Machine Intelligence Based Health Condition Prediction for Industrial Machinery in Multilevel Fault Environments

Abstract

Industrial Machinery (IM) has a crucial contribution in current manufacturing processes and industrial production. Continuous condition monitoring is necessary for providing reliability during operations. In addition, it is very important for avoiding sudden malfunctions of machines and reducing maintenance expenses. Despite the fact that predicting the health of machines under different defect conditions is very complicated because of the variability of the defects. Therefore, this paper proposes a machine intelligence (MI)-based system for estimating the health of two specific industrial machines. These machines include the water pump (WP) and milling machines (MM). The proposed framework employs a hybrid MI model by integrating CatBoost (CTB), Decision Tree (DCT), and Naïve Bayes (NIB) to enhance classification performance. The effectiveness of the proposed framework is evaluated against several well-established MI models, including CTB, AdaBoost (ADB), XGBoost (XGB), DCT, Neural Network (NNW), Stochastic Gradient Descent (SGD), Random Forest (RNF), NIB, K-Nearest Neighbour (KNN), and Support Vector Machine (SVM). Experimental evaluation is conducted using both 5-fold cross-validation and 80%:20% training:testing strategies on WP and MM datasets. The performance is assessed using classification accuracy (CAC), F1-score (F1), precision (PRS), and recall (RL). For the MM dataset, the proposed framework achieved CAC, F1, PRS, and RL values of 93.8%, 93.7%, 93.8%, and 93.7%, respectively, using 5-fold cross-validation, which further improved to 95.9%, 95.8%, 95.9%, and 95.9%, respectively, under the 80%:20% training:testing strategy. Similarly, for the WP dataset, the proposed framework obtained CAC, F1, PRS, and RL values of 95.4%, 95.2%, 95.3%, and 95.4%, respectively, using 5-fold cross-validation, and 97.2%, 97.2%, 97.1%, and 97.1%, respectively, under the 80%:20% training:testing strategy. One-way ANOVA further demonstrated statistically significant differences among the competing models for both datasets under the two evaluation protocols (F = 28235.80, 23522.51, 24171.49, and 23199.06, p < 0.001), confirming the statistical significance of the proposed framework. Descriptive statistical analysis indicated stable classification performance with low standard deviation across all experimental settings, while the percentage improvement analysis showed that the proposed framework outperformed the second-best model (CTB) by 1.80%–2.29%. Overall, the proposed framework consistently outperformed the competing MI models across all evaluation metrics, demonstrating its effectiveness and reliability for accurate multi-level fault classification, intelligent health condition prediction, and predictive maintenance of industrial machinery. All simulations and experimental analyses are implemented using Python in Jupyter Notebook 7.