Machine Learning Strategy for Fault Classification Using Only Nominal Data

Abstract

Machine learning methods are increasingly used for rotating machinery monitoring. Usually at set up, only data associated to an engine in a good state, the so called nominal data, are available for the machine learning phase. Nevertheless a classifier requires faulty data to be trained at identifying the causes of the anomalies and this fact has generally limited the usage of data driven approaches to fault detection tasks. The paper suggests a strategy to use machine learning methods even for fault classification purposes and diagnostics. Within the proposed framework three different machine learning methods, Gaussian Mixture Model (GMM), Support Vector Machines (SVM) and Auto Associative Neural Networks (AANN) have been implemented, tested and compared. The idea is to take into account some ‘a priori’ knowledge about the faults to be classified, to drive the behavior of the machine learning methodology (SVM or AANN or GMM) to be more or less reactive to the different faults. The indicators (features) more sensitive to each kind of fault are firstly selected on the basis of expert knowledge. For each different fault, a set of indicators is defined and computed from nominal data only. Each set is then used to produce training data for one specific fault. Such data sets are then used to train one instance of each method for each different fault. The underlying logic is that fault tuned input data is able to produce fault tuned instances of the methods. For example the instance trained with the indicators associated to a fault ‘A’ reacts more powerfully in presence of the fault ‘A’ than the others. Once an anomaly is detected, the comparison among the reactions of the different ‘fault tuned’ instances allows classifying the fault, not just to detect it. The results show best detection performances for SVM whilst AANN outperforms the other two methods for classification.