Scholarlink

<p>Neural networks for time-series classification typically process all frequency components uniformly, ignoring the spectral structure that domain experts routinely exploit. We propose FSM-Net, an architecture that learns to partition the frequency domain into target and interference bands via trainable STFT-based gates, processes them through dual-pathway LSTMs, and recombines them via attention-based adaptive fusion. Evaluated against eight baselines spanning convolutional, recurrent, attention, and learnable-filterbank paradigms on CWRU and Paderborn University bearing diagnosis and MIT-BIH arrhythmia detection, FSM-Net achieves 97.74%, 92.61%, and 96.54% test accuracy respectively. While modern convolutional baselines (LiConvFormer, SincNet, WDCNN) attain marginally higher clean accuracy, FSM-Net offers two complementary advantages. First, the learned gates yield physically interpretable decompositions: on CWRU the network suppresses the kinematic fault bins (BPFO/BPFI gates&#xa0;<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\approx \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>≈</mo> </math></EquationSource> </InlineEquation>&#xa0;0.05) and concentrates activation in the 3–6&#xa0;kHz resonance band, autonomously rediscovering the envelope-analysis principle of bearing diagnosis. Second, FSM-Net is substantially more robust to noise: at 5&#xa0;dB SNR FSM-Net retains 59.3% accuracy while CNN-LSTM and SincNet collapse to 40.2% and 9.0%. Removing the interference pathway costs 6.1% on CWRU, confirming that explicitly modeling residual content is critical. FSM-Net is well-suited to safety-critical industrial monitoring, where verifiable behavior and noise tolerance carry weight alongside raw accuracy.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FSM-net: Dual-path signal classification with learned frequency decomposition

  • 期刊论文

  • 发表日期：

    2026年8月2日

查看全文
  • Jie Liu,
  • Weijian Qiu,
  • Zhuohong Wu

摘要

Neural networks for time-series classification typically process all frequency components uniformly, ignoring the spectral structure that domain experts routinely exploit. We propose FSM-Net, an architecture that learns to partition the frequency domain into target and interference bands via trainable STFT-based gates, processes them through dual-pathway LSTMs, and recombines them via attention-based adaptive fusion. Evaluated against eight baselines spanning convolutional, recurrent, attention, and learnable-filterbank paradigms on CWRU and Paderborn University bearing diagnosis and MIT-BIH arrhythmia detection, FSM-Net achieves 97.74%, 92.61%, and 96.54% test accuracy respectively. While modern convolutional baselines (LiConvFormer, SincNet, WDCNN) attain marginally higher clean accuracy, FSM-Net offers two complementary advantages. First, the learned gates yield physically interpretable decompositions: on CWRU the network suppresses the kinematic fault bins (BPFO/BPFI gates  \(\approx \) ≈  0.05) and concentrates activation in the 3–6 kHz resonance band, autonomously rediscovering the envelope-analysis principle of bearing diagnosis. Second, FSM-Net is substantially more robust to noise: at 5 dB SNR FSM-Net retains 59.3% accuracy while CNN-LSTM and SincNet collapse to 40.2% and 9.0%. Removing the interference pathway costs 6.1% on CWRU, confirming that explicitly modeling residual content is critical. FSM-Net is well-suited to safety-critical industrial monitoring, where verifiable behavior and noise tolerance carry weight alongside raw accuracy.