Early detection of malfunctions in industrial machinery is critical for minimizing downtime and ensuring production efficiency. Although deep learning approaches based on acoustic signals have shown promising results, many existing models rely on single-feature inputs or lack robustness in noisy environments—limiting their effectiveness in real-world factory settings. In this study, we propose AFT-Net (Acoustic Fusion Transformer Network), a deep learning framework that integrates convolutional, recurrent, and attention-based components to robustly detect machine faults from acoustic signals under varying noise conditions. To enhance feature diversity and improve discriminative capability, AFT-Net fuses three complementary audio representations: Mel spectrograms, MFCCs, and spectral contrast. These features respectively capture perceptual frequency energy distributions, spectral envelope structures, and harmonic–noise patterns. The fused multimodal representation is processed through a hierarchical architecture composed of convolutional encoders for local time–frequency pattern extraction, a bidirectional LSTM for sequential dependency modeling, and a Transformer encoder to capture global contextual relationships via self-attention. This design significantly improves the model’s ability to detect subtle and temporally extended fault anomalies. Experimental results demonstrate that the method achieves high classification accuracy and remains robust across diverse machine sound anomalies, confirming its practicality for real-world industrial applications.

acoustic signal processing,fault diagnosis,deep learning,industrial machines,health monitoring