Federated Deep Learning Approach for Predictive Failure Detection in Distributed Automotive Manufacturing Parts

Advanced technologies, such as automation, robotics, and sensor-enabled IoT devices, play a crucial role in the automotive industry. They provide essential predictive analytics and help identify issues to be addressed before they grow larger. While there are challenges like real-time decision-making, communication delays, and data security, these aspects present opportunities for improvement within centralised machine-learning frameworks. This work presents an exciting Federated Deep Learning (FDL) architecture designed for predictive maintenance in distributed manufacturing. This innovative setup features a predictive maintenance model that is collaboratively trained by a group of plants, ensuring that their sensor data remains confidential and privacy is upheld. It features exciting predictive maintenance innovations that blend various strategies within the realm of CNN, LSTM, and Autoencoder architectures. The study explores benchmark data sets, including the NASA Bearing and SECOM, along with data from real-world CNC spindles, robotic arms, and simulated automobile sensor data. The FDL approach reached an impressive 91.8% accuracy and a 0.91 F1 score, closely matching the performance of centralised predictive deep learning models, which achieved 92.5% accuracy and a 0.91 F1 score. Additionally, the FDL approach surpassed traditional models like random forest at 88.2% and SVM at 85.7%, showcasing its effectiveness. The FDL approach has truly excelled beyond traditional machine learning models, showcasing the impressive predictive performance of Federated Deep Learning.