Using Communication Networks in Control Systems: The Theoretical and Practical ChallengesView this Special Issue
Research Article | Open Access
Manran Wang, Li Guo, Jinhao Chen, Bingqi Lv, Yang Yang, "Superresolution Reconstruction of Electrical Equipment Incipient Fault", Journal of Control Science and Engineering, vol. 2018, Article ID 1630402, 11 pages, 2018. https://doi.org/10.1155/2018/1630402
Superresolution Reconstruction of Electrical Equipment Incipient Fault
With the rapid development of industry and technology, the electrical power system becomes more complex and the electrical equipment becomes more diverse. Defective equipment is often the cause of industrial accidents and electrical injuries, which can result in serious injuries, such as electrocution, burns, and electrical shocks. In some cases, electrical equipment fault may result in death. However, in some special situation, some fault is very small even invisible, such as equipment aging, holes, and cracks, so the detection of these incipient faults is difficult or even impossible. These potential incipient faults become the biggest hidden danger in the electrical equipment and electricity power system. For these reasons, this paper proposes a superresolution reconstruction method for electrical equipment incipient fault to ensure complete detection in electrical equipment, which aims to guarantee the security of electrical power system operation and industry production. Experimental results show that this method can get a state-of-the-art reconstruction effect of incipient fault, so as to provide reliable fault detection of electrical power system.
In the electrical power system, there is no doubt that the safety of electrical equipment is the basis for ensuring the stability and reliability . Aging and fine lines of electrical equipment components can be characterized as incipient fault of electrical equipment. In power system, fault of electrical equipment components may manifest themselves as abnormal deviations in system behavior and operation. However, due to their very slow evolution, their effects may be confused with noise and uncertainty, which constitutes the characteristic that incipient fault is difficult to detect. Because it is not easy to be discovered, incipient fault is often regarded as precursors in significant accidents. Earlier detected type and exact location of the fault can cause faster detection and repairing of the fault, which is very important for the stable operation of the power system . In recent years, there are many researchers who focus on fault detection of the electrical power system. Huang D et al. developed a transformer fault information pattern recognition and diagnosis model using the objective entropy weight method . Taha I B M et al. designed a conditional probability scheme to inspect transformer incipient fault . And Huang D et al. presented an improved hidden Markov model (HMM) algorithm for fault diagnosis of urban rail transit motors equipment; they also used a back-propagation neural network for multiple fault of complex equipment bearings [5, 6]. However, these methods can only be targeted at minor devices, failing to accurately determine potential incipient fault, which is a limitation in fault detection application. At present, it seems that few works are concerned with the topic of incipient fault detection [7–9]. In view of the above situation, this paper proposes a preprocessing method based on superresolution (SR) reconstruction, which can helpfully detect the incipient fault by improving the resolution of the electrical equipment image. In this method, a deep network structure is combined with sparse prior information, so as to obtain the high-resolution (HR) version by mapping from low-resolution (LR) input. Finally the SR reconstruction result of incipient fault can be obtained to improve the subjective visual effect, which can help detect and analyze the electrical equipment fault.
This paper is organized as follows. The basic framework and methods of SR are discussed in Section 2. In Section 3, we evaluate our model on some electrical equipment images containing electrical fault and give a detailed analysis of these results. Finally, Section 4 draws a conclusion.
2. SR Reconstruction Method of Electrical Equipment Incipient Fault
2.1. Basic Framework of SR
In some special electrical situations where it is difficult to produce HR monitoring video (images), SR reconstruction is an efficient method to improve the resolution of the captured monitoring video (images). SR is a problem of obtaining a HR image from multiple or single LR images , which is an inverse problem of imaging process. In imaging process, the LR image is acquired through various imaging devices which are corrupted by noise and other degraded effect [11–13], and the imaging process is shown in Figure 1(a). It is worthwhile to improve the resolution of LR images in some special situations. The observation model of imaging process is mathematized as (1) and shown in Figure 1(b).where represents the original real HR continuous natural scene, is the output digital LR image, is the point-spread function (PSF), which represents blurring matrices and downsampling matrices, and is the additive noise from different environment and device . From this observation model, SR is an ill-posed inverse problem of reconstructing a HR image from an observed LR image.
