Generating Thangka images using diffusion models faces challenges such as domain shift and insufficient semantic alignment. This paper proposes a cross-lingual Text-to-Thangka generation method based on PixArt-Σ. First, a high-quality Thangka dataset with precise Chinese and English annotations is constructed. Second, a bilingual text encoding scheme is designed to provide enhanced conditioning inputs to the Diffusion Transformer (DiT). English text is processed by the T5 model, while Chinese text is encoded by mT5 with a deep MLP adapter layer for feature mapping to the DiT input space. For the training strategies, a two-stage approach is employed for Chinese, involving initial fine-tuning of the mT5 encoder via LoRA, followed by joint training of the adapter layer and DiT model with the encoder frozen. Conversely, the English model is trained directly on T5-encoded features. Both languages adopt a Min-SNR weighting strategy to enhance detail generation. Experiments demonstrate that compared to the vanilla PixArt-Σ baseline, our method achieves significant improvements in semantic alignment and image quality, with LPIPS scores decreasing by 6% and FID scores decreasing by 40%. Both Chinese and English models achieve comparable generation quality, with English prompts showing slightly better semantic alignment while Chinese prompts achieve better distributional similarity.

ABSTRACT Increased awareness of Tibetan cultural preservation, along with technological advancements, has led to significant efforts in academic research on Tibetan. However, the structural complexity of the Tibetan language and limited labeled handwriting data impede advancements in Optical Character Recognition (OCR) and other applications. To address these challenges, this paper proposes an innovative Tibetan data augmentation technique, using Generative Adversarial Networks (GANs) to synthesise arbitrary handwriting images in variable calligraphic styles based on inputs. Moreover, our method leverages a Real‐Fake Cross Inputs Strategy during training to enhance generation diversity and improve model generalisability in generating handwritten text beyond the training set and pre‐defined corpus. The model was trained on three Tibetan handwriting datasets, including Umê style numerals, Uchen style consonants, and Khyug‐yig style words. Experimental results demonstrate that the model successfully generates realistic and recognisable Tibetan numeral and consonant handwriting, achieving Fréchet Inception Distance (FID) scores of 14.45 and 27.63, respectively. The proposed method's effectiveness in augmenting OCR models was validated as evidenced by a reduced OCR Word Error Rate (WER) on the augmented datasets.

This article investigates how to effectively defend load-frequency control(LFC) of power systems against Denial-of-Service (DoS) attacks. The core innovation lies in the synergy between T-S fuzzy modeling and an Adaptive Memory Event-Triggered Mechanism (AMETM). At the heart of this mechanism is the dynamic optimization of the triggering threshold, achieved by leveraging both historical and real-time data on attack patterns and system dynamics, which ensures stability is strictly maintained while network bandwidth is significantly conserved. The T-S fuzzy model plays a pivotal role in this framework by transforming the complex nonlinear system into a fuzzy subsystem structure amenable to theoretical analysis, enabling stability proof based on the Lyapunov-Krasovskii functional and Linear Matrix Inequalities (LMIs). Simulation results confirm the outstanding performance of the proposed method in enhancing communication efficiency and system resilience. Note to Practitioners-Load frequency is directly linked to power quality and is a critical indicator of stable grid operation. In practice, load-frequency control (LFC) faces new risks from Denial-of-Service (DoS) attacks, which can disrupt communication between control centers and field devices. To address this issue, this study proposes a unified control framework that integrates T-S fuzzy modeling with an AMETM. By dynamically adjusting communication thresholds using both real-time and historical system data, the method reduces bandwidth usage and enhances resilience under cyberattacks. For practitioners, this provides a practical solution for frequency stabilization in scenarios with limited communication resources, such as dispatch center regulation and cross-area grid coordination. Though validated in simulations, real-world deployment requires integration with supervisory control and data acquisition systems and extensive field testing. Future work will focus on delay compensation, high renewable penetration scenarios, and applications in other automation systems vulnerable to communication attacks.

Visually impaired individuals face significant challenges in environmental perception. Traditional assistive technologies often lack adaptive intelligence, focusing on individual components rather than integrated systems. While Vision-Language Models (VLMs) offer a promising path to richer, integrated understanding, their deployment is severely limited by substantial computational requirements, demanding dozens of gigabytes of memory. To address these gaps in computational efficiency and integrated design, this study proposes a dual technological innovation framework: a cross-modal differentiated quantization framework for VLMs and a scene-aware vectorized memory multi-agent system. The quantization framework implements differentiated strategies, reducing memory from 38GB to 11.3GB. The multi-agent system uses vectorized memory and perception-memory-reasoning workflows to provide environmental information beyond the current view, achieving 2.83-3.52s latency to initial speech output. Experiments show the quantized 19B-parameter model only experiences a 2.05% performance drop on MMBench and maintains 63.7 accuracy on OCR-VQA (original: 64.9), outperforming smaller models with equivalent memory. This research advances computational efficiency and assistive technology, offering comprehensive assistance in scene perception, text recognition, and navigation.

