面向滚动轴承故障诊断的轻量化多域特征融合方法——结合样本自适应注意力与紧凑型MLP

A Lightweight Multi-Domain Feature Fusion Method for Rolling Bearing Fault Diagnosis—Combining Sample-Adaptive Attention with a Compact MLP

摘要: 滚动轴承在小样本和变工况噪声条件下的故障诊断，仍是面向嵌入式状态监测系统的实际难题。本文提出一种轻量化多域特征融合框架：并行提取11维时域统计特征、13维基于包络谱的频域特征以及8维小波包子带能量，然后通过样本自适应多域注意力融合(Sample-Adaptive Multi-domain Attention Fusion, SAMAF)模块根据每个样本各分支的可信度对其进行重加权。融合后的32维特征送入仅含4324个参数的紧凑型两层MLP进行分类。基于SKF 6205物理模型生成的仿真数据集(四类工况、共1600段)上，10次随机划分的平均诊断准确率为95.48%，与简单拼接基线相当，但显著优于各单域基线，且在小样本场景中优势明显：每类仅10个训练样本时，SAMAF达到93.42%，比简单拼接高1.42个百分点。8组可靠性指标的消融实验进一步表明，在SNR = 4 dB强噪声下移除样本级注意力会使准确率下降3.5个百分点，证实了样本自适应加权机制的价值。该方法CPU单样本推理时间低于0.002 ms，适合嵌入式部署。

Abstract: Rolling bearing fault diagnosis under limited samples and noisy industrial conditions remains a practical challenge for embedded condition-monitoring systems. This paper proposes a lightweight multi-domain feature fusion framework that integrates an 11-dimensional time-domain branch, a 13-dimensional frequency-domain branch (including envelope-spectrum descriptors) and an 8-dimensional wavelet-packet time-frequency branch through a Sample-Adaptive Multi-domain Attention Fusion (SAMAF) module. SAMAF combines vector-level ANOVA F-value gating with sample-level reliability-driven branch weighting, introducing no additional trainable parameters; the fused 32-dimensional feature is then classified by a compact two-layer MLP with only 4324 parameters. A physically grounded simulated dataset based on SKF 6205 bearings with four fault conditions and 1600 segments is constructed for evaluation. Experimental results show that the proposed method achieves 95.48% ± 0.95% accuracy over ten random splits and 96.50% ± 0.85% under 5-fold cross-validation, with clear advantages in the small-sample regime: with only 10 training samples per class, SAMAF reaches 93.42%, exceeding simple concatenation by 1.42 percentage points. An ablation study over eight reliability-indicator variants confirms that removing sample-level attention causes a 3.5-percentage-point drop at 4 dB SNR, validating the value of sample-adaptive weighting. The total inference time is below 0.002 ms per sample on CPU, making the framework suitable for embedded deployment.