JP7739986B2 - Ultrasonic flaw detection method for round bar materials - Google Patents

Description

The present invention relates to an ultrasonic flaw detection method for round bar material, and in particular to an ultrasonic flaw detection method that can accurately distinguish and detect surface flaws from surface flaws that occur inside the round bar material just below the surface.

When detecting flaws near the surface of a round bar material, the shear waves of the ultrasonic beam Ub for flaw detection are used, with the refraction angle (sector scan angle) set to approximately 45 degrees, as shown in Figure 6. However, with this method, the flaw echo signals (Figures 7(1) and 7(2)) from the surface flaw M1 (Figure 6(1)) that opens on the surface of the round bar material M and the surface flaw M2 (Figure 6(2)) that occurs just below the surface of the round bar material M can appear at roughly the same magnitude and at the same time, making it often difficult to distinguish between the two.

For example, Patent Document 1 discloses a flaw detection method in which different sector scan angles are set and the presence of a surface flaw or subsurface flaw is determined when the magnitude of the flaw echo signal obtained at each sector scan angle exceeds a predetermined threshold.

Patent Publication No. 2012-225887

However, with the conventional method described above, it was necessary to change the sector scan angle and repeat the same inspection, which took time, making it difficult to quickly inspect round bar material flowing through the line.

The present invention aims to solve these problems by providing an ultrasonic flaw detection method for round bar material that can quickly and reliably distinguish between surface flaws and subsurface flaws.

In order to achieve the above object, in the first invention, an ultrasonic beam for flaw detection (Ub) is transmitted toward a round bar material (M), and the reflected ultrasonic waves reflected and returned by the surface flaws (M1) of the round bar material (M) or the surface flaws (M2) occurring just below the surface of the round bar material (M) are received. The flaw echo signal (Sk) in the received signal (Sa) is detected before and after its peak value (a) at a predetermined angle determined in advance for the neural network (3) used, taking into account the learning error and learning time. The flaw echo signal (Sk) is extracted over a time period corresponding to the wavelength conversion length of the flaw detection ultrasonic wave (Ub), and the signal data of the extracted flaw echo signal (Sk) is given to the neural network (3) for learning, and the signal data of the flaw echo signal (Sk) in the new received signal extracted over a time period corresponding to the predetermined wavelength conversion length is given to the trained neural network (3) to determine whether the flaw corresponding to the flaw echo signal (Sk) is a surface flaw (M1) or a subsurface flaw (M2).

In the first invention, the use of a neural network eliminates the need to change the sector scan angle and repeat the same flaw detection as in the past, allowing for rapid flaw detection that distinguishes between surface flaws and subsurface flaws. Furthermore, the signal data for the neural network is flaw echo signals extracted over a time corresponding to a predetermined wavelength conversion length, which is determined in advance based on a balance between learning error and learning time. This allows the neural network to be trained efficiently with the required learning accuracy. In addition, signal data for flaw echo signals extracted over a time corresponding to the predetermined wavelength conversion length from a newly received signal is provided to the trained neural network to determine whether the flaw corresponding to the flaw echo signal is a surface flaw or subsurface flaw. This shortens the neural network's inference time and makes it possible to handle fast line speeds for conveying round bar material.

In this second invention, the signal data used is two-dimensional waveform image data of the flaw echo signal extracted over a time period corresponding to the specified wavelength conversion length.

In this second invention, by using two-dimensional waveform image data as signal data, the accuracy of distinguishing between surface defects and subsurface defects is improved compared to using one-dimensional waveform data.

The symbols in parentheses above indicate, for reference purposes, the correspondence with the specific means described in the embodiments below.

As described above, the ultrasonic flaw detection method for round bar material of the present invention makes it possible to quickly and reliably distinguish between surface defects and subsurface defects on round bar material traveling through a production line.

1 is a diagram showing the configuration of an ultrasonic flaw detection device to which the method of the present invention is applied. 3A and 3B are waveform diagrams of received signals obtained by a probe and enlarged views of parts thereof; FIG. 2 is a conceptual diagram showing an example of a data string obtained by digitizing a received signal. FIG. 1 is a diagram illustrating a schematic configuration of a convolutional neural network. 10 is a graph showing the relationship between the average learning error and the average learning time with respect to the wavelength-equivalent length. FIG. 10 is a conceptual cross-sectional view of an angle beam flaw detection method according to a conventional example. FIG. 10 is a waveform diagram of a flaw echo signal showing a conventional example.

Please note that the embodiment described below is merely an example, and various design improvements made by those skilled in the art that do not deviate from the gist of the present invention are also included within the scope of the present invention.

Figure 1 shows the configuration of an apparatus for implementing the ultrasonic flaw detection method of the present invention. In Figure 1, a probe 1 having a single wideband ultrasonic vibrator is provided facing the outer peripheral surface of a metal round bar material M. In this embodiment, the probe 1 generates a shear wave ultrasonic beam Ub, which is an ultrasonic beam for angle beam flaw detection, with a sector scan angle of approximately 45 degrees and reaches the surface of the round bar material M and the area including the surface layer immediately below it. This ultrasonic beam Ub is typically a pulse wave with a length of 1.5 to 3 wavelengths. The ultrasonic beam Ub is reflected by each part of the round bar material M, including the surface flaw M1 or surface flaw M2, and the reflected ultrasonic waves generated are received by the probe 1 and input to the discrimination device 2 as a received signal Sa.

