majority of the variance of the data after the dimensionality reduction performed by RF can be retained.
The study employs a method for visualizing high-dimensional data, namely the t-distributed stochastic neighbor embedding (t-SNE) method [30]. Figure 14 presents two-dimensional clustering distributions for the original data, data after PCA dimensionality reduction, and data after RF-PCA dimensionality reduction. The results indicate that the clustering performance of the original data is moderate, with significant overlap between aging levels 1-3 and 5-6. While the results are somewhat improved after PCA dimensionality reduction, substantial overlap persists between aging levels 5 and 6. In contrast, the RF-PCA dimensionality reduction method exhibits superior classification performance, primarily due to its ability to address inherent data redundancy in the original data. After applying RF-PCA dimensionality reduction, the clustering result effectively separates sample points related to aging levels 1 to 4, and only a few data points exhibit overlap between aging levels 5 and 6.
Fig. 14. t-SNE results of SiR sheets. (a) t-SNE result for origin data. (b) t-SNE result for PCA. (c) t-SNE result for RF-PCA.
4. Aging state evaluation model and result analysis
4.1 Deep learning models
Convolutional neural network (CNN) is a common deep learning algorithm extensively employed in the field of image processing. CNN achieves feature extraction, abstraction, and classification of input data through the successive stacking and connection of components such as convolutional layers, pooling layers, and fully connected layers. The convolutional layer primarily serves the purpose of extracting local features from input data [31]. The pooling layer is employed to reduce the number of parameters and mitigate the issue of overfitting, with commonly utilized pooling operations including maximum pooling and average pooling. The fully connected layer represents the final layer of the CNN architecture, and it is utilized for the tasks of classification or regression. Additionally, incorporating activation functions such as Sigmoid, Tanh, ReLU, and others between the convolutional and fully connected layers enables the non-linear mapping capabilities that linear models cannot handle [32].
The one-dimensional CNN architecture utilized in this study is depicted in Fig. 15. The architecture primarily consists of an input layer, four convolutional layers, four max-pooling layers, and a fully connected layer. A total of 300 samples (50 samples per aging level) were extracted from the region of interest and partitioned into training and testing sets using different ratios for inputting into the CNN model. After each convolutional layer, a ReLU activation function and a batch normalization are applied. The ReLU activation function introduces non-linear mapping capabilities, enabling the neural network to learn more complex functions. Batch normalization normalizes the data across each feature channel, improving the convergence speed of the neural network. The parameters of the CNN model utilized in this study are shown in Table 2. The model comprises four convolutional kernels with sizes of 16, 32, 64, and 128, each with a dimension of 1 × 1 and a stride of 1. Following batch normalization and ReLU activation, the output is fed into a max pooling layer with a size of 2 × 1. The final max pooling layer has a size of 128 × 2. The resulting features are then transmitted to the fully connected layer to classify different SiR aging levels.
Table 2. CNN model parameters setting
The hyperspectral data was divided into training and testing sets using ratios of 5:5, 6:4, 7:3, and 8:2, respectively. The testing results of the CNN model are shown in Fig. 16. The results from the testing set indicate that the highest accuracy can be achieved when the data is divided in a ratio of 7:3. Nevertheless, across all four partition ratios, there were occurrences of misclassification between samples of aging level 5 and aging level 6, highlighting the difficulty in achieving a flawless classification. This challenge can be attributed to the subtle disparities in spectral reflectance between the two aging levels, posing a challenge for the CNN model to accurately distinguish them.
Fig. 16. The evaluation results of the CNN model. (a) The ratio is 5:5 (b) The ratio is 6:4 (c) The ratio is 7:3 (d) The ratio is 8:2.
4.2 Compared to other methods
To achieve the optimal fusion algorithm for evaluating the aging levels of hyperspectral data, this study compares various dimensionality reduction methods and classification algorithms. Specifically, PCA and RF are chosen as the comparative dimensionality reduction techniques, whereas support vector machine (SVM) and back propagation neural network (BPNN) are selected for the comparative classification algorithms.
