From Data to Diagnosis point condition monitoring through machine learning
Author: Richard Parkinson
Day: Introduction Day
Session: Digitisation and Maintenance

We live in an age where we can collect and store huge amounts of data about anything we want to monitor. A railway infrastructure network could easily have thousands of point machines each moving many times a day, potentially creating an overwhelming amount of data. The challenge is in using this to answer to a few important questions:

  1. Which machines need attention and when?
  2. What are the issues?
  3. How well were they maintained?

In this presentation, key steps in turning that data into meaningful, high-level answers are explored. Capturing the right data:What affects point machine condition? Some factors can be measured directly such as temperature variation, vibration and track movement. Others must be measured indirectly like machine wear, damage, set up and quality of maintenance. Indirect factors can be monitored through the dynamics, stresses and electrical characteristics of the machine. All these measurement techniques come with challenges such as safety risks, a harsh railway environment, cost, location and the maintenance regime.Current and voltage monitoring are the most popular as they are simple, safe and inexpensive. Looking at several types of point machine, I will explore what level of data quality is appropriate, how they relate to position and stress, and their limitations in providing a complete picture of the asset condition.Linking the data to the electro-mechanics of the machine:Most point machines follow a standardised sequence of operation consisting of unlock, traverse and lock. These sections can vary dramatically in both the magnitude and shape within the power profile. Segmenting the data allows component operation to be isolated for accurate analysis.Machine learning can be used to identify these sections robustly and consistently with no human guidance. Defining machine condition:Once the point move has been segmented, metadata can be derived to quantify the machine state. Condition or 'health' can then be calculated to determine the immediacy of response required and to prioritise maintenance.Machine learning can be trained using historic data to identify these different health states.Identifying the causes of poor condition:There are various ways the causes of potential failures can be spotted.

  1. Unusual behaviour can be related to the mechanics, indicating where to inspect first.
  2. Templates can be created by applying faults to machines under test conditions and recording their characteristics.
  3. Lessons learned from past maintenance can guide how to fix problems.

These methods can be used to increase the effectiveness and reduce the frequency of maintenance.Communicating the diagnosis:It is important to provide clear and relevant information to all stakeholders. A front line maintenance team needs to know the identified issues and how effectively they solved them; a maintenance manager needs the information to be able to prioritise maintenance and scheduled in future intervention; and a senior manager would like to know KPIs of the overall network system state. We have developed a system, built upon machine learning and data analysis, tailored to provide answers to the questions in the introduction.