Expert Systems最新文献

In industrial product manufacturing, the automated detection and localisation of surface defects are of significant importance for ensuring quality control. However, existing computer vision-based defect detection methods struggle to achieve both lightweight design and high accuracy on resource-constrained embedded platforms, which limits their application in practical industrial detection environments. To address this issue, we propose DSGNet, a lightweight surface defect segmentation model, which serves as a core defect detection and localisation method for industrial inspection systems. The proposed model adopts an asymmetric encoder-decoder structure to simplify the overall architecture. We designed an efficient feature extraction network by using four lightweight feature extraction units based on efficient convolutions. Furthermore, we introduce a hierarchical adaptive upsampling fusion (HAFU) mechanism and a lightweight bidirectional multiscale strip attention (LBMSA) feature refinement module to effectively fuse and refine the multilevel features extracted from the encoder. We conducted comprehensive evaluations of DSGNet on three typical surface defect datasets: Neu-Seg, MSD and MT. While maintaining an extremely low complexity with only 0.49 M parameters, DSGNet achieved impressive mIoU scores of 83.39%, 91.61% and 80.72% on three datasets, respectively. These results indicate that DSGNet is a promising solution that balances lightweight design and detection accuracy for industrial real-time detection systems, demonstrating strong potential for practical deployment. Our code is available at https://github.com/young-zyy/DSGNet.

Human Activity Recognition (HAR) using built-in sensors in wearable devices offers valuable insights for continuous behaviour monitoring and health assessment. However, existing HAR approaches face several key challenges: (1) limited capacity to capture multi-scale temporal dependencies; (2) difficulty in recognising low-motion or static activities when relying solely on inertial sensors such as 3D accelerometers (3D-ACC); and (3) insufficient utilisation of auxiliary physiological signals in conventional single-task learning frameworks. To address these limitations, we propose a Hierarchical-Split Multi-Scale Convolutional Network with Multi-Task Learning for wearable HAR. The proposed model integrates a hierarchical-split convolutional block to efficiently extract both local and global temporal features. It further employs a channel-wise attention mechanism to adaptively fuse photoplethysmography (PPG) and 3D-ACC signals, enabling the model to capture complementary physiological dynamics and improve the recognition of subtle or low-motion activities. In addition, we introduce a multi-task learning framework with an auxiliary heart rate estimation task, which enhances physiological representation learning without requiring additional annotations, thereby improving model robustness and generalisability. Extensive experiments on two public HAR datasets demonstrate that our method consistently outperforms existing approaches under both overall and cross-subject evaluation protocols, highlighting its effectiveness in realistic, user-independent settings.

We propose an adapter-regularised continual learning framework for dynamic financial sentiment encoding, addressing the dual challenge of retaining long-term domain knowledge while adapting to transient market sentiment patterns. The proposed method augments a pre-trained RoBERTa model with lightweight adapter modules and elastic knowledge consolidation, enabling parameter-efficient updates without catastrophic forgetting. Each transformer layer incorporates parallel adapters with gating mechanisms, selectively blending original and adapted features to maintain stability during incremental training. Moreover, the framework dynamically distills emerging sentiment lexicon terms into the encoder through an auxiliary distillation loss, ensuring sensitivity to evolving financial language. The resulting sentiment embeddings are fused with market data via cross-modal attention, where adapter gating outputs further modulate the contribution of textual features. Realised with a RoBERTa-large backbone and online Fisher matrix approximation, our approach demonstrates compatibility with high-frequency trading scenarios while outperforming static domain-adapted models. The key innovation lies in the tight integration of adapter-based continual learning with financial lexicon dynamics and modality-specific gating, a combination not explored in prior work. Experimental validation across multi-modal market datasets confirms the model's ability to simultaneously preserve financial domain knowledge and capture temporal sentiment shifts, outperforming both generic continual learning methods and specialised financial encoders. This advancement bridges the gap between stable representation learning and real-time market adaptation, offering a principled solution for dynamic sentiment analysis in algorithmic trading systems.

