Predictive maintenance of petrochemical equipment based on LSTM neural network and vibration signal analysis

Predictive maintenance (PDM) is vital in industrial settings to enhance efficiency, reduce downtime, and extend equipment life. Early fault detection is especially critical in petrochemical plants, where machines operate under severe and unstable conditions. Traditional approaches often focus on individual components, neglecting system-wide health monitoring. To address this, the study proposes a Deep Learning (DL)-based PDM framework adaptive to petrochemical equipment. It leverages multi-sensor data, including vibration, temperature, and electrical current from key rotating machines. Preprocessing involves Savitzky–Golay filtering to reduce noise, Z-score normalization for scaling, and Wavelet Packet Decomposition for rich time-frequency feature extraction. The core model, Dynamic Gravitational Search-tuned Intelligent LSTM Neural Network (DG-ILSTM-Net), combines gravitational search-based hyperparameter optimization with LSTM’s sequential learning capability. This enables real-time anomaly detection, supporting informed maintenance and uninterrupted production. Experiments show the proposed model achieves 99.50% accuracy and outperforms traditional Machine Learning methods in F1-score, recall, and precision. The results demonstrate that DG-ILSTM-Net is a scalable, accurate, and domain-specific PDM solution for complex petrochemical systems.