A data-driven cost-sensitive sorting method with active learning for ordinal risk classification of failure modes

With the advancement of technology and the complexity of service environments, more and more failure mode related data can be recorded and used in Failure Mode and Effect Analysis (FMEA). However, it is time-consuming and costly to invite experts to annotate and fully rank a large amount of failure mode data. In this case, the ordinal risk classification is required. On the one hand, the ranking results of traditional FMEA may be unreliable due to its sensitivity to the changes in risk factors. On the other hand, managers often pay extra attention to high-risk failure modes as they can cause greater losses. Therefore, we propose a data-driven cost-sensitive FMEA sorting (FMEA-SORT) method, which not only can select a small number of representative samples from massive failure mode data for expert annotation via active learning, but also obtains the weights of risk factors and ordinal risk classification model of failure modes based on data features and classification instances. Firstly, with the aid of clustering trees, we develop an active learning method to select high-quality representative points from preprocessed failure mode data for expert annotation, thereby getting the risk levels of representative points under various risk factors. Secondly, considering the importance of different risk levels, the semi-supervised cost-sensitivity decision trees are designed to obtain the local limiting profiles under each risk factor. Based on the available ordinal classification instances of failure modes, the weights of risk factors and the global limiting profiles of ordinal risk classes are further determined by constructing the minimum misclassification cost optimization model. Then, the ordinal risk classification rules for failure modes are established. Finally, the comparison and sensitivity analysis are designed to verify the effectiveness of FMEA-SORT based on the data from the existing literature and UCI datasets.