Abstract
It is still a challenging problem to detect small defects for eddy current array probes, which requires the probes to possess excellent sensitivity, as well as high spatial resolution. This paper presents a novel high-resolution magnetic field imaging probe with two rows of tunneling magnetoresistance (TMR) array sensors. The bare die sensors are integrated on a printed circuit board by golden wire bonding technology. The two rows of sensors are placed staggered with each other. The data of the two arrays are merged into a matrix, in which way the image pixel pitch is increased to 0.25 mm. The probe employs a differential scheme to suppress the noise, so as to detect the weak signal of small defects. To highlight the weak defect indications, feature extraction and segmentation algorithms are developed. The experimental results confirm that the proposed method can inspect a small defect with dimensions 1 mm (length) × 0.1 mm (width) × 0.1 mm (depth) on a stainless-steel sample.
Keywords
Introduction
The earlier a defect is detected, the better for maintaining the health and safety of the structure. Therefore, it is critical to detect small defects [1]. Eddy current testing (ECT) technology is sensitive to surface and near-surface defects. In addition, ECT has the characteristic of non-contact, fast speed and low surface treatment requirements. ECT is widely utilized in industry for metal structure inspection. Conventionally, induction coils are employed to pick-up the magnetic signal in ECT. To further increase the image spatial resolution and expand the operating frequency range of an ECT probe, magnetoresistance sensors, such as giant magnetoresistance (GMR) and tunneling magnetoresistance sensors, were developed to measure the magnetic field of the eddy current [2]. Array probes with multiple sensors can generate a C-scan image in a linear scan, which can image a large area efficiently [3,4].
Due to a small defect usually having weak signal in a small area, to effectively detect the signal, it not only requires the inspection system to have excellent resolution and sensitivity, but also needs advanced image processing algorithms to overcome the problem of the low signal-to-noise ratio of the original image. In previous studies, the moving average filter, spline interpolation and symmetry corrections were utilized to suppress noise, and then extracted features for supervised machine learning [5]. By combining low-rank information into the sparse matrix, small defects were recognized more effectively, the weak signal of a defect was efficiently extracted from noise and background [6]. Fisher linear discrimination was used to realize automatic feature extraction and selection for multilayer structural defect classification [7]. These methods increase the detectability for small defects to some extent, but the spatial resolution of the probe and the interference of the noise still hinder the accurate identification of small defects.
This paper presents a novel high-resolution magnetic field imaging probe with two rows of TMR sensors. The distance between each two adjacent sensors in an array is 0.5 mm. By interpolating the data of the two arrays, the image resolution is increased to 0.25 mm. A feature extraction method based on Gray-Level Co-occurrence Matrice is utilized to suppress the noise and recognize the indication of a small defect from the image.
The design of the tunneling magnetoresistance sensors array.
The photographs of the defects #1–#5.
The dimensions of defects on the samples
The raw experimental images of the defects #1 and #2.
The design of the probe is shown in Fig. 1(a). The probe contains two rows of TMR sensors, with 32 sensors in each row. The length, width and height of the TMR sensor are 0.45 mm, 0.45 mm and 0.2 mm, respectively. More details of the sensors can be found in [2]. The distance between two sensors’ centers in a row is 0.5 mm. The two rows are misaligned 0.25 mm along the horizontal direction, and the distance between the two rows is 1.6 mm. The sensitive axis of the TMR sensors are perpendicular to the excitation coil plane. The sensors are wire bonded on the bottom of a printed circuit board (PCB). A rectangular excitation coil is attached to the sensor arrays, which providing an almost uniform magnetic field in the sensor area. The shape of the excitation coil is similar as the coil shown in [2]. The blue dashed lines represent the excitation current of the coil. Here the arrows indicate the current direction of the AC excitation at a certain moment. The number of turns of the excitation coil is 30, and the dimensions of the coils are 60 mm (length) × 35 mm (width). The probe works in differential model, namely the difference between two sensors of the two rows is recorded as the output signal. Considering the combinations of the two sensors, the probe has two measurement modes, namely: Mode A and Mode B, as shown in Fig. 1(b). The red and blue boxes in Fig. 1(b) indicate sensors whose signals are differentiated. The data of the two modes are merged into a matrix during the signal processing, in which way the image spatial resolution is increased to 0.25 mm.
The raw experimental images of the defects #3–#5.
The process of algorithm.
