Research On Health Management Algorithm Of Elevator Servo System Based On Deep Learning

As an important part of the main control system of the aircraft,the elevator system controls the longitudinal movement of the aircraft and has practical significance for the flight safety of the aircraft.However,there is a strong coupling relationship between the internal systems of flight control system,which makes it difficult to accurately establish a physical model.In recent years,because of its multi-hidden layer structure,deep learning can extract information layer by layer,and can extract fault features for fault diagnosis and prediction.Therefore,this paper focuses on the fault diagnosis and prediction in the health management of elevator servo system based on deep learning.Aiming at the problem of poor fault diagnosis effect and weak generalization ability of traditional aircraft rudder surface,a convolutional neural network(CNN)algorithm combined with support vector machine(SVM)classifier is proposed to build a fault diagnosis model suitable for civil aircraft elevators.The elevator fault simulation model was built on the AMESim hydraulic simulation platform to perform fault simulation to obtain the corresponding fault data.Due to the high dimensionality of the acquired data,the principal component analysis method(PCA)was introduced to reduce the dimensionality of the data.PCA handling of the nonlinear system is difficult to extract characteristic value,so the kernel function is added for improvement,and the PCA algorithm improved by the radial basis kernel function is selected through experiments to reduce the dimensionality of the data.Data for dimension reduction after fault diagnosis with CNN-SVM model,and with the traditional network(CNN),deep belief network(DBN)model test,the results of the experimental results show that the proposed method for elevator fault recognition accuracy rate can reach more than 99%,in order to be able to directly observe the differences in the feature representation of these three models,the recognition results are dimensionality reduction visualization(T-SNE),and through the graph after visualization,it can be seen that the features extracted by the CNN-SVM model have obvious clusters.Finally,noise is added to the data set to verify that the model has good anti-noise ability,generalization ability and reinforcement learning ability compared with the other two models.Aiming at the complexity of elevator fault prediction,this paper proposes a rudder surface fault prediction method based on the combination of fault data training prediction model and classification model.After the dimensionality reduction of the KPCA method,the data training capsule network combined with the long-short-term memory network(Caps Net-LSTM)prediction model predicts the value of the feature value in the future single-step and multi-step changes in time,and uses the predicted output as the classification the input of the model predicts whether there will be failures and what kinds of failures will occur in the future.Experiments show that Caps NetLSTM’s single-step and multi-step prediction errors are reduced by 12.78% on average compared to Caps Net,and 25.13% on average compared with LSTM.Finally,it was verified on the real B777 flight data,selecting 21 parameters related to the flight path angle,predicting the change of the flight path angle,and fitting the curve between the predicted value and the real value to show the effectiveness of the selected model.