Transformer Vibration Forecasting for Advancing Rail Safety and Maintenance 4.0

Darío C. Larese, Almudena Bravo Cerrada, Gabriel Dambrosio Tomei, Alejandro Guerrero-López, Pablo M. Olmos, María Jesús Gómez García · 2025

Read Transformer Vibration Forecasting for Advancing Rail Safety and Maintenance 4.0 by Darío C. Larese, Almudena Bravo Cerrada, Gabriel Dambrosio Tomei, Alejandro Guerrero-López, Pablo M. Olmos, María Jesús Gómez García (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-11730.txt.

Introduction

Darío Larese and co-authors forecast train axle vibrations in 2025 with a transformer model, aiming to warn of faults before they cause a crash or a costly failure. Railways are moving from fixed inspection rounds to condition based care, but real fault recordings are expensive to gather.

Their model, named ShaftFormer, reads accelerometer streams from axles and predicts how the shaking will evolve under varied speeds, loads, and fault states. A rival version adds spectrum based methods. Simulated signals fill the gaps where real fault data are missing, hardening the detectors trained on them. The authors frame the work as a step toward maintenance that acts on prediction, not on schedule.

Catalog record

CREATORDarío C. Larese, Almudena Bravo Cerrada, Gabriel Dambrosio Tomei, Alejandro Guerrero-López, Pablo M. Olmos, María Jesús Gómez García
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE26 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-11730

