An identification for channel mislabel of strong motion records based on Siamese neural network

Strong motion records are first-hand data for studying the seismic response of sites or engineering structures, and their objectivity is crucial for the credibility of the results in earthquake engineering and engineering seismology. However, domestic and international earthquake data may be mislabeled between horizontal and vertical channels. This issue is typically addressed by manually comparing the similarity between the three components of strong motion records, which is inherently subjective and inefficient in identification. To achieve the intelligent recognition of massive records, this study used 14,983 sets of ground motion records with significant differences between horizontal and vertical components from the NGA-West2 database. A Siamese neural network preliminarily distinguished the similarity between the acceleration waveform and the three components of the Fourier amplitude spectrum (FAS) of ground motion records. Combined with manual identification, an efficient and accurate method for identifying vertical components in ground motion records was proposed, and applied to verify the channel directions of the strong motion records in Strong Motion Network in China. It was found that 308 sets of records from 170 stations were suspected of mislabeling vertical and horizontal components. This advancement significantly enhances the objectivity of strong motion records. This proposed method holds potential for remote maintenance of strong motion stations, verifying the channels of strong motion instruments, and mitigating the negative impact of channel confusion on research results. •The study proposes a method to verify the accuracy of vertical component labels in ground motion records by using Siamese neural network.•The method combines the similarity of acceleration waveforms and smoothed FAS. It improves efficiency while ensuring accuracy.•This method was applied to test and verify Chinese strong motion records.