Railway point-operating machines (POMs) are the electro-mechanical actuators that move turnout switch blades, and their actuation-current signatures are routinely logged by signaling condition-monitoring systems. Mistry, Lane and Allen (Sensors, 2020) showed that a fast-Fourier-transform-based curve-fit feature (the residual standard error, RSE) can flag anomalous actuations in large volumes of completely unlabeled Network Rail (UK) current data — but they had to impose one arbitrary global alarm threshold (RSE >= 0.3) on all instruments, even though their own per-machine baselines ranged from about 0.04 to 0.9; they did not separate the two actuation directions (normal NR out-swing vs. reverse RN return); and they could not validate detected candidates against actual failure or repair events because the maintenance history was unavailable to them. The 2023 Actuators review of point-machine fault detection still lists dynamically adjusted detection thresholds as an urgent, unresolved challenge, and the 2024–2026 one-class/semi-supervised literature on switch machines indicates that the problem remains open: a field-deployable methodology that learns alarm thresholds per instrument or per group of similar machines from unlabeled fleet data, quantifies the effect of treating NR and RN movements separately, and validates detected anomalies against maintenance records.

Background

A railway turnout (switch) lets trains pass from one track to another; the point-operating machine (POM, also called a switch machine or point mechanism) is the electro-mechanical actuator that slides and locks the movable switch blades. POM and turnout faults are among the largest contributors to railway infrastructure failure — the source study cites signaling-equipment and turnout failures at about 55% of all infrastructure component failures — so operators continuously log each POM's motor current draw during every actuation. Each actuation produces a characteristic current profile with two phases: a push-out phase (motor start-up and unlocking) and a swing phase (blade movement and locking). A POM can move in the normal (NR) direction, letting a train pass straight through, or the reverse (RN) direction, routing it to the alternate path; the two movements load the mechanism differently and earlier studies report direction-dependent current signatures. Because real fault labels are almost never available at fleet scale, detection must operate on unlabeled data, and any alarm rule ultimately rests on a threshold that separates normal variability from anomalous behavior. Early fixed-threshold alarms produced many false positives and missed failures, motivating statistical and learning-based approaches (Kalman-filter reference comparison, harmonic-regression prediction of the next current curve, one-class classifiers such as the one-class support vector machine, which learns a boundary around normal data when no fault examples exist, and outlier-detection methods). Mistry, Lane and Allen (2020) processed 23,389 daily log files of unlabeled single-channel current data (the pre-processed remainder of the 45,690 files supplied by Network Rail, UK, which record 331 POMs across three UK regions — London North-Eastern, London North-Western and Sussex — between April and June 2018; 293 POMs remain after cleansing): for each actuation they computed a fast Fourier transform (FFT — decomposition of the signal into its frequency components), fitted a line to the log-transformed FFT magnitudes, and used the residual standard error (RSE — the residual variability of that linear fit, a measure of how much the actuation's spectrum deviates from a clean power-law shape) as an anomaly feature, alongside the traditional peak current of the push-out phase and mean current of the swing phase. The method worked qualitatively — instruments with degraded behavior showed clear RSE shifts — but three limitations were explicit. First, the alarm rule used a single arbitrary threshold, RSE >= 0.3, for every instrument, even though sequential per-machine RSE plots showed instrument-specific baselines ranging from roughly 0.04 to 0.9, so a fixed global threshold can both miss faults on quiet machines and mischaracterize normal behavior on noisy ones; the authors proposed dynamic, machine-learning-based thresholding, suggesting clustering of machines in the peak-current versus mean-current scatter plots to set thresholds relative to groups of similar machines (e.g., of the same type). Second, NR and RN actuations were pooled, although the cited literature reports direction-dependent differences and other studies analyze the two directions separately. Third, although every record carries a timestamp, the absence of ground truth about when failures or repairs occurred prevented any history-aware validation — it remained unknown whether the flagged actuations correspond to actual faults, and how maintenance timelines could be incorporated into the monitoring methodology. Later literature bears directly on this problem. Guo, Wan and Ye (IEEE Trans. Instrum. Meas., 2020) developed an unsupervised method for Chinese high-speed-railway turnouts that mines unknown modes of the normal current curves by clustering and estimates per-mode detection thresholds, with one-class transfer learning for rare modes — demonstrating the feasibility of learned thresholds, but on three-phase current data with a labeled test set, and without addressing direction separation or maintenance-event validation. Reviews of data-driven point-machine fault diagnosis (Hu et al., Transp. Saf. Environ., 2022; Hu et al., Actuators, 2023) confirm the field-wide pattern: proximity- and one-class-based methods still require manually adjusted thresholds or hyperparameters, and the 2023 review names dynamically adjusted fault-detection thresholds as an urgent open challenge while pleading for shared labeled datasets — evidence that the ground-truth gap persists. An active 2024–2026 research program on switch-machine fault diagnosis with scarce labels (semi-supervised and one-class matrix/tensor machines — classifiers that operate directly on the matrix- or tensor-structured current-signal data rather than on flattened feature vectors — evaluated on ZDJ9 and S700K turnout datasets, two widely deployed Chinese point-machine types) continues to attack label scarcity, but none of these works provides a self-calibrating threshold scheme for fully unlabeled fleet data, a direction-separated comparison, or maintenance-aligned validation.

