Advanced engine diagnostics using artificial neural networks
Introduction
Gas turbines (GT) are mechanical devices operating on a thermodynamic cycle with air as the working fluid. The air is compressed in a compressor, mixed with fuel and burnt in a combustor with the gas expanded in a turbine to generate power used in driving the compressor and external loads (thrust or shaftpower) depending on requirements.
The main gas path components of the GT, which are compressor, combustor and turbines, are usually very reliable. but could result in low availability of the whole unit if a forced unexpected outage is encountered as it can take some considerable time to effect repairs on them. This is made worse if the breakdown occurred when the maintenance crew was unprepared for it. Improving availability and reducing life-cycle costs of the GT require maintenance schemes such as condition-based maintenance (CBM) that advocates maintenance only when it is necessary and at the appropriate time instead of after a fixed number of operating hours or cycles. For the operational health of the engine to be regularly monitored for gas path faults, such measurable parameters as shaft speed, pressures, temperatures, fuel flow and shaftpower thrust are required.
Gas path faults include fouling, erosion, foreign object damage, blade tip clearance and corrosion. These cause changes in the component performance parameters and because these are not measurable, the thermodynamic relationship between these performance parameters such as efficiencies and flow capacities and the measurements taken from the engine can be used to detect, isolate and assess the level of fault present on any component.
Traditional techniques for gas path fault diagnosis such as visual inspection, fault trees, fault matrixes and gas path analysis (GPA), have their limitations. Current research have thus been focussed on the application of such advanced techniques as artificial neural networks (ANN), genetic algorithms (GA), expert systems and fuzzy logics to engine diagnostic problems.
In this paper, we review the need for engine diagnostics and maintenance, present some aspects of the ANN application to diagnostic problems, highlight some features of ANN that make it amenable to GT diagnostics as well as its limitations and finally discuss its application to gas path fault diagnosis of a developed case study. The engine used for this analysis is a two-shaft aeroderivative gas turbine, thermodynamically similar to the Rolls Royce Avon. A schematic of this engine’s configuration is shown in Fig. 1.
Section snippets
Reliability, availability and economy
Operation and maintenance costs of a gas turbine contribute a major portion of the annual maintenance budget of a company. In view of the changes in world economy towards globalisation and openness of the market, any efforts that can reduce the total cost of ownership and life-cycle cost of the equipment will be added advantages.
The primary objectives of all maintenance strategies are to reduce equipment downtime, increase reliability and availability of the equipment which at the same time
Artificial neural networks and fault diagnosis
Information in the open literature shows the level of research being carried out in fault diagnosis using ANN. ANNs have been applied in the following areas amongst others:
- •
Single and dual sensor fault (SF) identification and assessment of industrial power plant [2], [3], [4], [5], soft failures of sensors and actuators in automobile engines [6].
- •
Fault diagnostics in nuclear power plants [7], [8], turbofan engines [9].
- •
Mechanical damage detection involving vibration signals measured from faults
Measurable parameters and measurement uncertainties/errors
It is obvious that the ability to accurately determine engine health largely depends on the accuracy of measurements available. Many sensors installed on the engine operate in very hostile environments at extremes of temperature and/or pressure. Unfortunately, sensor measurements are often distorted by noise and bias, thereby masking the true condition of the engine and leading to incorrect estimation results. This creates the situation where sensor reliability may be lower than component
Case study
Gas path faults can occur during the operation of a gas turbine and because they affect performance and life, it is necessary that they should be diagnosed and corrected. We have addressed this problem by applying a hierarchical neural network structure (Fig. 5). The procedures adopted include the following steps:
- 1.
Obtaining a thermodynamic model of the engine from which simulation data would be generated for training and testing the networks. This approach was applied because it is extremely
Network anatomy and results
In Table 3, a summary of the classification networks developed in this work are presented including results obtained in terms of correctly classified test patterns. The network type is probabilistic neural network (PNN) which can be set up in less than 2 min when data is available. This network requires no “training” but its hidden layer takes up processing units or neurons equal to the number of training patterns while the input and output layers are, respectively, equal, in terms of the number
Comparison of diagnostic approach with other techniques
Table 9, Table 10 show the comparison between the diagnostic results from two gas path analysis techniques with those of the trained networks for compressor and gas generator turbine faults, respectively. GPA as a tool for engine diagnostics was initially introduced by Urban [16] and involves the thermo-mathematical matching of engine measurements (dependent variables) to performance parameter (independent variables) changes. This is based on the premise that faults in the gas path of a gas
Conclusion
A hierarchical approach to gas path diagnostic for a two-shaft simple gas turbine involving multiple neural networks has been presented. The described methodology has been tested with data not used for training and generalisation is found to be appropriate for actual application of this technique. Also, the level of accuracy achieved by this decentralised application of ANN shows derivable benefits over techniques that require just a single network to perform fault detection, isolation and
References (17)
- et al.
Fault diagnosis in power plant using neural networks
Inform. Sci.
(2000) - et al.
Multiple-sensor fault-diagnoses for a two-shaft stationary gas-turbine
Appl. Energy
(2002) - et al.
Artificial neural network based fault diagnostics of rotating machinery using wavelet transforms as a pre-processor
Mech. Syst. Signal Process.
(1997) - R. Singh, Managing gas turbine availability, performance and life usage via advanced diagnostics, in: Proceedings of...
- T.-H. Guo, J. Nurre, Sensor failure detection and recovery by neural networks, in: Proceedings of the International...
- T.-H., Guo, J. Saus, C.-F. Lin, J.-H. Ge, Sensor validation for turbofan engines using an autoassociative neural...
- D.W. Dong, J.J. Hopfield, K.P. Unnikrishnan, Neural networks for engine fault diagnostics, in: Proceedings of the 1997...
- et al.
Use of artificial neural networks to analyze nuclear power plant performance
Nuclear Technol.
(1992)