Fault diagnosis of railway assets using acoustic monitoring

Inoue, Katsuhito ORCID: 0000-0002-3393-7372 (2024). Fault diagnosis of railway assets using acoustic monitoring. University of Birmingham. Ph.D.

Preview

Inoue2024PhD.pdf
Text - Accepted Version
Available under License All rights reserved.
Download (26MB) | Preview

Abstract

Machines such as point machines and contactors are called Single Throw Mechanical Equipment (STME), and the failure of these machines affects the operation of trains significantly and may cause accidents, resulting in many human casualties and significant economic damage.
Although condition monitoring methods for these machines have been studied, conventional methods using voltage sensors, load sensors, and displacement sensors need to be attached directly to the high-voltage circuit or mechanical portion of the equipment which can be dangerous to attach and maintain, and when a sensor fails, it can adversely affect the functioning of the equipment. In addition, the environment where railway equipment is installed is harsh and there is a high possibility of sensor failure due to exposure to iron powder, dust, oil, and vibration generated from the tracks, wheels, ballast and passing trains. Air-borne acoustic monitoring is attracting attention because these data can be measured in a non-contact way from a certain distance, which means there is less risk of the sensor having a negative impact on the equipment or of the sensor breaking due to vibration or natural environmental factors, reducing the risk of the workers involved in data acquisition and sensor maintenance. Furthermore, it may enable low-cost maintenance by monitoring several pieces of equipment from a single sensor as sound is spread around.
Condition monitoring method using acoustic data has attracted attention and a lot of research has been done but most research targets are rotating machinery such as axles and bearings. Few studies have applied it to STME, and even in a few cases, it focuses on failures that can be detected by existing fault detection methods.
This thesis proposes a condition monitoring method using acoustic data for STME, and faults to be studied are difficult to detect and diagnose by existing methods. In order to verify the effective diagnostic technique for STME, normal and simulated fault data are acquired from two types of STME, a point machine and a contactor, in this study. A fault diagnosis system is developed using features and classification techniques that are expected to be useful from previous studies. Using the developed method to diagnose the data of each STME, it was found that using the Mel Frequency Cepstral Coefficient (MFCC) as a feature is the most accurate for both STME. As for classifiers, both Support Vector Machine (SVM) and K-Nearest Neighbour (KNN) could diagnose faults with sufficiently high accuracy for both STMEs, but the accuracy using KNN was slightly higher than that using SVM.
The noise tolerance of the proposed method was also evaluated by adding noise artificially. In addition, multiple microphones are used to investigate which location is best for data acquisition. Finally, the transferability of the model, in which the model-trained data from one machine can be used for other machines, is verified using MFCC as a feature. The verification revealed that the combination of MFCC and SVM could improve the transferability performance of the model for several faults.
The verification of this study revealed that fault diagnosis is possible from the acoustic data of the STME by using the proposed method, and since this method is tolerant of noise and its performance of transferability is high, field deployment would be anticipated to be successful.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):