Cutting Downtime in Half with Deep Neural Networks

The base CatBoost model employed an autoencoder mod- el as a feature input, and to refine our approach, we lever- aged domain expertise from subject matter experts (SMEs) and conducted Physics-driven failure mode analysis. We further customized the model by incorporating informa- tion about each specific machine’s operating conditions and insights gathered from other machines to improve the model’s accuracy. Fractal deployed this solution on Azure, using Azure Data Lake for storage, Databricks for data analysis, and Azure Machine Learning for model training. Our offering: Engineered ESP success through predictive maintenance precision

TOOL

AIM

RESULT

To differentiate between normal and deviated equipment states

Improved accuracy in predicting the condition of equipment Enhanced capacity to predict equipment conditions An accurate representation of the intricate conditions that equipment may face Improved accuracy of the model in addressing ESP system conditions A significant boost in the precision and timeliness of equipment condition predictions

Deep Neural Network

An input to the primary CatBoost model

Autoencoder Model

To gain insights from SMEs and physics-driven failure mode analysis

Domain Expertise & Analysis

To learn from other similar machines to enhance model accuracy

Customized Machine Conditions

To increase precision and timeliness in predictions

Additional Model Enhancements

© 2023 Fractal Analytics Inc. All rights reserved

5

Made with FlippingBook - PDF hosting