Defect Size Estimation Research Articles

Magnetic flux leakage defect size estimation method based on physics-informed neural network.

Magnetic flux leakage (MFL) is a magnetic method of non-destructive testing for in-pipe defect detection and sizing. Despite the fact that recent developments in machine learning have revolutionized disciplines like MFL defect size estimation, the most current methods for quantifying pipeline defects are primarily data-driven, which may violate the underlying physical knowledge. This paper proposes a physics-informed neural network-based method for MFL defect size estimation. The training process of neural network is guided by the MFL data and the physical constraints that is mathematically represented by the magnetic dipole model. We use synthetic MFL data produced by a virtual MFL testing of pipeline defects to validate the proposed method through a comparison to purely data-driven neural networks and support vector machines. The findings imply that the physics-informed strategy can both improve predictive accuracy and mitigate physical violations in MFL testing, providing us with a better knowledge of how neural networks perform in defect size estimation. This article is part of the theme issue 'Physics-informed machine learning and its structural integrity applications (Part 2)'.

Ultrasonic guided waves through Coalesence of matrix cracks and delaminations in laminated composites

Composite fabric-matrix interface mismatch during manufacturing and applied loads cause matrix micro-cracks and their evolution toward interface delamination and stiffness reduction. Fatigue and impact loads can grow these matrix cracks further as early damage precursors, which are not easily detectable. We simulate ultrasonic guided wave propagation in laminated composite with matrix crack mixed delamination site to understand better detection possibilities for these damage precursors using the ultrasonic inspection method. Since extremely small wavelength ultrasonic waves attenuate rapidly in composites, whereas long wavelength guided waves may not produce high sensitivity to small-scale defects, therefore we attempt to analyze super-resolution quantitative scheme with suitable frequencies to correlate defect parameters. We employ a time-domain spectral finite element method for fast computation of high-frequency wavefield while accurately resolving the matrix crack-delamination geometric features. Its high-order field interpolation enables us to model the nearfield interactions of the waves with the transverse micro-cracks and further with an evolved macro-scale delamination. The computational scheme is aimed to scale to complex defect local field simulations and cover larger size composite structures. This is achieved due to a mathematically consistent diagonalized mass matrix property which makes the implicit time marching-based computation highly efficient for highly resolved and large-size finite element systems. The transient dynamic stress field around the cracks modulates the guided wave mode packets. We analyze the characteristics of the signal sensitivities due to the defect features, most notably, an interesting correlation between matrix crack density and frequency modulation or sub-band contribution. Wave interaction with the defect region creates weak but distinct reflection patterns while transmitting through the region. The study provides useful and new understanding towards defect size estimation and indicative measure of thickness wise layers of matrix cracks and its coalescence into delamination in terms of energy spectral density of the sub-bands.

Estimation of Defect Size and Cross-Sectional Profile for the Oil and Gas Pipeline Using Visual Deep Transfer Learning Neural Network

The magnetic flux leakage (MFL) defect detection of oil and gas pipelines faces two tasks, defect type identification and defect size and shape estimation. However, there are few pieces of research on defect shape estimation, especially fewer research works on defect cross-sectional profile estimation. The complex nonlinear relationship between the defect profile and the MFL signal makes the defect profile difficult to be estimated. In this article, we propose a novel visual deep transfer learning (VDTL) neural network, which not only predicts the defect size but also estimates the defect cross-sectional profile. VDTL network consists of a visual data transformation layer, a transfer learning convolutional neural network (CNN) layer, and a fully connected layer. In addition, we propose an augmentation data to figure (ADF) transformation method for one-dimensional MFL signals, and a fusion algorithm for two-dimensional radial and axial MFL images, which enriches the defect information in the images. Based on the Alexnet network, the multikernel maximum mean discrepancy (MK-MMD) transfer learning algorithm is introduced to improve the accuracy. Experiments are carried out on the data collected in the laboratory and on the data simulated by the finite element method. The results show that the prediction errors for defect length, depth, and defect cross-sectional profile are 0.67 mm, 0.97%, and 2.67%, which are the smallest among the other methods. The research provides a theoretical basis for accurate defect prediction and the safe maintenance of oil and gas pipelines.

