{"id":15064,"date":"2022-08-09T00:00:00","date_gmt":"2022-08-09T00:00:00","guid":{"rendered":"https:\/\/www.dmcinfo.com\/our-work\/machine-learning-and-telemetry-analytics\/"},"modified":"2025-05-29T22:47:09","modified_gmt":"2025-05-29T22:47:09","slug":"machine-learning-and-telemetry-analytics","status":"publish","type":"our_work","link":"https:\/\/www.dmcinfo.com\/our-work\/machine-learning-and-telemetry-analytics\/","title":{"rendered":"Machine Learning and Telemetry Analytics"},"content":{"rendered":"

The project consisted of three parts: unsupervised learning, supervised learning, and the creation of a telemetry analytics dashboard. The system was developed using Python<\/a> and Jupyter<\/a>.<\/p>\r\n\r\n

Part 1: Unsupervised Learning – Pattern Search<\/h2>\r\n\r\n

DMC preprocessed the data using K-Nearest Neighbor’s (KNN) imputation, feature scaling, normalization, and\u00a0Principal Component Analysis (PCA).\u00a0DMC then created clustering models to discover the underlying pattern in the raw data and decoded the pattern to produce meaningful data. Later,\u00a0evaluation matrices (Silhouette Score and Davies Bouldin Score) and cluster visualization\u00a0techniques (T-Distributed Stochastic Neighbor Embedding and Principal Component Analysis) were applied\u00a0to evaluate the results.<\/p>\r\n\r\n

Oil wells that differed in condition\u00a0(location, depth, equipment, etc.)\u00a0had\u00a0different telemetry performances. By clustering the well condition data, we could identify if there were similarities between certain well conditions and if\u00a0clusters were substantially different from each other. Analyzing\u00a0the corresponding telemetry performance\u00a0of the clusters\u00a0guided\u00a0us to set up\u00a0wells with\u00a0better telemetry performance.<\/p>\r\n\r\n

Part 2: Supervised Learning – Diagnosis and Prediction<\/h2>\r\n\r\n

After Phase I (Unsupervised Learning), DMC\u00a0created a\u00a0more powerful machine learning model capable of diagnosing\u00a0well conditions and providing guidance on ways to optimize telemetry performance. We designed\u00a0the\u00a0model to predict telemetry performance for newly-acquired sets of\u00a0well condition data.<\/p>\r\n\r\n

To train supervised learning models, the prediction goal (telemetry performance) must match up with the inputs (well conditions). The raw telemetry performance data consists of time traces, so to prepare\u00a0the training data,\u00a0DMC extracted features from the time traces and paired them\u00a0up with corresponding well condition data using Dataiku<\/a>.<\/p>\r\n\r\n

With the training data well prepared, DMC\u00a0preprocessed\u00a0the dataset using\u00a0one-hot encoding, imputation\u00a0on the missing values, and feature scaling. DMC then trained non-parametric models (K-Nearest Neighbors and Decision Tree Regression) and parametric models (Lasso regression, Kernel Ridge Regression) as baselines. Finally, a Deep Neural Network was developed to perform the diagnosis and prediction tasks.<\/p>\r\n\r\n

In order to use the neural network to\u00a0improve\u00a0telemetry performance, DMC designed feature\u00a0importance analysis methods targeting specific diagnostics. DMC used several statistical approaches to identify numeric\u00a0and categorical features, then\u00a0ranked the\u00a0conditions of a\u00a0well to optimize telemetry performance.<\/p>\r\n\r\n

Part 3: Telemetry Analytics Dashboard – Visualization<\/h2>\r\n\r\n

The telemetry analytics web interface\u00a0was designed using Plotly<\/a>\u00a0for the purpose of visualizing the massive dataset, running statistical analysis, and displaying the resulting graphs.\u00a0The dashboard provided a tool for the client to easily visualize the data and obtain information without being exposed to implementation details.<\/p>\r\n\r\n

\"Telemetry<\/figure>\r\n\r\n

The dashboard displayed the oil wells on a map based on their recorded geographic\u00a0locations, allowing users to click\u00a0on a well location to reveal\u00a0detailed information about the\u00a0well. Users could select multiple wells on\u00a0the map or use filters to select wells meeting certain criteria. Based on the\u00a0selected data, the dashboard can\u00a0run statistical analyses\u00a0and display a variety of visualizations (heat maps, word clouds, histograms). Users can also export selected data and graphs and save filter configurations for later use.<\/p>\r\n\r\n

Learn more about DMC’s Test and\u00a0Measurement<\/a> expertise and contact us <\/a>for your next project.\u00a0<\/strong><\/p>\r\n","protected":false},"excerpt":{"rendered":"

The project consisted of three parts: unsupervised learning, supervised learning, and the creation of a telemetry analytics dashboard. The system was developed using Python and Jupyter. Part 1: Unsupervised Learning – Pattern Search DMC preprocessed the data using K-Nearest Neighbor’s (KNN) imputation, feature scaling, normalization, and\u00a0Principal Component Analysis (PCA).\u00a0DMC then created clustering models to discover […]<\/p>\n","protected":false},"author":8,"featured_media":15062,"template":"","meta":{"customer":"Schlumberger","summary":"

DMC’s client in the oil and gas industry needed a tool capable of diagnosing, predicting, visualizing, and analyzing telemetry data. Whereas conventional physical models are limited by human knowledge of the system, machine learning models can learn from data without the intervention of human knowledge. DMC used machine learning to diagnose and predict telemetry performance. The accompanying data visualization and analysis dashboard allowed the client to visualize the data by geographical location and filter selection and display the statistical analysis graphs.<\/p>\r\n","description":"","customer_benefits":"