Distributed Machine Learning Training in Oil and Gas Industry via Federated Learning

Machine Learning (ML) solutions are becoming a cornerstone of reliable data interpretation and various task automation in Oil and Gas (O&G) exploration and production. The conventional approach in ML training requires centralizing the entire training dataset, which presents challenges like data privacy, residency, scalability, heterogeneity, and version control. To alleviate some of these challenges, we propose a Federated Learning (FL) approach for distributed ML model training in the industry, particularly in conveyance applications, enabling collaborative learning across geographically diverse datasets and ensuring adaptability to varied sources. Typically, the first step in ML model training is to obtain and centralize data to enable model training on specific computers. In protected industries, data often must reside locally to ensure privacy. In such cases, where data residency is paramount, one can deploy distributed training through FL. In FL approach, ML models are trained locally on diverse computers on their local, domain-specific data and they only periodically communicate their model parameters with the central server for model aggregation. This approach produces a consolidated and aggregated central model that never sees the full data but leverages data from all the sources for a globally generalized ML model. We implemented FL deployment on geographically diverse computers where central server initializes and distributes global model parameters to each of the client computers where the private data resides, and each local client computer trains only on their datasets. After local training, the parameters of each local model are transmitted to central server for aggregation over the internet through Virtual Private Network (VPN) access and aggregated global model is redistributed back to the clients for further training if needed. The Pythonic FL framework makes the training process model-agnostic, allowing a variety of ML models to be trained with minimal changes. We demonstrate two successful examples in the context of wireline conveyance automation: one involving the training of an XGBoost classifier to predict high-tension events in pre-job scenario using specific job parameters such as cable properties, well and fluid information, tool specifications etc. and the second focuses on cable tension prediction using multi-layer perception (MLP) as a regression problem.