(a) The images of gray-level co-occurrence matrix for defects #3–#5, (b) is the image after threshold filter of (a), (c) is the image after regional filter of (b).
The processed image of defects #1–#5.
The sensors are selected by a multiplexer circuit and then connected to a bridge circuit to transfer the changes of the magnetic field to a voltage signal. The conversion factor of the sensor and circuit from the magnetic field to the output voltage is 1 V/Gs. The output signal is connected to and digitalized by a data acquisition (DAQ) system (NI 5753). The waveform of the excitation voltage is digitalized simultaneously as a reference signal. Then the in-phase (real) and quadrature (imaginary) components of the 100 kHz signal are calculated by a digital lock-in amplifier in the NI 5753. The experiments are controlled by a three-axis gantry scanning device with scanning speed 10 mm/s. Two samples with machined defects are inspected by the probe, namely: (i) a 304 stainless steel sample with a smooth flat butt weld containing machined defects on the weld (defect #1 and #2) and (ii) a flat plate 430 stainless steel with relative permeability about 800, which contains three machined defects (#3–#5). The photographs of the defects are presented in Fig. 2, and the dimensions of the defects are specified in Table 1.
In the experiment, the frequency of the excitation voltage was set to be 100 kHz and the current amplitude was 0.14 A. The raw experimental images are presented in Fig. 3 and Fig. 4, where the “Re” and “Im” represent the real- and imaginary-components of the signals, respectively.
It is found that a small defect has weak signal in a small area. To highlight the defect indication, it is necessary to process the images to reduce the inevitable experiment noise. The experimental images indicate that the defect images have certain texture characteristics. Therefore, a grayscale co-occurrence matrix (GLCM) is developed to extract the texture features of the image, and the noise is reduced by fusing the areas retaining texture features. Grey-level co-occurrence matrix is a widely used descriptor of texture features that can be extracted from grey-level images, and reflects the combined information about the gray-level distribution of the image with regard to direction [8]. The flowchart of the image processing algorithm is shown in Fig. 5.
Before the texture feature extraction, the median filtering is applied to the raw image to effectively suppress the pepper noise. The gray-level co-occurrence matrix is calculated by taking the offset parameters in different directions. The formulation is shown in (1), where i is the pixel value of position (x, y), j is the pixel value of position (x + Δx, y + Δy) in the image S, d and 𝜃 are the distance and angle between the two pixels. L is the maximum grayscale of the raw image, N x and N y are the number of rows and columns of the image.
The gray-level co-occurrence is calculated based on the probability P (i, j, Δ, 𝜃) of pixels with (x, y) and (x + Δx, y + Δy). In following calculation, d = 1, and four angles (0° 45° 90° 135°) are calculated. The algorithm employs the average of the gray-level co-occurrence matrix in four directions and unifies its range to 0–9. The contrast is calculated as the feature value. Finally, the image with the largest feature value in the two differential modes is retained, as shown in (2).
The images after regional filter and the raw image are multiplied to obtain the final result that preserves the signal in the defect region. From the processed images shown in Fig. 7, it is seen that the defect signal of the smoothed weld and the flat are successfully identified. Although the defects are successfully detected, the algorithm can be more extensively optimized and tested. As the algorithm bases on the texture feature of the defect signal, when the signal of a defect connects with the background noise and has no obvious shape characteristics, it is difficult to separate the signal from the background noise. Considering the fact that noise is random, the fusion of multi-frequency images may be studied to increase the reliability of the method.
In this paper, a high-resolution TMR array probe for the inspection of small defects is developed and tested. The probe utilizes a differential setting to suppress the common model noise and can obtain magnetic field images with high spatial resolution. A GLCM-based imaging processing algorithm is developed to extract the texture information of the defects in the images, which can effectively detect small defect on the stainless-steel samples. The texture information of the image is calculated from different orientations by using the gray co-occurrence matrix. Then a threshold filter is applied, by setting the gray value below the threshold to zero to obtain a binary image. Eventually, the filtered image mainly retains the indication of the small defects, and the noise is filtered out. The experimental results show that the algorithm can successfully detect the small defect with dimensions 3 mm × 0.1 mm × 0.1 mm on the smoothed weld sample and a defect with dimensions 1 mm × 0.1 mm × 0.1 mm on the surface of the 304 stainless-steel plate. It seems that the method proposed in this paper has super sensitivity and resolution for a small defect.