Machine Learning–Enabled Predictive Maintenance for Industrial Sensor Networks: A Hybrid Deep Learning Approach for Real-Time Anomaly Detection and Asset Health Management

Abstract

This study shows machine learning-based predictive maintenance architecture for industrial sensor networks and the design of a hybrid CNN–LSTM model for real-time anomaly detection and equipment health monitoring. The study is based on the Condition Monitoring of Hydraulic Systems dataset consisting of 2205 operational cycles, each with 60 seconds of sampling and 43,680 measurements from various sensors. Pressure, flow, temperature, vibration and motor-power signals under various equipment states are measured and provided. The sensor data were pre-processed using the following stages: missing value treatment, noise and outlier removal, sensor data synchronization, sensor data normalization, time series segmentation and feature engineering before the model was developed. The CNN learns and extracts the local patterns from the sensor signals and the LSTM learns and captures evolutions over time. The proposed model precision, recall and F1 score were 97.1%, 96.5% and 96.2%, respectively. The accuracy of the component level classification was 94.36% for cooler, valve, pump and accumulator. Additionally, the model's anomaly scores of 0.87 and 0.68 were found on abnormal operating time, demonstrating that the internal model can detect faults early and help to make decisions for timely maintenance.