新页面
Non-destructive Testing of Insulating Materials Defects Based on Terahertz Time-domain Spectroscopy System
Xiaoyu Luoa, Boxin Yanb, Chao Wangb,*, Ming Niea
a Electric Power Research Institute of Guangdong Power Grid Co.,Ltd. Guangzhou 510080, China
b Wuhan University of Technology, Wuhan 430070, Hubei, China
Abstract - Insulating materials are widely used in fields such as power grids, electrical engineering, and automotive manufacturing. However, defects such as bubbles and inclusions often arise during their processing and service, which not only affect performance but may also pose potential safety hazards. Terahertz technology, with its low radiation and high penetration characteristics, has become a new method for the non-destructive testing of insulating materials. However, its imaging is limited by hardware and optical diffraction constraints, leading to issues such as low resolution, high noise, and low contrast. This study focuses on epoxy resin insulating materials and employs a reflective terahertz time-domain spectroscopy system to perform point-by-point sampling, obtaining waveforms and images for detecting internal defects such as bubbles and inclusions, and analyzing the differences in defect characteristics. To address the issue of inadequate image quality, a preprocessing method combining non-local means filtering and contrast-limited adaptive histogram equalization is proposed to suppress noise and enhance local contrast. Additionally, the SRCNN network is applied for super-resolution reconstruction to improve image details and resolution. Experimental results demonstrate that the proposed method effectively alleviates issues such as low resolution and low contrast in terahertz images, significantly improving defect detection accuracy and reliability.
Keywords: insulating material, terahertz, non-destructive testing, image enhancement, super-resolution reconstruction
Introduction
Insulation materials serve as the foundation for the development of electrical products and are widely used in numerous fields, including power grids, electrical engineering, automobile manufacturing, and aerospace. For instance, epoxy resin is commonly employed as a supporting insulation component in high-voltage equipment such as GIL/GIS and cable terminals[1], which can reduce installation time by 80% and maintain partial discharge levels below 5pC. However, internal defects such as bubbles, cracks, impurities, and aging inevitably arise during the production and service life of insulation material components. These defects not only compromise the material’s normal operation but may also lead to functional failures, severe accidents, and significant economic losses. In practical engineering, there is a demand to detect potential internal defects in insulation materials without damaging power grid equipment, thereby assessing the extent of degradation. Non-destructive testing of insulation materials not only ensures equipment safety and extends service life but also improves operational efficiency and reduces costs. For example, the cost of timely repair for partial discharge in cable joints is merely 10%-20% of the expense required for post-failure replacement.
The common non-destructive testing methods currently used in the power grid field include infrared inspection, ultrasonic testing, and X-ray detection. Infrared inspection can only detect surface or near-surface defects and is highly sensitive to environmental temperature, limiting its effectiveness. Ultrasonic testing faces difficulties when detecting curved components and has poor penetration ability for high attenuation materials. X-ray inspection generates ionizing radiation, posing certain safety risks. Terahertz waves refer to electromagnetic waves with frequencies between 0.1 THz and 10 THz, corresponding to wavelengths ranging from approximately 0.03 to 3 mm. Terahertz waves have good penetration capabilities, easily passing through most non-polar materials and dielectric materials. They are also highly safe, as the energy of terahertz photons is very low compared to X-rays and does not cause damage to the material being tested due to ionizing radiation. Furthermore, terahertz waves have unique fingerprint spectral characteristics, which can be used for the exclusive identification and analysis of materials. Therefore, terahertz non-destructive testing technology is suitable for internal non-destructive inspection of insulating materials.
Terahertz imaging is achieved through point-by-point scanning of a sample. By extracting the time-domain and frequency-domain characteristics of the terahertz reflected waves from the sample, pixel values are assigned and converted into grayscale image data to generate the terahertz image. However, due to limitations such as terahertz wavelength, signal attenuation, noise interference, and hardware constraints of imaging systems, current terahertz images still suffer from issues like low spatial resolution, poor contrast, and blurred contours. To address these challenges, two primary solutions exist: first, enhancing hardware precision to improve terahertz image resolution, and second, using a super-resolution reconstruction [2] algorithm post-processing to refine acquired terahertz images for higher resolution. Notably, the second approach proves more cost-effective. To enhance imaging quality, improve defect recognition accuracy, and precisely characterize internal defects in insulation materials, this paper proposes a preprocessing method combining non-local mean filtering with contrast-limited histogram equalization, alongside an improved SRCNN (Super-Resolution Convolutional Neural Network) for terahertz image super-resolution reconstruction. The effectiveness of this methodology is validated through experimental testing on epoxy resin insulation material defect samples.
