Member-only story
Colocation Data Centers: An Overview of Predictive Maintenance (Part 3 of 3)
There are numerous traditional and deep-learning models used to predict potential failures and detect anomalies. This post compares a few of them. You can also explore pre-trained models and research papers for a deeper understanding.
The above is a sample training pipeline. You can use any of the MLOps frameworks, such as MLflow, Kubeflow, or any cloud provider’s AI platform. Common MLOps practices ensure reproducibility and repeatability and involve storing and versioning artefacts (machine learning models, training configurations, and evaluation metrics).
Please also check Part 1 and Part 2 which discusses the data collection and data pipeline.
Model Selection
Traditional models like Random Forests and Isolation Forests help detect anomalies in the data.
Random cut forest /Isolation Forest
Unlabeled data: RCF can work with normal data only, but including anomaly data can improve performance. If RCF is trained on a mix of normal and anomaly data, it would require labelled data to differentiate between the two. Isolation Forest is an unsupervised model.