EMPIRICAL ANALYSIS OF MACHINE LEARNING MODELS FOR PREDICTING EQUIPMENT FAILURES USING IoT SENSOR DATA
Abstract
This article examines the problem of detecting and predicting industrial equipment faults using IoT sensor data through machine learning techniques. Sensor readings such as temperature, vibration, pressure, voltage, and current, as well as FFT-based features, were statistically analyzed. Class imbalance and low signal informativeness were identified as key factors limiting model accuracy. Results obtained from Logistic Regression, Random Forest, and XGBoost models were comparatively evaluated, showing that when ROC-AUC values remain around 0.5, distinguishing fault and non-fault states becomes challenging. Correlation and feature-importance analyses confirmed the absence of strong dominant indicators. The findings highlight the need to improve sensor architecture and apply targeted feature engineering techniques. This study demonstrates that in predictive maintenance, data quality is more critical than model complexity.