Optimising Hyperparameters of Artificial Neural Network Topology for SHM Damage Detection and Identification

Artificial neural networks (ANNs) are a powerful method for solving classification problems, particularly for clustered data. However, one of the main challenges in using ANNs for general classification problems is determining the best topology (number of layers and neurons per layer) for a given problem or dataset. This study proposes a novel approach to address this challenge by using a optimisation algorithm. The algorithm first solves an external optimisation problem to obtain the optimal topology that maximises the accuracy of supervised learning. After, an inner optimisation problem is used to train the network to detect damage, using the Frequency Response Function of vibration-based measurements as input. To test the proposed methodology, it is applied to three different mechanical systems: metallic beams, rolling bearings, and composite plates. Principal component analysis is used to reduce the number of inputs to the neural network, and the external optimisation problem is solved using Particle swarm optimisation. The results demonstrate that the proposed methodology can accurately assess damage. Through the results, a discussion is presented about the potentialities and limitations of the proposed methodology as a support tool for damage detection.