CVCM Track Circuits Pre-emptive Failure Diagnostics for Predictive Maintenance Using Deep Neural Networks

Debdeep Mukherjee, Di Santi Eduardo, Lefebvre Clément, Mijatovic Nenad, Martin Victor, Josse Thierry, Jonathan Brown, Saiah Kenza · 2025

Read CVCM Track Circuits Pre-emptive Failure Diagnostics for Predictive Maintenance Using Deep Neural Networks by Debdeep Mukherjee, Di Santi Eduardo, Lefebvre Clément, Mijatovic Nenad, Martin Victor, Josse Thierry, Jonathan Brown, Saiah Kenza (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-2508-09054.txt.

Introduction

Debdeep Mukherjee and co-authors predict railway track circuit failures with deep networks in a 2025 paper. Track circuits locate trains and anchor railway safety signaling. When they fail, the harm cascades into major service disruption. Many begin as subtle anomalies not seen in watched signals.

The target technology varies current in a smooth, ongoing scheme. A neural framework sorts anomalies long before failure. Ten failure cases across sites backed the method. Overall accuracy topped ninety-nine percent in tests. Detection arrived within one percent of when the anomaly began. Conformal prediction added doubt estimates at high confidence. Coverage stayed steady across failure classes. The approach follows a global standard for watching the state of gear. Wide use of the circuit type around the world aids scaling. Other railway assets could adopt the same template. Planning upkeep ahead of failure is the payoff in daily use.

Catalog record

CREATORDebdeep Mukherjee, Di Santi Eduardo, Lefebvre Clément, Mijatovic Nenad, Martin Victor, Josse Thierry, Jonathan Brown, Saiah Kenza
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE27 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-09054

