- Title
- Deep learning-based data-driven predictive maintenance for railway tracks
- Creator
- Zeng, Cheng
- Relation
- University of Newcastle Research Higher Degree Thesis
- Resource Type
- thesis
- Date
- 2023
- Description
- Research Doctorate - Doctor of Philosophy (PhD)
- Description
- Railway track is a very complex system that involves many interactions, and thus generates many potential areas where a track failure can occur, such as rail structure and track substructure. Due to its economic and safety impact on railway operations, a track failure is extremely undesirable in the railway network. Traditionally, railroads carry out corrective and preventive maintenance strategies to avoid a failure in their network. But corrective maintenance cannot prevent failure in the first place, whereas preventive maintenance might result in unnecessary maintenance actions and costs. A more desired strategy is to perform predictive maintenance, that is to develop a predictive model for predicting the occurrence of track failure so that maintenance can be taken to the right place at the right time. The performance of the predictive maintenance depends heavily on the appropriate choice of data set and predictive models. However, even automatic monitoring systems have been applied in railway networks to collect massive monitoring data. These monitoring data is rarely used for predictive maintenance. Besides, most predictive models used in the predictive maintenance of railway tracks still rely on traditional statistical models and conventional machine learning methods, which makes them difficult to deal with massive and complex high-dimensional data effectively. Thus, it is therefore of great significance to explore advanced approaches and effectively use collected track data for the predictive maintenance of railway tracks. This thesis developed various approaches using the collected track data and advanced deep learning methods for predictive maintenance focusing on two main perspectives including failure prediction and useful lifetime estimation. In this research, two practical track failures including rail break and mud pumping in railway track are considered, covering both rail structure and track substructure. All the proposed approaches are implemented on a real-world dataset from a section of railway network in Australia to show their effectiveness. The following primary contributions to the predictive maintenance of railway tracks have been made through the thesis: (1) A data-driven model for the prediction of mud pumping defects using daily in-service train data is developed. The data-driven model is based on long short-term memory (LSTM) networks. Bayesian optimization method is used to optimize the hyper-parameters in the proposed LSTM networks. Genetic algorithm is used to select the best feature subset. The results show that the proposed model can be used to predict the mud pumping defects in advance, leaving enough time for maintenance. (2) An improved approach for mud pumping prediction is developed based on Geographic Information System (GIS)-based hydrological variables and in-service train monitoring data. Through GIS analysis, the hydrological variables along the entire railway network are estimated and combined with monitoring data for model development. A dual-channel neural networks model is designed to separately mine the characteristics in data with different attributes so that hydrological variables and monitoring data can be effectively integrated into the prediction framework. The results confirm that integrating GIS-based hydrological variables can generate more accurate predictions and reduce the false prediction rate, compared to using monitoring data only. In addition, Shapley addictive explanations (SHAP) is applied to estimate the importance of hydrological variables and reveal the possible relationships between the variables and the probability of mud pumping. The explainable results can help infrastructure managers automatically identify the most vulnerable sections in railway tracks, which facilitates targeted maintenance planning and track substructure design. (3) An unsupervised framework for mud pumping detection and severity analysis using rare failure data is proposed. The framework is based on a long short-term memory (LSTM) autoencoder and deep embedding clustering (DEC). This study highlights the potential of utilizing deep learning-based unsupervised methods for track failure detection, especially when dealing with very limited failure data. The results also show that the proposed unsupervised framework can provide insight into the severity of mud pumping defects. (4) A rail break prediction approach using imbalanced daily monitoring data is developed. A time-series generative adversarial network (TimeGAN) is employed to mitigate the problem of data imbalance. A feature-level attention-based bidirectional recurrent neural networks (AM-BRNN) is proposed to enhance feature extraction and capture two-direction dependencies from sequential data. Compared with five current generation methods, the results prove that the proposed TimeGAN can generate the best quality rail break samples. Compared with four current predictive methods, the results show that the proposed AM-BRNN can achieve the highest prediction performance. Real-life validation shows that the proposed model can predict the location of potential rail breaks three months ahead of time with high confidence. (5) A deep Bayesian survival approach, named BNN-Surv, is proposed for rail useful lifetime estimation. The proposed BNN-Surv is based on the framework of survival model specifically designed to handle censored data. The proposed BNN-Surv model uses a deep neural network as the hazard rate in the survival model to capture the non-linear relationship between covariates and rail useful lifetime. To consider and quantify uncertainty in the model, Monte Carlo dropout, regarded as the approximate Bayesian inference, is incorporated into the deep neural network to provide the confidence interval of the estimated rail useful lifetime. The results show that the proposed approach has the potential to estimate rail useful lifetime with satisfactory accuracy. Besides, the results obtained by BNN-Surv can provide confidence interval for the estimated rail useful lifetime, which is more useful for decision making and maintenance planning.
- Subject
- predictive maintenance; railway tracks; deep learning-based model; Bayesian survival approach
- Identifier
- http://hdl.handle.net/1959.13/1510379
- Identifier
- uon:56399
- Rights
- This thesis is currently under embagro and will be available from 25.07.2025, Copyright 2023 Cheng Zeng
- Language
- eng
- Hits: 514
- Visitors: 502
- Downloads: 2