The convergence of Artificial Intelligence (AI) and the Internet of Things (IoT) has transformed predictive maintenance and operational efficiency in the manufacturing sector. This dissertation presents an advanced deep learning-driven methodology for...
The convergence of Artificial Intelligence (AI) and the Internet of Things (IoT) has transformed predictive maintenance and operational efficiency in the manufacturing sector. This dissertation presents an advanced deep learning-driven methodology for assessing, predicting, and monitoring the operational states of manufacturing equipment within the framework of smart manufacturing. As Industry 4.0 evolves, the role of AI in enabling real-time monitoring and predictive analytics has become critical. This study aims to develop a robust AI-powered system capable of predicting the real-time operational status of manufacturing equipment, identifying anomalies, and facilitating predictive maintenance.
To achieve these objectives, the research employs state-of-the-art deep learning models tailored for time-series data analysis, including Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU). These models classify operational states - ‘stopped,’ ‘idle,’ and ‘running’ - and are seamlessly integrated into a cloud-based IoT architecture to enable real-time data acquisition, processing, and visualization. The study systematically collects and analyzes time-series data from six distinct types of manufacturing equipment. The dataset, spanning from October 2023 to January 2024, was obtained using advanced power measurement devices (GEMS3520), LoRaWAN Gateway, and the Thingplus cloud-based IoT platform. The data encompasses 18 critical electrical parameters, including voltage, current, active power, power factor, and cumulative energy consumption.
A rigorous feature selection process identified key determinants of operational states, including active power (f1_watt), power factor (f1_PF), current unbalance rate (f1_unbal), and cumulative active power difference (f1_kwh_imp_diff). To address data imbalance, the study incorporates advanced optimization techniques such as Focal Loss and employs the ‘Optuna’ library for hyperparameter tuning. The AI modeling process leverages PyTorch as the primary deep learning framework and integrates Bayesian optimization, grid search, and random search methodologies to fine-tune model performance. Evaluation metrics - including Confusion Matrix, Accuracy, Precision, Recall, F1-score, and receiver operating characteristic–area under the curve (ROC-AUC) - were utilized to assess model effectiveness. The optimized models demonstrated exceptional robustness, achieving an average classification accuracy of 96.87%, underscoring their reliability in predicting equipment states with high precision and minimal error rates, outperforming traditional statistical and machine learning models such as Multinomial Logistic Regression and SVM.
A significant contribution of this research is the development of a cloud-based real-time equipment monitoring system, leveraging IoT sensors and LoRaWAN communication for efficient data transmission. This system processes real-time operational data, predicts equipment states, and displays actionable insights through an intuitive, user-friendly dashboard. The dashboard facilitates real-time anomaly detection and provides automated alerts to enable proactive interventions, mitigating potential equipment failures and minimizing downtime. A key innovation of this research lies in its emphasis on accessibility and cost-effectiveness, particularly for small and medium-sized enterprises (SMEs) without existing MES infrastructure. Unlike traditional predictive maintenance systems that necessitate extensive infrastructure integration, this solution relies solely on power sensor data, significantly reducing deployment complexity and costs.
The experimental results highlight the superiority of the developed deep learning models over conventional predictive approaches. The RNN-based architectures effectively capture the nonlinearities and temporal dependencies inherent in time-series manufacturing data, leading to more precise and reliable predictions. Furthermore, the integration of a cloud-based IoT infrastructure enhances the system’s scalability, making it a viable and cost-effective solution for broader industrial adoption.
This dissertation presents a comprehensive and scalable framework for implementing predictive maintenance in smart manufacturing environments through the fusion of deep learning-based AI models and cloud-based IoT systems. The research not only enhances the operational efficiency of manufacturing equipment but also establishes a versatile foundation applicable across various industrial domains. The findings of this study contribute significantly to the ongoing digital transformation of manufacturing, offering practical insights that facilitate the widespread adoption of AI-driven predictive maintenance systems across industrial sectors.