SVM is a binary classification supervised learning algorithm with the primary goal of finding an optimal hyperplane in the sample space to maximize the margin between different classes, thereby achieving accurate data classification [33,34]. In this study, the SVM model's kernel function is set to the radial basis function (RBF) kernel, and grid search is employed to determine the optimal parameters. The penalty coefficient C is tuned in the range of (1, 200), and the kernel function radius g is tuned in the range of (0.01, 1). The grid search results indicate a penalty coefficient C of 100 and a gamma value of 1. Subsequently, data preprocessed through RF-PCA dimensionality reduction, PCA dimensionality reduction, RF dimensionality reduction, and the full spectral band, respectively, are individually fed into the SVM model. The data is partitioned into training and testing sets at various ratios. The training set is then utilized to train the model through 10-fold cross-validation, and the trained model is subsequently evaluated on the testing set. The accuracy results of the testing set are presented in Table 3.
Table 3. Accuracy of different divisions under different models
The BPNN is a type of multilayer feedforward neural network trained using the backpropagation algorithm. The network consists of three components: the input layer, hidden layer, and output layer, enabling it to achieve continuous nonlinear mappings [35,36]. In this study, the various parameters for the BPNN are set as follows: the number of iterations is 400, the number of hidden layer nodes is 6, the learning rate is 1e-2, and the target training error is 1e-6. The classification results of combining the BPNN model with different dimensionality reduction methods are also shown in Table 3.
The results reveal that, employing a 10-fold cross-validation approach, when the training and testing sets are divided in a 7:3 ratio, the combination of RF-PCA dimensionality reduction with the CNN model performs the best in predicting pre-processed hyperspectral data. The RF-PCA-SVM model closely follows as the second-best combination. In terms of dimensionality reduction for hyperspectral data, RF-PCA demonstrates the most effective reduction, preserving more relevant information. On the other hand, PCA is susceptible to the influence of redundant information, while RF captures a greater number of informative spectral signals, making the processing more complex. Therefore, in terms of dimensionality reduction effectiveness, PCA and RF slightly lag behind RF-PCA. In accuracy, the full spectral band exhibits the lowest performance due to the interference of redundant information. Among the classification models, the CNN model demonstrates outstanding predictive capabilities by effectively capturing local features and spatial relationships within images. Leveraging feature extraction through convolutional and pooling layers, the CNN model benefits from unique advantages in the field of image processing.
5. Visualization
Based on the comparison of prediction models, the RF-PCA-CNN model was chosen to visualize the aging of insulators artificially brushed with an acidic solution. The process of visualization is shown in Fig. 17. The results show that the flow and dripping of the acidic solution lead to varying degrees of aging in different regions. The inclined lower edge of the insulator disc in the insulator shed creates a favorable area for liquid adherence during its rolling movement. Consequently, the aging level is found to be highest at the edge of the insulator shed, which aligns with the observed scenario. In addition, Fig. 17 illustrates the proportion of aging level areas in the insulator shed. The combined area of aging levels 1 to 3 exceeds 70%, while the area corresponding to aging level 5 is the smallest. These observations indicate that the insulator exhibits a favorable overall insulation performance.
Additionally, validation was conducted on the visualization results. Due to the challenging nature of accurately measuring the distribution of surface aging states on insulators, this study conducted cropping of the third piece of SiR shed from the visualization results. Surface roughness and static contact angle were measured in different regions to validate the accuracy of the insulator aging state distribution. The results indicate that the static contact angle for region #1 is 109.20°, with Sa and Sq values of 105.18 nm and 131.29 nm, respectively. Regions #2 to #5 exhibit noticeable variations in results, while the most severely aged area #6, as per the classification results, has a static contact angle below 90°. In this region, the SiR shed material has become hydrophilic, with Sa and Sq values increasing to 345.60 nm and 439.18 nm, respectively. The surface exhibits evident holes, indicating severe acidic corrosion of the SiR shed material. The results of roughness and static contact angle in different regions align with the classification results presented in this paper.
6. Conclusion
This study utilizes a non-destructive, non-contact HSI technique in conjunction with deep learning algorithms to evaluate the aging condition of insulators under acidic environments. The conclusions are as follows:
- (1) In this study, a combined dimensionality reduction approach called RF-PCA is proposed, which integrates traditional feature wavelength extraction using RF with PCA. The RF-PCA method identifies feature wavelengths within the hyperspectral spectrum that are predominantly concentrated between 650 nm and 750 nm. Subsequently, the dimensionality of the feature wavelengths is reduced, and the top 7 principal components, accounting for over 90% of the total variance, are extracted. This approach enables the visualization and clustering of the SiR's aging status.
