Detection of Electric Motor Damage Through Analysis of Sound Signals Using Bayesian Neural Networks

TL;DR

This work proposes to use a Bayesian neural network to detect and classify faults in electric motors, given its efficacy with imbalanced training data, and demonstrates its performance on real life signals.

Abstract

Fault monitoring and diagnostics are important to ensure reliability of electric motors. Efficient algorithms for fault detection improve reliability, yet development of cost-effective and reliable classifiers for diagnostics of equipment is challenging, in particular due to unavailability of well-balanced datasets, with signals from properly functioning equipment and those from faulty equipment. Thus, we propose to use a Bayesian neural network to detect and classify faults in electric motors, given its efficacy with imbalanced training data. The performance of the proposed network is demonstrated on real life signals, and a robustness analysis of the proposed solution is provided.

Figures (6)

• Figure 1: Acoustic data acquisition from blender and electric impact drill using a smartphone (CC BY 4.0, source glowacz2018)
• Figure 2: FFT transforms. The light blue color of the plot represents signals with small maximum amplitudes, while the dark blue color corresponds to signals with large maximum amplitudes
• Figure 3: The diagram describes the structure of the Bayesian neural network used in the experiments. The developed network takes the successive values of the calculated FFT signal as input arguments, $X1$,...,$X5$, which are passed to five hidden layers, $H1$,...,$H5$, whose activation function is $\tanh$. The output of the network, $Y$, uses the sigmoid function as an activation function. In contrast to traditional neural networks, the weights are represented by distributions
• Figure 4: Comparison of the distributions of output values of the designed BNN for all signal classes.
• Figure 5: Confusion matrix for one of the 100 trials of the developed BNN. The proposed structure flawlessly detects damages and correctly recognizes signals from undamaged devices in almost $70\%$ of cases. This is a desirable property because it accurately informs about the occurrence of a malfunction in the system.