Applied Intelligence最新文献

Brain stroke is a major cause of disability and death, and accurate lesion segmentation is essential for early diagnosis and treatment planning. Although CT and MRI provide critical diagnostic information, the large variations in lesion appearance and the noise introduced by manual annotations make precise segmentation challenging. To address these issues, we propose SFJD-Net, a novel Stroke lesion segmentation network that leverages joint spatial-frequency domain feature enhancement and differential learning. SFJD-Net consists of three core modules: Multi-Branch Convolution Attention Encoder (MBCAE), Spatial-Frequency domain Joint feature Enhancement (SFJE), and Differential Learning Decoder (DLD). Compared with the traditional U-Net architecture, SFJD-Net introduces shallow edge information into deep semantic features to enhance texture and boundary representation. The MBCAE module adaptively captures multi-scale lesion features to enrich representations. The SFJE module enhances feature representations from both the spatial and frequency domains, which integrates positional cues and structural details to guide the network in focusing more accurately on target regions. The DLD module uses the reconstructed convolution kernel to record the differences between encoder and decoder, and integrates it into the decoding process through convolution operation, which reduces the semantic gaps and the probability of misjudgment of decoder. Extensive experiments on the published Ischemic Stroke Lesion Segmentation (ISLES) 2022 and 2018 datasets demonstrate that our method achieves state-of-the-art performance. In addition, SFJD-Net is successfully migrated to the pancreas cancer segmentation task of the MSD Cancer dataset, which fully proves that the network has a certain generalization ability.

Convolutional neural networks are widely used for image reconstruction tasks such as pansharpening, low-light enhancement and super-resolution, which are often necessary preprocessing steps prior to downstream applications in remote sensing. However, a critical limitation of existing convolutional neural network architectures is the progressive loss of fine-grained spatial details as information propagates into deeper layers of the network. The degradation of these details restricts the network’s ability to restore high-fidelity images. To address this challenge, this paper introduces a novel Dense Residual Connection Mechanism (DRCM), which establishes multi-pathways for comprehensive feature reuse to effectively preserves more spatial details. We demonstrate the validity of DRCM by integrating it into several representative baseline networks, including PanNet, FusionNet, and DMDNet for pansharpening and super-resolution, and LLCNN, SICE, and RSCNN for low-light enhancement. Experimental evaluations on benchmark datasets, i.e., WorldView-3, WorldView-2 and SICE, reveal that our DRCM-based networks achieve enhanced performance, showing significant gains in spectral-spatial fidelity, structural detail preservation, and overall reconstruction accuracy. Crucially, these performance gains are realized with only a minor increase in model parameters and computational cost, underscoring DRCM’s high efficiency compared to increasing model parameters to reach similar accuracy. This work represents an architectural innovation for optimizing image enhancement accuracy without compromising efficiency in low-level vision applications.

Graphical Abstract

With the development of the data era, heterogeneous networks have attracted extensive attention. However, community discovery in such networks is often challenged by fuzzy community boundaries caused by structural or semantic sparsity, a problem frequently overlooked in the literature. To this end, this paper proposes FTCDH, a novel algorithm for contrastive community discovery in heterogeneous graphs with fuzzy boundaries. First, the algorithm features an adaptive structural and semantic augmentation module, designed to synergistically enhance key topological and semantic patterns while mitigating the representation ambiguity caused by edge noise or irrelevant information. Second, the algorithm introduces a heterogeneous graph contrastive learning module, which utilizes a triple collaborative constraint mode performing multi-level, same-scale representation comparisons, to guide the model in deeply integrating enhancement information. A contrastive loss function then maximizes the consistency of representations across views, yielding highly discriminative node representations. This approach enhances the model’s adaptability to sparse patterns and its ability to distinguish fuzzy community boundaries. Extensive experiments on heterogeneous networks demonstrate the algorithm’s effectiveness, achieving a improvement of over 14% in NMI compared to the next-best baseline.