2. Basic principles
2.1 Principle of Terahertz Time Domain Spectroscopy System
The working principle of the reflective terahertz time-domain spectroscopy system[3] is shown in Figure 1. A femtosecond laser generates laser pulses, which are split into two beams after passing through a half-wave plate and a polarizer. The reflected beam serves as the pump light, while the refracted beam is the probe light. The pump light passes through another half-wave plate and polarizer, then is delayed optically before being focused onto a photoconductive antenna to generate terahertz pulses. After being reflected by a set of parabolic mirrors, the terahertz pulses are directed towards the sample under test. The probe light is focused onto the opposite end of the detection antenna after passing through a mirror and a lens. The probe light generates photo-induced carriers in the focal region of the detector, which are used to detect the instantaneous electric field amplitude of the terahertz pulse, thereby enabling terahertz detection and obtaining the terahertz time-domain waveform.
Figure 1 Schematic diagram of the reflective terahertz time-domain spectroscopy (THz-TDS) system
Figure 2 illustrates a terahertz time-domain spectroscopy scanning imaging system. The epoxy resin sample is placed on a two-dimensional scanning translation stage, with the mainframe and PC turned on, and testing begins after the mainframe has warmed up. The height of the translation stage is adjusted to maximize the peak-to-peak difference of the time-domain signal[4], in order to prevent diffraction effects from adversely affecting the terahertz image. The stepper motor driver, controlled by the PC, drives the stepper motor to move the stage, and a reflected terahertz signal is obtained for each pixel movement. The information from the obtained terahertz signal is used to assign values to the pixels, and these values are then converted into grayscale to produce the sample’s original terahertz image[5]. When defects are present along the optical path of the terahertz wave, the received terahertz wave undergoes distortion due to reflection at the defect interface. In the time domain, this manifests as changes in the pulse peak size, flight time, and signal envelope area, while in the frequency domain, it primarily manifests as variations in the amplitude of frequency components and changes in spectral energy.
Figure 2 Reflective terahertz time-domain spectroscopy (THz-TDS) scanning imaging system
2.2 Principle of Terahertz Image Super-Resolution Reconstruction
The terahertz image super-resolution reconstruction model used in this study consists of two parts: the first part is the preprocessing module, and the second part is the SRCNN network. The first part aims to perform denoising and image enhancement on the original terahertz image, while the second part aims to improve the spatial resolution of the terahertz image. The training network is shown in the figure. The first part mainly implements denoising and contrast enhancement through non-local mean filtering and contrast-limited histogram equalization, while the second part uses the SRCNN network, which consists of feature extraction layers, nonlinear mapping layers, and high-resolution reconstruction layers to achieve resolution enhancement.
Due to the physical characteristics of terahertz waves and the point-by-point scanning imaging method, the resulting terahertz images have pixels that are relatively isolated from each other. This leads to issues such as high noise, low contrast, and blurry edges, requiring preprocessing. This paper primarily employs the non-local mean filtering algorithm to remove background noise and uses contrast-limited histogram equalization to enhance the contrast of the terahertz image.
To address the problem of isolated pixels and unclear textures in the original terahertz images, the non-local mean filtering method[6] is applied for denoising, with the algorithm schematic shown in Figure 3. This algorithm requires calculating the similarity between all pixels and the current pixel. In practice, a large search window and a small neighborhood window are first defined, with the neighborhood window sliding within the search window. The weights are determined based on the similarity between the neighborhoods. The formula for non-local mean filtering is as follows:
| (1) |
Here, represents the weight, indicating the similarity between pixels and in the original image; is the search window of pixel ; and is the filtered image.
