Analyzing internet of things (IOT) data using federated learning

With the impact of Industry 4.0 and digital transformation processes, the need for data-driven decision-making mechanisms in the manufacturing sector has significantly increased. Anticipating equipment failures on production lines, detecting quality deviations in a timely manner, and enhancing process efficiency have become fundamental objectives of modern industrial systems. In this context, big data analytics and AI (artificial intelligence)-based decision support systems contribute substantially to making production processes more predictable and adaptive. However, these technological advancements also bring forth challenges such as data security, privacy concerns, and compliance with regulatory frameworks. Traditional machine learning approaches typically require the aggregation of large datasets on centralized servers, which poses considerable security risks, increases communication overhead, and imposes heavy computational burdens—especially in environments dealing with sensitive data. As a solution to these limitations, FL (Federated Learning) has emerged as a decentralized learning paradigm that allows model training on local devices without transferring raw data to a central location. This thesis investigates the application of FL in analyzing sensor data collected through the OPC (OLE (Object Linking and Embedding) for Process Control) protocol on production lines. In the proposed system, parameters such as temperature, pressure, and speed are processed locally at the client side without being transmitted to a central server. Instead, only the parameters of locally trained models are aggregated on a central server to construct a global model. To enhance data privacy, the architecture integrates differential privacy techniques and secure model aggregation protocols. Furthermore, optimization techniques such as gradient compression and sparse updates are implemented to reduce communication costs. The model's performance is evaluated using conventional classification metrics such as accuracy, precision, recall, and F1-score, as well as system-level metrics including latency, network traffic, computational efficiency, and generalizability under data heterogeneity. Experimental results demonstrate that the proposed FL framework delivers consistent and reliable performance across both small- and large-scale client scenarios. Moreover, the asynchronous communication structure enables successful model training even on low-resource client devices. In addition, the system's data preprocessing module employs advanced techniques including missing data imputation, timestamp synchronization, and anomaly detection to ensure that sensor data is properly prepared for modeling. The developed architecture offers a flexible structure adaptable to various production processes and provides a secure, real-time analytics infrastructure. In conclusion, this thesis presents a comprehensive implementation of Federated Learning in the manufacturing domain, illustrating how distributed AI solutions can be applied without violating data privacy. The proposed system enables scalable, secure, and high-performance analytics in multi-location industrial environments without relying on centralized data collection. These findings highlight the potential of FL not only as an academic concept but also as a critical component of future digital transformation initiatives in the industry. Future studies will aim to extend the system's applicability through testing with larger datasets and real-world industrial conditions.