Aluminum die-cast products often exhibit surface defects that vary in type and severity depending on product function and design requirements. Current defect detection primarily relies on manual inspection, which demands significant expertise, raises health and fatigue concerns, and is prone to human error. Automated defect detection offers a promising solution to reduce costs, improve efficiency, resolve occupational safety and health concerns, and mitigate challenges in labor shortages and training. This paper presents an ensemble deep learning (DL) approach for detecting surface defects in aluminum die-cast lids for residential gas meters, a quality-critical component with stringent safety standards. Specifically, we implement and evaluate three state-of-the-art DL architectures: a convolutional neural network (CNN), residual networks (ResNet-18), and Vision Transformer (ViT). In addition, we develop an ensemble model to further enhance performance. We leverage grid search and cross-validation for hyperparameter tuning and train/test each model ten times for comprehensive performance evaluation. Experiments on a large real-world dataset demonstrate that all models achieve high accuracy, precision, and recall, with CNN and ResNet-18 slightly outperforming ViT. The ensemble model further improves prediction accuracy and robustness. The paired t-tests showed that the ensemble model significantly performed better compared to CNN and ViT model. In summary, this study contributes to the advancement of automated inspection of surface defects in die-cast products by systematically comparing state-of-the-art deep learning methods, discussing model selection criteria, and optimizing ensemble strategies. Centered on CNN, ResNet-18, and ViT architectures, it proposes a rigorous methodological framework for surface defect detection and provides a foundational basis for subsequent research in in-situ quality control. Our codes are available at https://github.com/Alexruoyun/Aluminum-Die-Casting-Surface-Defect-Detection.

Autonomous vehicles require effective prediction of potential future trajectories of surrounding agents. The current trajectory prediction methods have limitations, firstly, traditional feature fusion methods merge scene features sequentially in a simplistic manner, often overlooking the intricate interrelations among scene elements, leading to incomplete selection and insufficient utilization of useful features; secondly, in multimodal trajectory prediction, the mode collapse issue inherent to probabilistic approaches results in inadequate expression of agent intent uncertainty, while overly anchor-dependent proposal-based methods can generate implausible trajectories. To address these limitations, We present a Dynamic scene and Relational reasoning Transformer (DRTR), a novel multimodal trajectory prediction method based on dynamic scene encoding and relational reasoning. A pivotal aspect of DRTR is the dynamic closed-loop modeling framework that effectively combines scene features to output three dynamic features: dynamic traffic flow, dynamic agents, and interactions between agents. This innovative framework ensures a comprehensive capture of the dynamic scene and its intricate interrelations. Then, DRTR initializes a set of trajectory suggestions representing various modalities and carefully refines these suggestions by sequentially fusing and querying dynamic scene features, ensuring predictions are both accurate and reflect multimodality. To further enhance model expressiveness, we introduce a feature selection network based on relational reasoning, which can recognize deep relationships between scene elements and select beneficial contextual features. Experiments on the Argoverse 1 dataset indicate that DRTR exhibits superior performance, particularly in multimodal trajectory prediction.

Reliable gait phase classification is essential for wearable-based locomotion analysis. Although gait cycle percentage prediction and gait phase classification are biomechanically related, knowledge transfer across these distinct objectives remains underexplored. In this paper, we propose a regression-to-classification transfer learning framework that utilizes temporal representations learned from continuous gait cycle progression to improve discrete phase recognition. We pre-train neural backbones on a regression task and transfer the learned representations to the classification task through model transfer (fine-tuning backbone weights) and feature transfer (using the backbone as a fixed feature extractor). To identify the optimal configuration for resource-constrained environments, we compare a compact Deep Neural Network (DNN) with 0.3 M parameters and a Transformer model across multiple sliding window sizes. Our experimental results demonstrate that model transfer achieves a superior F1-score of 0.9788, outperforming the feature transfer baseline and models trained from scratch. Efficiency tests show that the compact DNN achieves a Central Processing Unit (CPU) latency below 0.07 ms, supporting real-time data processing. Validation on an independent dataset further confirms cross-population robustness, achieving a classification accuracy of 92.3%. These findings suggest that regression pre-training captures effective temporal features, providing a practical framework for wearable-based gait analysis.