Condition monitoring and inspection are core activities for assessing and evaluating the health of critical infrastructure spanning from road networks to nuclear power stations. Defect detection on visual inspections of such assets is a field that enjoys increasing attention. However, data-based models are prone to a lack of available data depicting cracks of various modalities and present a great data imbalance. This paper introduces a novel data augmentation technique by deploying the CycleGan Generative Adversarial Network (GAN). The proposed model is deployed between different image datasets depicting cracks, with a nuclear application as the main industrial example. The aim of this network is to improve the segmentation accuracy on these datasets using deep convolutional neural networks. The proposed GAN generates realistic images that are challenging to segment and under-represented in the original datasets. Different deep networks are trained with the augmented datasets while introducing no labelling overhead. A comparison is drawn between the performance of the different neural networks on the original data and their augmented counterparts. Extensive experiments suggest that the proposed augmentation method results in superior crack detection in challenging cases across all datasets. This is reflected by the respective increase in the quantitative evaluation metrics.
23 August 2017 in IEEE Transactions on Reliability
Gas circulator (GC) units are an important rotating asset used in the advanced gas-cooled reactor design, facilitating the flow of CO 2 gas through the reactor core. The ongoing maintenance and examination of these machines are important for operators in order to maintain safe and economic generation. GCs experience a dynamic duty cycle with periods of nonsteady state behavior at regular refueling intervals, posing a unique analysis problem for reliability engineers. In line with the increased data volumes and sophistication of available technologies, the investigation of predictive and prognostic measurements has become a central interest in rotating asset condition monitoring. However, many of the state-of-the-art approaches finding success deal with the extrapolation of stationary time series feeds, with little to no consideration of more complex but expected events in the data. In this paper, we demonstrate a novel modeling approach for examining refueling behaviors in GCs, with a focus on estimating their health state from vibration data. A machine learning model was constructed using the operational history of a unit experiencing an eventual inspection-based failure. This new approach to examining GC condition is shown to correspond well with explicit remaining useful life measurements of the case study, improving on the existing rudimentary extrapolation methods often employed in rotating machinery health monitoring.
The creation of unwrapped stitched images of pipework internal surfaces is being increasingly used to augment routine visual inspection. A significant challenge to the creation of these stitched images is the need to estimate the pose and position of the camera for each frame, which is often alleviated through the use of a mechanical centralizer to ensure the camera is held in the center of the pipe. This article proposes a novel method for image centralization and pose estimation, which is particularly beneficial to circumstances where mechanical centralization is impractical. The approach involves post-inspection centralization of the captured video, by first estimating the probe’s position relative to the pipe, using an integrated laser ring projector combined with the camera sensor, and then using this position to unwrap the image, so it produces an undistorted view of the pipe interior (equivalent to unwrapping a centralized view). These unwrapped images are then stacked to produce a stitched image of the pipe interior. In this paper pose estimation was successfully demonstrated to have a 90% confidence interval of ±0.5 mm and ±0.5° in translation and rotation changes. This pose estimation is then used to create stitched images for both a visual test card image mounted inside a pipe and an aluminum pipe sample with artificial defects, in both cases demonstrating near equivalent results to those obtained using traditional mechanical centralization.
Integrity of industrial pipework is ensured through routine inspection. Internal visual inspection tools are capable of characterising degradation in the form of corrosion, pitting, erosion and cracking. The accuracy of such inspection systems has a direct impact on decisions regarding the remaining lifetime of the asset. By minimising error margins, the asset may be operated with confidence for longer, with less uncertainty. This paper considers a probe system consisting of a laser profiler and camera that produces a textured 3D model of the internals of 2 – 6 inch pipework. The accuracy of the system is defined by the ability to extract laser projections from an image as it travels down the pipe, to accurately reconstruct these projections into 3D and to estimate the probe trajectory as it travels through the pipe. This paper presents an error model of the laser profiler. It then presents a novel calibration routine to reduce the error caused by misalignment and tolerances during fabrication of the system. A key advantage of the proposed calibration technique over alternatives is that we can calibrate for errors without manually adjusting the probe, which enables fabrication of a smaller more robust measurement system. In lab-based trials our calibration technique reduced peak sizing errors from 2.7mm to 0.14mm in 120mm diameter pipes.
This paper considers factors which influence the visual motion estimation of a sensor system designed for visually mapping the internal surface of pipework using omnidirectional lenses. In particular, a systematic investigation of the error caused by a non-uniform 2-D spatial distribution of features on the resultant estimate of camera pose is presented. The effect of non-uniformity is known to cause issue and is commonly mitigated using techniques, such as bucketing; however, a rigorous analysis of this problem has not been carried out in the literature. The pipe's inner surface tends to be uniform and texture poor driving the need to understand and quantify the feature matching process. A simulation environment is described in which the investigation was conducted in a controlled manner. Pose error and uncertainty is considered as a function of the number of correspondences and feature coverage pattern in the form of contiguous and equiangular coverage around a circular image acquired by a fisheye lens. It is established that beyond 16 feature matches between the images, that coverage is the most influential variable, with the equiangular coverage pattern leading to a greater rate of reduction in pose error with increasing coverage. The application of the results of the simulation to a real world data set is also provided.