In recent years, significant studies have focused on monitoring the track geometry irregularities through measurements of vehicle dynamics acquired onboard. Most of these studies analyse the vertical irregularity and the vertical vehicle dynamics since the lateral direction is much more challenging due to the non-linearities caused by the contact between the wheels and the rails. In the present work, a machine learning-based fault classifier for the condition monitoring of track irregularities in the lateral direction is proposed. The classifiers are trained with a dataset composed of numerical simulation results and validated with a dataset of measurements acquired by a diagnostic vehicle on the straight track sections of a high-speed line (300 km/h). Classifiers based on decision tree, linear and Gaussian support vector machine algorithms are developed and compared in terms of performance: good results are achieved with the three algorithms, especially with the Gaussian support vector machine. Even though classifiers are data driven, they retain the essence of lateral dynamics.

Condition monitoring of track geometry irregularities from onboard measurements of vehicle response is a cost-effective method for surveilling qual-ity of track irregularities on daily basis. The monitoring of Alignment Level (AL)and Cross Level (CL) track irregularities is challenging due to the nonline-arities of the contact between wheels and rails. Recently, the authors proposed a signal-based method in combination with a machine learning (ML) fault classi-fier to monitor AL and CL track irregularities based on bogie frame accelerations. The authors concluded that the Support Vector Machine (SVM) fault classifier outperformed other traditional ML classifiers. Thus, an important question arises: Is the previously reported decision boundary an optimal boundary? The objective of this research investigation is to obtain an optimal decision boundary according to theory of probabilistic classification and compare the same against the SVM decision boundary. In this investigation, the classifiers are trained with results of numerical simulations and validated with measurements acquired by a diagnostic vehicle on straight track sections of a high-speed line (300 km/h). A fault classi-fier based on Maximum A Posterior Naïve Bayes (MAP-NB) classification is developed. It is shown that the MAP-NB classifier generates an optimal decision boundary and outperforms other classifiers in the validation phase with classifi-cation accuracy of 95.9±0.2% and kappa value of 80.4±0.6%. Moreover, the Lin-ear SVM (L SVM) and Gaussian-SVM (G SVM) classifiers give similar perfor-mance with slightly lower accuracy and kappa value. The decision boundaries of previously reported SVM based fault classifiers are very close to the optimal MAP-NB decision boundary. Thus, this further strengthens the idea of imple-menting statistical fault classifiers to monitor the track irregularities based on dynamics in the lateral plane via in-service vehicles. The proposed method con-tributes towards digitalization of rail networks through condition-based and pre-dictive maintenance.

Generally, the equivalent conicity function (ECF) is denoted by equivalent conicity at 3mm (λ_3mm) and a Nonlinearity Parameter (NP). NP describes the nonlinearity of the ECF and its influence on a vehicle design is explored thoroughly, however, NP’s role in vehicle and track maintenance is not researched yet. This paper investigates the influence of track maintenance actions on vehicle dynamics with help of NP vs λ_3mm scatter plots of ECF database. The ECF database is constructed by combining measured worn wheel and rail profile pairs of the Swedish high-speed vehicle and rail network, respectively. The ECF database revealed an inverse relationship between λ_3mm and NP, i.e., NP is negative for larger λ_3mm values. The combination of negative NP and high λ_3mm causes reduction in the vehicle’s nonlinear critical speed and vehicle often exhibit the unstable running on the Swedish rail network. Thus, the occurrence of ECF with negative NP and high λ_3mm is undersirable and the undesirable ECF can be converted into desirable ECF by grinding the rail, which converts ECF’s into positive NP and low λ_3mm combinations. Thus, the NP parameter along with the λ3mm must be considered in track maintenance decisions.

Intelligent fault identification of rail vehicles from onboard measurements is of utmost importance to reduce the operating and maintenance cost of high-speed vehicles. Early identification of vehicle faults responsible for an unsafe situation, such as the instable running of highspeed vehicles, is very important to ensure the safety of operating rail vehicles. However, this task is challenging because of the nonlinear dynamics associated with multiple subsystems of the rail vehicle. The task becomes more challenging with only accelerations recorded in the carbody where, nevertheless, sensor maintenance is significantly lower compared to axlebox accelerometers. This paper proposes a Temporal Convolution Network (TCN)-based intelligent fault detection algorithm to detect rail vehicle faults. In this investigation, the classifiers are trained and tested with the results of numerical simulations of a high-speed vehicle (200 km/h). The TCN based fault classification algorithm identifies the rail vehicle faults with 98.7% accuracy. The proposed method contributes towards digitalization of rail vehicle maintenance through condition-based and predictive maintenance.

Intelligently identifying rail vehicle faults instigating running instability from carbody floor acceleration is essential to ensure operational safety and reduce maintenance costs. However, the vehicle-track interaction's nonlinearities and scarcity of running instability occurrences complicate the task. The running instability is an anomaly in the vehicle-track interaction. Thus, we propose unsupervised anomaly detection and clustering algorithms based iVRIDA framework to detect and identify running instability and corresponding root cause. We deploy and compare the performance of the PCA-AD (baseline), Sparse Autoencoder (SAE-AD), and LSTM-Encoder-Decoder (LSTMEncDec-AD) model to detect the running instability occurrences.

Furthermore, we deploy a k-means algorithm on latent space to identify clusters associated with root causes instigating instability. We deployed the iVRIDA framework on simulated and measured accelerations of European high-speed rail vehicles where SAE-AD and LSTMEncDec-AD models showed 97% accuracy. The proposed method contributes to smart maintenance by intelligently identifying anomalous vehicle-track interaction events.

Identifying faults contributing to unsafe conditions, such as a high-speed railvehicle running instability, is crucial to ensuring operational safety. But the occurrence of vehicle running instability during regular operation across the whole vehicle fleet is a rare anomaly. An unsupervised anomaly detection (AD) based iVRIDA-fleet framework is therefore proposed to detect vehicle running instability and identify its root cause. The performance of Principal Component Analysis (PCA-AD, baseline model), Sparse Autoencoder (SAE-AD),and LSTM Encoder Decoder (LSTMEncDec-AD) models are evaluated to detect the occurrence of vehicle running instability. A k-means algorithm is then applied to latent space representations to identify various clusters associated with different root causes of observed vehicle running instability.The effectiveness of the proposed iVRIDA-fleet framework is demonstrated using onboard accelerations measured on a Swedish X2000 vehicle fleet. The probability of vehicle running instability occurrence is observed to be only 0.35% of onboard accelerations corresponding to 827,467 km travel distance.

Furthermore, the root causes identified by the iVRIDA-fleet framework are validated by investigating the maintenance records of the vehicles and track. It is identified that heavily worn wheels were the primary root cause of observed vehicle running instability, but the track (actual gauge and rail profiles) was also a contributing factor. The proposed algorithm contributes towards the digitalisation of vehicle and track maintenance by intelligently identifying anomalous events of the vehicle-track dynamic interaction.