The accuracy of recommendation algorithms is key to improving expert fault repair systems. However, current prediction methods using single-feature graph convolution still face challenges in adequately representing the embeddings of expert attributes and repair components. Additionally, graph convolutional network (GCN)-based feature encoders struggle to capture the potential interactions between experts and repair components. Moreover, the direct combination of different types of graph features can lead to an imbalance in the contribution of each feature type to the recommendation model, ultimately affecting the effectiveness of the recommendations. This paper integrates recommendation systems from the e-commerce domain with the Industrial Expert Repair Knowledge Graph (IERKG) in the industrial sector, addressing the problem of recommending maintenance experts. Specifically, this paper proposes the Knowledge Infused GraphFuseRec (KGFR), which adopts a dual-channel model structure. The KGFR includes the Knowledge Graph Heterogeneous Feature (KGHF) encoder, the Knowledge Graph Isomorphic Feature (KGIF) encoder, and the Knowledge Graph Feature Fusion (KGFF) encoder. Each module collaborates and communicates, dynamically fusing the two types of features in the knowledge graph, achieving more accurate embeddings of experts/users and industrial faults/projects. KGFR enhances the efficiency of industrial fault resolution and reduces the downtime costs incurred during maintenance. We conducted extensive experiments on industrial fault datasets and two public datasets of movies and music. The results demonstrate significant improvements in the performance of our proposed KGFR compared to the state-of-the-art recommendation algorithms.

Semantic segmentation is a critical task in LiDAR point cloud processing. Nevertheless, current approaches frequently rely too much on preceding frames, which might result in drift or cumulative mistakes due to unrestricted frame-by-frame stacking. This work offers a dynamic alignment of previous frame memory information with observations of the present frame. This ensures a more accurate capture of the features of the current frame and attempts to avoid errors due to changes in viewpoint or object motion. A new multi-scale feature fusion method was also shown. This method uses the spatiotemporal (ST) method to get the ST features, which lowers the differences between the 2D range image coordinates and the 3D Cartesian outputs. Through channel feature alignment and fusion, this method improves feature representation. SensatUrban, nuScenes, and SemanticKITTI datasets were used to test this approach. According to the experimental results, it performs more accurately than current state-of-the-art techniques.

Heterogeneous graphs can represent real-world problems in a way close to reality, supporting diverse types of vertices and edges. However, their inherent heterogeneity poses challenges in interpreting problem semantics. To address this, heterogeneous graph embedding, aiming to map graph elements to low-dimensional vectors, simplifies subsequent machine learning analysis. This approach has gained prominence in machine learning, fueling classification, recommendation, and similarity search applications. Embedding diverse data is essential for efficient data processing. Incorporating language models, like BERT, into heterogeneous graphs enhances semantic context capture, which is particularly useful when one vertex type represents text. Language models stand out in contextual representation, enriching graph vertex embeddings for various tasks. This paper proposes a novel approach to enhancing heterogeneous graph embeddings by combining language models and task class data. Our approach increases vector quality, accounting for graph structure, semantic textual information, and task labels. We compared our proposal with a language model in the aspect-based sentiment analysis task, demonstrating competitive results and, in some cases, a slight superiority. Furthermore, we explore applications of embeddings from auxiliary vertices in another task, highlighting another advantage of the approach over the language model.

The Soft Actor-Critic (SAC) algorithm as a deep reinforcement learning (DRL) algorithm based on maximum entropy of off-policy. By combining off-policy with Actor-Critic and applying it to the autonomous navigation of mobile robots, SAC performs well in continuous state and action spaces. However, this maximum-entropy-based learning process often encounters instability issues. In this work, we introduced a multi-Q value strategy to enhance stability. Environmental drift is exacerbated by the interplay between inherent sensor errors and the constantly changing conditions of a dynamic environment, we emphasized the utilization of a neural plasticity learning mechanism to enhance model robustness. This mechanism selectively retains highly active neurons while pruning fewer active ones, ensuring dual-frequency updating and learning for both high- and low-activity neurons, thus improving the model’s adaptability to environmental changes. Through comparative experiments, we evaluated the performance of MQSF-AC against the SAC, TD3, DDPG, PPO and REDQ algorithms. The results demonstrate that MQSF-AC, our proposed algorithm integrating neural plasticity learning and multi-Q techniques, achieves favorable outcomes for path planning tasks in deep reinforcement learning.

