This study presents a method of ultrasonic flaw identification using phased array ultrasonic inspection data. Raw data from each individual channel of the phased array ultrasonic inspection are obtained. The data trimming and de-noising are employed to retain the data within the boundary of the inspected object and remove the speckle noise components from the raw data, respectively. The resulting data are passed into a sequence of signal processing operations to identify embedded flaws. A shape-based filtering method is proposed to reduce the intensity of geometric noise components due to the non-uniform microstructures introduced in the manufacturing process. The resulting data matrices are integrated to obtain the intensity matrix of the possible flaw regions. Thresholding is applied to the intensity matrix to obtain the potential flaw regions, followed by a connected component analysis to identify the flaws. The overall method is demonstrated and validated using realistic phased array experimental data.
Ultrasonic bulk waves have excellent propagation properties for metal materials, and their interaction with material discontinuities has been utilized as an efficient tool for damage detection since 1940s [1–5]. An ultrasonic transducer can emit high-frequency bulk waves and simultaneously acquire the reflected echoes from the surface of the materials . The amplitude of the reflected echoes and the time of its occurrence are used to interrogate the internal states of the materials. In particular, this pulse-echo configuration is suitable for in-situ, intermediate, and overhaul maintenance via manual and automated procedures [7–12] and has become one of the most widely used configurations for product inspection, structural health monitoring, and safety assurance [13–15]. A phased array transducer has a group of individually controllable actuators, allowing for electronically steering the beam direction and adjusting the depth of the focus . Using phased array bulk waves has several advantages over the conventional monolithic bulk waves. For example, it can sweep a sectorial domain at one time, and by moving the probe, the same internal location can be interrogated from multiple directions and locations [16–18]. Consequently, the reliability of damage detections can greatly be improved [19,20]. In addition, the efficiency and flexibility of the ultrasound phased array also allow for rapid inspections of material blocks, reducing the overall inspection time [18,21]. The ultrasonic data acquired by a conventional monolithic transducer at a specific location are usually presented in a two-dimensional plot, representing the echo amplitude versus time-of-flight [22–24]. Manual interpretations of such data are not difficult due to the simplicity of such plots. However, phased array transducers emit and acquire the data simultaneously from multiple angles of incidence, making the manual interpretation highly nontrivial and time-consuming.
Recent studies on flaw detection with phased array ultrasound have been focused on improving the imaging technology using various signal processing techniques. Camacho et al.  developed an ultrasound approach based on the total focusing method (TFM) and phase coherence imaging (PCI) to monitor crack size during the fatigue test. Li et al.  reported an improved TFM imaging technique by combining velocity anisotropy and optimizing the aperture angle and frequency filter to improve the quality of the resulting images. Meksen et al.  proposed a method based on sparse matrix representation instead of the whole data in the time-of-flight diffraction (TOFD) technique to improve the efficiency in dealing with large datasets. Sinclair et al.  developed a digital signal processing scheme based on the synthetic aperture focusing technique (SAFT). By combining with a variation of Wiener filtering and autoregressive spectral extrapolation, the image resolution and size quantification were improved in weld applications. Guan et al.  proposed a time-domain SAFT technique to improve the spatial resolution of the inspection and validated the method using artificial and natural flaws. Fan et al.  developed an ultrasonic imaging method for concrete-filled steel tube inspections using time-of-flight data interpolation and normalization. Brizuela et al.  reported an ultrasonic imaging method for phased array data with dynamic depth focusing, SAFT, and PCI to improve the image resolution. Zhang et al.  proposed a method for sizing crack-like defects with similar or less size than the wavelength by measuring the scattering coefficient matrix of defects. Prager et al.  compared two defect sizing techniques, SAFT and TOFD, on a reactor pressure vessel mock-up. Peng et al.  presented an ultrasonic image-based sizing technique that can measure the cracks larger than two wavelengths. Although SAFT-based methods are shown to yield better performance in terms of spatial resolutions, the methods rely on a very high precision during data acquisition to ensure the in-phase summation and achieve the high signal-to-noise ratio. This can be a practical challenge when the inspected surface is irregular, making it less attractive in realistic engineering applications. Therefore, a systematical method for reliable recognition of embedded flaws using phased array data is highly demanded, yet remains a great challenge.
