带式输送机输送带纵向撕裂可能引发重大安全事故。针对现有输送带撕裂检测算法精度低、抗干扰能力差的问题, 提出了一种基于多尺度特征融合的纵向撕裂检测系统。系统通过线性激光和高速相机实时捕获输送机胶带表面图像, 使用LoG算法对图像进行预处理, 提取图像关键区域、减少数据冗余, 并通过多尺度特征融合神经网络进行撕裂检测。在检测算法方面, 在神经网络主干网络引入ConvNeXt特征增强模块, 提高模型对细小撕裂纹理的特征提取能力, 在Neck部分使用双向特征金字塔网络(BiFPN)融合浅层细节纹理特征, 减少下采样过程中深层网络细节信息的丢失。实验结果表明, 改进后的算法对纵向撕裂故障检测的检测精度P和平均精度均值mAP分别达到了96. 34%、94. 36%, 优于其他主流的检测方法。

Longitudinal tearing of belt conveyor may lead to significant safety accidents. However, existing algorithms suffer from low detection accuracy and poor anti-interference capability. This study proposes a longitudinal tear detection system based on multi-scale feature fusion. The system captures belt images in real-time using linear lasers and high-speed cameras, preprocesses the images using the LoG algorithm to extract key regions, thereby reducing data redundancy, and finally detects tears using the multi-scale feature fusion neural network. In terms of the detection algorithm, the ConvNeXt feature enhancement module is introduced into neural network backbone network to improve the feature extraction ability of the model for minor tear texture. Additionally, a Bidirectional Feature Pyramid Network (BiFPN) is employed in the Neck part to fuse shallow detail texture features, reducing the loss of detail information in deep layers during down-sampling. The experimental results show that the detection accuracy P and average accuracy mean mAP for longitudinal tear fault detection of the improved algorithm reach 96. 34% and 94. 36%, respectively, which is superior to other mainstream detection methods.