In case-based reasoning (CBR), the risk of data leakage is on the rise. Meanwhile, the ever-expanding scale of case bases has led to markedly reduced prediction efficiency. To address these dual challenges of privacy protection and reasoning efficiency, this paper proposes a novel differential privacy CBR method by improving k-means clustering, aiming to balance data privacy and availability effectively. To overcome the drawbacks of traditional k-means clustering of strong randomness in iterations, this paper optimizes the clustering iteration process through distance information learning. This paper introduces differential privacy to CBR with the improved k-means clustering and explores the privacy budget assignment way, which balances the privacy protection and data availability. Specifically, in order to reduce the noise disturbance for cluster centroid, we fit the iteration process and utilize the integral to approximate the normalized intra-cluster variance (NICV) of cluster. Based on the approximated NICV, we further select the certain iterations to assign the privacy budget and use the rest of iterations to amend cluster centroids. Then, based on the differentially privacy cluster centroids, we design attractive force and repulsive potential force to replace traditional distance metrics in CBR, which significantly improves prediction accuracy. Finally, a series of experimental analyses are conducted on multiple datasets to verify the superiority of the proposed method in terms of privacy protection, reasoning accuracy, and computational efficiency.

Adopting appropriate process parameters is crucial to produce high-quality squeeze-castings. To improve the efficiency and transparency of machine learning in the design of squeeze-casting process parameters, this study proposes an automated and interpretable artificial intelligence approach. The framework employs a stacked integrated model strategy to improve prediction accuracy, automatically optimizes the base model and its generated hyper-parameters, and ultimately interprets the model locally and globally using the Shapley interpretation technique. Application examples show that the relationship between material composition and optimal process parameters is extracted using the framework to obtain the implied relationship between material composition and process parameters in the black-box model, and the process automatically and efficiently completes the modeling and prediction tasks, which significantly improves the modeling efficiency. In addition, the accuracy of the tested performance on the squeeze-casting process dataset is better than that of other machine learning models such as Random Forest, and the predictive behavior of the mathematical model of the squeeze-casting material compositions and process parameters is explained by looking at the two dimensions, globally and locally. The methodology proposed promotes the development of new models for predicting the properties of squeeze castings and provides new ideas for squeeze-casting applications and material-process analysis.

Fault diagnosis of railway point machines becomes an urgent task due to their high fault rate (about 40% of the total fault number of railway signaling systems). Different from the traditional motor current curve-based diagnosis methods, considering the advantages of high precision and easy acquisition of vibration signals, this paper presents a novel vibration signals-based fault diagnosis method, aiming to address the issue of high discrimination feature extraction and selection of complex signals, and realizing high-precision fault diagnosis. First, to address the issue that traditional wavelet packet energy entropy is insufficient to characterize the fault information contained in more complex monitored signals, considering the coarse-grained signals also contain many fluctuations and energy information under different scales, coarse-grain process is introduced into wavelet packet decomposition, forming multi-scale wavelet packet energy entropy. Second, to reduce information loss during the classical coarse-grain process, an improved coarse-grain method using sliding window strategy is proposed. Third, aiming to the problem that the existing feature extraction methods mainly pay attention to the monitored signal itself, considering there are also some important information contained in the multi-order derivatives of the monitored signals, improved multi-scale derivative wavelet packet energy entropy is developed by introducing multi-order derivatives into multi-scale wavelet packet energy entropy, further enriching the fault information. Then, to ensure good robustness of selected features, a two-stage feature selection method combining ReliefF and support vector machine-recursive feature elimination (SVM-RFE) is presented to select the optimal feature subset. Finally, the diagnosis effect is verified using radial basis function (RBF)-SVM and compared to some feature extraction and selection methods. The fault diagnosis accuracies under both normal-reverse and reverse-normal directions reach 100%, demonstrating its feasibility. Besides, the performance of the presented method is also verified on the CWRU data set. This paper can also provide references for feature extraction and fault diagnosis of other machine learning-based diagnosis fields. Besides, it also provide a new way for engineering application of fault diagnosis of railway point machines, and provide theoretical support for on-site maintenance staff.