Over the past two decades, many efforts have been made toward improving the quantality of static sectorial images acquired using the phased array ultrasound at one location. For industrial applications where automated and continuous data acquisition are mandatory, data fusion and flaw detection become extremely nontrivial when the same spot is interrogated from different angles of incidence at different locations . There is no universal method to fuse the data, and few studies have been reported on rapid and reliable identification of flaws incorporating the multi-angle phased array data. The objective of this study is to develop a systematical methodology for embedded flaw identification using multi-angle phased array ultrasonic data. The method consists of two major steps. In the data pre-processing step, the raw data of each individual channel are packed as two-dimensional matrices and speckle noise components are removed from the data. In the flaw identification step, the data are processed using two shaped-based filters to eliminate the horizontal and vertical banded noise components. Connected component analysis is employed to identify the potential flaws from the processed matrices. The rest of the paper is organized as follows. The proposed method is presented in detail where all the necessary data processing steps are discussed in detail. After that, the proposed method is demonstrated using experimental data acquired from a block object with artificial flaws. Following that an aluminum alloy with artificial flaws and a natural flaw are used to validate the method. Finally, conclusions are drawn based on the current study.
2 Methodology Development
The ultrasound data, particularly the pulse-echo data, are usually corrupted with various noise components such as speckle noise, electrical noise, interface noise, and the geometric echo noise, significantly limiting the visual identification of the flaws [34,35]. In addition, the near-surface region cannot be inspected effectively due to the near-field effect of the ultrasound . Figure 1 presents a typical phased array dataset captured by the channel with a 0-deg angle of incidence in the form of a two-dimensional scalar matrix, and is shown as a two-dimensional intensity image with a color look-up table. The near-surface region and interface noise components are the saturated horizontal banded regions at the top and bottom of the image, respectively. The speckle noise components are localized features shown as random point-type high-intensity pixels, and the geometric echo noise components are those global features shown as low-intensity horizontal stripes. The role of the data processing is to reduce the noise components and to enhance the damage features, so that the potential flaws can be differentiated from the background of the ultrasonic data. To make the data processing more generic, the raw data from each individual channel are stored as two-dimensional matrices. Each of the components in the matrices is an intensity scalar representing the echo amplitude of a spatial location in the object being inspected. As shown in Fig. 1, the columns of the matrix index the sampling location of the probe along its moving path and the rows of the matrix index the depth of the ultrasound beam along the wave propagation path. The ultrasonic data can be normalized to the range of [0,1] to ease the subsequent data processing.
This proposed method consists of three major components, namely, the data pre-processing, the flaw recognition, and the multi-angle integration. The flowchart of the overall method is illustrated in Fig. 2. The data matrix is first trimmed according to the sound speed and the physical boundary of the object, and irrelevant and out-of-range data are removed. The remainder of the data is future processed using de-noise filters, e.g., a Gaussian filter, to eliminate the electrical and speckle noise components. The resulting data are passed into the flaw identification procedure where a sequence of filtering operations is applied. Several shape-based filters are used to identify possible flaw regions. The filtering operations are performed using two-dimensional fast convolutions. The connected component analysis is employed to obtain the pixel groups of each of the flaw regions. Thresholding is used to exclude unrealistic flaw regions, and the remaining flaw regions are considered as the identified flaws. The above process is made in each of the channels, and a pixel-wise summation is performed to integrate all the multi-angle channels.
2.1 Data Pre-Processing.
The data pre-processing consists of three steps as shown in Fig. 2. The purpose of the data trimming step is to remove the interface noise components and the out-of-range portion of the data. The interface noise components are the saturated data of the back wall region, which can easily be removed by trimming the data matrix. The out-of-range data in the matrix can also be removed according to the sound speed, the sampling time interval, and the physical dimensions of the object. A predefined intensity threshold (Ath) is applied to the trimmed data matrix to reduce the random electrical noise due to the hardware and equipment. The matrix components with intensities being less than the predefined threshold are removed. The intensity threshold (Ath) is determined based on the background noise components of the data associated with the material and the test condition. Expert judgment and trial testing can be used to estimate the value. Once the threshold level is determined using one specimen, the threshold is applicable to all the specimens of the same material under the same test conditions. The Gaussian low-pass filtering [37–39] is employed to reduce the speckle noise components introduced by the interaction of the reflected waves. The filtering effect can be tuned using two parameters of filter size (n) and standard deviation (σ). The geometric noise components are related to the non-uniform microstructures of the material, which are often introduced by joining of two parts, forging, and tempering. Consequently, the geometric noise components usually have certain regular geometry patterns. Geometric noise components often appear as continuous high-intensity strips or lines over the entire scan path, and the intensities in the bands or lines are approximately uniform. It should be noted that the patterns of the geometric noise components are not always the same; however, for an inspection task where multiple datasets are obtained from the same object, the noise features are relatively similar among the datasets. After the data pre-processing, the resulting data matrix is passed to the flaw recognition step. The method of flaw recognition is presented in Sec. 2.2.