With the rapid advancement of generative artificial intelligence, large language models (LLMs) have become increasingly integrated into education, particularly for automated formative feedback and writing assessment. This study introduces and evaluates an AI-driven intelligent feedback system aimed at promoting sustainable and inclusive practices in higher education. It does so by delivering cost-effective feedback in under-resourced contexts and adaptive guidance for diverse learners. The system is built on transformer-based models (BERT and RoBERTa) to support personalised writing evaluation and feedback generation. The system aims to improve students’ self-assessment accuracy (SAA), a critical factor for self-regulated learning, while addressing the challenge of delivering high-quality feedback efficiently in under-resourced contexts. A quasi-experimental design was employed to examine the effects of LLM-generated feedback (LLMF) on students’ SAA and to investigate how these effects vary by initial ability. Results indicated no significant group-level difference in posttest SAA between the experimental and control groups. More importantly, interaction analysis revealed a significant moderating effect of Initial Self-Assessment Accuracy (ISAA). Students with lower baseline accuracy benefited substantially from LLMF, while those with higher baseline SAA showed limited change. This compensatory effect highlights the potential of LLMF to reduce inequities in self-regulated learning. These findings demonstrate the potential of AI-driven feedback systems to cost-effectively reduce calibration gaps and foster metacognitive development. By embedding adaptive and personalised mechanisms, such systems advance educational equity and promote scalable personalised learning. They also contribute to the broader agenda of intelligent and sustainable education.

Most graph summarization algorithms are tailored to a specific graph summary model and were designed for one-time computations only, that is, batch-based computations. We developed a universal approach for parallel graph summarization and three algorithms to compute graph summaries—a batch-based algorithm for static graphs, an incremental algorithm for evolving graphs, and a hash-based algorithm that scales to large graphs and large schema structures, that is, using paths of length up to $��k�$ to define vertex equivalence. Experimenting with benchmark and real-world datasets, we observe that the incremental algorithm almost always runs faster than batch computation, even when 50% of the graph changes, and even when using fewer cores; however, it only uses 8% more memory ($��\pm �1�\%�$). Furthermore, we show that the hash-based algorithm can compute 10-hop equivalent subgraphs on graphs with over 10 M edges within seconds, on graphs of 100 + M edges within a few minutes, and on graphs of 1 + B edges in less than an hour. We analyse the complexity of our algorithms in detail and prove that the incremental algorithm is correct. Overall, we show with these three algorithms that our parallel approach for graph summarisation is versatile and opens the path for various applications that require summaries of large-scale graphs.

Multivariate time series forecasting (MTSF) involves predicting future values of multiple interrelated variables based on historical observations. While existing models often struggle to capture complex temporal multi-scale dependencies and simultaneously modelling intraperiod and interperiod variations, thereby limiting their predictive accuracy. To address these limitations, we introduce a novel forecasting model named MTSPnet. This model employs a two-dimensional (2D) temporal multi-scale patching strategy, which converts one-dimensional (1D) time series data into 2D multi-scale Patch across different time periods. Additionally, MTSPnet incorporates two complementary modules: an interactive multilayer perceptron (MLPmix) module and a dynamic depthwise separable convolution (DDSC) module. These modules enable MTSPnet to efficiently extract both local and global temporal features, further enhancing its ability to model multi-scale dependencies and periodicity variations. Experimental evaluations on seven real-world datasets demonstrate that MTSPnet achieves superior performance in long-term forecasting, proving its effectiveness as a robust and efficient solution for accurate time series prediction.

The Internet of Vehicles (IoV) is revolutionising transportation by connecting vehicles, infrastructure and devices, enabling more intelligent and safer mobility. One key challenge is ensuring efficient and secure communication among vehicles with varying capabilities, including different sizes, speeds and sensor configurations. This research introduces a Federated Learning-Driven Deep Learning (FLDL) approach to intelligent collision avoidance, designed to address the heterogeneity of vehicles in the IoV ecosystem. The system integrates real-time data from vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communications, while considering factors like vehicle type, road conditions, driver behaviour and Digital Twins. Our approach leverages multiple Federated Learning strategies, which enhance privacy protection, reduce communication overhead and enable real-time decision-making without the need for centralised data storage. Experimental results show that the GNN + FedGC model achieves the highest performance with an accuracy of 98.8%, outperforming other models such as MLP with FedLU (98.5%), DRL with FedPPO (98.3%) and LSTM with FedSGD (97.65%). The integration of Digital Twins further enhances model accuracy by simulating real-time vehicle behaviour and environmental conditions. This FL-based system not only improves collision prediction but also enhances safety, reduces accident rates and supports scalable decision-making in smart city transportation systems.