雷神技术公司

Though type-Ia supernovae (SNe Ia) are found in all types of galaxies, recent local Hubble constant measurements have disfavored using SNe Ia in early-type or quiescent galaxies, aiming instead for better consistency with SNe Ia in star-forming, late-type host galaxies calibrated by Cepheid distances. Here we investigate the feasibility of a parallel distance ladder using SNe Ia exclusively in quiescent, massive (log M_*/M_⊙≥ 10) host galaxies, calibrated by tip of the red giant branch (TRGB) distances. We present TRGB measurements to four galaxies: three measured from the Hubble Space Telescope with the ACS F814W filter, and one measured from the JWST NIRCam F090W filter. Combined with literature measurements, we define a TRGB calibrator sample of five high-mass, early-type galaxies that hosted well-measured SNe Ia: NGC 1316 (SN 2006dd), NGC 1380 (SN 1992A), NGC 1404 (SN 2007on, SN 2011iv), NGC 4457 (SN 2020nvb), and NGC 4636 (SN 2020ue). We jointly standardize these calibrators with a fiducial sample of 124 Hubble-flow SNe Ia from the Zwicky Transient Facility that are matched in host-galaxy and light-curve properties. Our results with this homogenized subsample show a Hubble residual scatter of under 0.11 mag, lower than usually observed in cosmological samples of the full SN Ia distribution. We obtain a measurement of the Hubble constant, H_0 = 75.3 ± 2.9 km s^-1 Mpc^-1, including statistical and estimated systematic uncertainties, and discuss the potential to further improve the precision of this approach. As calibrator and supernova samples grow, we advocate that future cosmological applications of SNe Ia use subsamples matched in host-galaxy and supernova properties across redshift.

A novel method for receiving coherently encoded optical signals, without adaptive optics or other wavefront correction approaches, is demonstrated in atmosphere along with a fully-fade tolerant 10 Gbps modem. The receiver sensitivity using a DPSK signal is compared with OOK, and power penalties are evaluated. The experiment confirms the ability of the receiver to function without wavefront correction at high data rates with the ability to scale to higher-order modulation.

SUMMARY & CONCLUSIONSThis paper presents a comprehensive technical framework for the development and deployment of an advanced predictive maintenance capability that leverages large language models (LLMs) to transform traditional industrial asset management practices. The proliferation of sensor technologies and digital maintenance records has created vast repositories of both structured and unstructured data that remain largely underutilized by conventional predictive maintenance systems. Our approach addresses this challenge by extracting actionable insights from heterogeneous historical datasets—including structured sensor outputs, unstructured maintenance logs, technician reports, and operational documentation—to enable timely, context-aware predictions for maintenance interventions.We outline a novel system architecture that seamlessly integrates data engineering pipelines, domain-specific LLM fine-tuning methodologies, probabilistic inference engines, and human-in-the-loop feedback mechanisms. The proposed framework employs a GPT-style decoder-based architecture with 6.7 billion parameters, fine-tuned using a hybrid approach combining supervised learning with contrastive learning objectives to optimize performance on maintenance-specific tasks. The system processes multimodal inputs through sophisticated data fusion techniques that temporally align textual observations with corresponding sensor measurements, creating comprehensive equipment health assessments that incorporate both quantitative metrics and qualitative expert knowledge.Our methodology addresses critical limitations of existing predictive maintenance approaches, particularly their inability to effectively leverage the wealth of institutional knowledge captured in maintenance narratives and their reliance on handcrafted features that may not generalize across different equipment types or operational contexts. The LLM-based system demonstrates superior pattern recognition capabilities, identifying subtle semantic indicators in maintenance logs that correlate with specific failure modes while maintaining interpretability through attention mechanisms and confidence scoring.Empirical evaluations conducted on a comprehensive real-world industrial dataset comprising over 10,000 maintenance events from critical rotating and thermal equipment demonstrate substantial improvements in maintenance precision and system uptime. Compared to baseline approaches using GRU-based sequence models and random forest classifiers, our LLM-integrated system achieves a 23% increase in Mean Time Between Failures (MTBF), 15% reduction in unplanned downtime, and significant improvements in prediction precision (Precision@5 increasing from 0.41 to 0.76). Additionally, technician satisfaction scores improved from 3.2 to 4.4 on a 5-point Likert scale, indicating enhanced usability and trust in AI-assisted maintenance decision-making. The system's ability to process multi-modal inputs and generate interpretable outputs with associated confidence scores makes it particularly well-suited for deployment in safety-critical industrial environments where human oversight remains essential.

Provides society information that may include news, reviews or technical notes that should be of interest to practitioners and researchers.

Provides society information that may include news, reviews or technical notes that should be of interest to practitioners and researchers.