A Scale-Invariant Spectral Line Density Detector for Bearing Fault Detection on FPGA

Abstract

Condition monitoring of rolling-element bearings is regarded as a fundamental component of predictive maintenance in rotating machinery because unplanned down-time can be mitigated and equipment service life can be extended through early fault detection. High classification accuracy on benchmark vibration datasets has been achieved by neural fault detection methods; however, supervised training is required, and limited interpretability is provided. In contrast, energy- and RMS-based detectors demand low computational resources and suit embedded implementation, but assume fault progression raises vibration amplitude. This paper presents a Spectral Line Density (SLD) detector that counts spectral bins exceeding a fixed multiple of the in-band mean. A ratio statistic gives scale invariance and removes data-dependent multiplication from detection. It combines raw-spectrum and resonance-envelope branches via speednormalized fusion and identifies the damaged bearing element using a characteristic-frequency comb. We evaluated on three public datasets namely CWRU, Paderborn, and MAFAULDA. Seeded faults raise amplitude enough for an energy detector to reach ROC-AUC 1.00, but naturally evolved accelerated-lifetime damage drops its performance to chance level (0.59). In contrast, the fused SLD detector achieves a ROC-AUC of 0.90 under recording- and speed-independent leave-one-bearing-out evaluation across 32 bearings and retains a recall of 0.82 at 90% specificity. We implement the detector in fixed-point arithmetic and validate a time-multiplexed architecture on an Altera DE1-SoC (Cyclone V FPGA), using 4,450 ALMs at 74.7 MHz with a multiplier-free decision path.