Research On Aero-engine Sensor Failure Diagnosis Based On Neutral Network

The sensors of aero-engine control system are the most prone to failure. If a failure occurs, it will probably result in the engine control system failure and a terrible disaster will be followed. Building and improving the sensor Fault Detection, Isolation, Accommodation(FDIA) mechanism of the aero-engine control system is necessary which will greatly improve the system reliability. In this paper, the issue of sensor failure of aero-engine control system is researched and the neural network is used to design the FDIA system which has strong features of non-linear, fault tolerance and self-learning. This FDIA system can not only detection and isolation the engine sensor fault, but also can reconstruct the signal of the fault sensor. to ensure that the sensor failure control system can continue to operate safely and reliably. Thus, the aero-engine control system can work normally and reliably.Applications of neural network in the aero-engine sensor failure diagnosis system are researched in this paper. The mainly works were done as following:First, a new approach is proposed using partition flight envelope and two-stage empirical modeling strategy based on neutral network in the process of whole flight envelope multi-state modeling. Clusters of input parameter data are formed using GMMs and neutral network predictors are designed and trained for every flight envelope partition, so the destination of the whole flight envelope aero-engine sensor failure diagnosis can be realized in real-time and on-board.Second, for each sub-block of the flight envelope, dual redundant predictors based on BP neutral network are designed working on the sensor diagnosis system. The neutral network temporal redundant predictor and spatial redundant predictor were created over the time series redundant information of single sensor and the space redundant information of multi-sensor respectively. The threshold-value differentiate method was applied for real-time sensor failure detection. Isolate the fault sensor and use the outputs of spatial redundant predictor to accommodate it.Finally, a common simulation platform is established using Simulink software. Several simulation experiments were done to verify the effectiveness of the method proposed above.Stuck failure, constant-bias failure and Cconstant-gain failure are simulated and analyzed.After that, the various factors that impact the result of fault diagnosis are analyzed. The problem how to reduce the misdiagnosis rate and missed diagnosis rate is discussed too.Digital simulation results show that these methods are feasible and effective.