(a) The colorful imaging process of a natural scene
(b) The image degraded model
Figure 1(a) is the basic illustration of the colorful imaging process of a natural scene, which can obtain R, G, and B channel of an image, respectively, by CCD array. Figure 1(b) is the image degraded model; it corresponds to (1). SR is an inverse process of Figure 1(b), so it is an ill-posed inverse problem which estimates a HR version closed to an original real HR scene.
In recent image processing area, the type of SR reconstruction can be divided into single-image SR, single-video SR, and multiple-video SR. This paper focuses on single-image SR; it is more useful in the incipient fault detection of electrical equipment. Single-image SR also can be divided into classical method [15, 16] and learning-based method [17–19]. Figure 2(a) shows the classical multiframe image to achieve single-image SR; there are 4 LR images with subpixel translation; the complementary information can be fused to reconstruct a HR image with higher resolution. In Figure 2(a), all of the small circulars, rhombuses and triangles represent subpixel different sample points in HR grid, respectively.
(a) Classical SR
(b) Example-based SR
The other recently popular method is learning-based SR reconstruction, one of which is illustrated in Figure 2(b) . All of these methods use single-scale or multiscale information to get learning network, so as to achieve HR reconstruction from LR input. In this paper, we adopt the second SR framework to reconstruct detailed information of incipient fault on electrical equipment.
2.2. SR Reconstruction Method for Incipient Fault Detection
For what is detailed above, in order to accurately detect the incipient fault and guarantee safety of electricity power system, improving the resolution of monitoring images of electrical equipment is an effective pathway. A SR reconstruction method is introduced in this paper, which adopt the idea of deep learning network model  shown in Figure 3. This model combines the traditional sparse coding model into deep learning so as to obtain HR result, which extends the conventional sparse coding model using several key ideas from deep learning, and complements large learning capacity to improve SR reconstruction performance.
Firstly, we implement a feed-forward neural network whose layers strictly correspond to each step in the processing flow of sparse coding based image SR; the framework is shown in Figure 3. In this way, the sparse representation prior is effectively encoded in our deep network structure; all the coded components are trained jointly through back-propagation. This method includes three stages as shown in Figure 3. The first stage is a convolutional process, which is patch extraction and representation. LR image is the Bicubic-upscaled version of input real LR image, and this layer extracts feature for each LR patch of size 5 × 5. In this stage, the size of our input patch must be the same as the filter spatial size as . This process can be represented as ; here and are convolution filter and deviation, respectively. In this paper, we extract 100-dimension feature in this stage. The middle stage of this model is a high-dimension mapping process, which includes three linear layers shown in Figure 4 . The sparse code is multiplied with HR dictionary in the last linear layer to reconstruct HR patch. This stage is equal to a filter with simple spatial support 1 × 1; it also can be mathematized as . Here and are also filter and deviation, respectively, just like in stage . The final stage in this model is reconstruction, which is also a convolution similar to . In this stage, all the recovered patches are placed back to the corresponding positions in the HR grid through a convolutional filter, so as to get the HR output result. This model combined with the focal points of sparse coding and the strong points of deep learning to achieve SR reconstruction. As a result, HR version of the incipient fault on electrical equipment can be more visible and helpful in achieving easier fault detection. The recurrent middle stage adopted the idea referred in , which is mathematized as Eq. (2) where is the input signal and is an coordinate-wise shrinkage function defined as with positive threshold . is the transpose of the dictionary matrix , and is . The matrices and are learned so as to minimize the approximation error to the optimal sparse coding on a given dataset. The final convolution stage producing HR reconstruction result is described as , is a vector, and project the coefficients to image spatial domain to get averaged HR result [22, 23].
On the other hand, loss function is also needed in this model, which can be obtained by minimizing reconstruction result and corresponding original HR version . In this paper, mean squared error (MSE) is adopted as loss function in this model, which is shown in Here SSE is the sum of squared error, is the number of samples, weight , is the original HR image, and is the estimated SR reconstruction HR result.
In order to show the efficiency of the proposed method for incipient fault SR reconstruction of electrical equipment, we give two experiments to straightly verify the effective result. One is simulation incipient fault image SR, and some good electrical equipment is manually added and some incipient fault on the surface. The other is electrical equipment images with real incipient fault on the surface.
