| Piezoelectric ultrasonic guided wave-based structural monitoring technology is widely used in ocean engineering,aerospace,power systems and many other fields,and has great practical significance for structural safety assessment and operation maintenance.Deep learning provides a promising tool in signal characterization,feature extraction,anomaly detection for ultrasonic guided wave sensing,and avoids the application difficulties of dispersion characteristics and modal mechanism in ultrasonic guided wave sensing to the greatest extent.However,the biggest challenge of data-driven monitoring method in practical engineering is that,a complete data set covering multiple monitoring structures,multiple monitoring tasks and multiple types of damage is quite difficult to be constructed.When the structural state data set is incomplete,how to apply the knowledge acquired from the existing data set to related fields through knowledge transfer is an attractive research to solve such problems.Focusing on key transfer monitoring issues in ultrasonic guided wave structure state monitoring,this thesis mainly carries out researches on the coding characterization and feature extraction,multi-task learning,semi-supervised learning and cross-structure transfer detection.With the help of the signal feature learning ability of deep learning and the knowledge generalization ability of transfer learning,these works can make full use of the advantages of wide detection range and high detection efficiency of ultrasonic guided wave.The main contents are listed as follows:(1)To realize intelligent structure state recognition based on ultrasonic guided wave sensing,a signal characterization method based on deep encoding network is studied.In the aspect of feature extraction and features encoding,an ultrasonic guided wave feature coding method based on local time-frequency features and unsupervised feature fusion is proposed.In the aspect of original signal encoding and representation,a deep convolutional autoencoding method is proposed.The deep encoding network can realize the main features extraction of the structure state monitoring signals and dimensionality reduction of the original guided wave signal,and overcome the shortcomings of the current guided wave detection methods which still need to manually select the guided wave features and rely heavily on the expertise.(2)To realize the structural state recognition at the absence of damage state samples,a structural state imaging algorithm based on semi-supervised deep fusion features is proposed.The siamese network is employed as the semi-supervised outlier state recognition model,and the feature mapping distance of the original signal encoding is utilized as the quantitative damage index.At the same time,the applicability of simulated damage reconstruction signals in structural damage monitoring of different material structures is studied.In this part,the structural state probability imaging of multipath ultrasonic guided wave sensing data is realized.(3)To realize the structural state recognition at the absence of some monitoring task samples,the original guided wave signal encoding was implemented as a general feature extraction model,and a multi-task structural state monitoring model based on ultrasonic guided wave sensing was established.The intelligent multi-task identification was accomplished by outputting multiple monitoring labels through parameter hard sharing multi-task network.A novel alternate training mechanism is designed to obtain enhanced representation of multi-task transfer features of ultrasonic guided wave signals.When the target task data is insufficient,it shows a better structural state detection ability,and the accuracy of the shared feature in the damage location recognition task is improved by 15.24%.(4)To realize the structural state recognition at the insufficiency of training samples of the target structure,a cross-structure state monitoring method based on feature distribution adaptation was proposed.By adapting to the feature distribution of ultrasonic guided wave signals in different structures,the consistent representation of the damage characteristics of the source domain structure and target domain structure is realized.In terms of network structure,the convolutional long-short term memory network can combine the temporal features and local features to synthesize better ultrasonic guided wave features.When the training samples of the target structure is insufficient,the damaged area energy ratio is increased by 95.38%,which effectively improved the adaptation learning and transfer ability of the detection model.The whole research work focuses on the structural state transfer detection method when the complete standard database has not yet been built.Combining with the deep learning and transfer learning technology,a novel ultrasonic guided wave detection approach is proposed in insufficient training samples condition,which provided a new path for guided wave feature extraction,multi-task transfer detection,semi-supervised learning and cross-structure transfer detection. |
PDF Full Text Request