ANOMALY DETECTION IN IOT-BASED MANUFACTURING USING MACHINE LEARNING AND PROCESS MINING: AN INTEGRATIVE TECHNIQUE (FRAMEWORK)

Abstract

This study proposes and empirically evaluates a hybrid approach of Machine Learning (ML) and Process Mining (PM) to solve the problem of anomaly detection in IoT-enabled manufacturing systems in a way that optimises simultaneously the accuracy of the fault diagnosis, its interpretability, and its operational reliability. The framework combines data-driven statistical intelligence and process-aware behavioural reasoning to find deviations in real-time industrial workflows by using a dataset from the Tennessee Eastman Process (TEP) and a Kaggle oil-and-gas sensor proxy. The experimental pipeline was built in Python 3.11, with the use of scikit-learn, TensorFlow and PM4Py on a Google Colab GPU environment. The five ML models were trained and validated with the stratified 5-fold cross validation technique, namely, Support Vector Machine (RBF kernel), Random Forest, Multilayer Perceptron (MLP), Autoencoder and Isolation Forest. Partitioning models have significantly lagged behind the performance of Random Forest and MLP which classified 100% (Precision = Recall = F1 = 1.000) of the results, thus confirming their robustness in critical fault identification. Process mining, with Directly-Follows Graphs and conformance checking, provided contextual knowledge about the deviations in the process: 76.9% of process traces contained micro anomalies in otherwise normal process behavior. A hybrid fusion model that added the ML and PM outputs by a conservative logical AND rule had perfect precision (1.000) and no false positive but at the expense of lower recall (0.133) which is a trade-off suitable for safety-critical applications where false alarms have a high cost. The system's robustness and resilience were tested under the assumption of sensor noise following a Gaussian distribution and simulated data losses, and the results demonstrated resilience (F1 ≥ 0.95) for up to 10% missing data and end-to-end latency of inferences was kept under millisecond to sub-second, supporting the real time deployment of the system. The study presents an interpretable, precision optimised architecture that integrates the predictive power of ML with the behaviour transparency of PM, providing a methodological basis for digital twins and predictive maintenance in Industry 4.0 manufacturing systems that can be scaled.