In these experiments, we firstly perform incipient fault extraction on images captured by high-definition (HD) camera, and then a degraded processing is applied to change the resolution to be , which is referred to as original HR image. For SR model, the resolution of LR input is by Bicubic interpolation. In our SR reconstruction process, the upscaling factor is 2. In order to clearly show the efficiency of the detailed method in this paper, the result of Bicubic interpolation, SR reconstruction of SRCNN , SR result by proposed method in this paper, and the ground-truth HR image are all provided to give a complete subjective comparable effect. At the same time, we also provide the objective quantitative metric peak-signal-noise-ratio (PSNR) [25, 26] and structural similarity index metric (SSIM) [25, 27] to evaluate all of these results, so as to further validate efficiency of the detailed method in this section. PSNR is a classical objective metric for image quality evaluation, which is widely used in many real applications. SSIM is a popular image quality assessment based on computing the structure similarity; it also provides superior performance.
3.1. Simulation Incipient Fault Experiment
In the following simulation experiment, we simulate the incipient fault of common four types of electrical equipment in substations. In Figures 5–8, we simulate the gaps and the oil pillow’s leakage of insulator marked by gray part; we also simulate the 35kv transformer’s crack with a black broken line, and a painted brown part on the circuit breaker is the simulated rust stain. All of these artificially added parts are simulated fault on the electrical equipment, which are very small or invisible in natural scene, because the size of electrical equipment is usually large or huge. Note that, in order to better display the visual results, we select detailed information of the fault for each image and enlarge them to show better display. All the following experiments also follow this rule.
From the different SR reconstruction results in Figures 5–8, the details of Bicubic reconstruction in Figures a1-a4 are the most obscured and ambiguous. Although the results of SRCNN shown in Figures b1-b4 are slightly clear, the detailed information is not sufficient enough to make accurate fault judgments. It is clearly shown that the image c1-c4 has a fine edge structure; the image feature structure is closer to the HR image (d1-d4). In the image detected by our method, the edge portion of the electrical device has more zigzag texture, and the image sharpening effect is better, so that the shape and angle of the notch and crack can be clearly observed to help diagnose electrical equipment incipient fault. All of the objective evaluation values are included in Table 1, which also proves the efficient SR reconstruction of the proposed method in this paper.
3.2. Real Incipient Fault Experiment
Although the simulation incipient fault SR experiment is effective and credible, we further provide real incipient fault SR experiment to validate the proposed method. Similar to the above experiments, we detect the actual fault commonly found in substations. In the same way, we also give objective evaluation and subjective evaluation indicators based on human vision in this SR reconstruction experiment.
In the power system, once incipient fault such as cracks and gaps occur in electrical equipment, which are difficult to be discovered by human, they even lead to a huge accident crisis for the whole system. In Figures 9–12, the Bicubic interpolation reconstruction (a5-a8) has the most blurring fault edge information and the highest distortion, which is helpless for the detection of incipient fault. The SR reconstruction results (b5-b8) by SRCNN are relatively acceptable. In Table 2 and Figure 13 (average value of these three methods of PSNR and SSIM), the results of the proposed method in this paper (c5-c8) are absolutely superior to Bicubic and SRCNN, the edges and lines of the fault are more clear, and the crack length and shape are more obvious. So it is shown that the proposed method is the effective way to make the incipient fault more visible.
3.3. Low-Resolution Actual Incipient Fault Experiment
In this subsection, the incipient fault images are captured by LR imaging device (such as cellphone with LR shot camera); some of them are also captured in special conditions such as dim environments. In this case, the original LR image is , and the upscaled factor is 2. Because there is not a HR version for objective quality evaluation, the blind image quality metric (BRISQUE) [28, 29] is adopted in this experiment. BRISQUE is a widespread no-reference image quality metric; it extracts mean subtracted contrast normalized (MSCN) coefficients to fit asymmetric generalized Gaussian distribution (AGGD), so as to estimate the image quality.