Classification and characterization of coexisting defects from magnetic flux leakage data using deep learning method

Ferromagnetic materials are widely used in infrastructure, such as steam generators, storage tanks, and gas pipelines. During their service time, ferromagnetic materials are subject to deterioration and defects are prone to generate which could damage infrastructures and cause catastrophic accidents. Magnetic flux leakage (MFL) is one of the widely used nondestructive evaluation (NDE) methods to detect and characterize defects in ferromagnetic materials to ensure infrastructure safety. However, many research works have been carried out on the modeling, classification, and characterization of a single defect, while the scenario of coexisting defects is ignored. In practical field, the coexistence of surface and subsurface defects within an overlapping area can cause much earlier than expected deterioration or even penetration, the result of which is more damaging. Here, we propose a convolutional neural network (CNN) based deep learning method to differentiate between single defect and coexisting defects scenarios and estimate the defect sizes including length, width, and depth. Finite-element-method (FEM) simulation models are developed to investigate the effect of coexisting defects on the measured MFL data. The models with different defect parameters are calculated to generate 354 MFL data for the training and testing of deep learning method. The experimental results show that the classification accuracy of deep learning method is over 94% and higher than the traditional machine learning methods, and the defect size estimation errors are within 0.97 mm, 0.59 mm, and 3.67% of wall thickness, respectively, which are validated to be a good classification and characterization tool for the coexisting defects scenario.

FMD-Framework: A Size Estimation Method for Pipeline Defects in Weld-Affected Zones

Defect size estimation plays a crucial role in pipeline magnetic flux leakage (MFL) inspection. For the defects in the weld-affected zones, size estimation is a challenge due to the interference of weld signals. In this paper, Feature-Modeling-Decoupling-Framework (FMD-Framework), a size estimation method, is proposed aiming at obtaining the accurate sizes of defects in the weld-affected zones based on MFL measurements. First, a feature extraction module is established to extract relatively pure defect features without weld information. In this part, the feature extraction module searches the intersection and difference between abstract features, and the weld-related information can be modeled and eliminated on the feature level. Second, an inversion module is constructed to decouple the information of different size dimensions on the feature level and achieve accurate estimation. In this part, the designed training suites and novel loss function guide the network to capture the feature information of a certain size dimension. The evaluation experiments are conducted based on the collected MFL measurements. FMD-Framework reduces the size estimation errors by around 2mm(length), 3mm(width), and 3%OD(depth). The results indicate the effectiveness of the proposed method.

A Pipeline Defect Inversion Method With Erratic MFL Signals Based on Cascading Abstract Features

Defect inversion, as a key step in magnetic flux leakage (MFL) inspection widely used in nondestructive testing (NDT) systems, is critical to quantitative analysis of pipeline risk level. However, the conventional methods are difficult to achieve desirable results with complex and changeable MFL signals, mainly because the defect inversion procedure relies on prior knowledge and the features of MFL signals are not fully mined. To resolve the above problem, a novel direct inversion method based on cascading abstract features is proposed in this article, capable of handling feature extraction and defect size estimation problems effectively under complicated conditions. First, an adaptive abstract defect feature extraction network with stacked multipath denoising sparse autoencoder (sMPDS-AE) is constructed, and several extra feature transmitting channels are created, so that abstract features can be transferred across multiple layers. Second, multifeature fusion and sparsity rules are designed to strengthen the sMPDS-AE model and unify abstract feature dimensions. Third, a novel ensemble-backed predictor is developed to study the complex nonlinear relationship between abstract features and defect sizes. With above-mentioned structure, the proposed method can mining out abstract features of pipeline defects from MFL signals adequately and distinguish defects with different sizes more accurately. Finally, several comparison experiments are conducted on a pipeline network with MFL signals. Experimental results and comprehensive analysis with other mainstream inversion methods validate the superiority of this method.

Theoretical and experimental analyses of vibration impulses and their influence on accurate diagnosis of ball bearing with localized outer race defect

This work presents an in-depth analysis of vibration response of ball bearings using a modified 2-degree of freedom (DOF) lumped parameter model. It incorporates additional deflection and multi-impact theories to closely simulate the behavior of healthy and defective bearings, under different load and speed values. The varying compliance vibration has been studied in detail and it is found that the location and the number of impulses due to varying compliance strongly depend on multiple factors, and mainly on the values of applied load and shaft rotational speed. These impulses may merge with the actual defect induced impulses and can alter its characteristics. Moreover, the strengths of these impulses are comparable with those of defect induced peaks, especially for faults at incipient stage. It is illustrated that how the existence of these peaks due to varying compliance may build a misleading impression of the existence of outer race defect even when the bearing is totally healthy. Further, there are cases, in which the defective bearing does not even exhibit the dual vibration impulses. The fault identification and defect size estimation done directly on the ground of these impulses without a prior knowledge of their actual cause, could bring the operator to highly non-optimal decisions.The numerical simulations have been conducted at different speeds, loads, defect sizes and locations, for investigating the influence of these parameters on the bearing response characteristics. Experimental results and detailed analyses are reported in this paper to support the proposed model and validate the numerical results.

