Feature Extraction and Reduction Applied to Sensorless Drive Diagnosis

In this paper, it is proposed a feature selection procedure based on Linear Discriminant Analysis. The aim behind this approach is to obtain a minimal set of features still enabling a separation between a number of different classes. Additionally, the reduced number of features implies faster computation and enables resource-limited hardware implementations for real-time signal processing applications. Also, incorporating only a small number of features retains the application's interpretability as a feature space of maximum three features can be visualised directly. Due to this, an expert can directly follow a decision system's answer. The proposed method has been evaluated in the context of an electric drive diagnosis application. In this scope, the LDA feature selection approach is at least as good as the benchmarked feature selection methods. When regarding only a minimal number of features, LDA outperforms the other approaches in terms of classification accuracy. As a secondary result. one can see how important a sensible choice of features is. While some arbitrary combinations produce completely inseparable feature spaces. Three are still combinations that can separate the classes even linearly such that no sophisticated classification concept (e.g. SVM) is needed. The authors are aware of the fact that the findings are shown only in the context of one specific application. Based on the work elaborated here, further research towards generalisation of the proposed approach is intended to be carried out. Additionally, the findings shall be examined using classifier concepts different from SVM, such as Fuzzy Pattern Classifiers.