Figures 14 and 15 are electrical equipment LR images captured in a dim environment; most of the incipient fault is difficult to detect under such circumstance. The burr sparked by insulator in Figure 14 and the soldering position on the switch blade in Figure 15 are easy to be recognized by SR reconstruction. Comparable fine structure SR results in Figure c9-10 are more vivid than other results. Furthermore, the objective evaluation metric of BRISQUE also proves the proposed method is higher than the SR results by other methods (lower BRISQUE value means more quality or higher resolution of the image) in Table 3.
A SR reconstruction method combining sparse coding and deep learning is used in this paper, so as to achieve preprocessing for accurate incipient fault detecting of electrical equipment. It can be seen from the subjective and objective results compared with other methods that the proposed SR method in this paper achieved significant visual improvement in image resolution. Through the above-mentioned detection of a series of initial defect incipient fault, such as incompleteness, cracks, and blemishes, this method can obtain relatively clear effect and retain most of fine edge information. Applying this method for practical incipient fault detection can improve fault resolution and achieve accurate fault location and judgment, so as to ensure the safe operation of electrical equipment and electricity power system. For actual application, our further work will combine this method in fault diagnosis and other areas, to ensure the reliability of electrical system operation.
The data used to support the findings of this study are available from the corresponding author upon request.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
This work is supported by National Natural Science Foundation of China (61663008), Natural Science Foundation of Hubei Province (2015CFC781, 2014CFB612), Foundation of CSC (China Scholar Council), and PhD Technology Program (MY2014B018).
- E. Zio and L. R. Golea, “Analyzing the topological, electrical and reliability characteristics of a power transmission system for identifying its critical elements,” Reliability Engineering & System Safety, vol. 101, pp. 67–74, 2012.
- N. Ismail, F. H. Nordin, and A. A. Alkahtani, “Detection of the source of the incipient faults produced by single phase inverter using feed- forward back-propagation neural network,” Indian Journal of Science and Technology, vol. 9, no. 48, 2017.
- D. Huang, C. Chen, G. Sun, and L. Zhao, “Recognition and Diagnosis Method of Objective Entropy Weight for Power Transformer Fault,” Dianli Xitong Zidonghua/Automation of Electric Power Systems, vol. 41, no. 12, pp. 206–211, 2017.
- I. B. M. Taha, D.-E. A. Mansour, S. S. M. Ghoneim, and N. I. Elkalashy, “Conditional probability-based interpretation of dissolved gas analysis for transformer incipient faults,” IET Generation, Transmission & Distribution, vol. 11, no. 4, pp. 943–951, 2017.
- D. Huang, L. Ke, X. Chu et al., “Fault diagnosis for the motor drive system of urban transit based on improved hidden markov model,” Microelectronics Reliability, vol. 82, pp. 179–189, 2018.
- D.-R. Huang, C.-S. Chen, G.-X. Sun, L. Zhao, and B. Mi, “Linear Discriminant Analysis and Back Propagation Neural Network Cooperative Diagnosis Method for Multiple Faults of Complex Equipment Bearings,” Acta Armamentarii, vol. 38, no. 8, pp. 1649–1657, 2017.
- G. Rigatos, P. Siano, and A. Piccolo, “Incipient fault detection for electric power transformers using neural modeling and the local statistical approach to fault diagnosis,” in Proceedings of the IEEE Sensors Applications Symposium, SAS '12, pp. 32–37, Italy, February 2012.
- T. Escobet, V. Puig, J. Quevedo, and D. Garcia, “A methodology for incipient fault detection,” in Proceedings of the IEEE Conference on Control Applications, CCA '14, pp. 104–109, France, October 2014.
- W. Ge, J. Wang, J. Zhou, H. Wu, and Q. Jin, “Incipient fault detection based on fault extraction and residual evaluation,” Industrial & Engineering Chemistry Research, vol. 54, no. 14, pp. 3664–3677, 2015.
- K. Zhang, X. Gao, D. Tao, and X. Li, “Single image super-resolution with non-local means and steering kernel regression,” IEEE Transactions on Image Processing, vol. 21, no. 11, pp. 4544–4556, 2012.
- I. E. Mourabit, M. E. Rhabi, A. Hakim, A. Laghrib, and E. Moreau, “A new denoising model for multi-frame super-resolution image reconstruction,” Signal Processing, vol. 132, pp. 51–65, 2017.
