A new adaptable data-driven model approach for applying predictive maintenance in industrial systems through confidence sphere and the HMM-SVM method

This paper presents a novel strategy in predictive maintenance for an industrial system by integrating statistical models, stochastic methods, and machine learning techniques in condition monitoring and predicting equipment health. Most of these models rely exclusively on either supervised or unsupervised learning and remain static once trained, which limits their adaptability to new operating conditions. Moreover, many industrial datasets lack ground-truth labels, making the direct application of supervised methods challenging. This paper addresses these gaps by presenting a novel strategy that combines statistical feature extraction to enhance the information, Principal Component Analysis (PCA) for dimensionality reduction, unsupervised learning through K-means clustering and Hidden Markov Models (HMM) to identify abnormal patterns in sensor data, and Support Vector Machines (SVM) for real-time supervised classification of health states. The integration of K-means enables the segmentation of the data into groups that represent observable operating conditions of the equipment. The confidence sphere technique acts as a dynamic alert mechanism, enabling the model to evolve with new operational behaviors. Applied to an industrial case study of a drying press in the paper industry, the framework achieved an accuracy of approximately 95% in a supervised model used to classify the three equipment health states (Good Functioning, Alert, and Failure). By enabling continuous monitoring and automatic retraining through the Confidence Sphere, the methodology reduces downtime and maintenance costs while providing predictive insights. These results demonstrate the effectiveness, scalability, and adaptability of the proposed approach for predictive maintenance in Industry 4.0. • Classification and labeling of equipment health states without prior technical information. • Development of an adaptive data-driven model with confidence sphere retraining. • Real-time prediction of health states using a supervised AI model. • Generic model applicable without prior equipment knowledge for predictive maintenance. • Scalable framework usable across different equipment and sensors.