Uma nova abordagem de inteligência artificial baseada em autoaprendizagem profunda para manutenção preditiva em um ambiente de produção de petróleo e gás Offshore

Abstract

Currently, the industry is undergoing a transition from Industry 3.0 to Industry 4.0, also called the "Fourth Industrial Revolution". This change is associated with the integration between physical and digital systems in manufacturing environments. This integration allows the acquisition, storage, processing and analysis of a large amount of data from various industrial processes. For example, historical data from an industrial production line contains information about processes, equipment, events and alarms that are essential for assertive decision making. Therefore, when processed and analyzed, they can bring valuable information and knowledge about the state of integrity of the assets. In this way, Artificial Neural Networks (ANN) become key technologies for the development of predictive maintenance solutions for Industry 4.0, since they have the ability to process large volumes of data to support decision making. However, labeled data is scarce or even non-existent in the industry, as there is a time constraint on the ability of specialists to process such an amount of data. In this context, the objective of this thesis was to research, study and develop a deep self-learning model to classify failures in multivariate time series. Self-learning is a combination of unsupervised and supervised machine learning. The main advantage is no need for labeled data in the model training step. The proposed approach was developed and validated using real and public data from Petrobras, known as the 3W base. This data has the history of oil and gas (O&G) production failures from offshore wells. Therefore, the development of the model contains four main steps. First, a small portion of unlabeled data from the database was randomly separated. Subsequently, unsupervised machine learning was used to pseudo-label the data through the recognition of anomalous and normal patterns. Then the supervised deep machine learning iterative development cycle begins. In this cycle, the model is trained with pseudo-labeled data. After training, predictions are performed on new unlabeled datasets to generate new pseudo-labels. New unlabeled data is gradually added to each training iteration of the cycle. For that, through the probability data of the Softmax activation function, a Confidence Layer was designed for pseudo-labeling in the most reliable predictions. Finally, the model ends development when the stopping criterion is met and breaks the training loop. As a result, initially, we have a labeled dataset and a model that can be improved with more unlabeled data. Even, through new development cycles, Transfer Learning can be used to teach other types of failures to the model and transform it into a multiclass. Therefore, the proposed model was able to classify the imminence of multiple O&G failures in multivariate time series. The semi-supervised deep self-learning model performed on par with supervised deep learning. The semi-supervised and supervised models showed an accuracy of 96% and 99%, respectively. Thus, the proposed model presented resources to enable machine learning in an industrial operational environment, since the lack of annotated data was overcome.