Abstract: Rolling element bearings are the silent workhorses of modern industrial machinery. They operate continuously, often in harsh conditions, and their failure — when unannounced — can bring entire production lines to a halt. The financial and human cost of such unplanned downtime has driven decades of research into condition monitoring. Yet, the field still lacks a method that is simultaneously reliable across diverse operating conditions, computationally light enough for embedded deployment, and capable of detecting degradation before energy-based indicators register any concern. This paper presents such a method. We propose an Orthogonal Subspace Projection (OSP) detector, inspired by the principle of destructive interference in quantum mechanics. We model the vibration signature of a healthy bearing as occupying a low-dimensional subspace of the frequency domain. When a new measurement arrives, we cancel its healthy component by projecting it onto the orthogonal complement of this subspace. The residual energy — the component that cannot be cancelled — is a scalar anomaly score that rises as degradation alters the spectral texture of the vibration signal, often long before the overall energy level changes. The method is validated on all three run-to-failure datasets from the NASA IMS Bearing Benchmark, comprising 9,464 sequential vibration recordings spanning more than 1,577 hours of continuous operation. We compare against four baselines: Root Mean Square (RMS), Kurtosis, Crest Factor, and PCA with Mahalanobis distance. The OSP detector is the only method to achieve zero false alarms across all three datasets simultaneously. On Set 3, it detects incipient degradation 37 hours earlier than RMS and 53 hours earlier than PCA-Mahalanobis. The entire reference model requires 8 KB of memory, and detection operates in under 0.1 ms per window on standard CPU hardware, with no GPU, no labelled fault data, and no neural network.

Keywords: bearing fault detection, orthogonal subspace projection, quantum-inspired signal processing, SVD, anomaly detection, NASA IMS dataset, predictive maintenance, vibration analysis, one-class detection

Pages: 115-123