Applied Intelligence最新文献

Travel time prediction has important influence on the overall control of urban Intelligent Transportation Systems (ITS). Urban arterial networks are typically composed of links and intersections, where each link or intersection can be regarded as a spatial node within the network. However, existing researches predominantly focus on modeling spatial nodes in the link modality to predict travel times in urban arterial networks, neglecting the potential correlations among heterogeneous modal nodes. To overcome these limitations, we propose a Heterogeneous Multi-Modal Graph Neural Network (HMGNN) specifically tailored for travel time prediction in arterial networks. Specifically, we innovatively construct spatial correlation graphs that capture the unique traffic characteristics of intersection modal nodes. Furthermore, we design a cross-modal graph generator that captures the latent spatiotemporal features between spatial nodes of distinct modalities, resulting in the generation of heterogeneous modal graphs. Finally, our proposed HMGNN model incorporates tailored network structures for graphs of varying complexities, enabling targeted mining of their inherent information to derive the final prediction results. Extensive experiments conducted using real-world traffic data from Zhangzhou, China, demonstrate that our HMGNN model achieves significant improvements in prediction accuracy.

Unmanned boats will encounter many static and dynamic obstacles during navigation, and only real-time obstacle sensing can ensure safe navigation and long endurance of unmanned boats. In this paper, LGCGNet is proposed to perform real-time water surface semantic segmentation on the images captured by the on-board camera. In order to ensure that the model adapted to obstacles with extremely variable scales, a local-global module is proposed in this paper. The local-global module consisted of residual dense dilated module and context-enhanced separable self-attention. Residual dense dilated module enabled the enhancement of local detail information and context-enhanced separable self-attention enabled model receptive field expansion. In addition, the sub-pixel downsampling module is used to avoid the loss of feature information to improve segmentation accuracy. Experiments on the MaSTr1325 dataset showed that LGCGNet apprpached the segmentation accuracy of state-of-the-art semantic segmentation models with only 689,000 parameters and 9.068G floating point operations per second, with an mIoU of 84.14%. In addition, the processing speed of LGCGNet is 34.86FPS, which meets the frame rate conditions of commercially available photovoltaic equipment. The experiments demonstrated that the LGCGNet proposed in this paper strike a good balance between achieving high accuracy, reducing model size and improving real-time performance.

Accurate metro passenger flow prediction can provide data support for vehicle scheduling and personnel allocation by metro operation departments, ensuring the efficient utilization of related resources. In recent years, Graph Convolutional Networks (GCNs) have demonstrated excellent performance in spatial processing, making them an effective method for extracting spatiotemporal dependencies in metro passenger flow prediction. However, traditional GCN models focus solely on static relationships between stations, overlooking the dynamic changes in station relationships and typically concentrating on short-term temporal dependencies while neglecting longer-term temporal features. To fully consider the spatiotemporal relationships within the metro network, a Multi-Dimensional Temporal Dependency Graph Neural Network (MTDGNN) is proposed for metro passenger flow prediction. Specifically, 1D dilated convolutions are employed to initially extract multi-dimensional temporal dependencies, generating multiple spatiotemporal dependency extraction channels. Two correlation matrices combined with GCN are then proposed to extract spatial relationships between stations within the metro network. The extracted spatiotemporal features are further captured by a Gated Recurrent Unit (GRU) to enhance temporal feature extraction. Subsequently, a multi-head attention mechanism is utilized to integrate the extraction results from multiple channels to obtain the final prediction. Finally, the model is evaluated using metro ridership data from two cities in southwestern and central China. The results indicate that the proposed model exhibits superior predictive performance compared to other methods. The MAE values on the two datasets are 1.5% to 59.3% lower than those of other methods, and the RMSE values are 3.4% to 60.0% lower than those of other methods.

In the domain of semiconductor integrated circuit manufacturing, accurately identifying the root causes of defects is critical for enhancing yield rates. Traditionally, this analytical process has been both time-intensive and challenged by inaccuracies, primarily due to the intricate and varied morphology of wafer defects. While convolutional neural networks (CNNs) with encoder-decoder architectures have made significant strides in the segmentation of defects, they inherently struggle to capture distant interactions and achieve high performance in classification tasks. Conversely, recent advancements in transformers have showcased their proficiency in learning global image dependencies. However, transformers often lack the specific graphical priors and the adaptability typically associated with CNNs. Addressing these limitations, we introduce SCSNet, an innovative architecture that merges the strengths of transformers and CNNs. This fusion network is designed to enhance both segmentation and classification of scanning electron microscopy (SEM) images of wafer defects. SCSNet incorporates a conventional encoder-decoder framework, supplemented by shape flow branches and multi-cross-attention (MCF) modules within a skip connection architecture. Rigorous experimentation on a dataset of 4425 high-resolution wafer defects, sourced from our operational wafer fabrication facility, demonstrates SCSNet’s superior performance. Notably, SCSNet surpasses existing advanced CNNs, transformers, and their hybrid counterparts, achieving a classification accuracy of 97.62% and a segmentation Intersection over Union (IoU) of 84.09%. Currently implemented on our local server for engineering use, SCSNet represents a major advancement in semiconductor manufacturing, offering a more precise and efficient tool for wafer defect analysis.

