Improved ResNet50 Deep Learning Algorithm for Dust Explosion Risk Assessment of Dust Removal Pipelines

Abstract

ABSTRACT In production environments dealing with explosive dust, the associated hazards pose critical risks to employee health and compliance with production environment regulations. To mitigate these risks, prevent equipment cross-contamination, and enhance workplace safety, most production equipment is equipped with dust removal systems. Given the potential explosiveness of dust, dust explosions can initiate high-temperature and high-pressure processes, leading to severe damage. Consequently, preventive measures must be implemented at the source. Incidents of dust explosions caused by dust removal systems are relatively common. To enhance the accuracy of identifying dust explosion risks within dust removal pipelines, this study presents an improved ResNet-50 algorithm for assessing explosion risks in such pipelines. By incorporating the SE attention module and leveraging ImageNet pre-training, we developed the SEResNet50 model, trained on a proprietary dataset using transfer learning. The ablation experiment verified the contribution of the SE attention module and transfer learning to the algorithm. Furthermore, the algorithm is compared with four classic recognition algorithms including VGG16, ResNet18, AlexNet, and GoogleNet. The results show that the SE attention mechanism and transfer learning improve the explosion risk assessment accuracy by 1.81% and 2.33% respectively. More specifically, the algorithm achieves a performance of 97.50%, 95.00%, 100.00%, and 97.43% in terms of Accuracy, Precision, Recall, and F1 Score respectively, which is better than the four algorithms. The algorithm’s ability to detect dust of different colors was evaluated through heatmap visualization of depth features. This algorithm meets the need for accurate risk assessment of dust removal pipelines and improves the safety management level of enterprises.