Condition Monitoring and Fault Diagnosis of Fluid Power Systems : Approaches with Neural Networks and Parameter Identification
Abstract: This thesis deals with condition monitoring and fault diagnosis of fluid power pumps and systems. Neural networks and parameter identification techniques have been important in the work. Vibration measurements have also been of importance.There is a need to find deterioration in fluid power systems (as in any system) at an early stage. There are many reasons for that, mainly related to economy, but also to security and environmental problems. It is expensive to have a non-running system, waiting for an unplanned repair, but it is also expensive to change parts of the system before they are worn out. If a part of the system breaks down, it may also cause personal injuries and environment pollution. One way to find deterioration is to use condition monitoring. It is important that the condition monitoring system is inexpensive, simple and robust. Discussed in this thesis is the use of vibration measurements for condition monitoring of fluid power pumps. Condition monitoring and fault diagnosis of a more complex system are also discussed. The analysis was carried out on a simulated system of the hydraulic control system of the gear box of a large vehicle.Neural networks trained with different training algorithms are investigated. The main reason for the neural network training for condition monitoring is to produce as reliable results as possible; the training time is not of great importance. A methodology is proposed for condition monitoring by means of neural networks. A methodology based on identifying system parameters in a well verified simulation model to reach a model behaviour identical to the system behaviour is also discussed.The neural network approach requires advance knowledge about the system behaviour under faulty conditions and assumptions about which faults are likely to arise. The parameter identification approach requires a well verified simulation model but no actual knowledge or assumptions about probable faults. The neural networks require a rather lengthy training time, while the evaluation is immediate. The evaluation with parameter identification involves optimisation and will require some time. This means that the main computational requirements are in the starting phase for the neural network method, while it is in the evaluation phase for the parameter identification method.
This dissertation MIGHT be available in PDF-format. Check this page to see if it is available for download.