基于LSTM增量网络的载荷识别方法研究OA

Load identification problem includes load location and load time history reconstruction.In view of the ill-posedness of matrix inversion in load identification,the limitation of solution accuracy of model-driven methods by modeling accuracy,and the requirements of load location,a load location and time history reconstruction method driven jointly by Long Short-Term Memory(LSTM)neural networks and incremental learning was proposed.Firstly,the singular value decomposition of the acceleration signal was carried to dimensionality reduction to extract the core features and reduced noise and redundant information.Secondly,combined with the LSTM network,an implicit function model between the structural loading location,the magnitude of the load and the vibration response was gradually constructed,and an incremental learning mechanism was introduced to improve the generalization ability of the network model.Then,the network hyperparameters were automatically adjusted in combination with the Bayesian optimization method.Finally,numerical analysis was conducted on the rectangular cantilever plate model,and experimental verification was carried out using the aluminum alloy cantilever plate and the pantograph frame.The results show that the proposed method can accurately identify the load location and reconstruct the load time-history curve.The load location accuracy reaches 95.8%,and the peak errors are within 1.5%.This method has certain adaptability to the time-history reconstruction and load location problems of relatively complex structures subjected to impact loads.Only a small number of sensors need to be arranged in combination with the structural characteristics to effectively identify the impact loads.