Applied Intelligence最新文献

In multi-domain spoken language understanding (MSLU), intent detection and slot filling are crucial components. While prior research has shown improvement in MSLU model performance through the integration of intent and slot features, such works typically treat multi-domain tasks as a collection of independent single-domain tasks, neglecting both intra-domain and inter-domain correlations. In this paper, we propose Piece Attention and Position-aware Embedding with the Top-k Network (PAPET), which leverages fine-grained features to capture multi-domain correlations. Specifically, we segment intents and slots into fine-grained action, domain, and attribute pieces to capture the attention with utterances. In multi-domain tasks, piece attention can effectively model both intra-domain correlations through the utilization of domain pieces, as well as inter-domain correlations by leveraging action and attribute pieces. Moreover, we introduce the top-k network and relative position-aware embedding to effectively handle multi-intent and word-to-word correlations, respectively. We perform experiments on two publicly available MSLU datasets, CrossWOZ and RiSAWOZ. The main results indicate that PAPET enhances the performances of previous SLU models, achieving improvements in joint accuracy of 2.43% and 2.19% on the respective datasets. Ablation and multi-domain experiments validate the effectiveness of PAPET in tackling the challenges of MSLU. Additional experiments further validate the effectiveness of PAPET on the MSLU task from four key perspectives: compatibility with BERT, comparative performance with large language models, computational efficiency, and error analysis.

This paper addresses the issue of fixed-time neural adaptive event-triggered control for nonstrict-feedback nonlinear systems with full-state constraints, input dead-zone, and saturation. Radial basis function neural networks (RBFNNs) are used to identify the unknown nonlinearities. The paper considers both input saturation and dead-zone effects, approximating these non-smooth nonlinearities with a non-affine smooth function and then transforming them into an affine form using the mean value theorem. The approach integrates backstepping recursive design with a varying threshold event-triggered condition to create an event-triggered neural adaptive fixed-time control algorithm that employs barrier Lyapunov functions (BLFs) and RBFNNs. By applying the fixed-time stability criterion, the proposed controller ensures that the tracking error converges to a smaller region within a fixed time and that all variables in the closed-loop system remain bounded. Finally, two simulation examples are provided to demonstrate the effectiveness of the proposed method.

The scarcity of labeled data in real-world applications has sparked interest in semi-supervised learning (SSL) methods. However, traditional SSL models often rely on assumptions like the manifold or low-density separation, which may not hold in dynamic streaming environments. This challenge is further exacerbated by concept drift and high dimensionality, which impairs the effectiveness of SSL models. In this paper, we introduce DMReSSL, a pioneering deep metric framework for reliable semi-supervised learning under concept drift. Unlike conventional approaches, DMReSSL leverages a metric learning module to transform the original data features into a metric latent space, where it dynamically maintains a set of metric-embedded micro-clusters with evolving reliability attributes. The reliability of these micro-clusters is updated in an online fashion, based on prediction performance, time effects, and local label distributions, ensuring that the model adapts to changing data distributions. Extensive experiments conducted on eight real-world and six synthetic datasets demonstrate that DMReSSL outperforms the second-best algorithm by 1.32% in accuracy using only 5% of labeled data, significantly enhancing model robustness and efficiency in semi-supervised data streams learning.

Aspect-Based Sentiment Analysis (ABSA) focuses on understanding fine-grained sentiment information by analyzing the sentiment polarity corresponding to particular aspects in sentences. At present, graph neural networks are widely utilized to model the explicit relationships between aspects and opinions derived from the syntactic structures of dependency trees. However, these methods struggle to handle sentences with complex structures and multiple aspect–sentiment pairs. To solve this problem, we propose a Bilateral Synergistic Aggregation Network (BSAN) that integrates semantic and syntactic information to capture sentiment features that are specific to particular aspects. Specifically, within the Syntactic Distillation Module (SDM), we employ a Syntax View Graph Convolution (SynVGC) layer to encode the dependency-tree graph and extract syntactic features, while a Transformer layer is incorporated to capture sequential dependencies and refine the representations of aspect terms. Furthermore, the Semantic Optimization Module (SOM) utilizes Abstract Meaning Representation (AMR) as structured input and integrates attention mechanisms with graph convolutional networks to effectively model the semantic relations represented in the AMR. In addition, the Graph Cognitive Fusion Module (GCFM) is designed to facilitate the integration and interaction of syntactic and semantic representations. Finally, extensive experiments on four publicly available benchmark datasets demonstrate that our proposed BSAN model achieves competitive performance.

