Intelligent Condition Monitoring of Railway Signalling Lamps and Track Circuits Using LoRaWAN and Multi-Sensor Data Fusion

Abstract

Railway signalling lamps and track circuits are among the highest-consequence assets in rail operations because undetected degradation can precipitate signal-passed-at-danger (SPAD) events and collisions. On Tanzania’s Standard Gauge Railway (SGR), as on most Sub-Saharan African railways, maintenance of these assets remains predominantly reactive and time-based, with no continuous condition-monitoring instrumentation and long fault-detection latencies. This paper presents an intelligent condition-monitoring framework that couples a Long-Range Wide Area Network (LoRaWAN) sensing architecture with multi-sensor data fusion and a normalised Asset Health Index (AHI) to deliver predictive maintenance for colour-light signalling lamps and DC track circuits. Following a Design Science Research (DSR) methodology, the study proceeded through three phases: problem relevance, artefact construction and artefact evaluation. In the problem-relevance phase, a structured questionnaire survey of sixteen of the seventeen specialist SGR signalling practitioners on the Dar es Salaam–Morogoro corridor established the prevailing maintenance paradigm and identified the most informative fault indicators, namely light intensity and temperature for lamps, and track current and ballast resistance for track circuits, which directly shaped the sensor suite. Two fusion strategies, weighted-average and Dempster-Shafer evidential fusion, were formulated to compute a single interpretable AHI per asset, and four anomaly-detection algorithms were implemented: a statistical AHI-threshold detector, an Isolation Forest, a One-Class Support Vector Machine and a supervised Random Forest. On a physically consistent, synthetically generated multi-sensor proof-of-concept dataset of 8,064 readings, the supervised detector exceeded the F1 ≥ 0.90 design target for both classes under a stratified random hold-out (signalling lamp F1 = 0.962; track circuit F1 = 0.921), but not under leakage-controlled leave-one-asset-out validation (F₁ = 0.885 and 0.798, respectively). The AHI warning fired 27.0 and 36.2 hours ahead of a conventional threshold alarm on the two critical gradual-degradation faults and flagged warning-severity faults that the conventional alarm never detected. Modelled LoRaWAN link and energy performance met design targets (mean packet delivery ratio 98.85%; projected node battery life 5.95 years), although a long latency tail under traction-energisation events was observed. The results demonstrate that continuous multi-sensor health monitoring can convert a reactive maintenance posture into a predictive one for the two most safety-critical signalling asset classes on an electrified African SGR corridor.