A Lightweight Multi-Sensor Feature-Level Fusion Framework for Noise-Robust Fault Diagnosis of High-Precision Milling Machines

The operational reliability of milling machines is a critical factor affecting manufacturing quality and productivity in modern industrial systems. Although multi-sensor fusion can overcome the limitations of single-sensor monitoring, many existing methods rely on computationally intensive deep learning models, which restrict their applicability in resource-constrained industrial environments. To address this issue, this paper presents a multi-sensor feature-level fusion framework for fault diagnosis in high-precision milling machines. The proposed framework integrates vibration signals from three sensing directions through signal preprocessing, multi-domain feature extraction, random-forest-based feature selection, feature-level fusion, and lightweight neural classification. Within this framework, a lightweight multilayer perceptron (MLP) is adopted as the default classifier to achieve an effective balance between diagnostic performance and computational efficiency for edge-oriented deployment. The experiments consider five machine states, comprising two non-fault operating states (No_Load and Normal) and three abnormal or fault-related states (Eccentricity, Bearing_Housing_Shim_15, and Bearing_Housing_Loose). Experimental results show that the proposed framework substantially outperforms single-sensor approaches, improving diagnostic accuracy from 74.00–86.00% to 96.00%. Comparative analysis across different classifier backbones further demonstrates that the lightweight MLP provides the most favorable efficiency-performance trade-off, while the multi-domain fusion (MDF) mechanism exhibits the strongest robustness under noisy conditions, maintaining 94.38 ± 1.96% accuracy at 0 dB SNR and 88.88 ± 2.53% at −10 dB SNR. These results indicate that accurate, efficient, and noise-robust milling-machine fault diagnosis can be achieved through a coordinated combination of multi-sensor feature fusion, lightweight classification, and robust fusion design, thereby providing a practical solution for intelligent industrial condition monitoring.