Sustainable production proposed solution based on predictive maintenance using machine learning an‎d deep learning techniques

5/17/2027 12:00:00 AM

With Industrial Revolution 4.0ʹs emergence an‎d intelligent manufacturing systems, increasing amounts of industrial sensor data can be employed to predict machine failure an‎d to increase operational reliability. Therefore, predictive maintenance is an important strategy for minimizing downtime. However, traditional machine-learning models have limited effectiveness when applied to predictive maintenance because of various challenges, including severe class imbalance, infrequency of failure events, an‎d temporal complexity within multivariate industrial time series data. This dissertation proposes a deep-learning-based predictive maintenance model framework for predicting infrequent machine failures, utilizing the Microsoft Azure Predictive Maintenance dataset. A unified, multi-source dataset consisting of telemetry data, failure logs, maintenance records, error logs, an‎d machine metadata was consolidated, then processed through a series of preprocessing, temporal feature engineering, an‎d event-based sequence modeling procedures. The impact of rolling window size on temporal feature scope an‎d model performance was eva‎luated through a comparative analysis of two rolling window sizes (10-hour an‎d 20-hour). Experimental results indicated that while the Ran‎dom Forest model produced exceptionally high accuracy rates (0.9871–0.9903), it generated significantly more false negatives than the deep-learning models. The BiLSTM model was able to improve the detection of temporal failure events using sequential learning, while the hybrid CNN–LSTM model produced the most reliable performance overall, with high recall rates an‎d the lowest false negative rate for all experimental results. Additionally, the experimental results demonstrated that increasing the rolling window enhances feature stability for classical models, while the deep learning architecture provides robustness through its ability to capture long-term degradation trends directly from time series sequences. Thus, hybrid temporal deep-learning models that incorporate imbalance-aware learning an‎d event-based sequence constructs provide a robust, industry-relevant solution for early-warning failure detection in highly imbalanced multivariate industrial time-series environments.

يادگيري تقويتي عميق , يادگيري عميق , شبكه عصبي كانولوشن , انتقال يادگيري با نمايش , سيستم‌هاي سايبري-فيزيكي

Deep Reinforcement Learning , Deep Learning , Convolutional Neural Network , Transfer Learning with Demonstration , Cyber-Physical Systems

Nooruldeen Alkhulaifi

Dr. Hassan Naderi