Evaluation of plates in similitude by experimental tests and artificial neural networks

In the last century, the introduction of similitude theory allowed engineers to define the conditions to design a scaled-up or down version of the full-scale structure by means of a set of tools known as similitude methods: the scaled structure can be tested more easily, and then, by using the scaling laws, the prototype behavior can be recovered. However, such a response reconstruction may become hard for complex structure under incomplete or distorted similitude frameworks. Machine learning methods, with their automating characteristics, may help to circumvent these difficulties. This work is divided into two parts. First, five clamped-free-clamped-free plates in similitude are experimentally tested. In the case of complete similitude, these laws allow to accurately reconstruct the response. When the similitude is distorted, these laws are not always valid, failing to predict the dynamic behavior in some of the frequency ranges. Then, the experimental results are used to validate the prediction and identification capabilities of artificial neural networks. The artificial neural networks proved to be robust to noise and very helpful in predicting the response characteristics and identifying the model type, although an adequate number of training examples is needed. Further tests proved that the number of samples is drastically reduced by choosing accurately the features.

Automated summary

Similitude theory was introduced in the last century, allowing engineers to design scaled versions of structures using similitude methods. While traditional laws may fail in predicting dynamic behavior when similitude is distorted, artificial neural networks have proven to be robust in predicting response characteristics and model identification.