Interpretable acoustic fault diagnosis of industrial valves and pumps using a multi-feature fusion attention network

Abstract

Acoustic-based fault diagnosis provides a non-invasive solution for monitoring industrial valves and pumps, but its reliability is often limited by background noise, class imbalance, and limited interpretability. This study proposes a multi-feature fusion and attention network (MFA-Net) for interpretable machinery acoustic anomaly detection under noisy conditions. The framework fuses Mel spectrograms, Mel-frequency cepstral coefficients, and spectral contrast, and uses a hierarchical convolutional neural network-bidirectional long short-term memory-transformer architecture to model local time-frequency patterns, temporal dependencies, and long-range contextual relationships. Experiments were conducted on valve and pump subsets of the Malfunctioning Industrial Machine Investigation and Inspection dataset under −6, 0, and 6 dB conditions. To avoid information leakage, original recordings were split at the file level; offline augmentation was applied only to abnormal training recordings, and all test recordings remained original and non-augmented. The fixed feature configuration achieved accuracies of 0.9880–0.9988 for valves and 0.9703–0.9988 for pumps, with area under the Receiver Operating Characteristic curve values of 0.9822–1.0000. Comparisons with Support Vector Machine, convolutional neural network, and standalone Transformer baselines show that MFA-Net provides competitive classification performance and stronger pump diagnosis under several noisy conditions, while computational-cost analysis indicates an inference latency of 6.0401 ms per sample on an NVIDIA GeForce RTX 4060 Laptop graphics processing unit. Gradient-weighted class activation mapping and Transformer attention visualizations further localize discriminative acoustic evidence in time-frequency and temporal domains. Overall, MFA-Net provides a noise-resilient and interpretable acoustic diagnosis approach for industrial condition monitoring.