Problem Statement

For railway point-operating machines monitored only through unlabeled actuation-current logs, develop and validate a fault-detection methodology that removes the fixed global alarm threshold. Concretely, on real fleet data of the kind studied in the source (continuous single-channel current-draw records of every actuation, with timestamps, and no fault labels): (1) the anomaly threshold must be set dynamically per instrument or per group of similar machines, so that machines with different normal baselines are neither systematically over-alarmed nor under-alarmed, and the procedure by which thresholds are learned must be fully specified rather than hand-tuned per machine; (2) actuations in the normal (NR) and reverse (RN) directions must be analyzed separately, with a quantified comparison — expressed in explicitly defined detection metrics, for example detection and false-alarm rates — showing whether direction separation makes the detection results more conclusive than pooled analysis; and (3) the detected anomaly candidates must be brought into correspondence with maintenance or failure records (repair logs, failure reports, or an equivalent ground-truth timeline aligned via the record timestamps) to establish how many correspond to actual faults and how maintenance timelines can be incorporated into the monitoring methodology — established against whatever such records are available, and otherwise specified as a concrete protocol. The three strands serve one overarching objective — a self-calibrating, field-deployable condition-monitoring pipeline whose detections are verifiable against ground truth — established end-to-end for real fleet current data of the kind studied in the source (or for an explicitly identified, obtainable dataset of comparable actuation-current records), with the detection and false-alarm behavior quantified and the correspondence of the detected candidates with the maintenance or failure timeline established.

References

  1. P. Mistry, P. Lane, P. Allen, 'Railway Point-Operating Machine Fault Detection Using Unlabeled Signaling Sensor Data', Sensors 20(9):2692, 2020. DOI: 10.3390/s20092692, https://doi.org/10.3390/s20092692
  2. X. Hu, T. Tang, L. Tan, H. Zhang, 'Fault Detection for Point Machines: A Review, Challenges, and Perspectives', Actuators 12(10):391, 2023. DOI: 10.3390/act12100391, https://doi.org/10.3390/act12100391
  3. X. Hu, Y. Cao, T. Tang, Y. Sun, 'Data-driven technology of fault diagnosis in railway point machines: review and challenges', Transportation Safety and Environment 4(4):tdac036, 2022. DOI: 10.1093/tse/tdac036, https://doi.org/10.1093/tse/tdac036
  4. Z. Guo, Y. Wan, H. Ye, 'An Unsupervised Fault-Detection Method for Railway Turnouts', IEEE Transactions on Instrumentation and Measurement 69(11):8881-8901, 2020. DOI: 10.1109/TIM.2020.2998863, https://doi.org/10.1109/TIM.2020.2998863
  5. W. Li, Z. Xu, M. Mei, M. Lan, C. Liu, X. Gao, 'A Semi-Supervised Adaptive Matrix Machine Approach for Fault Diagnosis in Railway Switch Machine', Sensors 24(13):4402, 2024. DOI: 10.3390/s24134402, https://doi.org/10.3390/s24134402
  6. C. Chen, J. Yan, M. Mei, K. Huang, C. Liu, 'Early fault detection in railway point machine via adaptive-weight one-class support matrix machine', Measurement Science and Technology 37(4):046209, 2026. DOI: 10.1088/1361-6501/ae31ff, https://doi.org/10.1088/1361-6501/ae31ff
  7. C. Chen, X. Li, M. Mei, K. Huang, L. Yang, 'Incipient fault detection of railway point machines using flexible convex hull-based one-class tensor machine under weak-feature condition', Nondestructive Testing and Evaluation (advance online publication, pp. 1-24), 2025. DOI: 10.1080/10589759.2025.2583253, https://doi.org/10.1080/10589759.2025.2583253
  8. F. P. García Márquez, F. Schmid, 'A digital filter-based approach to the remote condition monitoring of railway turnouts', Reliability Engineering & System Safety 92(6):830-840, 2007. DOI: 10.1016/j.ress.2006.02.011, https://doi.org/10.1016/j.ress.2006.02.011 (whose experimental dataset of 151 recreated-fault point-machine movements splits 79 reverse-to-normal and 72 normal-to-reverse, analyzed per direction)