Figure 3 Schematic illustration of the non-local means (NLM) filtering principle
The method for determining the similarity between two neighborhood windows is to calculate the mean squared error (MSE) between them:
| (2) |
Here, is the normalization factor, which is the sum of all weights; is the mean squared error; is the filtering coefficient. The larger the value of , the better the denoising effect, but the image becomes more blurred. Conversely, the smaller the value of , the worse the denoising effect, but the less distortion after denoising.
The basic idea of histogram equalization is to perform a nonlinear mapping of image pixel values to make their grayscale distribution more uniform. However, traditional histogram equalization globally stretches the grayscale histogram of the entire image, which may excessively enhance noise in low-contrast areas and lead to the loss of details in high-brightness regions. When histogram equalization is applied to terahertz images, which are low signal-to-noise ratio (SNR) images, the noise is significantly amplified, resulting in salt-and-pepper noise or artifacts. Contrast-limited adaptive histogram equalization (CLAHE)[7] is an improved version of histogram equalization that prevents the issue of excessive noise enhancement. CLAHE limits the contrast within each local region after dividing the image into blocks, thereby avoiding the introduction of excessive noise caused by overly high contrast.
CLAHE sets a threshold for the original image’s histogram, and if the grayscale of a certain bin exceeds this threshold, it is clipped. The excess values are then evenly distributed across the available grayscale levels, as shown in Figure 4. Additionally, to avoid blocky discontinuities caused by image segmentation, bilinear interpolation is used to reconstruct the pixel grayscale values.
Figure 4 Schematic illustration of CLAHE principle
As shown in Figure 5, the background of the terahertz image processed by NLM and CLAHE is significantly cleaner, with a substantial reduction in noise. The readability of weak signal areas has been notably improved; low-contrast features such as bubbles, inclusions, and cracks are more pronounced. Additionally, edge and texture information is better preserved, avoiding the blurring issues typically associated with traditional filtering methods.
Figure 5 Preprocessing workflow diagram
The SRCNN network in the second part is primarily used to enhance the preprocessed images. Unlike the traditional SRCNN[8], the SRCNN in this study uses a 7×7×64 convolutional kernel and a Rectified Linear Unit (ReLU, Max(0, x)) in the first layer to extract features and obtain low-resolution features. In the second layer, a 5×5×32 convolutional kernel is employed to nonlinearly map the low-resolution features to 32 sets of high-resolution features. In the third layer, a 5×5×1 convolutional kernel is used to integrate the high-resolution features and generate the final high-resolution image.
In this study, low-resolution terahertz images with c channels and height and width dimensions of h and w, respectively, are used as the input to the SRCNN, as shown in Figure 6. The first convolutional layer (Conv1) is used to extract low-resolution features, with a convolutional kernel size of f1×f1, n1 filters, and a ReLU activation function. To ensure that the image size remains unchanged after convolution, padding should be applied during the convolution process, with the padding size set to padding = f1//2. After Conv1, the output feature map has a size of n1×h×w.
| (3) |
The second layer of the SRCNN network, Conv2, is used to establish the mapping relationship between low-resolution and high-resolution features. It maps the n1 low-resolution feature vectors from Conv1 to n2 high-dimensional feature vectors, which represent the features of the high-resolution image. The convolutional kernel size of Conv2 is f2×f2, with n2 filters, and the activation function is ReLU. Padding is applied with a size of padding = f2//2. After passing through Conv2, the output feature map has a size of n2×h×w.
| (4) |
The third layer of the SRCNN network, Conv3, is used to reconstruct the high-resolution image. Conv3 can be understood as a set of linear filters, and the average of the filtered results represents the predicted image, with the output being the final reconstructed image. The convolutional kernel size of Conv3 is f3*f3, with c filters, and padding is applied with a size of padding = f3//2. After passing through Conv3, the output is a high-resolution image with a size of c×h×w.
| (5) |
Unlike the traditional SRCNN, in this study, the convolutional kernel size in the first layer is f1×f1 = 7×7, with padding = 3. In the second layer, the convolutional kernel size is f2×f2 = 5×5, with padding = 2. The convolutional kernel size in the third layer is the same as that in the traditional SRCNN, f3×f3 = 5×5. This design is related to the particularities of terahertz images.
