Performance comparison of non-adaptive and adaptive optimization algorithms for artificial neural network training applied to damage diagnosis in civil structures

• Published in: Applied soft computing 2021-06, Vol.104, p.107254, Article 107254
• Main Authors: de Souza, Calebe Paiva Gomes, Kurka, Paulo Roberto Gardel, Lins, Romulo Gonçalves, de Araújo, José Medeiros
• Format: Article
• Language: English
• Subjects: Artificial neural network; Civil structures; Damage diagnosis; Non-adaptive and adaptive algorithms; Real-time monitoring
• ISSN: 1568-4946; 1872-9681
• DOI: 10.1016/j.asoc.2021.107254

The structural health monitoring of civil structures has been highlighted by the perception that the cost involved in the prevention of structural accidents is lower than the cost of correcting them. In addition, in the case of large structures (e.g. bridges, dams and industrial sheds), it is desired that the damage diagnosis should occur quickly from real-time monitoring without interrupting the use of the structure. The structural diagnosis may be performed based on vibration measurements, considering that the structural damages modify modal parameters of the structure (e.g. frequencies and mode shapes). With these dynamic responses, the application of appropriate computational tools to recognize and classify structural damage is intended, and artificial neural networks (ANN) have gained a lot of attention in achieving these goals. In this work, the methodology of diagnosing structures by real-time monitoring is originally developed and based on initial definition of the optimal topology, avoiding both the use of multiple hidden layers and the combination of several neural networks, and by applying of non-adaptive and adaptive first-order algorithms for agile network training, in order to be mathematically suitable for continuous and real-time monitoring, which allows updating both the dataset and neural parameters without greater computational effort. As some of these algorithms were not addressed in the diagnosis of civil structures with the aforementioned hypotheses, until this research, the performance of each algorithm was verified in case studies that simulate classic structural systems adopted in most civil buildings. Finally, the results endorse the feasibility of improving the structural diagnosis based on the training of a simple neural network, with one hidden layer, associated with non-adaptive or adaptive first-order optimizers that guarantee the agile assessment of structural integrity in real time. •Large structures may be monitored without interrupting their operation.•Damage diagnosis based on vibration measurements is an efficient practice.•Individual neural network with single hidden layer enable the damage diagnosis.•First-order learning methods are appropriate for real-time structural health monitoring.•The proposed approach allows continuous assessment of structural integrity.