This study analyzes the impact of different types of random noise applied in Denoising Autoencoder (DAE) training on fault diagnosis performance, with the aim of improving noise removal for vibration time series data. While conventional studies typically train DAEs using Gaussian random noise, such noise does not fully reflect the complex noise patterns observed in real-world industrial environments. Therefore, this study proposes a novel approach that uses high-frequency noise components extracted from actual vibration data as training noise for the DAE. Both Gaussian and high-frequency noise were used to train separate DAE models, and statistical features (mean, RMS, standard deviation, kurtosis, skewness) were extracted from the denoised signals. The fault diagnosis rates were calculated using One-Class Support Vector Machines (OC-SVM) for performance comparison. As a result, the model trained with high-frequency noise achieved a 0.0293 higher average F1-score than the Gaussian-based model. Notably, the fault detection accuracy using the kurtosis feature improved significantly from 26.22% to 99.5%. Furthermore, the proposed method outperformed the conventional denoising technique based on the Wavelet Transform, demonstrating superior noise reduction capability. These findings demonstrate that incorporating real high-frequency components from vibration data into the DAE training process is effective in enhancing both noise removal and fault diagnosis performance.

With increasing interest in smart factories, considerable attention has been paid to the development of deep-learning-based quality inspection systems. Deep-learning-based quality inspection helps productivity improvements by solving the limitations of existing quality inspection methods (e.g., an inspector’s human errors, various defects, and so on). In this study, we propose an optimized YOLO (You Only Look Once) v5-based model for inspecting small coils. Performance improvement techniques (model structure modification, model scaling, pruning) are applied for model optimization. Furthermore, the model is prepared by adding data applied with histogram equalization to improve model performance. Compared with the base model, the proposed YOLOv5 model takes nearly half the time for coil inspection and improves the accuracy of inspection by up to approximately 1.6%, thereby enhancing the reliability and productivity of the final products.

With the Fourth Industrial Revolution unfolding worldwide, technologies including the Internet of Things, sensors, and artificial intelligence are undergoing rapid development. These technological advancements have played a significant role in the dramatic growth of the predictive maintenance market for mechanical equipment, prompting active research on noise removal techniques and classification algorithms for the accurate determination of the causes of equipment failure. In this study, time series data were preprocessed using the denoising autoencoder technique, a deep learning-based noise removal method, to improve the accuracy of failure classification from mechanical equipment data. To convert the preprocessed time series data into frequency components, the short-time Fourier transform technique was employed. The fault types of mechanical equipment were classified using the vision transformer (ViT) technique, a deep learning technique that has been actively used in recent image analysis research. Additionally, the classification performance of the ViT-based technique for vibration time series data was comparatively validated against existing classification algorithms. The accuracy of failure classification was the highest when the data, preprocessed using a Denoising Autoencoder (DAE), were classified by a Vision Transformer (ViT).

An unsupervised anomaly-detection framework is proposed and field validated for thrust-bearing monitoring in the impulse turbine of a shoreline oscillating water-column (OWC) wave energy converter (WEC) off Jeju Island, Korea. Operational monitoring is constrained by nonstationary sea states, scarce fault labels, and low-rate supervisory logging at 20 Hz. To address these conditions, a 24 h period of normal operation was median-filtered to suppress outliers, and six physically motivated time-domain features were computed from triaxial vibration at 10 s intervals: absolute mean; standard deviation (STD); root mean square (RMS); skewness; shape factor (SF); and crest factor (CF, peak divided by RMS). A feature-based autoencoder was trained to reconstruct the feature vectors, and reconstruction error was evaluated with an adaptive threshold derived from the moving mean and moving standard deviation to accommodate baseline drift. Performance was assessed on a 2 h test segment that includes a 40 min simulated fault window created by doubling the triaxial vibration amplitudes prior to preprocessing and feature extraction. The detector achieved accuracy of 0.99, precision of 1.00, recall of 0.98, and F1 score of 0.99, with no false positives and five false negatives. These results indicate dependable detection at low sampling rates with modest computational cost. The chosen feature set provides physical interpretability under the 20 Hz constraint, and denoising stabilizes indicators against marine transients, supporting applicability in operational settings. Limitations associated with simulated faults are acknowledged. Future work will incorporate long-term field observations with verified fault progressions, cross-site validation, and integration with digital-twin-enabled maintenance.

We propose a phase-aware complex-spectrogram autoencoder (AE) for preprocessing raw vibration signals of rotating electrical machines. The AE reconstructs normal components and separates fault components as residuals, guided by an input-phasor phase-orthogonality regularization that defines parallel/orthogonal residuals with respect to the local signal phase. We use a U-Net-based AE with a mask-bias head to refine local magnitude and phase. Decisions are based on residual features—magnitude/shape, frequency distribution, and projections onto the normal manifold. Using the AI Hub open dataset from field ventilation motors, we evaluate eight representative motor cases (2.2–5.5 kW: misalignment, unbalance, bearing fault, belt looseness). The preprocessing yielded clear residual patterns (low-frequency floor rise, resonance-band peaks, harmonic-neighbor spikes), and achieved an area under the receiver operating characteristic curve (ROC-AUC) = 0.998–1.000 across eight cases, with strong leave-one-file-out generalization and good calibration (expected calibration error (ECE) ≤ 0.023). The results indicate that learning to remove normal structure while enforcing phase consistency provides an unsupervised front-end that enhances fault evidence while preserving interpretability on field data.