2.2 Flaw Recognition.
The flaws are loosely defined as embedded material discontinuities, such as voids, inclusions, cracks, and so on. The basic idea of the flaw recognition is to further reduce the intensity of the noise components while keeping the signal features associated with the flaws intact.
The discontinuities introduced by the flaws can interact with the propagation of the ultrasonic waves, causing reflections and refractions. The reflected waves can be picked up by the transducer and the resulting data corresponding to the reflected waves exhibit higher intensities, compared with the background intensity. The local variation of the signal intensities introduced by the material discontinuities can be used to identify potential flaws. The basic idea is to detect the boundaries of the high-intensity regions associated with potential flaws. To achieve that, a sequence of filtering operations based on the image gradient is used. The shape-based filtering is proposed in this study and is implemented using fast two-dimensional convolutions. Figures 3(a) and 3(b) present vertical and horizontal filters, respectively. Other directional pattern filters are shown in Fig. 3(c). The data matrix is processed separately using horizontal and vertical shape-based filters, and the resulting data matrices are combined. An optimal overall threshold (Bth) is chosen to identify the edges, i.e., flaw boundaries, from the background. Edge components in terms of pixels in the context of two-dimensional images with values smaller than Bth are set to 0 and the rest of the components are set to 1. A suitable value for the threshold Bth can be determined based on the maximum value of the data matrix resulting from shape-based filters. Note that the edges obtained by filtering are shown as ring-like patterns. The closed operation of morphological image processing can be used to fill these ring-like patterns .
The connected components analysis is employed to obtain individual flaws [34,41]. The operation generates a list of connected pixel groups each of which represents a potential flaw. It should be noted that the connected components analysis identifies the flaw regions in terms of the indices of the image. The data values of each flaw are subsequently retrieved from the corresponding raw data matrix using those indices. The parameters in each step of this method need to be carefully set according to the test specimen and the testing conditions to reduce noise components while preserving features of the actual flaws.
2.3 Multi-Angle Integration.
The above steps are performed for each of the channels of phased array data. Each channel of the phased array data differs in the beam angle. Therefore, a sectorial area can be inspected by different angles simultaneously. Some inclined cracks may be parallel to the wave propagation path. In this case, the flaw may not interact with the wave effectively to generate strong reflected waves. Consequently, relying on a single-angle ultrasound beam is not reliable. The integration of multi-angle channels can greatly improve the probability of detection. In this study, a direct pixel-wise summation at the physical location of the object is performed using all participating channels. The identified flaw regions are mapped into the actual physical locations according to the sampling location, the angle of incidence, the sound speed, and the length of the propagation. The identified flaws from each of the channels can be integrated to form a fused result.
3 Experimental Validations
Experimental validation of the overall proposed flaw recognition method is performed. Phased array ultrasonic inspections are performed on metal material blocks to acquire data. The data are used to demonstrate the overall procedure and validate the effectiveness of the proposed method. The implementation of ultrasonic flaw recognition with multi-angle phased array data is made using Octave which is an opensource alternative to the matlab package. The Olympus OmniScan MX2 phased array system with a 5-MHz linear phased array probe and a position encoder is used to perform the inspection and acquire data, as shown in Fig. 4(a). Before carrying out phased array ultrasonic testing on the specimen, the Time Corrected Gain calibration is made to eliminate the effect of distance on the echo amplitude such that the same reflector at different depth can produce the same echo intensity. In the testing, the probe is configured to sweep over angles from −20 deg to +20 deg with a step size of 1 deg and is attached directly onto the scanning surface using couplant gels. The beam with a 0 deg angle of incidence is considered as the one that is perpendicular to the scanning surface. The scan pattern of the testing is shown in Fig. 4(b). The probe is moved linearly along the surface of the specimen at one cross section of the object. An internal location of the tested block can be interrogated by the beams with different angles at different locations along the scan path.
3.1 A 4340 Steel Block With Artificial Flaws.
A testing block made of 4340 steel with four artificial side drill holes (SDHs) is inspected using the above experimental setup. The dimension of the block and the locations of features are shown in Fig. 5. The length and height of the block are 185.00 mm and 76.10 mm, respectively. The diameter of all SDHs is 1.27 mm. The horizontal and vertical distances for two adjacent SDHs are 50.80 mm and 7.62 mm, respectively. The probe is moved on the top surface of the block along the direction perpendicular to the axis of SDHs, and the step length along the scan path is 0.1 mm. The sampling time interval is 10−8 s, and the sound propagation speed in the 4340 steel is 5920 m/s. The acquired data are processed and analyzed using the proposed method for flaw recognition.
