Test-Training Leakage in Evaluation of Machine Learning Algorithms for Condition-Based Maintenance

Many articles have been published utilizing machine learning algorithms for condition-based maintenance through the analysis of vibration signals. One extensively researched topic is the classification of fault types in rolling bearings. There is a fairly widespread problem in the evaluation of these learning algorithms, where the separation of examples between the test and training sets is incorrect, leading to an optimistic conclusion about the algorithm's performance even when it is not the case. In this article, we will review this issue and explain how the data should be properly divided between the test and training sets to avoid this occurrence.