Industrial computed tomography based intelligent non-destructive testing method for power capacitor

Zhenxing Cheng, Peng Wang, Wei Qin, Zidi TangISCMI 2023

Abstract

Power capacitors are critical components in electrical grids, yet their internal structural integrity is often difficult to assess without destructive testing. This paper presents an Intelligent Non-Destructive Testing (I-NDT) framework that combines Industrial Computed Tomography (ICT) with deep learning to automate defect detection. By utilizing a Single Shot MultiBox Detector (SSD) with a ResNet50 backbone and a specialized image-splitting strategy, the system effectively identifies microscopic welding voids. Experimental results demonstrate a Mean Intersection over Union (MIoU) of 0.86, offering a robust, objective solution for industrial quality control.

Introduction

Power capacitors provide essential reactive power compensation and harmonic filtering for the electrical grid. However, they are prone to internal structural degradation due to aging or overheating—faults that remain invisible from the exterior. Traditional utility inspection relies on periodic manual testing or electrical differential methods, which often fail to detect internal issues until catastrophic failure occurs. While Industrial Computed Tomography (ICT) allows for internal visualization, the resulting high-resolution scans are massive and complex. Relying on human technicians to identify microscopic defects in thousands of scans is slow, subjective, and prone to error due to visual fatigue. This study proposes an automated framework to address these challenges.

Methodology

The proposed I-NDT framework integrates high-resolution imaging with a deep learning detection pipeline.

The proposed Intelligent Non-Destructive Testing (I-NDT) framework, illustrating the pipeline from ICT image acquisition to defect detection.

Detection Architecture and Strategy

The core of the system is a Single Shot MultiBox Detector (SSD) utilizing a ResNet50 backbone, chosen for its balance of speed and robustness. A critical challenge in industrial NDT is the massive size of ICT images (4096 × 4096 pixels). Simply resizing these images for a neural network would cause microscopic features, such as welding spots, to be lost. To prevent this, the researchers implemented a "Split-Detect" strategy: the original image is sliced into an 8×8 grid of 512 × 512 sub-images. Each sub-image is analyzed individually, and the results are merged to ensure spatial resolution is preserved for defect detection.

Data Acquisition and Augmentation

The study utilized 20 power capacitors of various types. High-resolution volumetric images were generated using a 450kV industrial X-ray source. To overcome the scarcity of faulty industrial samples, the team employed data augmentation techniques, including kernel filters and random cropping. This process expanded an initial set of 40 original scans into a robust training set of 3,000 sub-images. The data was split into a 90% training set (2,700 images) and a 10% testing set (300 images), with a focus on detecting "voiding spots"—microscopic gaps in internal welding points.

Key Results

The trained SSD model achieved a Mean Intersection over Union (MIoU) of 0.86 across the test dataset. The system successfully transitioned the inspection workflow from a manual, subjective process to an objective, automated "rank and rate" system.

Example of the SSD model identifying welding voids (voiding spots) within a power capacitor sub-image.

Qualitative analysis confirmed that the model could accurately localize microscopic voiding spots, marking them with bounding boxes in the final output. This level of precision is difficult for human inspectors to maintain consistently over large datasets.

Discussion and Conclusion

The research highlights that preserving image resolution through splitting is vital for detecting microscopic industrial defects. Furthermore, it demonstrates that intelligent data augmentation can make deep learning viable even when physical sample sizes are limited.

Future work will focus on expanding the dataset to include more diverse failure modes and optimizing the backbone architecture—potentially using attention-based mechanisms like Transformers—to further improve accuracy in noisy X-ray environments. This framework represents a significant step toward the "third revolution" in coordinate metrology, enabling non-destructive measurement of complex internal geometries.

Summary updated

BibTeX
@inproceedings{cheng2023industrial,
  author = {Zhenxing Cheng and Peng Wang and Wei Qin and Zidi Tang},
  title = {{Industrial computed tomography based intelligent non-destructive testing method for power capacitor}},
  booktitle = {ISCMI},
  year = {2023},
}
Figure