A novel machine-learning based on the global search techniques using vectorized data for damage detection in structures

With recent ground-breaking advances, machine learning (ML) has been applied widely in numerous fields in this day and age. However, because of the application of backpropagation algorithms based on gradient descent (GD) techniques, the network of ML may be trapped in local minima, especially if its starting point is not on the same side of the global best or the network contains too many local minima. This drawback may reduce the accuracy and effectiveness of ML. To transcend these limitations of ML, numerous researchers have employed algorithms based on global search techniques to eliminate initial local minima of the network by looking for a beneficial starting point. Nevertheless, those solutions are only valid under certain circumstances when the network only contains a few local minima and they are distributed on the same side. With complex problems such as structural health monitoring (SHM), the network always exists of different error surfaces with numerous widely distributed local minima. The approach of the selection of a good starting position for the network may no longer be useful. Therefore, this paper proposes a novel machine-learning based on an evolutionary algorithm, namely Cuckoo search (CS) to solve the local minimum problem of ML in the most radical way. CS algorithm based on the global search technique is employed to work parallel with ML during the process of training the network. This win-win approach has both advantages of GD techniques (fast convergence) and stochastic search techniques (avoiding being trapped in local minima). The core idea of the proposed method is recapped as follows: (1) ML using the GD technique is first applied to speed up convergence; (2) if the network gets stuck in local minima, CS with global search capability is applied to assist particles in escaping from local minima; (3) the GD technique is applied again to increase the convergence speed. Steps 2 and 3 are repeated until the target is achieved. Additionally, to handle the large amount of data used to train the network, we also apply a vectorization technique for the data of the objective function, which significantly reduces the computational cost. This is another contribution of this work. To assess the performance of the proposed approach, both numerical and experimental models with different damage scenarios are considered. The results showed that the proposed approach completely outperforms CS, ML, and other hybrid ML in terms of accuracy and considerably reduces calculational costs compared to CS.