Research on a Method for Classifying the Weight of Surplus Materials Based on an Improved CBAM-3LResNet18

In order to address the limitations of traditional machine learning methods in the weight-based classification of foreign objects within sealed chambers, such as insufficient feature extraction and limited model recognition rates, this paper proposes a weight-based classification method for foreign objects based on an improved CBAM-3LResNet18 architecture. Initially, sound signals generated by foreign object collisions are converted into time-frequency spectra, and manual features are extracted from the time-frequency plots. Secondly, the Convolutional Block Attention Module (CBAM) is integrated into the conventional ResNet18 model, and three fully connected layers (3 L) are designed to optimise the model, thereby enhancing its capacity to extract and fuse key features. In conclusion, a substantial amount of experimental data was compared and analysed in order to validate the hypothesis that the improved model is more accurate and robust than the original model and traditional methods when it comes to identifying the weight of foreign objects. The experimental results demonstrate that the model attains recognition accuracy rates of $80.3 \%, 78.6 \%$, and $80.3 \%$, respectively, under three distinct sample conditions, thereby exhibiting a marked superiority over conventional methodologies. Furthermore, it evinces a high degree of discriminative capability for weight features of extraneous objects in complex signal environments, indicating a promising trajectory for practical application.