Hybrid machine maintenance planning systems using case-based reasoning and evolution strategies

Abstract

Machine failure prediction involves the processing of massive amounts of data, expert's knowledge of various types of machines, problems of noisy data and operation variability. The chief advantages of effective machine failure prediction are that it improves system reliability, reduces maintenance costs, and improves response times in emergencies as well as the appropriateness and management of these responses. In recent decades, there has been much promising research into machine failure prediction that has involved neural networks, case-based reasoning and data mining, yet while all of these methods have proven useful, they differ in their suitability for application to machines of differing levels of reliability. Neural network approaches also require additional difficulties in that they must use specific feature extractors and that they learn slowly. This project has the objectives of integrating neural networks, case-based reasoning and data mining approaches into an Integrated Optimization Maintenance Planning System (IOMPS) that will create a machine failure prediction model which incorporates an efficient machine maintenance planning system including an automatic failure prediction and scheduling features. To determine the optimized application model, we first pre-process the machine maintenance case file data. This data is then hybridized through a neural network and using mining association rules for case-based reasoning. The output of these two processes is then evaluated for speed and accuracy. This knowledge is then applied to future similar problems. IOMPS has been used to handle complex machine failure predictions and associated maintenance planning on a number of the Hong Kong Mass Transit Railway Corporation Limited (MTRCL) systems, these systems, including the Automatic Fare Collection System, Passenger Escalator and Station Chiller Plant of (MTRCL), all operate differently as their maintenance planning processes are individually designed, yet IOMPS significantly outperformed the existing systems.