Accurate and interpretable delivery delay prediction is essential for effective decision-making in modern supply chains. However, most existing approaches focus on predictive accuracy while providing limited support for actionable decision-making, especially in resource-constrained environments such as Small and Medium-sized Enterprises. To address this gap, this paper proposes a Knowledge-Enhanced Small Language Model (KE-SLM) framework for AI-enabled decision support in delivery delay prediction. The framework transforms structured data into semantically enriched textual representations, integrates statistical context and auxiliary model signals, and employs teacher–student distillation to transfer reasoning capabilities from large language models to lightweight models. Experimental results on a real-world dataset show that KE-SLM achieves superior recall and F1-score compared to strong baselines, improving the identification of high-risk orders. More importantly, the model generates natural language explanations that provide actionable insights, enabling transparent and human-centered decision-making. This study demonstrates a practical and scalable approach to integrating prediction, interpretability, and decision support within a unified framework, contributing to the development of AI-enabled decision support systems in smart supply chains.
Due to the powerful feature learning and classification capabilities of convolutional neural networks, they have been widely applied in fields such as image recognition and classification, object detection, and semantic segmentation. Therefore, enhancing the efficiency and accuracy of convolutional neural networks in deep learning has become increasingly important. In convolutional neural networks, the initial design of the convolutional kernels significantly influences the efficiency of deep learning iterations. This paper focuses on the recognition and classification of image tables, proposing a method of customizing the initialization of convolutional kernel parameters based on several local elements of images and the similarity between their pixel distribution features and the initial parameters within the convolutional kernels. The deep learning of image tables through convolutional neural networks is then conducted on this basis, and compared experimentally with traditional initialization methods such as Normal and Xavier. The experimental results show that the parameter initialization method proposed in this paper achieves notably higher classification accuracy in the early stages of neural network training and overall higher classification accuracy for table recognition.