Deep Learning-based Blade Damage Detection Under Variable Conditions

High-speed rotating blades are widely used in modern industry,early identification of blade damage is very necessary.Blade tip timing(BTT)technology is an online monitoring method of blade vibration with great development potential at present.The blade vibration signals obtained by this method are severely under-sampled signals.In this thesis,VAMP reconstruction method based on deep learning is introduced to sparse reconstruction of signals,and multi-frequency vibration signals of blades are obtained.BTT method brings two inherent problems.On the one hand,the sensor data is highly uncertain due to the deviation of the blade.On the other hand,as BTT signal is a serious under-sampled signal,it is difficult to get the real vibration characteristics of the blade.In addition,variable working conditions often lead to obvious non-stationary nonlinear characteristics of blade vibration,so it is very difficult to extract sensitive blade damage characteristics.Therefore,how to overcome the influence of variable conditions and identify sensitive damage characteristics is a key problem in the application of BTT method in blade damage detection in engineering practice.Therefore,the deep learning method is introduced to solve the problems that the signal has strong uncertainty,the real vibration characteristics of the blade are difficult to be obtained by reconstruction,and the vibration signal has non-stationary nonlinear characteristics,which makes it difficult to extract sensitive features.This method can process complex nonlinear signals and extend the hierarchical representation of the original data by constructing a deep neural network with multi-layer nonlinear transformation.In addition,it can automatically learn and extract intrinsic feature information,independent of prior knowledge and experience,which solves the difficult problem of extracting sensitive features.Therefore,this thesis proposes a blade damage detection method based on deep learning,and obtains BTT vibration data through simulation and carries out experimental verification.The structure of the main content includes:1.Aiming at variable speed,due to the static angle position error,blade translational motion and reverse motion cause BTT vibration signal analysis produced deviation.Aiming at the problems,analyzed blade tip timing vibration measuring principle and factors that influence the mechanism analysis and their solutions,and reduced the deviation caused by the offset motion of the blade at variable speed.2.Aiming at the compressed sensing model in the order domain,the dynamic simulation model of vibration displacement data of damaged and normal blade was constructed,and the order analysis of displacement signal of BTT sensor is carried out to verify the validity of data.3.Aiming at the problem of multi-frequency vibration signal reconstruction under variable working conditions,a sparse reconstruction model based on VAMP algorithm is constructed,and a deep learning method is introduced to verify the reconstruction order and theoretical order of the model.The advantage of this method is that it has higher precision results and less noise interference.4.To evaluate crack of blade multi-frequency vibration signal,puts forward blade damage detection method based on the deep learning under variable speed,analysis the effect of blade vibration signal preprocessing on model accuracy,and compared different layer number of model precision and the length of the training time,deep learning method get well damage detection ability and generalization ability.5.Build a platform of blade vibration data,collected crack and normal blade data of BTT sensor.The data was reconstructed,then conducted blade damage detection,the results show that under the variable condition can accurately reconstruct order,its damage detection model can accurately distinguish crack blade data and normal blade data.