Modelagem de sinais vibracionais por sensores virtuais baseados em LSTM para análise modal

Abstract

Structural monitoring based on vibration signals has become a relevant tool for tracking the integrity of structures subjected to different operational loads, mainly in situations where the instrumentation of physical sensors becomes limited by several factors. In this context, virtual sensors act as a complementary alternative to traditional monitoring, allowing responses in regions without instrumentation to be estimated from known variables. Furthermore, for applications in modal analysis, it is not enough for the predicted signal to present only low statistical error indexes, it is also necessary for the estimated signal to preserve the modal characteristics of the structure, such as natural frequencies, damping and mode shapes. Therefore, this work aimed to evaluate the performance of different configurations of LSTM neural networks in the construction of virtual sensors for modeling vibration signals of a cantilever beam. The research was developed based on signals collected in a wind tunnel at velocities of 5 m/s and 10 m/s, in which a beam was instrumented with five accelerometers, with sensors A1 and A5 used as network inputs and sensors A2, A3 and A4 used as scanning points to be estimated. Three LSTM configurations were evaluated, varying the size of the temporal window, the number of neurons and the dropout rate, while the comparison was carried out using statistical metrics, considering MAE and RMSE. Then, the signals predicted by the best configuration were evaluated using the SSI and FDD methods and the Modal Assurance Criterion (MAC). The results showed that configuration C2, designed with a temporal window of 256 points, 64 neurons per layer and dropout of 0.2, presented the best overall performance, indicating that the simple increase in the temporal window contributed more effectively to the estimation of the signals than the increase in network complexity. In the modal analysis, the virtual signals maintained the modal characteristics, presenting low values of relative frequency error and damping ratios. In addition, MAC values equal to 1.00 on the main diagonal indicated a strong correspondence between the real and virtual mode shapes. Thus, the results allow concluding that the virtual sensors based on LSTM networks presented satisfactory performance in the modeling of vibration signals, as well as potential for applications in Operational Modal Analysis.