A comparative study of vibrational response based impact force localization and quantification using different types of neural networks / Wang Yanru
Impact force indentification plays an extremely important role to monitor structure health, where impacts can damage the structure, such as moving vehicle and industrial machines. Direct identification is not efficient and difficult due to environmental constraints and complex machines movements. Th...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
2018
|
| Subjects: | |
| Online Access: | http://studentsrepo.um.edu.my/9202/1/Wang_Yanru.jpg http://studentsrepo.um.edu.my/9202/11/wang_yanru.pdf http://studentsrepo.um.edu.my/9202/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Impact force indentification plays an extremely important role to monitor structure health, where impacts can damage the structure, such as moving vehicle and industrial machines. Direct identification is not efficient and difficult due to environmental constraints and complex machines movements. Therefore, a lot of indirect methods were proposed in past studies. ANNs is one of the methods, which has drawn attention by researchers in recent years due to its unique advantages compared with other indirect identification techniques. It can analyze complex relationship of nonlinear input-output by learning from datasets without any mathematical model. Past studies mainly used conventional network Multilayes Perceptron (MLP) for impact indentification and it was already successfully applied in soem fields. But there are some disadvantages for MLP, such as local minima and slow learning speed. Radial Basis Neural Network (RBFN) was proven more error-tolerant and better than MLP using input feature peak arrival time (PAT). iN the previous study, more accure input feature (i.e minimum arrival time (MAT)) was proposed to compare the accuracy of RBFN and MLP. In the study, the effort to find better neural network algorithms in the application of impact force indetification continues. Another two Adaptive Neuro-fuzzy Inference System (ANFIS) and Generalized Regression Neural Network (GRNN) are proposedto study their effectiveness in solving impact indentification problem. This is because GRNN can avois local minima like MLP. Inaddition, it has similiar architecture with RBFN. MLP & RBFN use gaussian function as activation fuction, but GRNN has one extra special linear layer, where outputs are considered in this layer, thus GRNN's performance in impact force indetification is expected to be good. Moreover, GRNN evaluates each output independently from the other outputs. It may be ore accurate than MLPwhen there are multiple outputs. In addition, ANFIS uses hybrid learning algorithm. It is mixed with least mean square and gradient descent method, which cause many advantages, such as much better learning ability and less computational time. Therefore, this study will compare the accuracy and the effectiveness of ANFIS and GRNN with the conventional RBFN and MLP algorithms throught experimental verification. The result showed that the proposed neural networks GRNN and ANFIS were effective to identify impact force. The most appropriate neural network for impact force localization was GRNN, which improved the accuracy by 66.11% than previous algorithm RBFN. In addition, the most proper neural network for impact force quantification was ANFIS, which improved the accuarcy by 42.35% than the commonused neural network MLP. |
|---|