Inverse analysis of low frequency electromagnetic signals for sizing local wall thinning using a multivariate probabilistic model

Local wall thinning is a common form of degradation in carbon steel pipes, and a low frequency electromagnetic method is proposed in this study to inspect such wall thinning defects. In addition, an appropriate method needs to be developed to solve the inverse problem, which is to estimate defect size based on the inspection signal. Resolving the inverse problem of nondestructive inspection usually involves a machine learning algorithm and training data should be large and realistic to enable the algorithm to produce an accurate and reliable estimate of defect size. However, the acquisition of such training data is sometimes time-consuming and costly. Therefore, this study aims to estimate the size of local wall thinning based on the inspection signal by developing a new method to generate training data for a machine learning algorithm. With the aid of signals obtained from numerical simulation, a multivariate probabilistic model is proposed to infer a joint distribution over features extracted from measured multi-frequency signals from the low frequency electromagnetic inspection method. The joint distribution is subsequently leveraged to quickly generate sufficient training data for a Gaussian process regression algorithm that uses the signal features as the input to estimate defect size. The multivariate probabilistic model is proved able to reasonably characterize the joint distribution over the features. Moreover, the trained algorithm has been validated by experimental data and it is confirmed it can be used to estimate the residual thickness of a pipe wall with errors within tolerance and high reliability even when the lift-off is changed.

Machine Learning-Based Corrosion-Like Defect Estimation With Shear-Horizontal Guided Waves Improved by Mode Separation

Shear Horizontal (SH) guided waves have been extensively used to estimate and detect defects in structures like plates and pipes. Depending on the frequency and plate thickness, more than one guided-wave mode propagates, which renders signal interpretation complicated due to mode mixing and complex behavior of each individual mode interacting with defects. This paper investigates the use of machine learning models to analyse the two lowest order SH guided modes, for quantitative size estimation and detection of corrosion-like defects in aluminium plates. The main contribution of the present work is to show that mode separation through machine learning improves the effectiveness of predictive models. Numerical simulations have been performed to generate time series for creating the estimators, while experimental data have been used to validate them. We show that a full mode separation scheme decreased the error rate of the final model by 30% and 67% in defect size estimation and detection respectively.

Investigation on dual-domain data processing algorithm used in thermal non-destructive evaluation

ABSTRACT The paper presents the results of research on the data processing algorithm used to detect material defects using active thermography. The algorithm allows the analysis of thermogram sequences in both time and image domain. In the first stage of the algorithm operation, mathematical morphology or filtered contrast methods are used to remove the uneven heating from the sample, as well as to segment and detect defects using local and global thresholding methods. In the next stage, it is possible to determine the number of defects as well as automatically estimate their depth and characteristics (insulator/conductor) in relation to the background material (material without defect). The presented algorithm was tested on two material samples, i.e. PMMA and Expanded PVC, for two phases of the thermal process, i.e. heating and cooling. The study found that the best defect detection and characterisation results are obtained when processing thermographic data from the cooling phase in combination with a Top Hat morphological transformation, local thresholding (for defect detection), and relative incremental filtered contrast (for defect size estimation).

A mechanics and signal processing based approach for estimating the size of spall in rolling element bearing

Rolling element bearing (REB) is a critical component in any rotating machinery. It is subjected to surface damage during operation due to localised fatigue. Any prediction about damage severity helps in effective maintenance planning. Defect size estimation is one of the methods of assessing damage severity in REB and is the subject of this study. Several researchers have formulated analytical as well as signal processing based methods to estimate the defect size in REB. The study proposes a method that is a blend of a mechanics-based approach and signal processing approach for estimation of the size of the defect. The mechanics-based approach is based on the Hertzian contact theory and considers bearing operational parameters along with the effect of the load zone. The analogy of correlating peak acceleration at impact with force in the load zone is proposed and validated experimentally. The signal processing approach for locating entry and exit events encompasses concepts of cross-correlation, Variational mode decomposition (VMD), curve fitting, and root location. Parameter optimization of VMD is done using non dominated sorting particle swarm optimization. The proposed method is used to estimate the size of the defect, which is in the form of an artificially generated spall. Experiments are performed at various speeds and sizes of spall on outer as well as inner raceways. The results advocate the improved efficacy of the proposed method in estimating the size of artificially generated spalls.