Relational database querying in Tibetan remains a markedly low-resourced endeavor. Existing Text-to-SQL models have only been evaluated on English or Chinese corpora, leaving Tibetan entirely unexplored. To address these issues, TSpider as the first Tibetan Text-to-SQL dataset is constructed by translating the CSpider corpus and comprising 11,840 Tibetan natural-language queries. Furthermore, the Input Feature Enhancement (IFE)-LLM framework is employed: its cross-encoder is adapted to Tibetan by integrating CINO, and the downstream open-source large language model is then prompted to generate executable SQLite SQL statements. Experimental evaluation on TSpider reveals that our framework attains 31.8

Independent indoor mobility remains a critical challenge for individuals with visual impairments, largely due to the limited capability of existing assistive systems in detecting fine-grained hazardous objects such as chairs, tables, and small obstacles. These perceptual blind zones substantially increase the risk of collision in unfamiliar environments. To bridge the gap between monocular 3D vision research and practical assistive deployment, this paper proposes an Adaptive Multi-scale Attention Aggregation (AMAA) framework for monocular 3D semantic scene completion using only a wearable RGB camera. The proposed framework addresses two major limitations in 2D-to-3D feature lifting: noise diffusion during back-projection and structural instability in multi-scale fusion. A parallel channel–spatial attention mechanism is introduced to recalibrate lifted features along semantic and geometric dimensions, while a hierarchical adaptive gating strategy regulates cross-scale information flow to preserve fine-grained structural details. Experiments on the NYUv2 benchmark demonstrate that AMAA achieves an overall mIoU of 27.88

Recently, most Thangkas are preserved in the form of digital images. These Thangkas have lost detail and reduced resolution due to oversights in preservation techniques. This paper proposes to apply super-resolution to low-quality Thangkas to retain more semantic information while maintaining image fidelity. Specifically, a high-resolution Thangka dataset is constructed for model training. Subsequently, the semantic information of images are injected in hidden feature based on SinSR, which extracts semantic information from low-resolution images during the original downsampling process, then integrates the semantic masks into the residual blocks of the Unet framework. At the same time, we conducted an in-depth analysis of multiple fusion locations and methods, selecting the most suitable solution. Experiments on the Thangka dataset illustrate that, compared with other models, our method not only strengthens semantic fusion but also preserves finer local details, yielding an 8

Tibetan is a low-resource language with limited parallel speech corpora spanning its three major dialects (Ü-Tsang, Amdo, and Kham), limiting progress in speech modeling. To address this issue, we propose TMD-TTS, a unified Tibetan multi-dialect text-to-speech (TTS) framework that synthesizes parallel dialectal speech from explicit dialect labels. Our method features a dialect fusion module and a Dialect-Specialized Dynamic Routing Network (DSDR-Net) to capture fine-grained acoustic and linguistic variations across dialects. Extensive objective and subjective evaluations demonstrate that TMD-TTS significantly outperforms baselines in dialectal expressiveness. We further validate the quality and utility of the synthesized speech through a challenging Speech-to-Speech Dialect Conversion (S2SDC) task.

This paper presents an automatic design for a fuzzy adaptive periodic-disturbance observer (APDOB) to suppress multi-changing periodic disturbances in robotic and industrial systems. While the conventional APDOB relies on fixed parameters, its performance is limited under signals with multiple, time-varying fundamental frequencies. Our proposed method introduces a fuzzy logic system to enable the online adjustment of the six key APDOB parameters. To eliminate the need for manual, expertise-dependent tuning, the entire fuzzy controller is automatically designed by a novel differential evolution algorithm named SaMoCpDE, which features a hybrid mutation strategy and self-adaptive control parameters to enhance search robustness. Experimental results validate that the SaMoCpDE designed fuzzy APDOB significantly outperforms the conventional APDOB and demonstrates superior compensation performance. Furthermore, comparative studies with seven evolutionary algorithms confirm the efficacy and stability of the SaMoCpDE in this automatic design framework.

