To solve the problem that transmission line icing data are affected by environmental and climatic factors and exhibit complex spatiotemporal characteristics

this paper proposes an automatic icing detection model named HCNNs based on heterogeneous collaboration and cross-scale parallel attention fusion. The model consists of three heterogeneous sub‑classifiers working collaboratively. According to the classification error rate of the current classifier

a dynamic weight adjustment mechanism is used to guide subsequent sub‑classifiers to focus on the misclassified samples of previous components. By leveraging heterogeneous characteristics to extract diverse features

the model achieves collaborative coupling among components and improves recognition accuracy. In addition

a parallel fusion attention mechanism (PFA) is proposed and embedded into the HCNNs model. Global pooling and local pooling operations are introduced to decouple convolutional layer features and achieve cross‑scale fusion

thereby optimizing the feature extraction capability of the model. Based on the field monitoring data of transmission line icing states obtained by using distributed optical fiber acoustic sensing technology

including no icing

first-level icing

and second-level icing conditions

the performance of the proposed algorithm is analyzed. The results demonstrate that the HCNNs model achieves a recognition accuracy of 97.14%

precision of 98.26%

recall of 99.30%

and F1‑score of 98.78%

outperforming the compared models. The proposed method provides a valuable reference for icing severity detection of long-distance overhead transmission lines.