Defect size estimation in rolling element bearings with angled leading and trailing edges

Rolling element bearings eventually become worn and develop surface defects, such as spalls, dents, and pits. Previous researchers have used bearings with defects that have sharp rectangular edges to develop methods for the estimation of the size of a defect. However, these idealized rectangular defects do not occur in the real world, they are in fact sloped. Tests on bearings with sloped edges on the entry and exit of the defect exhibited similar vibration characteristics as bearing with sharp 90° edges. A new defect size estimation method was developed to improve the accuracy in estimating the size of a defect that has sloping leading and trailing edges. A large dataset was obtained and has been made publicly available.

An Appearance Inspection with Defect Size Estimation based on Deep Learning Algorithm

An Appearance Inspection with Defect Size Estimation based on Deep Learning Algorithm

Bayesian optimized deep convolutional network for bearing diagnosis

The successful diagnosis of the faulty signal in rolling element bearings relies on the accurate evaluation of the early fault present within the components of bearings. Because of system imperfection and interference between the data acquisition devices, the fault signal is heavily masked by noise. Hence, to extract the fault information, signal processing techniques are widely used in bearing diagnosis. Although numerous research have dedicated on finding representative features to indicate the damage of the bearing elements, the correlation between defect size and acquired vibration signal has not been properly established. In the recent few years, deep learning has been widely used in bearing diagnosis. In general, the unprocessed signal is directly input into the deep learning model and the neural network extracts useful features during the optimization process. Until now, the selection of features from signals is arbitrary which does not yield much insights into the diagnosis and prognosis. In addition, the features could contain noise information, which could possibly deteriorate diagnostic or prognostic results, while the effect of using preprocessed data has not been fully explored. In this paper, we present an innovative diagnosis model using the deep convolutional network with Bayesian optimization to diagnose the defect severity of bearings. The acquired signal is initially processed using the complementary ensemble empirical mode decomposition method to extract the frequency band containing the fault signature. An experimental based defect size estimation equation is implemented to calculate the defect size based on the signal and experimental setup. After the estimated defect size is obtained, the deep convolutional neural network is implemented to categorize the defect severity. The parameters of the network are optimized by the Bayesian algorithm. The proposed algorithm can be used for diagnosis of the health condition of various rotating machinery.

Estimating the size of naturally generated defects in the outer ring and roller of a tapered roller bearing based on autoregressive model combined with envelope analysis and discrete wavelet transform

• Accurate estimation of size of natural defects in rolling element bearings. • Use of envelope analysis and discrete wavelet transform of pre-whitened signal. • Comparison of size estimation at different speeds and loads. Vibration analysis has been widely used for condition monitoring and fault diagnosis of rolling element bearings. The next step is to determine defect size from the measured vibrations which can be used in the decision making about the bearing replacement. This paper presents an experimental investigation of the defect size estimation from the vibration signals of two types of naturally generated defects, on the outer race and roller of two tapered roller bearings. Vibrations are measured at different shaft speeds and loads. The signatures generated by the entry of the rolling elements into and exit from the defective area are enhanced by the autoregressive model and envelope analysis. The duration between these two events is used to estimate the size of natural defects. The accuracy of the results is compared with the results obtained by the autoregressive model and discrete wavelet transform.

Experimental study on enhancement of defect detection and defect size estimation in deep groove ball bearing

Rolling elements bearings are widely used in rotating equipment and machines to support load and to reduce friction. The presence of micron sized defects on the mating surfaces of the bearing components can lead to failure through passage of time. The large size defects on bearing elements can be detected/identified by time domain and frequency domain analysis of its vibration signals. However, it becomes difficult to detect local bearing defects at their initial stage either due to their smaller size or presence of noise. In the present experimental study, detection of local defects like crack and pits on bearing races have been carried out. Vibrations generated by healthy bearing and bearing having circular or rectangular defect on either race of bearing have been analyzed using MATLAB software. Signal to noise ratio of vibration signal has been enhanced through self- adaptive noise cancellation. Moreover, width of rectangular defects and diameter of circular defects have been estimated through ‘db5’ wavelet. The estimated defect sizes have been compared and validated through measured actual crack width or pit diameter. Accuracy of results proves that wavelet analysis of time domain signal can be used with confidence to estimate the width of fatigue crack and pit of the bearing races.

