Predictive Maintenance Experiences on Imbalanced Data with Bayesian Optimization Approach

Abstract

Predictive maintenance solutions have been recently applied in industries for various problems, such as handling the machine status and maintaining the transmission lines. Industrial digital transformation promotes the collection of operational and conditional data generated from different parts of equipment (or power plant) for automatically detecting failures and seeking solutions. Predictive maintenance aims at e.g., minimizing downtime and increasing the whole productivity of manufacturing processes. In this context machine learning techniques have emerged as promising approaches, however it is challenging to select proper methods when data contain imbalanced class labels.

In this paper, we propose a pipeline for constructing machine learning models based on Bayesian optimization approach for imbalanced datasets, in order to improve the classification performance of this model in manufacturing and transmission line applications. In this pipeline, the Bayesian optimization solution is used to suggest the best combination of hyperparameters for model variables. We analyze four multi-output models, such as Adaptive Boosting, Gradient Boosting, Random Forest and MultiLayer Perceptron, to design and develop multi-class and binary imbalanced classifiers.

We have trained each model on two different imbalanced datasets, i.e., AI4I 2020 and electrical power system transmission lines, aiming at constructing a versatile pipeline able to deal with two tasks: failure type and machine (or electrical) status. In the AI4I 2020 case, Random Forest model has performed better than other models for both tasks. In the electrical power system transmission lines case, the MultiLayer Perceptron model has performed better than the others for the failure type task.