Abstract
Industrial pump units require condition recognition models that remain reliable under measurement noise and can reject health states that are not observed during training. Existing methods usually focus on noise-robust feature extraction, limited-sample classification, or unknown-state detection separately, leaving their joint interaction insufficiently explored. This paper proposes NMEKNet, a noise-robust open-set condition recognition framework for industrial pump units under limited samples. The feature extractor employs a multi-scale depthwise residual convolutional neural network, in which parallel depthwise separable convolutions and residual fusion capture short impact transients, modulation patterns, and wider temporal variations from multi-channel vibration segments. To improve robustness to controlled signal-to-noise ratio (SNR)-induced noise variation, online signal-to-noise-ratio jitter and Noise-domain MixStyle expose the network to input perturbations and feature-statistic shifts during training. For open-set recognition, a known-only calibrated Energy-kNN scoring rule combines classifier energy with feature-neighbour distance to integrate global classifier evidence and local feature-manifold support. Experiments on the Royal Netherlands Navy Electric Pump test bench and a self-built laboratory pump-loop platform under few-shot and controlled low-SNR settings show that NMEKNet provides stable known-state recognition and effective unknown-state rejection. Sensitivity and cross-SNR calibration analyses further demonstrate the robustness of the main parameter settings while revealing the limitations of transferring a fixed threshold across severe noise shifts. The results indicate that multi-scale representation learning, noise-domain regularisation, and validation-calibrated Energy-kNN scoring are suitable for controlled industrial pump-unit monitoring, while validation under non-Gaussian interference and broader field conditions remains future work.
Funding Information
  • National Natural Science Foundation of China (52172367)

This publication has 35 references indexed in Scilit: