The industrial application of artificial intelligence (AI) has witnessed outstanding adoption due to its robust efficiency in recent times. Image fault detection and classification have also been implemented industrially for product defect detection, as well as for maintaining standards and optimizing processes using AI. However, there are deep concerns regarding the latency in the performance of AI for fault detection in glossy and curved surface products, due to their nature and reflective surfaces, which hinder the adequate capturing of defective areas using traditional cameras. Consequently, this study presents an enhanced method for curvy and glossy surface image data collection using a Basler vision camera with specialized lighting and KEYENCE displacement sensors, which are used to train deep learning models. Our approach employed image data generated from normal and two defect conditions to train eight deep learning algorithms: four custom convolutional neural networks (CNNs), two variations of VGG-16, and two variations of ResNet-50. The objective was to develop a computationally robust and efficient model by deploying global assessment metrics as evaluation criteria. Our results indicate that a variation of ResNet-50, ResNet-50224, demonstrated the best overall efficiency, achieving an accuracy of 97.97%, a loss of 0.1030, and an average training step time of 839 milliseconds. However, in terms of computational efficiency, it was outperformed by one of the custom CNN models, CNN6-240, which achieved an accuracy of 95.08%, a loss of 0.2753, and an average step time of 94 milliseconds, making CNN6-240 a viable option for computational resource-sensitive environments.
Rolling-element bearing faults are a primary mechanical failure mode in rotating systems. In Permanent Magnetic DC (PMDC) motor applications operating under variable torque, vibration-based diagnosis is affected by load-dependent excitation and commutation-induced disturbances, which introduce amplitude bias and reduce the reliability of conventional statistical features. This study proposes Cross-Reference Energy Attention (CREA), a physics-guided dual-sensor feature framework for three-class bearing states in PMDC motor systems. CREA isolates fault-relevant content within a hardware-agnostic, empirically selected mid-frequency carrier band and incorporates a spatially separated reference sensor to evaluate transmission consistency. This design suppresses disturbances generated locally by the motor while retaining structurally transmitted bearing signatures. Experiments were conducted on a PMDC motor dynamometer with seeded bearing defects under controlled torque variation. GroupKFold cross-validation was implemented using the acquisition run as the grouping variable to prevent data leakage across runs. Under per-run normalization designed to eliminate amplitude memorization, conventional motor-side baseline features degraded to 0.495 ± 0.110 window-level accuracy, whereas the four-feature CREA representation maintained 0.999 ± 0.002. Systematic ablation and SHAP analysis demonstrate that carrier-band energy features provide the dominant discriminatory contribution, while cross-sensor interaction metrics supply complementary transmission validation consistent with the underlying mechanical model.
The industrial application of artificial intelligence (AI) has witnessed outstanding adoption due to its robust efficiency in recent times. Image fault detection and classification have also been implemented industrially for product defect detection, as well as for maintaining standards and optimizing processes using AI. However, there are deep concerns regarding the latency in the performance of AI for fault detection in curvy and glossy products, due to their nature and reflective surfaces, which hinder the adequate capturing of defective areas using traditional cameras. Consequently, this study presents an enhanced method forcurvy and glossy surface image data collection using a Basler vision camera with specialized lighting and KEYENCE displacement sensors, which are used to train deep learning models. Our approach employed image data generated from normal and two defect conditions to train eight deep learning algorithms: four custom Convolutional Neural Networks (CNNs), two variations of VGG-16, and two variations of ResNet-50. The objective was to develop a computationally robust and efficient model by deploying global assessment metrics as evaluation criteria. Our results indicate that a variation of ResNet-50, ResNet-50224, demonstrated the best overall efficiency, achieving an accuracy of 97.79%, a loss of 0.1030, and an average training step time of 839 milliseconds. However, in terms of computational efficiency, it was outperformed by one of the custom CNN models, CNN6-240, which achieved an accuracy of 95.08%, a loss of 0.2753, and an average step time of 94 milliseconds, making CNN6-240 a viable option for resource-sensitive environments.