| Author: | Jiang, Gaofeng |
| Title: | Intelligent condition assessment of maglev system by data-driven and deep learning algorithms |
| Advisors: | Ni, Yi-qing (CEE) |
| Degree: | Ph.D. |
| Year: | 2025 |
| Department: | Department of Civil and Environmental Engineering |
| Pages: | xxvi, 330 pages : color illustrations |
| Language: | English |
| Abstract: | The maglev system has emerged as one of the most advanced rail transport modes, primarily due to the minimal friction between its rail and vehicle. However, the sophisticated components of this system make it susceptible to levitation and motion stability issues, necessitating precise coordination control and meticulous adjustments. Consequently, effective maglev condition assessments are urgently needed. Traditionally, maintenance of these systems has relied on regular inspections by operators with specialized expertise, time-consuming, labor-intensive, and costly practice. In contrast, data-driven and deep learning algorithms offer a more efficient and accurate alternative by enabling direct and automatic feature learning from raw data. This thesis proposes several advanced solutions for achieving intelligent maglev condition assessments. First, the performance of maglev rail joints is investigated. Maglev rail joints are essential components that connect adjacent F-type rail sections in a maglev guideway. Defects in maglev rail joints may result in large suspension gap fluctuations, a failure of suspension control, and even sudden clashes between the electromagnet and F-type rail. Monitoring of the maglev rail joint conditions is therefore necessary to maintain safe maglev system operation. Motivated by this need, in this thesis, an approach to multiple damage detection in maglev rail joints is developed based on convolutional neural network (CNN) and time-frequency spectrogram (TFS) characterization. Acceleration record data, acquired from rail joints of a test line by a condition monitoring system, were used for maglev rail joint condition evaluation. The results show that three conditions of rail joints could be identified successfully: bolt-looseness-caused rail step, misalignment-caused lateral dislocation, and normal. Although deep learning algorithms have been used to assess the conditions of maglev rail joints, most methods rely on data collected in controlled scenarios, such as trains running at constant speeds. Therefore, a domain adaptation (DA) approach is proposed to diagnose the conditions of maglev rail joints under complex operational conditions. DA is performed by integrating sample moments with different orders into the transfer loss of a neural network. DA can be used to reduce the domain shift caused by differences in operational conditions, and the feature knowledge is transferred. The proposed approach is validated using a dataset of TFSs derived under two operation modes: stable passing and braking. This proposed approach can successfully identify the conditions of maglev rail joints, despite changes in the maglev train operation mode. Levitation control performance is another crucial aspect of maglev operation because some external interference causes control instability and failure. However, data-driven diagnosis is challenging due to difficulties in feature learning, such as massive amounts of data, reduced label availability, and possible experimental restrictions. Hence, a novel framework for diagnosing maglev levitation control failure is formulated based on a contrastive self-supervised learning (SSL) algorithm. This approach uses inherent co-occurrence patterns in the data to model class invariance. The discriminative features of normal and failure conditions caused by load changes, rail irregularity, and resonance are accurately extracted using the proposed framework. The contrastive SSL algorithm also provides insight into the features underlying the data and guides the model in classification, yielding an accuracy comparable to that of the fully supervised model. This algorithm can therefore be used to activate feature learning from large-scale unlabeled data. The condition assessment can provide accurate and robust predictions of the maglev rail joint condition and levitation control performance, but it just indicates whether the components are normal or abnormal and in which condition, and it does not comprehensively explain the situation. Therefore, an expert visual question answering (VQA) model for presenting details and findings from the data is proposed; this model summarizes the damage and failure discriminative evidence accumulated from prior knowledge. The dataset is organized in the format of an image–question– answer tuple, with images generated from TFSs and questions and answers formulated with reference to the structural dynamic characteristics in a maglev system. The results show that the proposed VQA model reliably answers different questions correctly, enhancing human recognition of damage and failure features and contributing to the achievement of a more intelligent data-driven condition assessment. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13928