Industrial micro-CT demands reconstruction methods that reliably preserve fine internal details while keeping scanning cost manageable. Analytical and iterative algorithms remain widely deployed. However, fidelity degrades as projection counts decrease, while higher projection counts inevitably increase scanning cost. To address this trade-off, implicit neural representations (INRs) have been explored as promising approaches to improve memory and computational efficiency while enhancing reconstruction quality. Nevertheless, current INRs converge slowly, suffer from spectral bias, and lack integrated rendering capabilities. We therefore introduce HR-NAF, a high-resolution neural attenuation field that integrates ray-based sampling (conforming to CT geometry and densely sampling regions of interest), space-variant Fourier encoding (SVFE) to mitigate spectral bias, and a bin scheduler that accelerates convergence by progressively emphasizing high-frequency details. Furthermore, HR-NAF supports direct 3D visualization without external engines. Our method achieves a substantial gain in reconstruction quality with varying numbers of projections, and this performance gap became more pronounced with limited-view settings. This superior fidelity directly enhances the reliable identification of critical high-frequency defects (e.g., pores and cracks) in industrial components, underscoring HR-NAF’s potential to significantly improve the accuracy and efficiency of non-destructive testing (NDT) procedures.

Unsupervised anomaly detection is widely studied in industrial applications where anomalous data is difficult to obtain. In particular, reconstruction-based anomaly detection can be a feasible solution if there is no option to use external knowledge, such as extra datasets or pre-trained models. However, reconstruction-based methods have limited utility due to poor detection performance. A score-based model, also known as a denoising diffusion model, recently has shown a high sample quality in the generation task. In this paper, we propose a novel unsupervised anomaly detection method leveraging the score-based model. The proposed method shows promising performance without requiring external knowledge. The score, a gradient of the log-likelihood, has a property that is available for anomaly detection. The samples on the data manifold can be restored instantly by the score, even if they are randomly perturbed. We call this score-based perturbation resilience. On the other hand, the samples that deviate from the manifold cannot be restored in the same way. The variation of resilience depending on the sample position can be an indicator to discriminate anomalies. We derive this statement from a geometric perspective. Our method shows superior performance on three benchmark datasets for industrial anomaly detection. Specifically, on MVTec AD, we achieve image-level AUROC of 97.7% and pixel-level AUROC of 97.4% outperforming previous works that do not use external knowledge.

Robustness of neural network models is important in fault diagnosis (FD) because uncertainty in operating conditions varies the power spectral densities of vibration data; however, it is unknown to users due to the limited explainability of the models. This article proposes an FD framework with a power-perturbation-based decision boundary analysis (POBA) to explain the decision boundaries of vibration classification models. In the POBA, perturbed data are obtained from training data by power perturbation on frequency bands centering on dominant class-discriminative frequencies. The decision boundary of a model is then evaluated and visualized to users by testing the model on the perturbed data. Furthermore, the decision boundary information can be used to define a robustness score per class, and a robust model can be obtained by ensembling trained models using their robustness score per class. Demonstration using two vibration datasets verifies the explainability and robustness of the proposed FD framework.

Transmission line (TL) inspection is important for ensuring a stable supply of electricity to rural areas. Currently, there are several TL detection approaches based on computer vision; however, they have limitations owing to background clutter in visible light images. This article presents a novel multimodal image feature fusion module that utilizes both visible light and infrared images to enhance the TL-detection performance. The proposed module consists of a multibranch feature extraction (MFE) block followed by a channelwise attention (CA) block. The first block extracts the representative features of each modal input using multiple branches. The outputs of the MFE block are jointly aggregated into an attention vector in the CA block. Finally, the attention vector recalibrates each input feature of the proposed module. To reduce the number of additional parameters due to the insertion of the module, we introduced a channel-shrink factor in the MFE block and utilized a$1\times {1}$convolution in the CA block. Comparison experiments with various augmented conditions of day, night, fog, and snow were conducted on a real-world dataset, which we constructed by visible light and infrared images. The results showed that the proposed module outperformed not only the case of single modal input but also the state-of-the-art fusion methods, regardless of the baseline networks. Additionally, the proposed module showed effectiveness in terms of capacity when the baseline network has a large number of weight parameters.

This article proposes a new end-to-end deep learning model for fault diagnosis using three-axis vibration signals measured from facilities. The three-axis vibration signals measured in the time domain are used without domain transformations to train the end-to-end model. The proposed model is designed to effectively extract feature maps of eachX-,Y-, andZ-axis individually from the three-axis vibration signals using a grouped 1-D convolution. The feature maps extracted from each axis are composed of specific frequencies of each axis. Accordingly, the proposed model classifies faults based on the frequency characteristics of each axis. In addition, this article proposes a method to visualize the decision criteria of the proposed model in the frequency domain. Using the proposed model and the proposed visualization method, it is possible to grasp the degree to which specific frequencies of eachX-,Y-, andZ-axis affect the decision criteria of the proposed model. The experiment is conducted by applying the proposed model to an open dataset with three-axis vibration signals measured from a rotary machine. Experimental results demonstrate that the proposed model achieves high accuracy and provides the explainability in eachX-,Y-, andZ-axis using the visualized decision criteria in the frequency domain.

A linear motion (LM) guide is a mechanical tool for requiring linear motion in a system. Repeating linear movements can cause cracking and deterioration of the LM guide, which can lead to a decrease in productivity. Therefore, predicting the status of the LM guide and diagnosing faults are essential for systems including the LM guide. In this article, we propose a novel framework of fault diagnosis model based on deep learning using a vibration sensor signal mounted on the LM guide. This framework contains the learning vibration signal in the time domain using the proposed 1-D convolutional neural network model and the visualization of the classification criteria in the frequency domain using the learned model in the time domain. To utilize the visualization in the frequency domain, the proposed model is designed to maintain the frequency information in the learning process. With the learned model, we propose a frequency domain-based grad-CAM to visualize the classification criteria in the frequency domain to help to explain the characteristics of normal and fault data. Using LM guide data under various conditions, we visualize the classification criteria of the learned model in the frequency domain.