On the other hand, the features that were not extracted directly from the waveform, such as features derived by a subsequent model fitting process, were not as sensitive as previous two features. The parameters set of a decaying function fit to the data is an example of such features (see Table 1). It was considered that this low-sensitivity was caused by the difference of error components between experimental and numerical responses. The experimental data generally includes the measurement noise and environmental variability. However, numerical responses are free from such noise components. The result of fitting process was greatly influenced by variability in the experimental data; therefore, it was not considered to be appropriate to use such features for comparing experimental and numerical outputs. High dimensional features such as frequency response function were not considered because of difficulties in making quantified comparisons of such high-dimensional quantities. For validating nonlinearity modelling parameters; Gap and ', we found that some time-series model parameters; e.g., AR model parameters as shown in Table 2, had sensitivity to the accuracy of response. This point was also indicated in the previous work that summarized response features for analysing nonlinearity in dynamic responses for the application of structural monitoring [2]. Table 1. Examples of response features for validating damping parameters (Blue: experiment, Red: numerical with all 1% damping ratios, Green (accurate); numerical with damping ratios from experimental modal analysis) Feature description and definition: output xi Response feature plot Standard deviation (Basic statistics) Energy (Temporal moments) ¦ N i i k i k t x M 1 2 in k=0 Decaying function fitting y(t)=Y0exp( Dt) 157
RkJQdWJsaXNoZXIy MTMzNzEzMQ==