A Low-Cost, Scalable Approach for Compressor Fault Monitoring Using Deep Learning on Acoustic Signals
2026-02-22
Sumana Roy, Pratyush Kumar Pal, Narottam Behera, Sandip Kumar Lahiri
ABSTRACT In chemical and process industries, reciprocating air compressors are critical single-line equipment whose unexpected failure can trigger plant-wide shutdowns.Legacy compressors often lack built-in monitoring systems, posing significant challenges for early fault detection.This study proposes a non-intrusive, deep learning-based framework for detecting compressor faults through acoustic signal analysis, aiming to retrofit predictive maintenance capabilities into aging assets. A publicly available dataset of air compressor acoustic recordings was utilized, encompassing healthy and seven fault conditions.Sequential models based on Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM) networks were first developed using manually extracted spectral features.Subsequently, a Convolutional Neural Network (CNN) was trained directly on mel-spectrogram representations of the sound signals.Data augmentation techniques were employed to improve model generalization. Performance was evaluated through per-class precision, recall, F1-score, confusion matrices, and cross-validation.The LSTM model achieved a validation accuracy of 92%, which improved to 94% with the BiLSTM architecture.The CNN model achieved 96.6% validation accuracy, further increasing to 98.3% after augmentation, with a macro-F1 score of 98.6%.Cross-validation demonstrated stable performance (±0.4% deviation). A real-world proof-of-concept test on 20 new compressor recordings achieved 95% accuracy, validating the model’s practical deployment capability.The proposed deep learning framework provides a scalable, cost-effective solution for sound-based fault diagnosis in compressors, eliminating the need for physical sensor installations.The CNN model trained on mel-spectrograms proved particularly effective, offering near-real-time prediction performance with minimal hardware requirements.