On Wednesday, June, 3rd ,2026 the department of Mechatronics engineering at Al-Khwarizmi College of engineering held a master thesis defense titled “Distributed Split Learning Architecture Based on Industrial Internet of Things for Predictive Maintenance)),
for the master candidate (Hanen Hussein Adel) at the thesis defense hall, where a degree of passing with distinction was awarded for her research findings.
The committee is composed of the following members:
- Professor Dr. Ameer Hussein Murad, Chair
- Professor Dr. Waleed Khalid Khairi, Examiner
- Assistant Professor Dr. Azzah Abdel Razzaq Abdelkarim, Examiner
- Professor Dr. Osama Fadhil Abdullatif, 1st Supervisor
- Assistant Professor Dr. Ali Hussein Hamad, 2nd Supervisor
This thesis aimed to develop advanced procedure for predictive maintenance using data in industrial internet of things, since it represents a key enabler for improving equipment reliability and operational efficiency in Industrial Internet of Things (IoT) environments. However, centralized prediction-based artificial intelligence (AI) requires raw data to be uploaded to the server, raising privacy, scalability, and communication concerns. Also, a deep learning model on IoT nodes requires substantial computation.
This thesis proposes a distributed PdM framework based on split learning (SL) and split federated learning (SFL). It compares the performance with that of a centralized baseline deep neural network (DNN) and SCINet models.
The experimental setup consists of three AC motors equipped with vibration, current, and temperature sensors, connected to Raspberry Pi devices via the I2C protocol to enable efficient, synchronized data acquisition. The collected dataset includes normal operating conditions and multiple simulated fault types, such as high vibration, overcurrent, overheating, overload, and stop-rotation, to emulate real industrial scenarios. Data preprocessing techniques such as normalization and labeling were applied before model training.
Experimental results showed that SCINet achieved 99.35% with the lowest test loss while SL model achieved a testing accuracy of 99.46%, as well as the proposed SL model demonstrated significant generalization and best performance. SFL model, on the other hand, offered distributed scalability and improved privacy despite a minor decrease in the accuracy, achieving testing accuracy (97.89%).
Overall, the results confirmed that SL provides near-optimal predictive performance, while SFL offers a practical, privacy-preserving solution for distributed predictive maintenance in industrial environments.
