JP7823091B2 - A method for identifying carbon steel based on its emission spectrum using a steel plate cutting laser - Google Patents

Description

Patent Law Article 30, Paragraph 2 applies. Optronics Co., Ltd. Monthly OPTRONICS 2024.1 vol. 43 No. 505 pp. 133-137 Published January 10, 2024

The present invention relates to a method for determining carbon steel based on the emission spectrum of carbon steel using a steel plate cutting laser. In particular, it relates to a method for determining carbon steel based on emission spectrum analysis using a small spectrometer and AI data processing.

Laser cutting equipment is a tool used to cut carbon steel. Traditionally, lasers could only cut carbon steel plates approximately 20 mm thick. However, in recent years, with the increasing power output of multi-mode fiber lasers, laser cutting machines capable of cutting carbon steel plates 50 mm or thicker have emerged. However, if the wrong processing conditions are selected for the steel plate being processed and the wrong type of carbon steel is processed, the suboptimal processing conditions can not only result in a decrease in processing quality, but can also cause excessive spatter, which can damage parts of the processing equipment, including the laser head. As a preventative measure, it is effective to identify the type of carbon steel and measure the element ratio before processing.

For this reason, laser-induced breakdown spectroscopy (LIBS), an analytical method that irradiates the surface of the sample to be analyzed with high-energy pulsed laser light, converting the sample's atoms and ions into plasma, and then spectroscopically measures the element-specific light emitted when the sample's atoms return from an excited state to their ground state, is attracting attention.

The mainstream laser light source for LIBS is a nanosecond pulse laser (Non-Patent Document 1), (Non-Patent Document 2), but recently there have been reports of a double-pulse type that combines microsecond and nanosecond pulses (Non-Patent Document 3), as well as a UV wavelength (380 millisecond pulse laser) and artificial intelligence (AI) used to measure the iron element in aluminum alloys (Non-Patent Document 4), and the spectrometers used in this research are multi-channel and have extremely high wavelength resolution.

However, these methods have drawbacks, such as the high specifications of the laser light source or spectrometer, making them expensive to offer as an option for each laser processing machine in India and other developing countries where there is high demand for carbon steel.

Yamamoto, K. Y., et al., Laser-induced breakdown spectroscopy analysis of solids using a long-pulse (150 ns) Q-switched Nd: YAGlaser. Applied spectroscopy, 2005. 59 (9): p. 1082-1097. Sturm, V., et al., Carbon analysis of steel using compact spectrometer and passively Q-switched laser for laser-induced breakdown spectroscopy. Optics Express, 2019. 27 (25): p. 36855-36863. Cui, M., et al., Improved analysis of manganese in steel samples using collinear long-short double pulse laser-induced breakdown spectroscopy (LIBS). Applied spectroscopy, 2019. 73 (2): p. 152-162. Shuang, Q., et al., The Accuracy Improvement of Fe Element in Aluminum Alloy by Millisecond Laser Induced Breakdown Spectroscopy Under Spatial Confinement Combined With Support Vector Machine. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2022. 42 (2): p. 582-586.

The present invention aims to provide a method for determining carbon steel that can reliably identify the type of carbon steel before processing using a simple method without using a high-spec laser light source or spectrometer, and that can measure the element ratios specific to carbon steel.

As a result of extensive research, the inventors have come to the following invention.
[1] Provided is a method for distinguishing carbon steel, characterized in that in LIBS measurement of carbon steel using a near-infrared multimode fiber laser light source, the laser pulse width and oscillation signal of which are controlled by a pulse controller, an elemental analysis of the carbon steel is performed using a cutting laser head in a wavelength range of 200 nm to 600 nm, and the type of carbon steel is determined from the relative ratio of the peak heights using the spectral peaks of iron, chromium, and manganese.
[2] Provided is a method for distinguishing carbon steel according to [1], characterized in that the LIBS measurement generates a spectrometer trigger signal from a laser oscillation signal to a pulse generator, transmits the signal to the spectrometer, and the spectrometer acquires data synchronized with the laser oscillation and performs synchronous spectroscopy.
[3] In [1], there is provided a method for discriminating carbon steel, characterized in that AI using a convolutional neural network is used to determine the type of carbon steel, and peak heights included in the wavelength range of actual spectrum measurement data are used as learning data for training, thereby discriminating the type of carbon steel.

FIG. 1 is a diagram showing the configuration of a measuring instrument according to the present invention. FIG. 10 is a diagram showing the results of measuring the emission spectrum of S15C carbon steel obtained by setting the laser pulse width to 25 microseconds and synchronizing the spectrometer with a trigger signal. FIG. 1 shows part of the spectral data of each typical sample and a standard substance. FIG. 2 is a schematic diagram of a reading flow of processing according to the present invention. FIG. 2 is a schematic diagram of a learning flow of processing according to the present invention. FIG. 1 is a schematic diagram of a flow of determination processing according to the present invention. FIG. 1 is a conceptual diagram showing the structure of a CNN used in the present invention. FIG. 10 is a diagram illustrating an AI determination process in which database model generation is always performed as a separate process by cooperation between a CPU and a GPU. FIG. 10 is a diagram showing the results of the transition of accuracy (individual judgment) over time (number of days) when judgment is made from the spectral data of each carbon steel. FIG. 10 is a diagram showing the results of the transition of accuracy (by concentration) over time (number of days) for each of the carbon steel concentration ranges of low, medium, and high (low concentration: S15C and S20C, medium concentration: S35C, high concentration: S45C and S55C) when determining the accuracy from the spectral data of each carbon steel. FIG. 10 is a diagram showing the correlation between the individual determination accuracy rate and precision. FIG. 10 is a diagram showing the correlation between the accuracy rate of concentration-specific determination and precision.