This paper investigates the finite-time synchronization (FTS) of Takagi-Sugeno (T-S) fuzzy multiplicative stochastic coupled memristive neural networks (CMNNs) with probabilistic delayed impulsive effects. Initially, a novel CMNNs model is developed, incorporating mismatched parameters and adjustable coupling strength to enhance the accuracy of the CMNNs. Multiplicative noise is unavoidable in information transmission. Therefore, this type of noise is taken into account in the CMNNs. To enhance the system's robustness, a probabilistic time delayed impulsive controller with fuzzy impulsive strength is introduced. Moreover, finite-time stability theory and impulsive control theory are utilized to establish sufficient conditions for FTS with synchronizing and desynchronizing impulses on CMNNs. Meanwhile, the upper bounds for the settling time of synchronization are effectively estimated. Finally, a numerical example is provided to illustrate the effectiveness of the main results.

This paper proposes Multi-Layer Heterogeneous Memristive Neural Network Genetic Algorithm (MHMNNGA) for intelligent construction site safety monitoring. Physical measurement models are established for 16 sensing channels across five semantic dimensions, and a mutual-information-based heterogeneity index is introduced to justify the heterogeneous network design. Three functionally distinct node types are hierarchically allocated with intra-layer memristive connections, and a four-component joint loss function optimizes classification accuracy, conductance constraints, modal decorrelation, and sparsity simultaneously. A tightly coupled genetic algorithm with block-wise simulated binary crossover and layer-aware adaptive mutation ensures provable almost-sure global convergence. Experimental results demonstrate that MHMNNGA achieves 94.1% accuracy and 93.2% Macro-F1, outperforming GA-MLP by 6.9 percentage points and BP-MLP by 16.8 percentage points, while the 16-channel scheme surpasses the 4-channel variant by 10.5 percentage points.

This paper proposes an intelligent control framework for precision aquaculture, integrating a heterogeneous multi-layer memristive neural network (HMMNN), improved ant colony optimization (IACO), and event-triggered impulsive control (ETIC). A normalized mutual information heterogeneity metric quantitatively justifies the three-layer HMMNN architecture, in which memristors encode long-term inter-parameter coupling dynamics. IACO incorporates adaptive pheromone evaporation and truncated Gaussian sampling to overcome sequentialdependency topology and physical box constraints in training. HMMNN inference dynamically determines the impulsive gain matrix at each triggering instant, with closed-loop exponential stability rigorously proved via the Lyapunov method. Simulation results demonstrate improvements of 65.3% and 66.3% in water temperature and dissolved oxygen control accuracy over conventional PID, a 36.6% reduction in control energy consumption, and an event-triggering rate of only 3.07% relative to periodic sampling, confirming the effectiveness and efficiency of the proposed framework.

Large Language Models (LLMs) have achieved remarkable success in high-resource languages, yet progress in Tibetan remains severely constrained. While recent efforts have begun to address pre-training data scarcity for Tibetan, a more fundamental gap persists: no existing resource supports the complete LLM development pipeline, spanning pre-training, instruction tuning, safety alignment, preference optimization, and reasoning supervision. We introduce the Tibetan Foundation Dataset (TFD), the first structured, large-scale, and expert-curated dataset covering all key stages of Tibetan large language modeling. TFD comprises TIBSTC, a unified corpus of over 11 billion tokens with curated sub-datasets for instruction tuning, safety alignment, and preference optimization, and TIBSTC-CoT, the first large-scale Tibetan chain-of-thought dataset. We demonstrate its utility by training the Sun-Shine family of Tibetan LLMs, achieving substantial improvements over strong baselines on understanding, safety, reasoning, and generation benchmarks. These results underscore that advancing low-resource language modeling requires not only scale, but a structurally complete data ecosystem. We release TFD to facilitate reproducible research and the development of robust, culturally aligned Tibetan LLMs. Code and data are available at https://github.com/Vicentvankor/sun-shine.

This paper focuses on the exponential stability of switched memristive neural networks (SMNNs) with leakage and discrete delays subject to false data injection (FDI) attacks. First, this paper designs the SMNNs model with leakage and discrete delays based on the decay process of neurons and the limitations of signal transmission. Second, an admissible edge-dependent average dwell time (AED-ADT) switching law is introduced, which can utilize the mode information and reflect the switching direction between modes. Then, a mode-dependent impulsive controller is proposed to stabilize the SMNNs subject to FDI attacks. By constructing Lyapunov-Krasovskii functionals and using inequality techniques, the exponential stability criterion is derived. Finally, a numerical example is given to illustrate the validity of the obtained results.