Low-rank representation (LRR) has been proved to be effective in exploring low-dimensional subspace structure embedded in the observations. However, existing LRR algorithms often pay no attention to data redundancy, easily leading to performance decay. In addition, the LRR characterizes data global inter-connections, from which some latent similarity features should be further learned and exploited to improve the performance of clustering. Therefore, a novel method termed Graph Regularized Independent Latent Low-Rank Representation (GRI-LLRR) is presented to address the above issues. As we know, Hilbert–Schmidt Independence Criterion (HSIC) measures the independence between two distributions. In the proposed method, it is introduced and developed to another novel graph regularization independent term to remove the uncorrelation between vectors and to preserve the data local geometry. With other constraints, including the sparse, nonnegative and symmetric, the LRR is obtained from the observations. Then, the proposed method further learns the cosine features as latent representation of the LRR for final clustering. Massive experiments have been conducted on eight benchmark data sets. Experimental results show that the proposed GRI-LLRR outperforms some state-of-the-art (SOTA) approaches with improvements of 2.24%, 2.73%, and 2.65% on average for CCA, NMI, and Purity, respectively.

Neighbourhood rough set (NRS)-based feature selection has been extensively applied in data mining. However, the effectiveness of the NRS model is limited by its reliance on the grid search method to determine the optimal neighbourhood parameter, insensitivity to data distribution under different features, and consideration of uncertainty measures from only one single perspective. To address the aforementioned issues, this study first defines a spatial function that can obtain information about the distribution of samples in space according to the change in the feature subset. On this basis, three perspectives of dynamic neighbourhoods are proposed: pessimistic, neutral, and optimistic. Next, the concept of the dynamic neighbourhood rough set (DNRS) model is developed. The most significant feature of this model is its adaptive ability to dynamically update the neighbourhood radius of samples on the basis of the information of their distribution in space, without the necessity of setting neighbourhood parameters artificially. Then, algebraic and information-theoretic views are introduced to propose multi-perspective dynamic neighbourhood entropy measures, which effectively measure the uncertainty of the data. In addition, a nonmonotonic feature selection algorithm based on mutual information is designed to overcome the limitations of feature selection algorithms that rely on monotonic evaluation functions. This algorithm utilizes multi-perspective dynamic neighbourhood entropy measures from a neutral perspective. Finally, to mitigate the high time complexity in feature selection for high-dimensional datasets, the Fisher score is introduced in an initial dimensionality reduction method. The results of the experiment show that the algorithm effectively eliminates redundant features and improves accuracy.

As an effective alternative model to recurrent neural network (RNN), echo state network (ESN) has garnered more attention due to its efficiency in handling time series data. Despite the simple training process and rapid convergence speed of ESN, appropriate parameter settings and a concise network structure are crucial for optimal model performance. Therefore, many optimization algorithms have been proposed to obtain the optimal parameters of ESN. Among these methods, the Pigeon-Inspired Optimization (PIO) has gained attention due to its fast search speed, strong evolution capability, and excellent optimization ability. However, the main drawbacks of PIO are that it may easily get trapped in local optima and achieve lower precision results. To address these issues, this paper proposes a hybrid algorithm combining adaptive improved pigeon-inspired optimization with tabu search (TS-APIO) algorithm. By combining the improved PIO and the tabu search (TS), it not only enhances the global search capability but also strengthens its robustness. Additionally, the adaptive adjustment mechanism can improve the generalization ability. Through theoretical analysis and simulation examples, the TS-APIO algorithm can adaptively select the optimal ESN parameters and structure based on different scenarios. It can effectively enhance the ability to capture the dynamic features and reduce the prediction error.

Forecasting hog prices is an important and challenging task for pig producers and managers as it plays a crucial role in decision-making processes. Given the significant impact of raw pork supply, public concern, animal diseases, and international markets on hog prices, this study proposes a comprehensive and explainable hybrid model for hog price forecasting by combining principal component analysis (PCA), variational mode decomposition (VMD), weighted average algorithm (WAA) algorithm, and temporal fusion transformers (TFT). To improve the quality of input variables, search engine data reflecting public concern about live pig prices are dimensionally reduced using PCA. This reduction process helps in eliminating unnecessary information and enhancing the input’s relevance. Additionally, VMD is applied to decompose raw pig futures prices, enabling the capture of their underlying trends over time. Subsequently, all the input variables, including the processed search engine data and the decomposed pig futures prices, are fed into the WAA-TFT model. WAA algorithm optimizes the parameters of the TFT model, resulting in accurate predicted values. The interpretable nature of the TFT model provides valuable decision-making insights for practitioners in the agricultural products market. The experimental results show that the proposed model achieves a mean absolute percentage error (MAPE) of only 1.76% on the Chinese hog price prediction dataset, demonstrating the excellent predictive performance of the proposed model.

In Large Scale Group Decision Making (LSGDM), the differences in decision-makers’ professional backgrounds and attitudes often lead to high-quality decisions being overshadowed by numerous low-quality decisions, thus affecting the accuracy of the final decision. This study proposes a new decision-making method to address this challenge. First, a few experts are invited to make decisions as cluster centers, followed by obtaining decisions from a large number of ordinary decision-makers. The ordinary decisions are then generated and modified using a Discrete Conditional Variational Autoencoder (DCVAE) to enhance decision quality while maintaining consistency with expert decisions. Finally, the normalized prediction selection rate (NPSR) and the Borda Count consensus method are integrated to obtain the final result. Experimental results demonstrate the effectiveness of this method in improving the quality of LSGDM, providing a new solution to the coexistence of high- and low-quality decisions.