Lightweight VGG Rolling Sound Classification Model Based on Combinatorial Features

In response to the intelligent transformation of the steel manufacturing industry, many teams are exploring the application of artificial intelligence (AI) algorithms to optimize the steel rolling process. Intelligence brings advantages to the steel industry in terms of labor reduction, high-quality development, and data-driven decision-making compared to traditional models. The basic direction of the development of intelligence in the steel industry is to use acoustic and visual features for real-time monitoring of steel rolling engineering applications and to provide a visual representation of the field state, but due to the objective nature of the production methods in the steel industry, the quality of the visual features is more inconsistent. Therefore, this study analyzed the available sounds of critical processes generated in the rolling shear zone. It was observed that these sounds exhibit few samples and a high degree of similarity, complexity, and partial synchronization. To address this issue, feature fusion methods were proposed, furthermore, there were challenges in generalization and accuracy improvement when training with limited datasets, therefore data augmentation techniques were applied to improve the performance. To address the problem that Convolutional Neural Networks (CNNs) with high computational complexity cannot be efficiently applied in the field, the model is optimized accordingly. Finally, the effectiveness of the proposed method is validated on the proposed rolling line shear zone sound dataset. The experimental results show that the study can efficiently classify the sound in the steel production site, providing a research basis for the development of real-time sound monitoring applications in the steel production site.