Predictive Maintenance Application on Machine Overstrain Failure with Node-Red and Isolation Forest Anomaly Detection

Predictive maintenance on machines includes the detection of many failures, including machine overstrain failure. However, applying anomaly detection in machine predictive maintenance is still a research opportunity. Our research aim is to evaluate the performance of several anomaly detection methods in detecting overstrain failures. We obtained the dataset from Kaggle, where its features include machine temperature and rotational speed. We pre-processed the dataset, among others, by applying a mean decrease in impurity (MDI) feature selection. We then implemented and compared three advanced anomaly detection methods from the last three years: one-class support vector machine (OCSVM), local outlier factor (LOF), and isolation forest (iForest). We use several metrics for the comparison, namely the receiver operating curve (ROC), the area under the curve (AUC), geometrical mean (G-Mean), sensitivity, and specificity. Lastly, we designed and built a predictive maintenance application using Node-Red. The test results prove that the MDI feature selection selects two features: “Torque” and “Tool Wear;” for the anomaly detection model. iForest is the most optimal anomaly detection method with an AUC of 0.96. Through ROC optimization, it is found that iForest’s optimum threshold is 0.56. At this optimum threshold, the sensitivity, specificity, and G-Mean of iForest are 0.99, 0.90, and 0.95, respectively. Finally, our Node-Red dashboard displays “Torque,” “Tool Wear” and anomaly detection data. The application can also respond to anomaly detection.