Anomaly detection in multivariate time series of drilling data

Abstract

Different clusters of abnormal activities often arise within same temporal domain of drilling operations. This contrasts with employing simplified scenarios, such as anomaly detection models for specific issues like stuck pipe or loss circulation. The dynamic nature of drilling environments demands a holistic framework that can adapt to evolving conditions and detect anomalies beyond predefined categories. There exists a need to evaluate the performance of various data-driven models and achieve a broader applicability of anomaly detection instead of focusing on a single abnormal activity type. This study presents a generalized framework designed to detect anomalous events in drilling operations. Three different unsupervised learning algorithms, principal component analysis (PCA), isolation forest (IF) and lstm-autoencoder (LSTM-AE), are used to determine abnormal patterns and irregularities in the drilling data. PCA is chosen for its interpretability and efficiency in handling high-dimensional data by reducing the dimensionality. IF is selected for its effectiveness in isolating anomalies, making it robust for scenarios where anomalies exhibit distinct patterns. LSTM-AE is employed due to its capability to capture temporal dependencies and nonlinear patterns in time series data. A multivariate time series of drilling data from an offshore well is collected to use in this work. The dataset comprises approximately six months of drilling operations, including drilling and non-drilling periods. The input variables are chosen from the common controllable parameters continuously monitored by the drilling crew. The pre-processed drilling data with input features is organized into groups based on each drilling section. These unlabeled data is split into train, validation and test sets. Cross-validation (CV) is applied into train and validation sets to prevent overfitting and tune the hyperparameters. The test data is used to evaluate the generalizability and robustness of each model. Evaluation metrics such as precision, recall, F1 score, and receiver operating characteristic (ROC) curve are used to select the best-performing model. The thresholds for anomaly identification are determined by analyzing the reconstruction errors from PCA and LSTM-AE models, as well as anomaly scores from the IF algorithm. Subsequently, evaluation metrics are calculated using these thresholds, assigning true and false predictions to each data point. These predictions are then compared with the ground truth data labels obtained from the daily reports. Experimental results show that the proposed approach can effectively predict anomalous events in drilling operations. It is observed that there is a causal link between irregularities in time series drilling data and the occurrence of anomalous events in drilling operations. Implementing the proposed approach in real-time monitoring systems may improve overall efficiency and minimize risks in drilling operations. The presented methodology can be tested with diverse drilling datasets to enhance the generalizability and reduce bias, leading to a more comprehensive and reliable assessment across various scenarios.