Existing wheelset measurement methods primarily rely on one or several single-line lasers, which provide only limited information. Their measurement reliability and accuracy are insufficient to meet the requirements for high-speed train wheelset measurements. To accurately measure a wheel’s parameters online, we propose a novel 3-D reconstruction framework driven by a multiline laser vision sensor array. For local 3-D profile reconstruction, we propose multiline laser calibration, local joint calibration, and moving multitarget joint calibration to achieve the local stitching of multiline laser profiles. For global 3-D profile reconstruction, a global dynamic calibration method is proposed to concatenate local wheel 3-D profiles with varying positions and poses to a complete wheel 3-D profile. Field experimental results show maximum errors of 0.15 mm for flange height, 0.21 mm for flange thickness, and 0.38 mm for diameter. To our knowledge, this is the highest accuracy achieved in the measurement of wheelset geometric parameters.

In harsh and complex aerospace environments, stress concentration causes fatigue cracks to easily form at the edges of rivet holes in an aircraft fuselage, threatening overall aircraft safety. Although testing methods based on high-order focusing ultrasonic guided waves are sensitive to cracks, there remains a lack of high-precision quantification methods. Existing signal processing and machine learning methods have shortcomings. For example, the former cannot achieve intelligent detection and the latter has black box problems with inexplicability. To solve the above-mentioned problems, we propose a space-time cross-fusion bin-to-bin distance-theory-hybrid deep neural network (DNN) called SpatNet that combines the physical knowledge of defect quantification with a DNN to effectively detect and quantify rivet hole cracks in an aluminum alloy fuselage. In addition, we analyze the attenuation coefficientskR,G,Band window length$w_{l}$, which can describe the degree of physical knowledge introduction, and compare the proposed SpatNet with other classical methods. The results show that the reasonable introduction of physical knowledge can effectively improve the prediction accuracy of a DNN and reduce network dependence on large-scale datasets. Related research on SpatNet based on the fusion of physical knowledge and neural network space-time cross-fusion has the potential to promote the further development of the intelligent ultrasonic testing field.

To detect the wheelset deformation and wear, an intuitive method is to first collect wheelset multi-line laser stripe images by monitors, and then extract the centerlines to construct 3D contours that are transferred to the cloud data center by 6G communication. The images, however, contain flairs and fractures due to the influence of environmental interference and reflected light on the smooth surfaces. The image defects affect the accurate extraction of stripe centerlines. To segment the defects and inpaint them, we propose a new lightweight U-shaped segmentation model URS. A target-shaped receptive field is designed to efficiently extract the details of the local, the ring-shaped, and the cross-shaped context around the local, which facilitates segmenting various defects. A scale-select sub-module is designed to adjust the weights of features from different receptive fields. To train the model, a multi-line laser image defect segmentation dataset MLIDSD is constructed. Experiments demonstrate that the performance of our model surpasses twelve SOTA models explicitly and can achieve a balance between the accuracy and the lightweight requirement.

Wheelset fault detection with high accuracy is challenging due to poor image quality. Specifically, the wheelset images are collected dynamically outdoors and suffer from diffuse reflection and environmental interference. Thus, the images contain light stripe adhesions (light flairs) and local fractures to be inpainted. The existing inpainting models are inapplicable to restore grayscale wheelset images. They are also too heavy to be deployed in an embedded wheelset monitoring equipment. In this paper, we propose a lightweight high-precision inpainting model that consists of a recurrent similarity network with the ghost convolution (RSG-Net) to remove light flairs and repair local fractures. RSG-Net replaces standard Pconv (partial convolutional) layers with soft-coding ones that can improve the feature representational ability. To reduce the influence of the background region features on image restoration, an asymmetrical similarity measure is designed to calculate not only the angle difference between the target and the source feature vectors but also the activation of the source ones. The multi-scale structural similarity (MS-SSIM) loss term is introduced to precisely guide the structural information restoration, such as the stripe edges. Moreover, the ghost convolution is introduced in RSG-Net to realize the model compression that can retain the core features of wheelset images and remove the redundant features. We conduct three groups of experiments that demonstrate the accuracy superiority of the proposed RSG-Net over the baseline methods, and the number of parameters is reduced by about 50%.