Browse the full library · About the Tower of Babel

Full text

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

CVCM Track Circuits Pre-emptive Failure Diagnostics for Predictive Maintenance Using Deep Neural Networks Debdeep Mukherjee, Di Santi Eduardo, Lefebvre Clément, Mijatovic Nenad, Martin Victor, Josse Thierry, Jonathan Brown, Saiah Kenza Abstract Track circuits are crucial for railway operations. They are the main signalling sub-system used to locate trains within track segments. Continuous Variable Current Modulation (CVCM) is one of such track circuit technologies. As with any equipment deployed on field, this safety critical component may experience failures, provoking cascading blockages of operations. Most such failures develop first from an anomaly into a critical state over time. These anomalies are usually reflected in the monitored signals, though not always visually distinguishable. Conventional approaches, which typically rely on such prominent changes in signals, therefore fail to identify failures beforehand. Identifying failure types at early stages, when they are still only anomalies in the signal, is of high importance. It allows for improved maintenance planning, minimizing time and revenue loss from operational disruptions. Leveraging deep neural networks, we propose our new methodology for predictive maintenance to classify anomalies sufficiently in advance into the type of failures in future. We demonstrate our method’s efficacy on 10 failure cases for CVCM, independent of installation location or configurations. Our method is ISO-17359 compliant and especially successful in improving early-stage anomaly classification compared to conventional approaches, which is paramount for effective predictive maintenance. We achieve solid operational performance with promising results: 99.31% overall classification accuracy across all failure types with early-stage detection averaging below 1% into start of anomaly development in signals. Finally, utilizing conformal prediction techniques, we provide accompanying uncertainty quantification metric for maintainers to measure model confidence in a particular prediction, where we reach 99% confidence, with sufficient coverage for each failure class. Given CVCM is deployed in various urban settings worldwide, our research holds significant relevance for maintenance personnel. Our methodology is scalable and can also be generalized across different track circuits and railway systems for improving reliability of railway operations through predictive maintenance. Keywords CVCM Track Circuits, Railway Predictive Maintenance, Anomaly Classification, Deep Learning, Conformal Prediction 1 Introduction 1.1 Continuous Variable Current Modulation (CVCM) Track Circuit Track circuits are the predominant technology in use today for locating trains along the track. For each track segment, there is a transmitted electrical signal that travels through each rail and is then received on the other end. When a train passes, it shunts the electrical signal due to wheel contact, stopping reception. This indicates track occupancy for that segment, which is critical information for other trains to avoid possibility of collisions. CVCM ( schon:hal-00924934 ) is one of such track circuit technologies that has a transmitted signal (TX) from the centre of a track segment, going out into both directions. This results in 2 received signals (RX), namely CAT and CAL, in the upstream and downstream sections (Figure 1). Physically, the track is a continuous piece of hardware, so CVCM divides it into distinct, electrically isolated segments, by frequency of transmission and reception used. Frequencies typically range from around 8.2 – 11 kHz and adjacent segments always have different frequencies. Locating trains is imperative for operations, making track circuits safety critical. In the advent of any failures, the track circuit defaults to an occupied state, halting further operations. Figure 1 : CVCM transmitted (TX) and received (RX) signals for a track segment 1.2 Problem Definition Failures in the CVCM hardware is therefore a significant problem faced by railway operators. The components typically undergo some degradation with use or due to unforeseen issues leading to different failures. This causes unwanted delays to trains and incurs costs such as revenue loss. Failures in the hardware are reflected in the characteristics of the received signals as they pass through the defective hardware components. The changes to the electrical signals are anomalous in nature from the usual nominal state of operation but could still be within operational tolerances, making them hard to distinguish. Additionally, a failure, before it reaches a critical state, will usually result from a progression of such anomalies that develop early on. Not all anomalies necessarily lead to an eventual failure so it is also necessary to discern this. It is vital to classify anomalies as early as possible during their initial formation to predict the failure types and time expected, so that maintenance can be planned in advance to minimize disruptions. Specifically for CVCM, due to its design, depending on which part (upstream or downstream), failures would affect each of the received signals differently. Figure 2 : Example of ’Nominal’ vs ’Anomaly’ received signal (RX) profiles As an anomaly progressively degrades, compared to the nominal behaviour, there is a point in time where it deviates sufficiently to visibly tell the signal is unusual. However, at this stage the failure is already critical, so diagnosis relying on visual separation (Figure 2) cannot help to identify a failure before it happens. Early stage classification of anomalies, when it has not yet developed into a failure (when it is difficult to visually distinguish from nominal conditions) is a key business requirement for effective predictive maintenance that ensures safe and reliable operations. (a) ’Early Stage’ nominal vs anomaly (b) ’Early Stage’ 2 different anomalies Figure 3 : CVCM ’Early Stage’ signal comparisons However, early stage anomalies look very similar to nominal signals (Figure 3(a)). Moreover, different anomalies are also not easily distinguishable between one another at these stages (Figure 3(b)). The main idea therefore is to successfully classify anomalies into possible failure types in future early enough (when changes may not yet be visible or evident to human eyes) before it has actually progressed into that failure, with high confidence. We propose our methodology (section 3), utilizing a set of deep neural network models to solve this problem. 1.3 Use Cases and Data Use Cases First we define the interested failure use cases for CVCM that require prediction at early stage. These use cases are based on operational experience and domain expert knowledge and are labelled as numbered anomaly classes for reference (in no particular order): Use case Description Anomaly 1 Decrease of the ballast resistance simulated with a 1-ohm resistance Anomaly 2 Degradation of the LC downstream contact Anomaly 3 Degradation of the LC upstream contact Anomaly 4 Degradation of the track transformer contact Anomaly 5 Degradation of the wheel-rail contact Anomaly 6 Intermittent degradation of the railway bonding between the remote amplifier (RA) and the sensor Anomaly 7 Progressive degradation of the railway bonding between the remote amplifier (RA) and the receiver Anomaly 8 Intermittent degradation of the railway bonding between the remote amplifier (RA) and the receiver Anomaly 9 Progressive degradation of the railway bonding between the remote amplifier (RA) and the sensor Anomaly 10 Broken rail downstream Anomaly 11 Broken rail upstream Table 1 : Anomaly use cases labels and descriptions Data In order to build and test our solution, we utilize signals data on the different failure use cases (Table 1) for CVCM. The proprietary dataset consists of 10 csv files containing the received signals of CVCM. There is one file per experiment of a failure use case produced in lab environment (except anomaly 6). It starts with fixed duration of nominal signal followed by

Read the complete text