基于脉冲数据重构的高速铁路关键零部件检测算法

Detection Algorithm for Key Components on High-speed Railways:Reconstruction Based on Spike Data

对运行中的高速铁路(简称“高铁”)关键零部件进行快速准确定位是高铁故障诊断的基础之一。然而,利用普通工业相机对运行中的高铁进行数据采集会存在运动模糊或关键零部件丢失问题,而利用超高速脉冲相机采集的数据进行目标检测模型训练又会存在数据难以标注以及模型难以训练等问题。对此,文章提出一种基于脉冲数据重构的高铁关键零部件检测方法。其首先基于脉冲触发原理,利用脉冲间隔还原不同位置的光强,并通过计算全局光强调整局部光强,且对调整后的结果进行滤波与局部增强处理,重构出高清分辨率灰度图像;在此基础上,对重构后的灰度图像进行数据清洗与标注,并利用标注后的数据进行高铁关键零部件检测模型训练;最后,基于TensorRT对训练后的模型进行加速。试验结果表明,经过加速后的高铁关键零部件目标检测模型能够实现98.7%的平均精确度以及平均3.5 ms的单帧推理速度,为算法的工程化应用奠定基础。

The accurate and fast localization of key components on high-speed railways in operation serves is one of the foundations for fault diagnosis of high-speed railways.However,utilizing ordinary industrial cameras for data collection on operational high-speed railways often results in motion blur or misses key components.On the other hand,training object detection models with data collected by ultra-high-speed spike cameras encounters challenges,such as difficulties in data annotation and model training.To this end,this paper proposes a detection method for key components on high-speed railways that leverages reconstruction based on spike data.Firstly,based on the principle of spike triggering,the light intensity was restored at different points following the inter-spike interval.Then,local adjustments to light intensity were made through global light intensity calculations.The adjusted results were further filtered and locally enhanced to reconstruct high-resolution grayscale images.Subsequently,after data cleaning and annotation based on the reconstructed images,the annotated data were used to train the detection model for key components on high-speed railways.Finally,the trained model was accelerated using TensorRT.Experimental results showed that the accelerated model achieved an accuracy of 98.7%and a single-frame inference rate of 3.5 ms both on average.The study findings lay a foundation for the engineering applications of the proposed algorithm.