Unsupervised Domain Adaptation by Statistics Alignment for Deep Sleep Staging Networks

Deep sleep staging networks have reached top performance on large-scale datasets. However, these models perform poorer when training and testing on small sleep cohorts due to data inefficiency. Transferring well-trained models from large-scale datasets (source domain) to small sleep cohorts (target domain) is a promising solution but still remains challenging due to the domain-shift issue. In this work, an unsupervised domain adaptation approach, domain statistics alignment (DSA), is developed to bridge the gap between the data distribution of source and target domains. DSA adapts the source models on the target domain by modulating the domain-specific statistics of deep features stored in the Batch Normalization (BN) layers. Furthermore, we have extended DSA by introducing cross-domain statistics in each BN layer to perform DSA adaptively (AdaDSA). The proposed methods merely need the well-trained source model without access to the source data, which may be proprietary and inaccessible. DSA and AdaDSA are universally applicable to various deep sleep staging networks that have BN layers. We have validated the proposed methods by extensive experiments on two state-of-the-art deep sleep staging networks, DeepSleepNet+ and U-time. The performance was evaluated by conducting various transfer tasks on six sleep databases, including two large-scale databases, MASS and SHHS, as the source domain, four small sleep databases as the target domain. Thereinto, clinical sleep records acquired in Huashan Hospital, Shanghai, were used. The results show that both DSA and AdaDSA could significantly improve the performance of source models on target domains, providing novel insights into the domain generalization problem in sleep staging tasks.<br>

Seismic Hazard Assessment for a Wind Farm Offshore England

Offshore wind has become a major contributor to reducing global carbon emissions. This paper presents a probabilistic seismic hazard analysis for the Sofia Offshore Wind Farm, which is located about 200 km north-east of England in the southern North Sea and will be one of the largest offshore wind farms in the world once completed. The seismic source characterization is composed of two areal seismic source models and four seismic source models derived using smoothed gridded seismicity with earthquake catalogue data processed by different techniques. The ground motion characterization contains eight ground motion models selected based on comparisons with regional data. The main findings are (1) the variation in seismic hazard across the site is negligible; (2) the main source controlling the hazard is the source that includes the 1931 Dogger Bank earthquake; (3) earthquake scenarios controlling the hazard are Mw = 5.0–6.3 and R = 110–210 km; and (4) the peak ground accelerations on rock are lower than for previous regional studies. These results could help guide future seismic hazard assessments in the North Sea.

Ray Optical Scattering from Uniform Reflective Cylindrical Metasurfaces using Surface Susceptibility Tensors

A ray optical methodology based on the uniform theory of diffraction is proposed to model electromagnetic field scattering from curved metasurfaces. The problem addressed is the illumination of a purely reflective uniform cylindrical metasurface by a line source, models the surface with susceptibilities and employs a methodology previously used for cylinders coated in thin dielectric layers [1]. The approach is fundamentally based on a representation of the metasurface using the General Sheet Transition Conditions (GSTCs) which characterizes the surface in terms of susceptibility dyadics. An eigenfunction description of the metasurface problem is derived considering both tangential and normal surface susceptibilities, and used to develop a ray optics (RO) description of the scattered fields; including the specular geometrical optical field, surface diffraction described by creeping waves and a transition region over the shadow boundary. The specification of the fields in the transition region is dependent on the evaluation of the Pekeris caret function integral and the method follows [1]. The proposed RO-GSTC model is then successfully demonstrated for a variety of cases and is independently verified using a rigorous eigenfunction solution (EF-GSTC) and full-wave Integral Equation method (IE-GSTC), over the entire domain from the deep lit to deep shadow.

Ray Optical Scattering from Uniform Reflective Cylindrical Metasurfaces using Surface Susceptibility Tensors

A ray optical methodology based on the uniform theory of diffraction is proposed to model electromagnetic field scattering from curved metasurfaces. The problem addressed is the illumination of a purely reflective uniform cylindrical metasurface by a line source, models the surface with susceptibilities and employs a methodology previously used for cylinders coated in thin dielectric layers [1]. The approach is fundamentally based on a representation of the metasurface using the General Sheet Transition Conditions (GSTCs) which characterizes the surface in terms of susceptibility dyadics. An eigenfunction description of the metasurface problem is derived considering both tangential and normal surface susceptibilities, and used to develop a ray optics (RO) description of the scattered fields; including the specular geometrical optical field, surface diffraction described by creeping waves and a transition region over the shadow boundary. The specification of the fields in the transition region is dependent on the evaluation of the Pekeris caret function integral and the method follows [1]. The proposed RO-GSTC model is then successfully demonstrated for a variety of cases and is independently verified using a rigorous eigenfunction solution (EF-GSTC) and full-wave Integral Equation method (IE-GSTC), over the entire domain from the deep lit to deep shadow.