The National Equities Exchange and Quotations (NEEQ) in China is a key platform for small and medium-sized enterprises (SMEs) to access public capital markets. However, their credit evaluation is challenging due to financial opacity and information asymmetry. Given that conventional credit evaluation methods mainly rely on financial statements, news reports, and transaction history, often overlooking the complex relationships that affect SMEs’ performance, we propose a KG-AttRGCN-XGBoost model to evaluate enterprise credit effectively. This model uses a knowledge graph (KG) to construct enterprise relationship networks, utilizes contextual embeddings of a relational graph convolutional network (RGCN) with a constrained attention mechanism to model complex inter-enterprise connections, and transmits the extracted features to XGBoost for credit evaluation. Experimental results show that our model significantly outperforms several popular graph neural network-based credit evaluation methods, better handling multi-relational data and achieving higher precision.

The formation quality of welded joints is influenced by the dynamic behavior of the molten pool, but existing visual sensing technologies are unable to directly measure its internal physical properties. Although numerical simulations play an important role in characterizing the welding formation process, their low computational efficiency and long processing time create a bottleneck in application. To address this, the study proposes an approach that integrates digital twins with numerical simulations to achieve precise and efficient prediction of this process. Firstly, a high-fidelity cold metal transfer welding digital twin model was developed on a finite element simulation platform, encompassing the geometry, physics, behavior, and rules of the welding formation process, and its reliability was verified through simulations of stainless steel welding joints. Subsequently, an improved conditional deep convolutional generative adversarial network (C-DCGAN) was designed as a lightweight surrogate model, incorporating affine mapping and residual structures, which efficiently replaced traditional numerical simulations while ensuring high accuracy and stability. The results show that the improved C-DCGAN model achieves higher predictive accuracy and efficiency than the comparative models at all three welding speeds. At a welding speed of 10 mm/s, the mean value of the MSE for molten pool image generation is as low as 0.0059, with good robustness. Meanwhile, the prediction time of the model requires only 5.44 s, significantly reducing the computation time compared to traditional numerical simulations and the comparative models. This study provides a foundation for the implementation of digital twin systems and advanced manufacturing quality control in industrial settings.

Multivariate multi-step time series prediction (MTSP) aims to predict the future time steps based on the historical data from multiple variables, but faces challenges such as complex dependencies between variables and error accumulation, leading to poor prediction accuracy. In this study, an attention-based multi-column neural network (AMCNN) for multivariate MTSP is proposed to improve the prediction accuracy. First, an attention mechanism is developed to reconstruct the inputs, highlighting the contribution of important variables to the prediction. Second, to obtain the optimal task allocation for MTSP, an improved binary particle swarm optimization (BPSO) algorithm is designed with the fitness function comprehensively evaluating the MTSP results in terms of both prediction error and shape difference. Finally, a direct-connected adaptive radial basis function neural network (DCA-RBFNN) is constructed as the subnetwork of AMCNN to improve the accuracy with fewer training iterations. Several experiments are performed on two benchmark datasets and two real-world datasets to verify the effectiveness of AMCNN in MTSP. The results demonstrate that AMCNN achieves the best modeling accuracy compared to other models, especially when the prediction horizon is larger, highlighting its superiority for long-term prediction. Positive effects on prediction accuracy have been further demonstrated from the perspectives of the attention-based reconstruction of inputs, the improved BPSO algorithm, and the design of the subnetwork, respectively. Besides, the subnetwork in ACMNN, that is, DCA-RBFNN, contributes to the accurate prediction with fewer training iterations.

