基于Mamba的井下皮带异物无监督检测模型研究  

Research on an unsupervised detection model for underground belt foreign objects based on Mamba

作　　者：马莉[1] 吴伟雪 代新冠[1] MA Li;WU Weixue;DAI Xinguan(College of Communication and Information Engineering,Xi’an University of Science and Technology,Xi’an 710054,China)

机构地区：[1]西安科技大学通信与信息工程学院,陕西西安710054

出　　处：《西安科技大学学报》2025年第2期372-382,共11页Journal of Xi’an University of Science and Technology

基　　金：陕西省重点产业链项目(2021ZDLGY07-08)。

摘　　要：为了解决井下皮带异物无法被精准定位、计算成本过大等问题,提出了一个基于Mamba的无监督运煤皮带异物检测模型,该模型由预训练编码器和基于Mamba的解码器组成。在Mamba解码器中,FHSS混合状态空间模块将Hilbert扫描位置编码、傅里叶变换、Einstein对角矩阵计算引入Mamba网络来增强通道建模及特征序列建模,并结合了基于重构方法和多类无监督异常检测的优点,解决井下异常数据集匮乏、难以采集的问题。结果表明:该模型精度比经典的4个异常检测模型分别提升了22.2%,10.9%,5.9%,2.1%,其参数量和FLOPs仅为26.109 M,8.497 G;与传统检测方法相比,不仅有效应对由于噪声、遮挡等因素导致的检测不确定性,确保了异物检测的鲁棒性和可靠性,且具备更小的模型体积,显著降低了模型在推理过程中的计算复杂度。研究对于煤矿井下的实际应用具有重要意义,能够更好地保障输送系统的安全性和稳定性。To solve the problems such as the inaccurate positioning of foreign objects on underground conveyor belts and excessive computational costs,an unsupervised foreign object detection model for coal conveyor belts based on Mamba was proposed.This model consisted of a pre-trained encoder and a Mamba-based decoder.In the Mamba decoder,the FHSS hybrid state space module incorporated Hilbert scan position encoding,Fourier transform,and Einstein diagonal matrix computation into the Mamba network to enhance both channel modeling and feature sequence modeling.By combining the advantages of reconstruction-based methods and multi-class unsupervised anomaly detection,it addressed the challenges of scarce and difficult-to-collect underground anomaly datasets.The results indicate that the accuracy of this model has improved by 22.2%,10.9%,5.9%,and 2.1%,respectively,compared to four classical anomaly detection models.Its parameter counts and FLOPs are only 26.109 M and 8.497 G,respectively.Compared to traditional detection methods,it not only effectively handles detection uncertainties caused by factors such as noise and occlusion,ensuring the robustness and reliability of foreign object detection,but also features a smaller model size,significantly reducing the computational complexity during the inference process.This improvement is of great significance for practical applications in coal mines,as it can better ensure the safety and stability of the conveying system.

关 键 词：井下皮带异物检测 Mamba 无监督训练 异常检测 空间状态模型 

分 类 号：TD528.1[矿业工程—矿山机电]