Failure and degradation prediction by artificial neural networks

Passenger and freight railway demand has grown significantly during the last decades. Railway operators have increased service frequencies and extended network connectivity to serve this demand, but these adjustments have increased system complexity and decreased buffering capacities. As a result, railway systems are now operating closer to their stability boundaries. Consequently, single failures or delays are more likely to propagate through the railway network and reduce operational reliability. One way railways are addressing these problems is by installing advanced monitoring and diagnostic devices in many components. These devices assist operating and maintenance personnel in handling faults and failures and thereby help reduce down time. While these monitoring systems advance conventional maintenance procedures, their main potential is providing input data for proactive predictive maintenance programs. In a predictive maintenance program, activities are performed based on the predicted failure and degradation behaviour of individual systems, made using the actual condition of these systems. This is in contrast to standard preventive maintenance programs based on the condition of an average system. Predictive maintenance programs increase operational reliability, help reduce delays, improve the efficiency of resource use and reduce overall maintenance costs, compared to standard preventive maintenance programs. However, these advanced monitoring and diagnostic systems generate a very large amount of high dimensional data that is difficult to analyse using standard approaches. In several similar cases, self-adaptive data-based algorithms, such as artificial neural networks, have been shown to be a promising method for predicting failures or degradation behaviour when applied alone or in combination with other prediction methods. This research evaluated the potential of artificial neural networks for predicting failures and degradation behaviour in railway systems. The research considered an extensive range of neural network types, characteristics and combinations, as well as different railway systems and problem types. The research developed a framework for designing and applying neural networks for predicting failure and degradation behaviour of railway systems. Several approaches were derived based on this framework, and then validated in eight case studies covering different railway systems, problem and data types, different algorithms and their characteristics. The main practical goal was to identify approaches for predicting the occurrence of operational disruption events and critical degradation conditions based on diagnostic data. The datasets used in the case studies were derived from two real railway systems: a rolling stock fleet and a turnout infrastructure system. As part of this research, seven types of neural networks were tested to diagnose and predict failure and degradation behaviour of railway systems: echo state networks, restricted and conditional restricted Boltzman machines, growing neural gas, extreme learning machines (a combination of artificial neural networks and support vector machines), deep belief networks, and multilayer feedforward networks based on multi-valued neurons (a special type of complex valued neural network). The neural networks were applied in three ways: as stand-alone algorithms, in combination with other neural networks or with other machine learning and soft computing techniques, such as principal component analysis and fuzzy sets. The approaches developed as part of this research were also compared to an alternative prediction approach, multilayer perceptrons trained with different types of learning algorithms.