Towards autonomous condition monitoring sensor systems

Abstract: Rolling element bearings are used to carry load and reduce friction between moving parts in rotating machines, which play a central role in society and industry, for example in the transportation and energy sectors. It is essential to monitor and maintain the condition of bearings such that machines can operate efficiently and any failures resulting in unplanned stoppages are avoided. Therefore, bearings with embedded sensing capabilities are becoming increasingly common, which makes it possible to consider bearings to be sensor systems that can monitor the condition of rotating machines. However, the task of automatically analyzing the signals is challenging because machines are different and evolve over time; moreover, the complexity of the signals, machines and possible failure modes is high and costly to accurately predict and model. Therefore, the use of unsupervised machine learning methods for the automated analysis of such signals and the detection of abnormal operational conditions is an interesting subject worth further exploration.Previous work has strongly depended on static features defined by human experts and thresholds that characterize abnormal operational conditions. Furthermore, machine learning methods typically depend on such static features to classify the faults and various operational conditions of the machine. This approach is challenging when reusing a method for different applications and environments, wherein similar features and thresholds can have different meanings. This problem is typically solved by reconfiguring or redesigning the condition monitoring system, thereby constraining the applicability and efficiency of the method.In this licentiate thesis, I investigate unsupervised methods for feature learning and anomaly detection. In particular, I focus on vibration signals, which contain information about both the bearing condition and the condition of the machine.The considered model represents the signal as a linear superposition of noise and atomic waveforms of arbitrary shape, amplitude and position. The atomic waveforms are adapted to each signal and machine using an unsupervised probabilistic optimization method and are considered features of the machine and physical processes exciting the signal. This model can automatically adapt the features to different environmental and operational conditions, thereby forming the basis for the development of a condition monitoring system that requires a minimum of manual configuration. Additionally, the model produces sparse codes that decrease the sensor data rate and, in principle, simplify the task of analyzing and communicating complex sensor information in resource-constrained embedded sensor systems.The thesis outlines an implementation of a sparse representation and dictionary learning method that is applied to vibration signals. I describe how signal analysis is performed using typical static pre-defined features and contrast this analysis with an analysis based on features that are automatically derived from the signal. In particular, the analysis focuses on the evolution of the vibration signal and the features when a fault develops within the ball bearing of a rotating machine. The evolution rate of learned features is defined and proposed as an interesting quantity for an autonomous condition monitoring process, and a first step towards an FPGA implementation of the method is presented.