- (2) Based on the CNN deep learning algorithm, a model for insulation aging assessment was constructed, comprising four convolutional layers. Comparative results with SVM and BPNN models suggest that, when the training and testing sets are divided in a 7:3 ratio, the RF-PCA-CNN model achieves the highest classification accuracy, reaching 94.44%.
- (3) Based on the evaluation method presented in this study, a visual analysis of the aging status of composite insulators was performed, effectively showcasing the distribution and proportion of aging areas. This method establishes a theoretical foundation for detecting potential defects in composite insulators used in transmission lines, thereby mitigating the occurrence of flashover accidents and facilitating the secure and stable operation of the power grid.
Funding
State Grid Corporation of China (No.5108-202299270A-1-0-ZB).
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper may be available from the corresponding author upon reasonable request.
References
1. C. Yin, Z. Xiao, Y. Guo, et al., “Method for detecting the pollution degree of naturally contaminated insulator based on hyperspectral characteristics,” High Volt. 6(6), 1031–1039 (2021). [CrossRef]
2. H. Ji, X. Zhang, Y. Guo, et al., “Evaluation method for the UV aging state of composite insulators based on hyperspectral characteristic,” IEEE Trans. Instrum. Meas. 72, 1–9 (2023). [CrossRef]
3. B. Gou, C. Xie, H. Xu, et al., “Cracking mechanism and degradation performances of HTV silicone rubber with interfacial defects under acid and thermal stress,” Eng. Fail. Anal. 127, 105468 (2021). [CrossRef]
4. S. Li, S. Yu, and Y. Feng, “Progress in and prospects for electrical insulating materials,” High Volt. 1(3), 122–129 (2016). [CrossRef]
5. L. Zhao, C. Li, S. Zhang, et al., “Online hydrophobicity measurement for silicone rubber insulators on transmission lines,” IEEE Trans. Power Deliv. 24(2), 806–813 (2009). [CrossRef]
6. M. M. Werneck, D. M. Santos, C. C. Carvalho, et al., “Detection and monitoring of leakage currents in power transmission insulators,” IEEE Sensors J. 15(3), 1338–1346 (2015). [CrossRef]
7. I. Ahmadi-Joneidi, A. Majzoobi, A. A. Shayegani-akmal, et al., “Aging evaluation of silicone rubber insulators using leakage current and flashover voltage analysis,” IEEE Trans. Dielect. Electr. Insul. 20(1), 212–220 (2013). [CrossRef]
8. R. A. Osornio-Rios, J. A. Antonino-Daviu, and R. J. Romero-Troncoso, “Recent industrial applications of infrared thermography: a review,” IEEE Trans. Ind. Inf. 15(2), 615–625 (2019). [CrossRef]
9. L. Jin, Z. Tian, J. Ai, et al., “Condition evaluation of the contaminated insulators by visible light images assisted with infrared information,” IEEE Trans. Instrum. Meas. 67(6), 1349–1358 (2018). [CrossRef]
10. C. Xia, M. Ren, B. Wang, et al., “Infrared thermography-based diagnostics on power equipment: state-of-the-art,” High Volt. 6(3), 387–407 (2021). [CrossRef]
11. R. Chakraborty and S. Reddy, “Performance of silicone rubber insulators under thermal and electrical stress,” IEEE Trans. on Ind. Applicat. 53(3), 2446–2454 (2017). [CrossRef]
12. Y. Qiu, X. Zhang, Y. Guo, et al., “Detection method of insulator contamination grades based on hyperspectral technique,” High Volt. Eng. 45(11), 3587–3594 (2019).
13. C. Yin, Y. Guo, X. Zhang, et al., “A novel method for visualizing the pollution distribution of insulators,” IEEE Trans. Instrum. Meas. 70, 1–8 (2021). [CrossRef]
14. Q. Hao, Y. Pei, R. Zhou, et al., “Fusing Multiple Deep Models for In Vivo Human Brain Hyperspectral Image Classification to Identify Glioblastoma Tumor,” IEEE Trans. Instrum. Meas. 70, 1–14 (2021). [CrossRef]
15. F. Zhang, F. Zhang, S. Wang, et al., “Hyperspectral imaging combined with CNN for maize variety identification,” Front. Plant Sci. 14, 1254548 (2023). [CrossRef]
16. Z. Khan, F. Shafait, and A. Mian, “Joint group sparse PCA for compressed hyperspectral imaging,” IEEE Trans. on Image Process. 24(12), 4934–4942 (2015). [CrossRef]
17. N. Falco, J. A. Benediktsson, and L. Bruzzone, “Spectral and spatial classification of hyperspectral images based on ICA and reduced morphological attribute profiles,” IEEE Trans. Geosci. Remote Sensing 53(11), 6223–6240 (2015). [CrossRef]
18. J. Xia, L. Bombrun, T. Adali, et al., “Spectral–spatial classification of hyperspectral images using ICA and edge-preserving filter via an ensemble strategy,” IEEE Trans. Geosci. Remote Sensing 54(8), 4971–4982 (2016). [CrossRef]
19. “Insulators for overhead lines-composite suspension and tension insulators for a.c. systems with a nominal voltage greater than 1000 V-Definitions, test methods and acceptance criteria,” IEC Standard 61109: 2008.
