AI-DRIVEN PREDICTIVE MAINTENANCE AND INTELLIGENT MONITORING OF MECHANICAL SYSTEMS USING MACHINE LEARNING AND IOT

Abstract

It is widely recognized that modern industrial automation depends on real time condition monitoring of the equipment in order to avoid catastrophic failure and ensure that maintenance resources are used in the most efficient manner while minimising unscheduled downtime. In this research paper, an end-to-end predictive maintenance system is proposed for high stress mechanical drivetrain and rotating machinery. The architecture proposed combines an industrial Internet of Things (IoT) edge sensory network and hybrid machine learning and deep learning pipelines. These data include multi-modal sensor telemetry, including tri-axial vibration profiles, acoustic emissions, thermal imaging data, operational load metrics and electrical motor current signatures, all acquired at high sampling rates continuously. Advanced Wavelet Packet Decomposition and fast empirical mode extraction are used to remove signal artifacts and high frequency noise. These statistical, temporal and spectral attributes are reduced to a few-dimensional vector each and each is individually judged by a predictive suite of models including Random Forest regressors, Extreme Gradient Boosting (XGBoost), Support Vector Machines, Long Short-Term Memory (LSTM) recurrent networks, and Deep Residual Convolutional Neural Networks (ResNet-1D). Experimental validations conducted on standard bearing and gear testbeds prove the accuracy of optimized hybrid CNN-LSTM model in multi-class fault classification of 99.14% and Root Mean Square Error (RMSE) in Remaining Useful Life (RUL) prediction of 4.18 operating cycles. The edge-to-cloud telemetry infrastructure has been proven to provide an inference latency of less than 12 milliseconds per monitoring window, making real-time autonomous prognostics feasible in Industry 4.0 applications.