Abstract:
Faults in bearings usually manifest as marginal defects that intensify over time, allowing for well-informed preventative actions with early Fault Detection and Diagnosis (FDD) protocols. Detection of the fault begins with capturing, for example, acceleration signals from a machine. Traditionally, handpicked descriptive statistical features (mean, RMS, skewness, kurtosis, etc.) or spectral diagrams obtained from these signals are then used for FDD. However, machine signals are often generated under non-stationary operating conditions of varying loads and speeds, requiring further intervention. More advanced signal processing techniques (spectral kurtosis, or cyclostationary analysis) are hence used to account for the non-stationarity of the signal. This is usually done by separating acceleration signals into deterministic and random components. Fault detection in bearings is possible by observing the random components of the signal.
A wealth of research has been invested in machine learning-based techniques to circumvent the problems associated with non-stationary signals. Many of these methods require vast amounts of historical data to train. Machines typically spend most of their life operating in a healthy condition, therefore, most historical data is occupied with data that comes from a healthy machine condition, training these methods is difficult, due to the shortage of data from a machine running in an unhealthy condition. Furthermore, well-performing machine learning algorithms still require a domain expert to extract features that are known to be fault sensitive. Deep learning is a recent approach in data analysis whereby feature extraction is incorporated within the training of the algorithm. The algorithm is given the ability to find and extract its features. The architecture of the algorithm allows for the extraction of complex hierarchical non-linear features. To the author’s knowledge, no attempt has been made to make full use of the power of deep learning together with the known structure of bearing acceleration signals to perform FDD.
In this work, a bearing FDD methodology is developed using deep learning approaches. A model based on Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) is used to learn a lower-dimensional representation of an acceleration signal. A regularization strategy based on information maximization is used, which allows deterministic and random components of the signals to be learned separately. This representation is subsequently used to perform bearing FDD. The algorithm is trained in a completely unsupervised manner on exclusively healthy data and requires no preprocessing of that data. Furthermore, no auxiliary signals such as a shaft encoder, which contains information about the machine operating condition, is required for the algorithm to work. The methodology was tested on well-known benchmark datasets, and it was shown to be robust against non-stationary operating conditions. The algorithm can learn its fault metric and by observing the trajectory of the signal representation, it is also able to diagnose the type of fault.