20. “Artificial pollution tests on high-voltage ceramic and glass insulators to be used on a.c. systems,” IEC Standard 60507: 2013.
21. “Guidance on the measurement of hydrophobicity of insulator surfaces,” IEC/TS Standard, 62073: 2016.
22. “Geometrical product specifications (GPS)–surface texture: areal, part 1: terms, definitions and surface texture parameters,” ISO Standard 25178-2: 2012.
23. W. Yu, H. Huang, and G. Shen, “Deep spectral–spatial feature fusion-based multiscale adaptable attention network for hyperspectral feature extraction,” IEEE Trans. Instrum. Meas. 72, 1–13 (2023).
24. F. Melgani and L. Bruzzone, “Classification of hyperspectral remote sensing images with support vector machines,” IEEE Trans. Geosci. Remote Sensing 42(8), 1778–1790 (2004). [CrossRef]
25. K. Lin, Y. Guo, Y. Liu, et al., “Outdoor detection of the pollution degree of insulating materials based on hyperspectral model transfer,” Measurement 214, 112805 (2023). [CrossRef]
26. N. D. Roberts, R. D. Kortschak, W. T. Parker, et al., “A comparative analysis of algorithms for somatic SNV detection in cancer,” Bioinformatics 29(18), 2223–2230 (2013). [CrossRef]
27. S. Liu, H. Li, Y. Zhang, et al., “Random forest-based track initiation method,” J. Eng. 2019(19), 6175–6179 (2019). [CrossRef]
28. X. Kang, X. Xiang, S. Li, et al., “PCA-based edge-preserving features for hyperspectral image classification,” IEEE Trans. Geosci. Remote Sensing 55(12), 7140–7151 (2017). [CrossRef]
29. J. Chen, P. Jönsson, M. Tamura, et al., “A simple method for reconstructing a high-quality NDVI time-series data set based on the Savitzky-Golay filter,” Remote Sens. Environ. 91(3-4), 332–344 (2004). [CrossRef]
30. R. D. Nimmo, G. Callender, and P. L. Lewin, “Methods for wavelet-based autonomous discrimination of multiple partial discharge sources,” IEEE Trans. Dielect. Electr. Insul. 24(2), 1131–1140 (2017). [CrossRef]
31. A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,” Commun. ACM 60(6), 84–90 (2017). [CrossRef]
32. J. Cheng, J. Sun, K. Yao, et al., “Multi-task convolutional neural network for simultaneous monitoring of lipid and protein oxidative damage in frozen-thawed pork using hyperspectral imaging,” Meat Sci. 201, 109196 (2023). [CrossRef]
33. Y. Zhang, S. Xiang, Y. Han, et al., “BP-based supervised learning algorithm for multilayer photonic spiking neural network and hardware implementation,” Opt. Express 31(10), 16549–16559 (2023). [CrossRef]
34. J. Li, X. Yao, X. Wang, et al., “Multiscale local features learning based on BP Neural Network for rolling bearing intelligent fault diagnosis,” Measurement 153, 107419 (2020). [CrossRef]
35. J. Hu, B. Wang, K. Di, et al., “Self-healing FBG sensor network fault-detection based on a multi-class SVM algorithm,” Opt. Express 31(25), 41313–41325 (2023). [CrossRef]
36. Z. Xu, Y. Jiang, J. Ji, et al., “Classification, identification, and growth stage estimation of microalgae based on transmission hyperspectral microscopic imaging and machine learning,” Opt. Express 28(21), 30686–30700 (2020). [CrossRef]
