Research Of Aero-engine Sensor Fault Diagnosis Based On Extreme Learning Machine

In view of the problems that traditional gradient-based learning algorithms faced, such as parameters selection difficulty, easily falling into local minimum, over-fitting and so on, the method of aero-engine sensor fault diagnosis based on Extreme Learning Machine is studied. It covers the advantages and improvements to Extreme Learning Machine algorithm, the establishment of double redundancy sensor fault diagnosis mechanism and the simulation of sensor fault diagnosis.Firstly, the advantages of Extreme Learning Machine algorithm have been analysed. As Extreme Learning Machine algorithm only needs to set the number of hidden layer neuron and select activation function, so it reduces the training time. Simulation shows that in the case of the similar training model accuracy builded, the rate of Extreme Learning Machine algorithm is above two hundred times faster than the rate of BP neural network algorithm. Their testing accuracy and testing time are close too.Secondly, the more realistic Online Sequential Extreme Learning Machine algorithm is analyzed,then Weight Online Sequential Extreme Learning Machine algorithm is proposed to improve the algorithm. By assigning different weights to the old and new data, no-equal weight process of those data is done. Simulation shows that Weighted Online Sequential Extreme Learning Machine algorithm has a higher model’s testing accuracy than traditional Online Sequential Extreme Learning Machine algorithm.Finally, the double redundancy models of aero-engine sensor fault diagnosis based on Weighted Online Sequential Extreme Learning Machine algorithm are established, and time redundancy and spatial redundancy are used to locate the fault sensor. And the idea of using residual changes amplitude of space redundancy to detect soft errors is proposed, which makes soft fault detection faster. Simulation results show that the program is feasible. It is able to diagnose the soft fault and hard fault of a single sensor accurately, and it can also display the results more intuitively.