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Transformer Vibration Forecasting for Advancing Rail Safety and Maintenance 4.0 Darío C. Larese, Almudena Bravo Cerrada, Gabriel Dambrosio Tomei, Alejandro Guerrero-López, Pablo M. Olmos, María Jesús Gómez García Abstract Maintaining railway axles is critical to preventing severe accidents and financial losses. The railway industry is increasingly interested in advanced condition monitoring techniques to enhance safety and efficiency, moving beyond traditional periodic inspections toward Maintenance 4.0. This study introduces a robust Deep Autoregressive solution that integrates seamlessly with existing systems to avert mechanical failures. Our approach simulates and predicts vibration signals under various conditions and fault scenarios, improving dataset robustness for more effective detection systems. These systems can alert maintenance needs, preventing accidents preemptively. We use experimental vibration signals from accelerometers on train axles. Our primary contributions include a transformer model, sf ( sf ), designed for processing time series data, and an alternative model incorporating spectral methods and enhanced observation models. Simulating vibration signals under diverse conditions mitigates the high cost of obtaining experimental signals for all scenarios. Given the non-stationary nature of railway vibration signals, influenced by speed and load changes, our models address these complexities, offering a powerful tool for predictive maintenance in the rail industry. Journal: Engineering Applications of Artificial Intelligence 1 Introduction Train operations rely heavily on the reliability of their wheelsets, with axles being crucial for supporting train weight and transmitting power. Malfunctions can lead to severe consequences such as accidents, derailments, and safety risks, often due to fatigue or cracks in the axles, which can result in loss of life and property damage ( Vagnoli et al., 2018 ) . To address these challenges, we propose a novel transformer model for predictive maintenance that forecasts a set of axle vibration signals obtained experimentally under different conditions of load, speed and flaws. In the field of predictive maintenance, forecasting is a crucial requirement for detecting faulty signals and simulating new signals which cannot be obtained easily. Through monitoring relevant variables, it is a natural solution to axle reliability issues. Previous research has explored using various sensors to detect wheel flaws ( Alemi et al., 2018 ) and predict failures in critical rail car components ( Li et al., 2014 ; Wang et al., 2022 ) . For railway axles, studies have utilized model-based signals from analytical ( Gómez et al., 2016b ) or numerical models ( Hassan et al., 2016 ) , as well as experimental signals from scaled rigs ( Gómez et al., 2016a ) or accelerometers on real systems. Despite the potential of these methods, bogie test rigs are rare and costly, limiting the scope and generalizing capabilities of the results ( Castejón et al., 2016 ; Gómez et al., 2016b ; Gómez et al., 2016a ; Rolek et al., 2016 ; Gómez et al., 2023 ) . The current trend in maintenance automation, known as Maintenance 4.0, leverages Machine Learning (ML) techniques. Recently, there has been growing interest in applying dl ( dl ) for condition monitoring and predictive maintenance in the rail industry. Studies have demonstrated the use of neural networks to analyze vibration signals and predict axle failures ( Gómez et al., 2016c ; Gómez et al., 2018 ; Gómez et al., 2020 ) . For example, Galdo et al. (2023) explored classifying crack conditions on railway axles using a differential convolutional classification network. However, existing studies often overlook the nonstationary nature of railway vibration signals, which are influenced by speed, load, and environmental changes ( Ye et al., 2022 ; Krummenacher et al., 2018 ) . This oversight hampers the development of models that can generalize across diverse operating conditions and structural variations. Moreover, many current methods rely on traditional ML algorithms or simplistic neural network architectures that may struggle to capture the complex relationships present in railway vibration data ( Serin et al., 2020 ) . Some research has proposed dl models for machine diagnosis, but these often lack domain-specific adaptations tailored to railway systems ( Zhao et al., 2019 ; Wang et al., 2021 ) . Simple transformer models have been applied to wheel defect classification tasks ( Wang et al., 2022 ) , yet they remain plug-in solutions rather than tailored DL models. This study aims to address these gaps by proposing the sf model and the ssf ( ssf ) model, which are variants of the original Transformer ( Vaswani et al., 2017 ) . These models implicitly consider the non-stationarity of the data and effectively model the signals. By incorporating domain knowledge from railway engineering, our methodology is designed to fit within the monitoring and maintenance tasks in the industry. Moreover, both models are able to generate signals under different environmental conditions such as load and speed. Potential applications of this study include outlier detection, missing data imputation, signal forecasting, and data augmentation. 2 Methodology 2.1 Data collection: experimental setup The experimental acquisition of vibration signals from railway axles involved investigating the dynamics of cracked rotors using a specially designed experimental setup, as detailed in previous studies ( Gómez et al., 2018 ; Gómez et al., 2020 ) . The experimental system was constructed within a bogie test rig, consisting of a fixed bench and a drive system to roll the tested axle, as shown in Figure 1 . Figure 1: Test rig with Bogie Y21 set, including tested wheelset (1), fixed wheelset (2), lhs ( lhs ) axle box (3), rhs ( rhs ) axle box (4), motor and driving system (5) Six accelerometers were placed, with three in each axle box ( lhs and rhs ) of the tested axle, measuring vibrations in three orthogonal directions: tangential (in two perpendicular directions), vertical (aligned with the gravitational force) and axial (parallel to the direction of the track). In addition, the rig featured a loading system that employs hydraulic actuators to apply a consistent vertical load to the bogie via a chain mechanism. Throughout testing, the selected load remained constant to accurately replicate real-world operating conditions. The experiments involved testing three distinct Wheelset Assemblies (WAs), labeled WA1, WA2 and WA3, resulting in the creation of three individual datasets. Each WA was subjected to testing under two different combinations of speed and load, for clockwise and counterclockwise rotations. The applied load and speed were kept constant throughout the tests and are outlined in 2 , offering a comprehensive overview of all test conditions. Monitoring these signals on a bogie instead of a wheel set produces very rich data, as the full bogie includes not only the wheels and axle, but also the suspension systems and load, reproducing with high fidelity the reality. Test condition Values Flaw condition Healthy shaft (D0) 5.7 mm (D1) 10.9 mm (D2) 15 mm (D3) Load 4t 10t Speed 20 km/h 50 km/h Wheelsets rotation Counterclockwise Clockwise Axle box accelerometer orientation Lengthwise Vertical Axle box accelerometer place rhs lhs Table 1: Conditions tested for each WA. The test rig comprises a fixed bench and a drive system for rolling the tested axle and also incorporates a loading system that uses hydraulic actuators that apply a vertical load to the bogie through a chain. During the test, the load remains constant, replicating in real-world situations. Table 1 shows the tested configurations. First,

Read the complete text