Interval sets, proposed by Yao, consist of the upper and lower-bound sets which are finite sets. However, in practical situations, the upper and lower-bound sets may be infinite sets. In this paper, we first propose the concept of generalized interval sets and define the distance between generalized interval sets. Then, we introduce a generalized interval set information system (GIS), in which the conditional attribute values are generalized interval sets and the decision attribute values are set-valued or single-valued. We mine the effect of different conditional attributes on decision making in the GIS in order to assign weight to each conditional attribute. Subsequently, we give the definition of generalized weighted neighborhood rough sets (GWNRS) in the GIS. Dependency degree is proposed based on GWNRS to evaluate the significance of attribute subsets. Furthermore, we use greedy search algorithm to perform attribute reduction in the GIS and evaluate the classification performance with SVM and KNN classifiers. In this process, we find the optimal neighborhood threshold by isometric search. Finally, we conduct experiments on six UCI datasets to validate the performance of the attribute reduction algorithm proposed in this paper.

Sleep staging is used to assist diagnosis of sleep disorders. Polysomnography (PSG) provides multiple channels of signal recording that reflects different working patterns of each sleep stage and can be utilized for automatic sleep staging. However, previous deep learning based sleep staging researches ignored the dependencies between local and global temporal sequences of multi-modal PSG signals. To solve this problem, this work proposed a multi-modal fusion network via temporal sequence for sleep staging (MMTSleepNet), consisting of local and global feature extraction (LGFE) block, adaptive modal feature recalibration (AMFR) block, multi-modal feature fusion (MFF) block and sleep staging block. LGFE block extracts both local and global contextual features to learn differentiated characteristics across sleep stages. AMFR block adaptively adjusts channel-wise weights on the concatenated feature maps to highlight more contributing PSG modalities, where each channel corresponds to a specific PSG signal or sub-signal. MFF block models feature dependencies to quantify the correlations between multi-modal temporal sequences. The proposed MMTSleepNet outperforms the state-of-the-art methods on Sleep-EDF-20, Sleep-EDF-78 and Sleep Heart Health Study (SHHS) datasets, with the accuracy rates of 87.5%, 82.9% and 87.8%, respectively.

Socially aware navigation is a fast-evolving research area in robotics that enables robots to move within human environments while adhering to the implicit human social norms. The advent of Deep Reinforcement Learning (DRL) has accelerated the development of navigation policies that enable robots to incorporate these social conventions while effectively reaching their objectives. This survey offers a comprehensive overview of DRL-based approaches to socially aware navigation, highlighting key aspects such as proxemics, human comfort, naturalness, trajectory and intention prediction, which enhance robot interaction in human environments. This work critically analyzes the integration of value-based, policy-based, and actor-critic reinforcement learning algorithms alongside neural network architectures, such as feedforward, recurrent, convolutional, graph, and transformer networks, for enhancing agent learning and representation in socially aware navigation. Furthermore, we examine crucial evaluation mechanisms, including metrics, benchmark datasets, simulation environments, and the persistent challenges of sim-to-real transfer. Our comparative analysis of the literature reveals that while DRL significantly improves safety, and human acceptance over traditional approaches, the field still faces setback due to non-uniform evaluation mechanisms, absence of standardized social metrics, computational burdens that limit scalability, and difficulty in transferring simulation to real robotic hardware applications. We assert that future progress will depend on hybrid approaches that leverage the strengths of multiple approaches and producing benchmarks that balance technical efficiency with human-centered evaluation. By reviewing the state of the art and highlighting these challenges, this work provides researchers and practitioners with a roadmap for advancing DRL-based socially aware navigation toward robust, real-world deployment.