Study On Structural Response Virtual Sensor Techniques Based On Deep Learning

With continuous development of engineering structure,response-based Structure Health Monitoring(SHM),damage detection and active vibration control are widely applied to various fields.However,in the area of Mechanical Engineering and Aerospace Engineering,situations,where physical sensors are not able to installed or implement at target location,may occur,due to the severe working condition or the wiring limit.Thus,it is of great importance to develop response virtual sensors using partial response measurements.This thesis studies on the response virtual sensor models,the main contents of the work are carried out as follow:(1)Engineering applications and research background of virtual sensors and deep learning are introduced to support the development of the structure response virtual sensor.(2)The transmissibility functions are summarized as the prior knowledge and main inspiration of the response virtual senor for time-invariant structures;deep forward network models,convolutional neural networks models and recurrent network models,particularly the Long-Short Term Memory(LSTM)models and Convolutional Long-Short Term Memory(Conv LSTM)models,are summarized as the theoretical fundamentals of the virtual sensors.(3)Inspired by the transmissibility functions,a four-layer neural network model,consisting two convolutional layers,one hidden layer and one output layer,is introduced as the response virtual sensor for time-invariant structures.Numerical examples indicate that this method is capable to perform accurate response prediction.(4)Inspired the Time-varying Autoregressive model with e Xogenous inputs(TVARX),a four-layer neural network model,consisting two Conv LSTM layers,one hidden layer and one output layer,is developed based on the novel response virtual sensor for time-invariant structures.This model is introduced as the response virtual sensor model for periodically time-varying structures.Numerical examples verifies its performance in response prediction.(5)A time-invariant simply supported beam experiment system and a periodically time-varying experiment system,consisting a simply supported beam with a moving mass sliding on it,are employed to verify the performance of the proposed response virtual sensors for time-invariant structures and periodically time-varying structures respectively.Then,the comparison between the proposed response virtual sensor for time-invariant structures and an existing response virtual sensor is carried out to further validate its performance.Experimental results show that the proposed response virtual sensors are capable of predicting the responses of the time-invariant structures and periodically time-varying structures with high accuracy.Moreover,the comparison indicates that the proposed virtual sensor for time-invariant structures contains higher prediction accuracy and the ability in dealing with situations where limited response measurements are available.