Concept drift, where the statistical properties of a target variable change over time, can degrade the performance of predictive models in dynamic environments. Traditional drift detection methods, often based on hypothesis testing, struggle with subtle changes in data distribution. Recent meta-learning frameworks improve drift detection but still face challenges in accurately classifying drift types due to high intra-class variation. This paper proposes a method to enhance drift type classification by reducing intra-class variation. Following feature extraction, we use a neuron network to embed the features, and then a fully convolutional network for classification, incorporating cosine similarity enhanced with an angular margin. This approach encourages tighter and more distinct class clusters, therefore helps improving classification accuracy. Extensive experiments on synthetic datasets show that our method significantly outperforms the existing meta-learning frameworks in drift type classification. This work also contributes to more reliable concept drift detection in real-time data streams.