Defect size estimation method for magnetic flux leakage signals using convolutional neural networks

A method for defect size estimation of magnetic flux leakage (MFL) signals using convolutional neural networks (CNNs) is proposed to overcome the problem of quantitative identification of pipeline MFL testing. The model mainly includes two modules: defect classification and defect size regression. The former is used to realise data fusion, feature extraction and tasks of three components (axial, circumferential and radial) of the MFL defect signal. The defect size regression module includes seven CNNs, which realise the size estimation of different types of defect. The input is a different type of defect in the defect classification module and the output is the length, width and depth information of the defect. Finally, the training and prediction are conducted using a defect dataset. The results show that the proposed method can effectively identify the MFL defect size of the pipeline.

Improvement of defect size and morphological estimation in shearography inspection by wavelet transform

Speckle shearography is often used in the aerospace industry to detect defects in composite materials because it allows noncontact, full field, and fast measurements. However, despite this advantage, this nondestructive technique is still not widely established, due to its limitations in performing defect quantitative estimations. The main issue lies in the extrapolation of phase profile boundaries, which represent defect edges. In this paper, this was achieved by studying the multimodal distribution of a characteristic parameter, called structural intensity, which is the probability of finding a denser population of local maxima of wavelet coefficients at a given position and adding an adaptive threshold selection in the structural intensity distribution. Once the damage boundaries were detected, the defect dimensions were computed by subtracting the local shearing amount, which was defined by a circle detection code, from each coordinate in both Cartesian directions. The paper shows an improvement to the algorithms able to detect defect area and perform morphological estimation previously developed by the authors and an in-depth study of the parameters involved in these algorithms. This resulted in both a high accuracy (0.04%) revealed by the calibration process on a PVC test panel and a close morphological agreement with the ultrasound reconstruction of unknown defects in composite materials. Thanks to this improvement, shearography becomes a quantitative inspection technique that takes few minutes compared to ultrasound inspection of the same area, which may require up to an entire day.

An Estimation Method of Defect Size From MFL Image Using Visual Transformation Convolutional Neural Network

In most current nondestructive testing systems, a magnetic flux leakage (MFL) method is widely used in various industry fields, where the structural integrity of specimens is of vital importance. The estimation of defect size in specimen from the MFL measurements is a key and difficult problem. The traditional methods have low precision because feature extraction procedure relies on prior knowledge and the ability of designer. Inspired by the idea of convolutional neural network (CNN), a novel visual transformation CNN (VT-CNN) is proposed in this paper to overcome the limitation of traditional method in a feature extraction procedure. By adding a visual transformation layer according to the characteristics of the MFL measurements, the VT-CNN can distinguish the defect feature with different sizes more accurately. Moreover, since the VT-CNN method is designed based on the deep learning theory, more industrial big data with accurate label should be used to train the network. Due to the difficulty of making real industrial big data, a novel mesher magnetic dipole model is designed to simulate this industrial process. A large simulated MFL measurement of irregular defects produced by this model can increase the number of training samples and improve the robustness of the network. Experiments to estimate natural corrosion defects on real industrial pipelines are performed to validate the proposed framework. The experimental results are illustrated in detail, which highlights the superiority of the proposed method in industrial applications.

Ultrasonic spectroscopic analysis of impact damage in composites by using laser vibrometry

Contemporary composite materials are continuously being modified and improved in accordance with growing quality requirements of aerospace systems. New manufacturing technologies stimulate development of appropriate (novel) nondestructive testing (NDT) methods and respective hardware for detecting and characterizing hidden defects. One of the recent improvements in this area is related to combining acoustic stimulation of test objects and scanning laser vibrometry. This paper describes an ultrasonic spectroscopic approach to laser vibrometry applied to hybrid composite materials subjected to impact damage. The meander-shaped signal for resonance stimulation was applied to enhance the detection of multi-component defects. It has been shown that wide-band frequency stimulation provides some advantages in regard to mono-frequency excitation because it activates more zones of multi-component defects thus allowing more accurate estimation of defect size.