Exploring the Effects of Emplacement Conditions on Explosion P/S Ratios across Local to Regional Distances

High-frequency (∼&gt; 2 Hz) seismic P/S amplitude ratios are well-established as a discriminant to distinguish between natural earthquakes and underground explosions at regional distances (∼200–1500 km). As research shifts toward identifying lower-yield events, work has begun to investigate the potential of this discriminant for use at local distances (&lt;200 km), in which initial results raise questions about its effectiveness. Here, we utilize data from several chemical explosion experiment series at the Nevada National Security Site in southern Nevada in the United States to study explosion Pg/Lg ratios across the range of local to regional distances. The experiments are conducted over differing emplacement conditions, with contrasting geologies and a variety of yields and depths of burial, including surface explosions. We first establish the similarities of Pg/Lg ratios from chemical explosions to those from historic nuclear tests and conclude that, as previous data have suggested, chemical explosion ratios are good proxies for nuclear tests. We then examine Pg/Lg ratios from the new experiment series as functions of distance, yield, depth of burial, and scaled depth of burial (SDOB). At far-local and regional distances, we observe consistently higher ratios from hard-rock explosions compared to ones in a weaker dry alluvium medium, consistent with prior regional distance results. No other trends with yield, depth of burial, or SDOB are strongly evident. Scatter in the observed ratios is very high, particularly at the shortest event-to-station distances, suggesting that small-scale path effects play a significant role. On average, the local distance explosion Pg/Lg ratios show remarkable consistency across all the variations in emplacement. Explosion source models will need to reproduce these results.

MODELING THE ANTENNA INPUT IMPEDANCE BY THE METHOD OF MOMENTS

The paper presents a comparison of excitation source models when modeling antennas by the method of moments. By using a set of adjacent edges when specifying the impact, it is possible to obtain correct results when the computational grid of the antenna model is more frequent. This is shown on the example of a symmetric electric vibrator.

A Unification of Heterogeneous Data Sources into a Graph Model in E-commerce

The incorporation of heterogeneous data models into large-scale e-commerce applications incurs various complexities and overheads, such as redundancy of data, maintenance of different data models, and communication among different models for query processing. Graphs have emerged as data modelling techniques for large-scale applications with heterogeneous, schemaless, and relationship-centric data. Models exist for mapping different types of data to a graph; however, the unification of data from heterogeneous source models into a graph model has not received much attention. To address this, we propose a new framework in this study. The proposed framework first transforms data from various source models into graph models individually and then unifies them into a single graph. To justify the applicability of the proposed framework in e-commerce applications, we analyse and compare query performance, scalability, and database size of the unified graph with heterogeneous source data models for a predefined set of queries. We also access some qualitative measures, such as flexibility, completeness, consistency, and maturity for the proposed unified graph. Based on the experimental results, the unified graph outperforms heterogeneous source models for query performance and scalability; however, it falls behind for database size.

Unsupervised Domain Adaptation by Statistics Alignment for Deep Sleep Staging Networks

Deep sleep staging networks have reached top performance on large-scale datasets. However, these models perform poorer when training and testing on small sleep cohorts due to data inefficiency. Transferring well-trained models from large-scale datasets (source domain) to small sleep cohorts (target domain) is a promising solution but still remains challenging due to the domain-shift issue. In this work, an unsupervised domain adaptation approach, domain statistics alignment (DSA), is developed to bridge the gap between the data distribution of source and target domains. DSA adapts the source models on the target domain by modulating the domain-specific statistics of deep features stored in the Batch Normalization (BN) layers. Furthermore, we have extended DSA by introducing cross-domain statistics in each BN layer to perform DSA adaptively (AdaDSA). The proposed methods merely need the well-trained source model without access to the source data, which may be proprietary and inaccessible. DSA and AdaDSA are universally applicable to various deep sleep staging networks that have BN layers. We have validated the proposed methods by extensive experiments on two state-of-the-art deep sleep staging networks, DeepSleepNet+ and U-time. The performance was evaluated by conducting various transfer tasks on six sleep databases, including two large-scale databases, MASS and SHHS, as the source domain, four small sleep databases as the target domain. Thereinto, clinical sleep records acquired in Huashan Hospital, Shanghai, were used. The results show that both DSA and AdaDSA could significantly improve the performance of source models on target domains, providing novel insights into the domain generalization problem in sleep staging tasks.<br>

Unsupervised Domain Adaptation by Statistics Alignment for Deep Sleep Staging Networks

Deep sleep staging networks have reached top performance on large-scale datasets. However, these models perform poorer when training and testing on small sleep cohorts due to data inefficiency. Transferring well-trained models from large-scale datasets (source domain) to small sleep cohorts (target domain) is a promising solution but still remains challenging due to the domain-shift issue. In this work, an unsupervised domain adaptation approach, domain statistics alignment (DSA), is developed to bridge the gap between the data distribution of source and target domains. DSA adapts the source models on the target domain by modulating the domain-specific statistics of deep features stored in the Batch Normalization (BN) layers. Furthermore, we have extended DSA by introducing cross-domain statistics in each BN layer to perform DSA adaptively (AdaDSA). The proposed methods merely need the well-trained source model without access to the source data, which may be proprietary and inaccessible. DSA and AdaDSA are universally applicable to various deep sleep staging networks that have BN layers. We have validated the proposed methods by extensive experiments on two state-of-the-art deep sleep staging networks, DeepSleepNet+ and U-time. The performance was evaluated by conducting various transfer tasks on six sleep databases, including two large-scale databases, MASS and SHHS, as the source domain, four small sleep databases as the target domain. Thereinto, clinical sleep records acquired in Huashan Hospital, Shanghai, were used. The results show that both DSA and AdaDSA could significantly improve the performance of source models on target domains, providing novel insights into the domain generalization problem in sleep staging tasks.<br>