Probabilistic Uncertainty-Aware Decision Fusion of Neural Network for Bearing Fault Diagnosis

Reliability is a central aspect of machine learning applications, especially in fault diagnosis systems, where only an accurate and reliable diagnosis system is economically justifiable, considering that any false diagnosis would lead to an increase in maintenance costs or a reduction in system efficiency. Recent advances in machine learning (ML) techniques have encouraged condition monitoring researchers to focus their efforts on finding suitable ML-based solutions for system condition assessment. However, to address the reliability issue, it is crucial to consider a larger amount of data measured by heterogeneous sensors on the system together with non-sensor information. The trend of data fusion has already started in other areas of ML application, and many of today's state-of-the-art models benefit from various types of fusion techniques to improve their accuracy. However, traditional classifiers do not provide any information about the prediction uncertainty, and they tend to show falsely high confidence when encountering low-quality data or previously unseen classes. Fusion of different data sources without considering the epistemic or aleatory uncertainty can lead to a deterioration of the result. Bayesian frameworks have traditionally been used to quantify uncertainty of systems; however, only recent advances made it possible to successfully implement Bayesian ML models.
The research methodology was investigated using the MAFAULDA dataset generated by SpectraQuest's Machinery Fault Simulator. This simulator experimentally simulated various bearing conditions, including normal operation and inner and outer ring bearing failures, at variable speeds. The dataset consists of 1951 instances measured using two triaxial accelerometers, a microphone, and a tachometer.
Diagnosis has been done via two multi label 1D Convolutional Neural Networks - each for a selected sensor - and their prediction along with their associated uncertainty quantity has been fused utilizing Bayesian model averaging. The methodology is capable of fusion of various decisions made based on different data sources and generate a unified decision with associated confidence level. Fusion process is uncertainty aware and application of 1D networks reduce the amount of data needed.