MATE Notifications
This project corresponds to AI Platform's Generative AI & Knowledge AI MLOps journey.
Project Goals
MATE is acronym for Maintenance Analysis Tool for Equinor, which aims to increase safety, production, barrier control and reduce maintenance costs. It extracts the free text from notifications held in Omnia classify such as sentence descriptions, equipment descriptions and how they failed. The Natural Language Processing (NLP) model is then trained to make predictions on their failure class or failure modes. There are total 5 failure classes and 70+ failure modes. The predictions are then loaded back to the database.
Two main types of identification tasks to be performed are:
(1) Machinery Failure Modes: The purpose of this algorithm is to analyze the text written (and form fields selected) by maintenance engineers in the event of a machine failure and correctly labelling the failure mode based on the description. The key objective is to optimize accuracy.
(2) Machinery Failure Classes: this is a multilabel analysis, with a focus on identifying failures of class DU (Dangerous Undetected). The key objective is to optimize recall. The priority of maintenance engineers is to understand and predict when a dangerous failure has gone undetected in order to take relevant action.
You can find more information about this tool at:
https://statoilsrm.sharepoint.com/sites/OMNIA.aurora/Shared%20Documents/General/1%20-%20Projects/Maintenance%20Analysis%20Tool%20for%20Equinor.pdf#search=mate%20notification%20nlp
Summary of Results
Training examples were found manually to use for labels. The text was then converted to features using tfidf. The XGBoost was used to build classifiers and Optuna was used for Hyperparameter tuning. Various modes and objetive functions were tried to during the tuning process. Predictions were obtained by running a Cronjob in AI Platform's Kubernetes cluster at midnight to pickup new notifications.
Project Team
Jennifer Sampson (Data Scientist)
Bjarte Johansen (Data Scientist)
Peter Koczca (Data Scientist)
Terje Elde (K8 Developer)
MLOps Challenges
Hyperparameter Tuning
The team needed an ability to perform hyperparameter tuning of XGBoost classifier by trying out various objective functions, weights and validation matrices.
Run Predictions on Schedule
During deployment, new notifications needed to be fetched every midnight to be processed by the model to generate predictions on classification of failure classes and failure modes.
MLOps Solutions
Optuna for Hyperparameter Tuning
The XGBoost classifier was trained and refined by using the Optuna Hyperparameter tuning framework.
K8 Chronjobs for Daily Predictions
The Cronjob on AI Platform's K8 cluster is scheduled to fetch new notifications everynight and generate predictions.
GitHub Repos
https://github.com/equinor/mate-notifications-nlp
https://github.com/equinor/mate-notifications-nlp-k8s