Unsupervised Learning for Bearing Fault Identification with Vibration Data

Machine learning methods are increasingly used for rotating machinery monitoring. Usually at system set up, only data of the machinery in healthy conditions, the so-called nominal data, are available for the machine learning phase. This type of training data enables fault detection capabilities and several methods such as Gaussian Mixture Model, One Class Support Vector Machines and Auto Associative Neural Networks (Autoencoders) have been already proved successful for this task. However, in some predictive maintenance applications, information on the type of defect may represent a key element for producing actionable information, e.g. to reduce diagnostic burden and optimize spare procurement. This requires to define classification strategies based on machine learning even in absence of data representing the behaviour of the system with defects. In this study we present an approach that uses only nominal data to train an autoencoder which will enable at same time fault identification and fault classification tasks. An autoencoder is a network which learns to duplicate the input at the output.  Even if this replication task may seem trivial, the presence of a “bottleneck” in the hidden layer forces the network to learn the significant features of the input data. As faulty data are expected to possess information content which is structured differently from the healthy ones their reconstruction at output will result inaccurate.  In conventional anomaly detection approaches, the module of the reconstruction error, defined as the difference between output and input, is uses to determine an unusual input such as faults.   The proposed approach represents a step forward as here a single autoencoder is used both for detection and classification. The underlying idea is that the components of the reconstruction error vector whose module is used to trigger fault identification in classical autoencoder approaches contain the information of the fault type. This way the analysis of the different components of the reconstruction error allows to differentiate the different types of fault. Two methods to analyse the components of the reconstruction error vector will be presented. In the first strategy autoencoder input features more sensitive to each kind of fault are identified on the basis of a-priori knowledge and an expected ranking pattern associated to each fault type is defined. Once a new input signal is presented to the autoencoder, in case the module of the reconstruction error detects an anomaly, the ranking of the magnitudes of reconstruction error components is established and the Rank Biased Overlap (RBO) algorithm is used to measure the similarity between the computed data and the expected fault templates.  This method can be easily generalised to other cases, since it is not the amplitudes of the errors that serve as a reference but only the ranking. The latter method plots the reconstruction error components on a polar diagram and uses their position to compute the position of the barycentre of resulting star-shaped figure. The angular position of the barycentre is then used to classify the fault type. A machine fault simulator has been used to generate 3 different types of bearing defects with different load, speed and noise conditions and a dataset of about 10000 signals has been employed to test the classification algorithms. The results obtained using the autoencoder method do not achieve the same performances as the conventional supervised learning algorithms. However, they proved to be 88% accurate in classification when SNR is above 0dB with the ranking based method overperforming the barycentre one.