The multidimensional long-term prediction of seismic time-history in dam areas holds significant im-portance for rapid damage analysis.Virtual sensors,as complementary sensing mechanisms to seismic physical sensors,facilitate seismic time-history predictions.However,existing virtual sensors face challenges in effec-tively predicting long-term sequences for multiple signals,leading to delays in analyzing dam seismic damage.Addressing the aforementioned issue,a multi-output seismic time-history long-term sequences prediction model based on TFA-Seq2Seq virtual sensors is proposed.This model enhances the Seq2Seq virtual sensors using multi-task learning,restructuring them into an"Encoder-3 Decoder"architecture.This structure establishes the map-ping relationship between multiple dam physical sensor signals and long-term seismic time-history in three free-field directions.Additionally,an attention mechanism is integrated to capture temporal dependencies among mul-tiple input signals,resolving synchronous multi-output prediction issues and enhancing prediction accuracy.Fur-thermore,Time-Frequency transform(TF)layers and their inverse transformation layers are introduced to improve the Encoder and Decoder,shortening the temporal length of seismic signals and extracting frequency domain fea-tures.Correspondingly,a model training strategy involving stochastic forced learning is proposed to overcome the limitations of existing virtual sensors in effectively predicting long sequences.Case studies demonstrate that the proposed method achieves a virtual sense of 10 seconds ahead for seismic signals in three directions within dam free field.Compared to models without attention mechanisms and single outputs,the proposed method exhibits an enhanced prediction accuracy of 6.88％and 3.32％,respectively.This research presents novel insights and ap-proaches for advancing the anticipatory sense of seismic information during seismic events.
free field seismicvirtual sensorsmulti-output long-term sequences predictionTFA-Seq2Seqmulti-task learning
国家科技期刊平台
NSTL
万方数据
