Predictive Maintenance AI Model in Industrial Automation

Abstract

This paper provides a framework for delivering an artificial intelligence (AI)-based Predictive Maintenance (PdM) solution in the context of industrial automation, by encompassing traditional machine learning (ML) as well as state-of-the-art deep learning (DL) techniques, to perform real-time fault prediction. The research adopted a 10-phase model, from data acquisition through to deployment, and worked with sensor data from motor-fan assemblies in the cement and manufacturing industries. The dataset included machine and process parameters such as vibration, temperature, pressure and shear, with preprocessing, such as in-situ feature engineering and normalization, applied as phases in the analytical synthesis. Traditional ML models (Logistic Regression, Support Vector, and Random Forest) were compared to DL models (Long Short-Term Memory, and Convolutional LSTM models) and overall, the Conv1D-LSTM model imposed the best solution, returning the best combination of performance, with 99.8% classification (diagnostic) accuracy, near perfect F1-score and ROC-AUC. Visualization was incorporated to incorporate functionality in a Plotly Dash dashboard, to allow real-time analysis of the sensor data, proactive fault detection, and trend tracking. This study concludes that DL models can effectively enable fault prediction applied to industrial machinery with operational, explainable, and deployable solutions offering the best opportunities to improve maintenance decision making process and enhance the fault prognosis.