Research On The Machine Learning Method And Its Application In Damage Assessment Of Plate Frame Subjected To Underwater Explosion

Underwater explosion(UNDEX)can cause shock wave,bubble pulsation and jet loads,which can result in different degrees of damage to the ship structure and seriously threaten the safety of the ship.At present,the research on damage assessment of ships subjected to UNDEX is mainly carried out by numerical simulation method,which generates massive data.Mining and analyzing the existing simulation results by machine learning(ML)methods can rapidly and accurately predict ship damage characteristics,which will bring unprecedented changes to the field of UNDEX and ship damage assessment,.and has extremely important military and engineering significance.Firstly,the basic theory and current study of physical phenomena,load characteristics and structural damage in the process of UNDEX were summarized.Then,deep neural networks(DNN)method and its applications in the field of ship damage assessment were introduced.At the same time,the basic principle of Arbitrary Lagrangian Eulerian(ALE)method was introduced,and the effectiveness of ALE method in damage assessment of ship structure subjected to UNDEX was verified by comparison with empirical formulas and experiments.Besides,the applicability of the DNN method in UNDEX problems were also verified.Secondly,taking the final plastic deformation and plastic strain distribution of plate frame structure under near-field UNDEX as an example,the influence of the number of hidden layers and neurons,the type of optimizer and activation function on the prediction accuracy and training efficiency of the DNN model was studied,and the optimal DNN model is determined.On this basis,the influence of input parameters in the DNN model was further studied,and the phenomena of under-fitting and over-fitting in the training process were discussed.Finally,a callback function was added to the model,which improved the training efficiency on the premise of ensuring the prediction accuracy and avoided the influence of under-fitting and overfitting on the prediction results.Thirdly,three different DNN models were constructed to predict the final plastic deformation,plastic strain distribution and displacement-time curve of typical positions of plate frame structures with different number and thickness of stiffeners under near-field UNDEX with different explosive equivalent and detonation distance.By taking the final plastic deformation prediction results as the input variables,the prediction accuracy of plastic strain distribution considering the influence of the boundary and the stiffeners was improved.In addition,by increasing the number of hidden layers and neurons,the displacement-time curves of the typical positions on the plate frame structures under UNDEX was predicted.Finally,a support vector machine(SVM)was established to identify the damage modes of the structures with different thickness under different intensity underwater contact explosion.On this basis,a DNN model was established to predict the breach size,crack shape,plastic deformation and strain distribution of the structures under underwater contact explosion.In addition,by inputting the time and the displacement and strain of the structures at the corresponding time into a new DNN model for training,the development process of the fracture of the frame structure under underwater contact explosion was predicted.