Information encryption based on DNA data archiving, referred to as DNA encryption, has been advocated for decades and has become highly appealing owing to its remarkable advantages, e.g. high storage capacity, complexity and programmability. Early DNA encryption schemes primarily leveraged the natural four-letter genetic alphabet for data storage, with message-storing DNA sequences easily decrypted by routine DNA sequencing, which is consequently vulnerable to attack and faces severe security challenges. Here, an unnatural base pair (UBP), dNaM-dTPT3, was introduced into the message and/or index DNA sequences, which can be stored either in vitro or in vivo; this approach achieved the bioorthogonal encryption of 'secret' messages, where message DNAs could be selectively, faithfully and readily retrieved or read exclusively in the presence of unnatural bases. Furthermore, a separative computational algorithm, named IM-Codec, was developed to encrypt the data into a 'key sequence' and an 'information sequence' through UBP insertion. Finally, a UBP-based multilevel DNA encryption approach was developed and validated for data encryption and decryption. The employment of the UBP expanded genetic system for data encryption should provide valuable solutions for archiving highly confidential data and thus usher in a new era of DNA encryption.

This paper focuses on the issues of fuzzy actor-critic learning architecture, including insufficient interpretability, lack of stability guarantee, and neglect of historical error information. A novel actor-critic learning architecture based on interval type-2 Takagi-Sugeno-Kang fuzzy neural networks (ISAC-IT2-TSK-FNN) is proposed, comprising an interpretable IT2-TSK fuzzy actor (IT2-TSK-FA) and a stability-informed IT2-TSK fuzzy critic (IT2-TSK-FC). In the structure learning of interpretable IT2-TSK-FA, this paper proposes a fuzzy set classification and aggregation method, which reduces the number of fuzzy rules and the complexity of the model. For parameter learning, a value function that concurrently considers control performance and interpretability is designed. To enhance the transparency of fuzzy set partitioning, this paper proposes an iteration-based adaptive learning rate adjustment method. In the parameter learning of stability-informed IT2-TSK-FC, the Lyapunov theorem is introduced. The constraint on the learning rate is derived based on the Lyapunov stability condition to ensure the stability of the control system. Additionally, a weighted historical error mapping method is proposed, which improves the sensitivity of stability-informed IT2-TSK-FC to error changes, enhancing the control strategy evaluation capability. Finally, an algorithm is designed to implement the learning process of the ISAC-IT2-TSK-FNN architecture, with simulation results validating its effectiveness and robustness in various control tasks and under conditions with noise and disturbance.

Memristors enable designing and deploying swarm intelligence algorithms. Here, the firefly algorithm (FA) is binarized and implemented via a memristor crossbar array for feature selection. To address FA’s slow convergence, a novel cosine similarity-based binary firefly algorithm (CS-BFA) is designed, which significantly boosts convergence speed compared to Euclidean distance-based BFA. CS-BFA is implemented using a validated circuit design and control scheme supporting parallel computing. Feature selection for logistic regression (LR) on the Sonar dataset shows that LR with memristive CS-BFA achieves 7.82

Soft network materials (SNMs) incorporating curved microstructures within lattice architectures have emerged as critical components in flexible electronics and tissue scaffolds. These materials demonstrate distinctive nonlinear mechanical behavior under tensile loading, replicating J-shaped stress-strain curves observed in soft tissues such as skin and tendon/ligament. While numerous SNMs featuring diverse microstructural and topological designs have been engineered over the past decade, and each characterized by a distinct design domain, the existing design methodologies present inherent limitations. Although various computational strategies have been proposed, their implementation typically requires multi-stage processes and subjective parameter selection, exhibiting narrow applicability to specific SNM. To address these design constraints, we developed an innovatively competitive coevolution-based differential evolution algorithm with multi-population architecture (CCMPDE). After integrating mathematical modeling of SNMs and finite element analysis (FEA), the CCMPDE-based strategy enabled concurrent optimization of multiple SNMs featuring different curved microstructures. Notably, the proposed strategy demonstrates exceptional compatibility with diverse topologies and mechanical relationships. Through computational and experimental case studies, designed SNMs (fabricated by biopolymers) composed of horseshoe, sinusoidal, and arbitrary curved microstructures successfully replicated the tensile responses of soft tissues. Furthermore, the CCMPDE-based strategy facilitated inverse design of biocomposite SNMs, achieving appropriate replication of target mechanical behaviors.