The phased array data captured by the beam with different incident angles at different locations on one cross section are presented in the first column of Fig. 6. It can be observed that the dataset images exhibit various noise components. In the data pre-processing stage, an intensity threshold Ath = 0.3 and a 5-by-5 pixel-wise Gaussian low-pass filter with a standard deviation of σ = 2 are used to reduce the noise components. The pre-processed data are shown in the second column of Fig. 6. After processing the data by several shape-based filters, the data obtained by filtering operations are integrated, and then a threshold Bth is applied to extract defect edges. The value of Bth is 5% of the maximum value in the data. The disk structuring elements with a radius of R = 4 are adopted in morphological image processing. The final results of damage identification for the raw data are presented in the third column of Fig. 6 where the four SDHs are clearly seen in each of the channels. The data matrices with identified flaws obtained from a total of 21 channels are combined, and the result is shown in Fig. 7. The four artificial SDHs are clearly identified. The numerical operations are carried out on a computer with an Intel(R) Core(TM) i5-4200U CPU to quantify the efficiency of the method. The computational time in this case is about 25 s. An aluminum block is used to further validate the performance of the proposed method on natural flaws.
3.2 An Aluminum Alloy With Natural Flaws and Artificial Flaws.
An aluminum alloy specimen made of ER2319 aluminum welding wires is manufactured by Wire and Arc Additive Manufacturing (WAAM) with cold metal transfer arc welding process. The process-induced flaws, such as inclusions and porosity, can be produced with particular deposition parameters and path patterns. Five artificial defects (SDHs) are manufactured in the aluminum alloy specimen. The dimension of the specimen and the locations of SDHs are shown in Fig. 8. The diameters of the SDHs are 2.00 mm. The scan pattern is the same with the phased array ultrasonic experiment used for the steel block and the step length along the scan path is also 0.1 mm. The sampling time interval is 5 × 10−8 s, and the sound propagation speed in the specimen is 6190 m/s. The same ultrasonic inspection setup is used to acquire the phased array data.
The raw data captured by different beams on one cross section are presented in the first column of Fig. 9. In addition, the result of multi-angle raw data integration is presented in Fig. 10(a). It is difficult to identify the natural flaws directly from the raw data due to the noise components. An intensity threshold Ath = 0.1 is applied to reduce noise components in the data pre-processing stage, and other parameters are the same as those used in the case of the steel block. The second column of Fig. 9 presents the pre-processed data. Five artificial SDHs and a natural flaw are identified as shown in the third column of Fig. 9, where the natural flaw is shown in the rectangular box. It can be noticed that the natural flaw is not identified in some of the individual channels in Fig. 9. However, with the multi-angle integration, both the natural and artificial flaws are reliably identified, as shown in Fig. 10(b). The proposed method is shown to be effective for flaw recognition by comparing Figs. 10(a) and (10b). The processing time for identifying flaws of the aluminum block using this method is 28 s. The performance of this methodology is considered as acceptable for a post-processing method.
This study presents a systematical method for ultrasonic flaw recognition using phased array data. The method consists of a data pre-processing step to remove noise components, a flaw recognition step to identify potential flaws in each of the channels, and a multi-angle channel integration step to fuse the results from all channels. The overall procedure is demonstrated using realistic experimental data acquired on metal blocks with artificial flaws. It is further validated using an aluminum block with both artificial and natural flaws. Based on the current results, the following conclusions are drawn: (1) the proposed data pre-processing method can effectively eliminate speckle noise components, (2) the proposed shape-based filters, implemented as convolution operations, can efficiently reduce the geometric noise components, and (3) the multi-angle integration can improve the reliability of the flaw recognition. This method can be used for the recognition of internal flaws in metal materials. The parameters in each step of this method need to be carefully set according to the test specimen and the ultrasonic testing conditions.
The authors would like to thank the anonymous reviewers for their constructive comments.
Science Challenge Project (Grant No. TZ2018007).
National Natural Science Foundation of China (Grant No. 51975546; Funder ID: 10.13039/501100001809).
the filter size of the Gaussian low-pass filtering
the radius of the disk in the close operation of the ring-like features
an intensity threshold for speckle and electrical noise components
an overall threshold for flaw boundaries
the standard deviation of the Gaussian low-pass filtering