Figure 6: SRCNN Network Diagram
3. Material and method
The terahertz time-domain spectroscopy (THz-TDS) imaging system employed in this study, provided by Qingdao Qingyuan Fengda Terahertz Technology Co., Ltd., consists of four primary components: a computer, main unit, imaging module, and a two-dimensional scanning platform. The system operates within a frequency bandwidth of 0.1-4.5 THz, with three selectable pulse scanning speeds (120ps@25Hz, 90ps@30Hz, and 300ps@10Hz). It demonstrates a lateral resolution of 1 mm, thickness detection range of 30 μm-9 mm, thickness measurement accuracy of ±2 μm, and an imaging field of view (FOV) of 100mm×100mm.
Epoxy resin, recognized for its high dielectric strength and corrosion resistance, has found extensive industrial applications as a critical insulating material in power grid systems. This study selected epoxy resin as the target insulating material. Specimens containing artificially induced defects (gas bubbles and embedded inclusions) were prepared through controlled gas injection during the curing process and deliberate inclusion of foreign particles. The specimens were fabricated with dimensions of 50mm×50mm×3mm. The experimental setup involved positioning samples on a lifting platform and acquiring terahertz signals through transmission-mode THz-TDS. A two-dimensional scanning platform performed raster scanning with a step size of 0.2mm, yielding 272×272 pixel images.
The acquisition of terahertz images presents notable challenges, including prolonged imaging durations (requiring no less than 40 minutes per specimen in this study). Furthermore, terahertz nondestructive testing for insulating materials remains in its nascent stages with insufficient data accumulation. These constraints make the establishment of large-scale terahertz image training datasets impractical. To address this limitation, we utilized the public DIV2K dataset for network training while reserving 10 acquired terahertz images for testing.
Following model architecture development and dataset preparation, network training was implemented using PyTorch 2.4.0 framework under Python 3.7 environment. Training computations were executed on an Intel Xeon 6133 processor coupled with an NVIDIA RTX 4080S GPU. The super-resolution reconstruction network was configured with 4× magnification factor, initial learning rate of 10-4, and batch size of 16. The model underwent 500 training epochs using Adam optimization algorithm for parameter updates.
4. Results and analysis
The following image demonstrates the experimental results of image super-resolution reconstruction. Figure 7(a) shows the original terahertz image, Figure 7(b) displays the image processed by the classical SRCNN [7] network, Figure 7(c) presents the image processed by the SRGAN [9] network, and Figure 7(d) shows the image processed by the SRCNN network proposed in this paper. As shown in Figure 7(a), the original terahertz image exhibits noticeable artifacts, but these artifacts are significantly improved in Figures 7(b) and (d). In Figure 7(c), the sharpness and contrast at the image edges are visibly enhanced, but some distortion and noise are introduced after SRGAN processing. Subjectively, the difference between Figures 7(b) and 7(d) is small, but from the zoomed-in view, the texture in Figure 7(d) is clearer than in Figure 7(b). In contrast, Figure 7(b) exhibits some mosaic effects, which negatively impact its overall quality.
Figure 7: Terahertz Images of Epoxy Resin Samples
(a) Original Image(b) SRCNN(c) SRGAN(d) Ours
The aforementioned three algorithms have each demonstrated a degree of enhancement on terahertz images. In addition to subjective evaluation[10], a quantitative comparison of their effectiveness is required. To objectively compare the performance of these algorithms, Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) were selected as evaluation criteria for assessing enhancement quality. PSNR, one of the most widely used metrics in image processing, represents the ratio between the maximum possible signal power and the power of corrupting noise that affects fidelity. A higher PSNR value indicates superior image quality. SSIM, another prevalent image similarity metric, emphasizes structural information in quality assessment, aligning with human visual perception. Thus, SSIM better approximates human judgment of image quality, where values closer to 1 indicate lower distortion.
