The foundational step in effective predictive maintenance involves a rigorous analysis of collected data, aimed at isolating significant sensor variables that are instrumental for the accurate assessment of the health of conveyor components (Bayoumi and McCaslin, 2017). The Pearson product moment correlation test and the Kendall rank correlation test are utilized due to their appropriateness for the continuous ordinal nature of the dataset (Khamis, 2008). Equation 1 represents the Pearson product moment correlation coefficient, a statistical metric that indicates the linear correlation between two variables. Equation 2 defines the Kendall rank correlation coefficient, reflecting the ordinal association between two variables. r x y = n ∑ x i y i − ∑ x i y i n ∑ x i 2 − ∑ x i 2   n ∑ y i 2 − ∑ y i 2 . (1) r x y = Correlation coefficient n = Sample size x i = Individual input sample points indexed with  i y i = Individual output sample points indexed with  i

Kendall rank correlation: τ = 2 n c − n d n n − 1 . (2) τ = Correlation coefficient n = Sample size n c = Number of concordant pairs n d = Number of discordant pairs

Subsequent to the identification of significant variables through correlation tests, our research explores various machine learning models for predictive maintenance, with each model presenting unique attributes and capabilities.

The industrial chain conveyor at the MODU facility, with its unique design and dimensions, is depicted in Figure 1. The conveyor has an elliptical track layout, materials, and dimensions, including the specifics of the ML2 plastic chain and the three-phase asynchronous AC motor. Its primary components are the gearbox, sprocket, and chain, and their design and critical importance is shown in Figures 2, 3. The common failure points include the depletion of gearbox lubricant and the wear of sprocket teeth, which set the stage for a predictive maintenance approach.

The structure of the predictive maintenance (PdM) system is illustrated in Figure 4. The integration, placement, and role of each sensor device—the Schneider inverter, IFM vibration sensor, and Bosch XDK sensor—are shown in Figure 5. The system features a data collection process, frequent collection, a communication protocol (MQTT), and storage (SAS H2 database). An Automation Industrial Computer NIFE 200 is the central controller and IoT gateway, seamlessly integrating sensor data into the predictive maintenance framework.

The Pearson product moment correlation test revealed a robust correlation between the gearbox’s lubricant oil level and various parameters, including temperature readings from the vibration sensor, A_peak, A_rms, V_rms, power, and drive thermal (Figure 8). Concurrently, the analysis showed a moderate correlation between the gearbox’s lubricant oil level and the temperature data collected by the Bosch XDK sensor, alongside line voltage, noise, and crest—a relationship further elucidated in Figure 9.

Conversely, the Kendall rank correlation test demonstrated a pronounced correlation between the lubricant oil condition in the gearbox and both A_rms and drive thermal. Additionally, this test indicated a moderate correlation with V_rms and power, detailed in Figure 10.

Based on these results, the final variables selected for the machine learning model building are A_rms, drive thermal, V_rms, and power. The random forest model demonstrated superior performance, achieving the highest accuracy and the lowest average squared error compared to other models such as decision tree, logistic regression, and gradient boosting. The random forest model’s accuracy reached 100%, significantly outperforming decision tree and logistic regression, which achieved 90% and 88%, respectively. These results were validated through ten-fold cross-validation, with a standard deviation in accuracy of less than 0.02, further highlighting the robustness and generalizability of random forest. The use of multiple models, including random forest, also aids in assessing model robustness and reducing the risk of overfitting (Table 4). The high performance model possessed low average squared error and high accuracy of classification. The accuracy and ROC plots are illustrated in Figures 11, 13 respectively. Figure 12 illustrates the accuracy through the proportion correctly classified versus cutoff point. Figure 13 shows the true positive rate (sensitivity) versus the false positive rate (specificity) for all possible cut-off values. The diagnostic test of the model showed a near-perfect test in which the ROC curve was almost vertical from (0,0) to (0,1) and horizontal to (1,1). Figure 15 indicates a decent separation from random classification, especially for the random forest model. As a result of the analysis, random forest was selected as the superior model based on its outcomes of highest accuracy and lowest average squared error. Although the random forest model achieved the best overall performance, the decision tree model was more interpretable, making it suitable for cases where model transparency is critical, despite its lower accuracy and higher susceptibility to overfitting. Logistic regression, while computationally efficient, struggled with the complexity of multiclass classification in this predictive maintenance task, showing lower accuracy than random forest. These results emphasize the trade-offs between model accuracy, interpretability, and computational cost in predictive maintenance tasks.

In Experiment 2, the Pearson product moment correlation test revealed a significant correlation between the sprocket teeth condition and several parameters, notably the temperature readings from the vibration sensor, motor thermal, drive thermal, temperature data from the Bosch XDK sensor, and A_peak (Figure 13). Simultaneously, the test indicated a moderate correlation of the sprocket’s teeth condition with A_rms, line voltage, V_rms, and current, with these relationships detailed in Figure 14.

Concurrently, the Kendall rank correlation test revealed a substantial correlation between the sprocket teeth condition and both the temperature data from the vibration sensor and motor thermal. Moreover, this analysis highlighted a moderate correlation with A_peak, A_rms, line voltage, V_rms, and current—relationships comprehensively illustrated in Figure 15.

The final variables selected for machine learning model building were temperature data from the vibration sensor, motor thermal, A_peak, A_rms, line voltage, V_rms, and current. The outcome and accuracy of the six machine learning models are summarized in Table 5. The accuracy and ROC plots are illustrated in Figures 16, 17, respectively. From the evaluation of machine learning models, the decision tree and random forest recorded the same outcome of highest accuracy and lowest average squared error. Random forest was selected as the best model based on its feature of randomized feature selection and ability to better generalize the data, which is believed to perform better during deployment.