NPLNet: A Noise-Prior Lightweight Network for On-Device Acoustic Fault Diagnosis of Railway Point Machines From Noisy Measurements

Despite the remarkable success of deep learning in fault diagnosis for Railway Point Machines (RPMs), existing models often overlook the non-ideal conditions inherent in manual field maintenance, such as environmental noise and signal instabilities stemming from handheld acquisition. This leaves a critical gap: achieving high-performance diagnosis while maintaining the lightweight design required for end-to-end deployment on handheld diagnostic instruments remains a formidable challenge. We address this gap by proposing the Noise-Prior Lightweight Network (NPLNet), an end-to-end lightweight fault diagnosis network. To improve robustness to amplitude fluctuations, we introduce the Random Amplitude Perturbation (RAP) strategy, an empirical augmentation designed to mimic gain variations commonly observed in imperfect handheld acquisition. The architecture features a backbone that integrates a Noise-Prior Spectral Filtering (NPSF) module for robust noise suppression and stacked Attentive Hierarchical Redundancy-reduction Cell (AHRC) units for multi-scale feature extraction. This is followed by a Multi-granularity Semantic Aggregation (MSA) head that performs robust classification by fusing semantic information from the hierarchical feature maps. Experimental results on the noisy ZDJ9 RPM acoustic dataset demonstrate that NPLNet achieves a peak diagnostic accuracy of 99.59% with a high inference throughput of 228.9 samples/s on a single-thread ARM Cortex-A53 CPU. This performance demonstrates a favorable trade-off along the accuracy-throughput Pareto frontier, outperforming several lightweight architectures in accuracy while maintaining competitive CPU throughput. The proposed network provides a lightweight and noise-prior-guided solution for handheld RPM acoustic diagnosis in noisy environments.