Figure 2 shows an example of a received signal Sa. The received signal Sa includes a flaw echo signal Sk reflected by a surface flaw M1 or a subsurface flaw M2, which is a signal within the time domain enclosed by a square in the figure. The received signal Sa is input to an AD conversion circuit (not shown) provided within the discrimination device 2. The AD conversion circuit has an input range that can cover the maximum positive and negative amplitude range of the received signal Sa, and the received signal Sa is converted by the AD conversion circuit into a digital signal Sd (Figure 3) consisting of a sequence of numbers (one-dimensional data sequence) corresponding to the amplitude at each specified sampling time. Note that Figure 3 is an example of a digital signal Sd when an 8-bit AD conversion circuit is used.

In the discrimination device 2, data corresponding to peak value a (Figure 2) of the flaw echo signal Sk is detected from the data sequence of each digital signal Sd corresponding to the received signal Sa, and a predetermined number of pieces of data (learning data) are extracted around the peak data, as described in detail below. This learning data corresponds to the flaw echo signal Sk within the time region enclosed by the square in Figure 2.

The training data is provided to a convolutional neural network (CNN) 3 of known configuration configured within the discrimination device 2. As shown in Figure 4, the CNN 3 has an appropriate number of convolutional layers 31 (e.g., nine layers), pooling layers 32, and a fully connected layer 33, and learns by providing multiple training data based on the flaw echo signals Sk in each received signal Sa obtained by irradiating a round bar M having multiple known surface flaws M1 and subsurface flaws M2 with an ultrasonic beam Ub as described above.

As mentioned above, the training data is a set number of data points extracted before and after the peak data. However, the inventors' experiments have shown that increasing the number of training data points does not necessarily improve the training accuracy. This is explained in Figure 5.

Figure 5 is a graph showing the wavelength (λ)-converted length of the extracted flaw echo signal Sk on the horizontal axis, and the average learning error (Crossentropy loss ave.: line x) and average learning time (learning time ave.: line y) on the vertical axis. Each boxplot in the figure was obtained by creating 10 CNN3 models by, for example, shuffling the order of multiple learning data sets, and measuring the learning error for each model. The wavelength (λ)-converted length (equivalent to the number of data sets in the learning data) of the extracted flaw echo signal Sk ranges from 2.0 wavelengths to 9.0 wavelengths in 0.5 wavelength increments.

As is clear from Figure 5, the average learning time (line y) increases as the wavelength (λ) converted length increases, i.e., the number of pieces of training data increases. However, the average learning error (line x) decreases as the wavelength (λ) converted length increases when the wavelength (λ) converted length is 5.0λ or less. However, when the wavelength (λ) converted length is 5.0λ or more, the average learning error does not change significantly even if the wavelength (λ) converted length increases further, i.e., even if the number of pieces of training data increases. Therefore, in this example, balancing learning time and learning error, the optimal condition for shortening learning time and minimizing learning error is to train CNN3 with a number of pieces of training data corresponding to flaw echo signals extending 5.0λ around the peak value of the flaw echo signal Sk, with the peak value at the center. However, if minimizing learning error is a priority, it may be possible to choose a longer wavelength (λ) converted length, i.e., a larger number of pieces of training data, even if it increases the average learning time.

By providing the trained CNN3, which has been trained through this process, with a number of flaw echo signals equivalent to 5.0λ, the same number as the training data, the probability of determining whether the flaw is a surface flaw or a subsurface flaw can be sufficiently increased, and the determination threshold can be raised to accurately identify whether the flaw corresponding to the flaw echo signal is a surface flaw or a subsurface flaw.

In the above embodiment, it was assumed that the optimal number of pieces of training data corresponds to a signal with a length of 5.0λ. However, if the number of layers, such as the convolutional layer 31, of the CNN 3 used changes, the graph in Figure 5 will change, so the optimal number of pieces of training data corresponds to a signal with a length of 5.0λ is not always the case.

Furthermore, in the above embodiment, a one-dimensional data string of the flaw echo signal Sk was used as the learning data, but the waveform of the flaw echo signal Sk may also be used as two-dimensional image data.

1...ultrasonic transducer, 2...discrimination device, 3...neural network, M...round bar material, M1...surface flaw, M2...surface flaw, Sa...received signal, Sk...flaw echo signal, Ub...ultrasonic waves for angle beam flaw detection.

Claims (2)

An ultrasonic flaw detection method for round bar material, comprising: transmitting ultrasonic waves for angle beam inspection toward the round bar material; receiving reflected ultrasonic waves reflected from surface flaws or surface flaws occurring inside the round bar material just below the surface; extracting flaw echo signals from the received signal around their peak value for a time period corresponding to a predetermined wavelength equivalent length of the ultrasonic waves for angle beam inspection, which is determined in advance based on the learning error and learning time of the neural network used; providing the signal data of the extracted flaw echo signals to the neural network for learning; and providing the signal data of flaw echo signals from a new received signal extracted for a time period corresponding to the predetermined wavelength equivalent length to the trained neural network to determine whether the flaw corresponding to the flaw echo signal is a surface flaw or a subsurface flaw. The ultrasonic flaw detection method for round bar material described in claim 1 uses, as the signal data, two-dimensional waveform image data of the flaw echo signal extracted at a time corresponding to the specified wavelength conversion length.