Cost-effective fault diagnosis in bearing grinding machines: a robust sae framework using minimal industrial sensor sets and early-cycle detection

Abstract Predictive Maintenance (PdM) in grinding processes is often constrained by the high cost of sensor infrastructures and the limited robustness of data-driven models in noisy industrial environments. This gap limits the effective adoption of advanced diagnostic techniques on the shop floor. This work presents a fault diagnosis framework based on a supervised autoencoder (SAE), designed for accurate and cost-effective industrial deployment using a reduced set of standard sensors. The approach integrates temporal segmentation, feature extraction, and a two-stage classification strategy for both fault detection and fault classification. The paper mages a comparison of SAE technique with other AI methods, making a detailed analysis of accuracy metrics. The main contributions are threefold: (i) the development of an SAE architecture capable of learning a highly discriminative latent representation of machine health; (ii) the demonstration that reliable diagnosis can be achieved using only two standard industrial sensors (force and acoustic emission), significantly reducing instrumentation requirements; and (iii) the introduction of an early fault detection strategy based on pre-contact data. Experimental validation on a bearing grinding machine dataset shows that the proposed method achieves accuracy above 92% with two sensors and up to 99.44% using only pre-contact signals. These results enable a zero-scrap strategy by allowing early process interruption before defective parts are produced.