Recognizing symbols in architectural CAD drawings is critical for various advanced engineering applications. In this paper, we propose a novel CAD data annotation engine that leverages intrinsic attributes from systematically archived CAD drawings to automatically generate high-quality annotations, thus significantly reducing manual labeling efforts. Utilizing this engine, we construct ArchCAD-400K, a large-scale CAD dataset consisting of 413,062 chunks from 5538 highly standardized drawings, making it over 26 times larger than the largest existing CAD dataset. ArchCAD-400K boasts an extended drawing diversity and broader categories, offering line-grained annotations. Furthermore, we present a new baseline model for panoptic symbol spotting, termed Dual-Pathway Symbol Spotter (DPSS). It incorporates an adaptive fusion module to enhance primitive features with complementary image features, achieving state-of-the-art performance and enhanced robustness. Extensive experiments validate the effectiveness of DPSS, demonstrating the value of ArchCAD-400K and its potential to drive innovation in architectural design and construction.

Engine health management is a critical technology for enhancing the safety, reliability, and economic efficiency of aero-engines. However, the severe scarcity of fault data presents significant challenges for intelligent fault diagnosis of inter-shaft bearings in aero-engines. Existing research on few-shot fault diagnosis of aero-engines have primarily focused on constant-speed conditions, without fully considering the speed fluctuations during aircraft takeoff and landing. To address these challenges, this study proposes a Multi-source Elastic Meta-Transfer Learning (MEMTL) method for few-shot cross-domain fault diagnosis of aero-engine inter-shaft bearings under unstable speed conditions. First, to handle the non-stationary nature of signals under variable-speed conditions, a wavelet attention network is designed for feature extraction under unstable speed conditions. This method employs continuous wavelet transform for time-frequency feature extraction from vibration signals, and integrates hybrid attention modules—channel attention and spatial attention—to adaptively capture multi-scale features across different frequency bands, thereby improving the detection of subtle fault signatures. Second, to address the information deficiency caused by limited fault data, a multi-source elastic relation network is developed. This model incorporates data from multiple source domains into the training process, fully leveraging the diversity of information from different operating conditions. To effectively deal with potential feature distribution discrepancies across source domains, a transfer learning strategy based on inter-domain elastic factors is introduced. This strategy dynamically adjusts the contribution of each source domain during training to reduce negative transfer effects in cross-domain learning. Comparative experiments with state-of-the-art methods demonstrate that the proposed method achieves higher accuracy, more stable performance, and stronger adaptability for few-shot cross-domain fault diagnosis under unstable speed conditions.

In federated learning(FL), clients involved in training initialize their local models with the latest global model at the start of each communication round. This initialization discards previous local knowledge, leading to the loss of personalized insights. To tackle this issue, personalized federated learning (PFL) methods based on self-knowledge distillation aim to preserve effective personalized knowledge while reducing communication overhead and ensuring data privacy. However, current studies on knowledge distillation(KD) for PFL often focus solely on transferring label knowledge, neglecting other latent aspects. Meanwhile, introducing self-knowledge distillation into PFL may lead to unreliable teacher models, thereby misleading local training. To address these concerns, we propose PFedSKD, a novel self-knowledge distillation method for PFL. PFedSKD stratifies neural network intermediate layers based on their structural position and network characteristics. It further introduces three distinct auxiliary classifiers for these layers to effectively transfer their respective features. During the self-knowledge distillation training process, PFedSKD selectively discards negative personalized knowledge to mitigate issues with unreliable teacher models. Extensive experiments across multiple datasets and diverse heterogeneous settings demonstrate that PFedSKD achieves higher accuracy by 0.09%-7.7% and improved personalized learning capabilities compared to existing advanced FL methods.

The paper presents an approach to improve the menu interface of mobile computers by supporting the user with audio information. The experiment that compared the standard menu with the sonically-enhanced menu on mobile computers was conducted. The results confirm the improvement of the sonically-enhanced menu in term of the performance and usability.

This paper presents a user study on interaction with a mobile device. We investigate the use of non-speech sound in mobile interfaces and design a sonically-enhanced interface. The sonically-enhanced interface is compared to a visual interface when users are walking. The results show that the sonically-enhanced interface can improve the mobile interaction as compared to the visual interface.