Novelty class detection in machine learning-based condition diagnosis

Industrial plant machines have a significantly lower frequency of defective data than the frequency of normal data. For this reason, machine learning is often applied using only some obtained state data. However, the low frequency of defect data does not guarantee that novel data occur, which is why detection of novelty class is required. This paper studies the novelty class detection method in multi-classification. Multi-class support vector machine was used for multi-classification. Cluster-based local outlier factor, histogram-based outlier score, outlier detection with minimum covariance determinant, isolation forest, and one-class support vector machine applied novelty class detection. Anomaly detection algorithms used the hard voting ensemble method. A feature engineering method that is advantageous for novelty class detection was confirmed by comparing the genetic algorithm (GA)-based feature selection and principal component analysis (PCA). Findings show that creating a model using GA-based feature selection for multi-classification and independent PCA for each class for novelty class detection is advantageous. With the use of an independent PCA, the problem was simplified to perform detection on a novelty class with a condition similar to the trained class.