As shown in Figure 8, the proposed algorithm achieves smoother textures and more pronounced edges. In contrast, images processed by the conventional SRCNN algorithm exhibit mosaic artifacts that degrade visual quality. Meanwhile, the SRGAN network, with its more complex architecture, was trained exclusively on open-source optical image datasets, making it more suitable for optical image reconstruction. Comprehensive analysis of Figure 8 and Table 1 leads to the conclusion that the improved SRCNN achieves optimal enhancement performance.
The superior performance of the proposed algorithm stems from two architectural innovations: First, the adoption of a 7×7 convolutional kernel in the initial layer reduces overfitting risks while enabling focused extraction of medium-scale features. Second, the implementation of a 5×5 convolutional kernel in the subsequent layer expands the receptive field[11], facilitating reconstruction of intricate textures. This dual optimization effectively mitigates checkerboard artifacts and improves edge continuity compared to conventional approaches.
Figure 8: Magnified Terahertz Image Details of Epoxy Resin Samples (a) Original Image(b) SRCNN(c) SRGAN(d) Ours
Table 1: Comparison of Reconstruction Results for Three Networks
| Project | PSNR/dB | SSIM |
| SRCNN | 27.62 | 0.925 |
| SRGAN | 21.32 | 0.754 |
| Ours | 27.39 | 0.928 |
5. Conclusions
This paper proposes a novel super-resolution reconstruction network that integrates traditional image processing algorithms with SRCNN to address the challenges of high noise, low contrast, and limited resolution in terahertz images, thereby enhancing the precision and accuracy of non-destructive testing for insulating materials in engineering applications. Terahertz images of epoxy resin samples containing bubble and inclusion defects were employed as experimental inputs. The framework first preprocesses the images using NLM and CLAHE algorithms. The preprocessed images are subsequently fed into the modified SRCNN network for super-resolution reconstruction. Experimental results demonstrate that the proposed algorithm effectively reduces artifacts, enriches image details, and improves spatial resolution compared to conventional approaches.
Acknowledgement
This work was supported by the Science and Technology Project of China Southern Power Grid Co., Ltd. (Grant No. GDKJXM20222546).
References
[1] Yufan W. Research on the Impact of Temperature Gradient on the Insulation Breakdown Characteristics of Epoxy Resin and Its Phase Field Simulation [D]. Tianjin University, 2022. (in Chinese).
[2] Koch M, et al. Terahertz time-domain spectroscopy. Nat Rev Methods Prim 2023;3 (1):48.
[3] Zhenwei Z. Research on Pulse THz Time-Domain Spectral Imaging and Continuous THz Wave Imaging Correlation Techniques [D]. Beijing: Capital Normal University, 2006. (in Chinese).
[4] Huaiyuan J, Hongwei M, Xingming B, et al. Detection of double⁃layer air gap defects based on terahertz imaging method[C]. IEEE International Instrumentation and Measurement Technology Conference(I2MTC). Glasgow: IEEE, 2021：1⁃5.
[5] Xiaobo Z. Improved Non-local Means Using Structural Similarity for Image Denoising[J]. Circuits, Systems, and Signal Processing,2025, Vol.44(4): 2706-2736.
[6] Malik Braik, Mohammed Azmi Al-Betar, Mohammed A. Mahdi, et al. Enhancement of satellite images based on CLAHE and augmented elk herd optimizer[J]. Artificial Intelligence Review,2025, Vol.58(2).
[7] Koch M, et al. Terahertz time-domain spectroscopy. Nat Rev Methods Prim 2023;3 (1):48. (in Chinese).
[8] Dong C，Loy C C，He K，et al． Learning a deep convolutional network for image super-resolution[A]． European Conference on Computer Vision[C]． Springer, Cham，2014． 184-199．
[9] Ledig C, Theis L, Huszar F, et al. Photo-realistic single image super-resolution using a generative adversarial network[J]. arXiv Preprint, 2016, arXiv:1609． 04802.
[10] SHEIKH H R, BOVIK A C. Image information and visual quality[J]. IEEE Transactions on Image Processing，2006：430-444．
[11] Qian L, Fengyu Z. Remote Sensing Image Super-Resolution Reconstruction Based on Multi-Path and Hybrid Attention Fusion [J]. Journal of Computer Applications, 2023, 44(7): 1508-1513. (in Chinese).