This thesis aims to study and test the federated learning approach to predict the remaining usefullife of operating machines. Federated learning is a method associated with a machine learningprocess to distribute the training across diferent devices and over time. The first applicationof federated learning was the “next word prediction” feature of mobile phones’ keyboards. Thismethodwas used to trainwith user’s typing data at the local device, butwithout sharing the originalinformation, keeping them private. All users benefit from the prediction, without sharingany personal text.
Machine learning processes like artificial neural networks need significant historical data.Remaining useful life prediction usually demands sensors’ data over time and information whenfailures occur. Once the method is adequately trained, the system should be able to estimatewhen the failure mode is developing and when the failure is likely to happen.Federated Learning stands out from traditional machine learning methods by enabling thetraining of machine learning models on multiple clients while keeping the original data private.The results are then aggregated on a central server, without data sharing. One motivationfor companies to ensure data privacy is the General Data Protection Regulation (https://ec.europa.eu/info/law/law-topic/data-protection_en), while limited or intermittentnetwork access of certain assets, such as ships traveling around the globe, further supports thecase for this approach.
The data available for model training is crucial for achieving accurate predictions. For instance,in the case of predicting the remaining useful life of machinery, a group of companiesowning similar equipment could utilize data from each other without sharing the original data.All companies could benefit from more precise predictions by sharing only the training parametersTo test the performance of the proposed Federated Learning approach, publicly availabledata from turbines will be utilized. The goal is to evaluate the model’s accuracy and how differenttraining configurations can affect the predictions. The method could then be expanded totrain on real data across different operators of similar equipment.