Fault diagnosis in modern industrial and information systems is critical for ensuring equipment reliability and operational safety, but traditional methods have difficulty in effectively capturing spatiotemporal dependencies and fault-sensitive features in multi-sensor data, especially rarely considering dynamic features between multi-sensor data. To address these challenges, this study proposes DyGAT-FTNet, a novel graph neural network model tailored to multi-sensor fault detection. The model dynamically constructs association graphs through a learnable dynamic graph construction mechanism, enabling automatic adjacency matrix generation based on time–frequency features derived from the short-time Fourier transform (STFT). Additionally, the dynamic graph attention network (DyGAT) enhances the extraction of spatiotemporal dependencies by dynamically assigning node weights. The time–frequency graph pooling layer further aggregates time–frequency information and optimizes feature representation.Experimental evaluations on two benchmark multi-sensor fault detection datasets, the XJTUSuprgear dataset and SEU dataset, show that DyGAT-FTNet significantly outperformed existing methods in classification accuracy, with accuracies of 1.0000 and 0.9995, respectively, highlighting its potential for practical applications.

Environmental sound classification (ESC) is crucial for applications such as intelligent surveillance, urban acoustic monitoring, and human-computer interaction. Although deep neural networks (DNNs) have significantly improved ESC performance, these methods often rely on large models and extensive pretraining, making them difficult to deploy in resource-constrained environments. Some existing lightweight models, while having fewer parameters, still suffer from limited representational capacity, leading to suboptimal generalization, especially in low-data scenarios. To address these challenges, we propose SpectroMaskNet, a compact dual-branch architecture. This design integrates global-local attention mechanisms with block-masked spectrogram augmentation, allowing the model to capture both long-term temporal dependencies and fine-grained spectral features. This enhances robustness and generalization, particularly in data-scarce situations. Experimental results on four benchmark datasets–ESC-10, ESC-50, UrbanSound8K, and SpeechCommandV2–demonstrate that SpectroMaskNet achieves accuracies of 97.50%, 95.50%, 96.32%, and 96.52%, respectively, outperforming existing lightweight baselines without requiring large-scale pretraining. Furthermore, the model maintains low computational complexity, making it well-suited for real-world ESC applications that demand efficiency and scalability.

Generative models based on Variational Autoencoders (VAEs) represent an important area of research in Controllable Text Generation (CTG). However, existing approaches often do not fully exploit the potential of latent variables, leading to limitations in both the diversity and thematic consistency of the generated text. To overcome these challenges, this paper introduces a new framework based on Hierarchical Latent Modulation (HLM). The framework incorporates a hierarchical latent space modulation module for the generation and embedding of conditional modulation parameters. By using low-rank tensor factorization (LMF), the approach combines multi-layer latent variables and generates modulation parameters based on conditional labels, enabling precise control over the features during text generation. Additionally, layer-by-layer normalization and random dropout mechanisms are employed to address issues such as the under-utilization of conditional information and the collapse of generative patterns. We performed experiments on five baseline models based on VAEs for conditional generation, and the results demonstrate the effectiveness of the proposed framework.