- P. Purkait and B. Chanda, “Super resolution image reconstruction through Bregman iteration using morphologic regularization,” IEEE Transactions on Image Processing, vol. 21, no. 9, pp. 4029–4039, 2012.
- L. Yue, H. Shen, J. Li, Q. Yuan, H. Zhang, and L. Zhang, “Image super-resolution: The techniques, applications, and future,” Signal Processing, vol. 128, pp. 389–408, 2016.
- R. Kolte and A. Arora, “Image super-resolution,” Lap Lambert Academic Publishing, vol. 3, no. 10, pp. 7195–7199, 2013.
- A. R. A. Nazren, S. N. Yaakob, R. Ngadiran, M. B. Hisham, and N. M. Wafi, “Improving iterative back projection super resolution model via anisotropic diffusion edge enhancement,” in Proceedings of the International Conference on Robotics, Automation and Sciences, ICORAS '16, Malaysia, November 2016.
- X. Liu, D. Song, C. Dong et al., “MAP-based image super-resolution reconstruction,” International Journal of Computer and Information Engineering, vol. 27, pp. 208–211, 2008.
- W. T. Freeman, T. R. Jones, and E. C. Pasztor, “Example-based super-resolution,” IEEE Computer Graphics and Applications, vol. 22, no. 2, pp. 56–65, 2002.
- W. T. Freeman, E. C. Pasztor, and O. T. Carmichael, “Learning low-level vision,” International Journal of Computer Vision, vol. 40, no. 1, pp. 25–47, 2000.
- J. Yang, J. Wright, T. S. Huang, and Y. Ma, “Image super-resolution via sparse representation,” IEEE Transactions on Image Processing, vol. 19, no. 11, pp. 2861–2873, 2010.
- X. Du, X. Qu, Y. He, and D. Guo, “Single Image Super-Resolution Based on Multi-Scale Competitive Convolutional Neural Network,” Sensors, vol. 18, no. 3, p. 789, 2018.
- C. Dong, C. Loy, K. He, and X. Tang, “Learning a deep convolutional network for image super-resolution,” in Computer Vision-ECCV, vol. 8692 of Lecture Notes in Computer Science, pp. 184–199, Springer, New York, NY, USA, 2014.
- Z. Wang, D. Liu, J. Yang, W. Han, and T. Huang, “Deep networks for image super-resolution with sparse prior,” in Proceedings of the 15th IEEE International Conference on Computer Vision, ICCV '15, pp. 370–378, Chile, December 2015.
- K. Gregor and Y. LeCun, “Learning fast approximations of sparse coding,” in Proceedings of the 27th International Conference on Machine Learning, ICML 2010, pp. 399–406, Israel, June 2010.
- K. I. Kim and Y. Kwon, “Example-based learning for single-image super-resolution,” in Pattern Recognition, vol. 5096 of Lecture Notes in Computer Science, pp. 456–465, Springer, Berlin, 2008.
- A. Horé and D. Ziou, “Image quality metrics: PSNR vs. SSIM,” in Proceedings of the 20th International Conference on Pattern Recognition (ICPR '10), pp. 2366–2369, Istanbul, Turkey, August 2010.
- Q. Huynh-Thu and M. Ghanbari, “Scope of validity of PSNR in image/video quality assessment,” IEEE Electronics Letters, vol. 44, no. 13, pp. 800-801, 2008.
- P. Zhao, Y. Liu, J. Liu, A. Argyriou, and S. Ci, “SSIM-based error-resilient cross-layer optimization for wireless video streaming,” Signal Processing: Image Communication, vol. 40, pp. 36–51, 2016.
- A. Mittal, A. K. Moorthy, and A. C. Bovik, BRISQUE Software Release, 2011, http://live.ece.utexas.edu/research/quality/BRISQUE_release.zip.
- A. Mittal, A. K. Moorthy, and A. . Bovik, “No-reference image quality assessment in the spatial domain,” IEEE Transactions on Image Processing, vol. 21, no. 12, pp. 4695–4708, 2012.
Copyright © 2018 Manran Wang et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.