BaseLine — Explainable Rail Anomaly Triage

Rail-maintenance teams receive large volumes of inspection, vibration, and acoustic evidence, but the information can arrive in formats that make prioritization slow and difficult. BaseLine explores a narrow question: how might a maintenance triage engineer understand the evidence behind a possible anomaly before deciding what deserves inspection?

This public case study does not claim predictive accuracy, operational validation, or automatic control of rail infrastructure.

The opportunity framing was based on desk research into rail inspection workflows, explainability, anomaly triage, and safety-critical decision support. No railway operator sponsored this concept, and no live infrastructure data or operator interviews are presented as evidence here.

The concept deliberately starts narrow:

One user: a maintenance triage engineer

One job: decide where to look first

One example defect class

Human review before every operational decision

No automated alerts to drivers, speed restrictions, or infrastructure actions

The interface exposes the evidence and reasoning associated with a detection before asking the engineer to accept a severity judgment.

The concept distinguishes between single-source and corroborated evidence so confidence is not communicated as a binary certainty.

Detection thresholds are visible and adjustable in the concept. Changes are intended to be logged rather than hidden in an administrative layer.

BaseLine is decision support. It does not issue automatic infrastructure commands or replace inspections.

A real pilot would require an operator partner, representative historical data, maintenance-triage participants, and an agreed defect class. Success criteria should be defined with the operator before testing and should measure decision usefulness, false-positive burden, comprehension, and whether the evidence changes inspection priorities.

A working interaction prototype, a narrowly defined MVP hypothesis, and a pilot framework. It demonstrates explainability and human-control principles; it is not evidence that the underlying detection model has been validated.