Laser Source, Spectrometer, and Synchronization Figure 1 shows the configuration of the measuring device of the present invention. A 1000 W, 1080 nm near-infrared multimode fiber laser was equipped with a cutting laser head manufactured by HSG Laser, and the laser pulse width and oscillation signal were controlled by a pulse controller. Typically, in LIBS measurements of carbon steel, spectrometer measurements quantitatively analyze elemental concentrations using the elemental ratio of iron to carbon present in the 190 nm to 195 nm range. However, in this invention, a compact spectrometer with a wavelength range of approximately 200 nm to 600 nm was used. This wavelength range was adopted because measurements in the 190 nm to 200 nm wavelength range are difficult to distinguish from noise, even when the laser pulse width is set to 10 microseconds.

The spectrometer generates a spectrometer trigger signal from the laser oscillation signal and sends it to the pulse generator to acquire data synchronized with the laser oscillation. As such, synchronization between the laser oscillation and the spectrometer signal is key to measuring the emission spectrum. When setting up the pulse generator, the laser oscillation signal generated by the burst signal is input to the external trigger of the pulse generator, and the first laser oscillation signal is sent to the spectrometer with a pulse width of 10 to 300 microseconds determined by the spectrometer. The trigger signal to the spectrometer can be specified anywhere from one shot, and can be delayed by 0 to 1000 microseconds using the delay function.

The surface of each carbon steel sample (S15C, S20C, S35C, S45C, and S55C) purchased from Standard Test Piece Co., Ltd. was degreased with alcohol and then fixed to the stage. The spectrometer's input fiber cable was installed near the center of the laser head and the sample. Furthermore, during spectrum measurement, an appropriate amount of compressed air was released from the laser head in synchronization with the laser irradiation to prevent spatter from ablation from entering the laser head. The laser head was angled 30 to 90 degrees, preferably 45 to 90 degrees, relative to the sample surface, and the head was positioned 10 to 15 mm from the sample surface. The input fiber cable's angle range was 45 to 90 degrees relative to the sample surface, and it was positioned 5 to 50 mm from the sample surface.

Figure 2 shows an example of the emission spectrum of S15C carbon steel, obtained by setting the laser pulse width between 10 and 300 microseconds and synchronizing the spectrometer with the trigger signal. Compared to using a nanosecond laser, the overall spectral intensity was very low and proportional to the pulse width. However, by referring to known literature and the NIST database, peaks of iron (Fe) at 372 nm, 373 nm, 386 nm, and 527 nm, a peak of chromium (Cr) at 358 nm, and a peak of manganese (Mn) at 403 nm were observed. Furthermore, the spectral intensity tended to increase with wavelength toward 600 nm. This is presumably due to the effects of the emission of elements in the compressed air during laser irradiation and the emission due to laser heating of the sample surface. Furthermore, the emission spectra of other carbon steels (S15C, S20C, S35C, S45C, and S55C) were similarly measured.

A Python-based program for determining carbon steel using spectral data and AI was created, incorporating AI for numerical calculations and AI for neural networks (NN). The AI for numerical calculations was based on the LIBS simulation data for the carbon steel used, and the relative ratio of peaks contained in specific wavelength ranges of the actual measurement data was used to calculate the peak shape most similar to that of several types of carbon steel simulation data. The following is the website from which the "NIST LIBS Database" was derived.
https://physics.nist.gov/PhysRefData/ASD/LIBS/libs-form.html
For the AI numerical calculations, the spectral data was preprocessed and used as input for the program. The preprocessing involved reading spectral data saved in text or CSV format, and extracting the peak position, intensity, and wavelength of peaks within a specified wavelength range (arbitrarily set between 200 and 600 nm) using Python library functions. A specified number of peaks were extracted from the acquired peak group in descending order, and then normalized to the smallest peak among them.
The same procedure was performed on the simulation data, changing the wavelength range (arbitrarily set between 200-600 nm) and the number of peaks, and the simulation data and the spectral data were compared in descending order of intensity. The comparison was carried out by calculating the squared error of the normalized peak intensity between each simulation data and the spectral data, and the simulation data with the smallest squared error was presented as the discrimination result.

NN-AI, on the other hand, is an AI that uses a convolutional neural network (CNN), a type of NN. In this case, peaks contained in specific wavelength ranges of actual measurement data were used as learning data for training and material identification. Specifically, CNN is a deep learning algorithm that mimics the "neurons" in the human brain, and by learning from data, it is able to extract features from visible light spectrum data and distinguish between them. Deep learning was applied to material identification.

Here, the spectral data (wavelength range (set arbitrarily between 200-600 nm)) was preprocessed and used as input for CNN. Preprocessing involved reading spectral data saved in text or csv format, and extracting the peak position, intensity, and wavelength of peaks in the specified wavelength range using a function from the Python library. A specified number of peaks were extracted from the acquired peak group in descending order, and normalized to the smallest peak among them. Specifically, peaks were obtained by loading the spectral data of known metals into the library and used as a standard for comparison. The SciPy find_peaks library, which detects peaks, was used. Two peaks were searched for within each specified wavelength range; if there were three sets of specified wavelength ranges, a total of six peaks would be found. A data error was detected if there were fewer than two peaks.

The CNN was constructed using an open-source library and consists of an input layer, convolutional layer, pooling layer, and output layer. Normalized peak intensity is used as input, and the input layer is one-dimensional, corresponding to the number of specified peaks. The convolutional layer is one-dimensional, ranging from 32 to 256 (32, 64, 128, 256), with a kernel size of 4 to 25 (2*2, 3*3, 5*5), and the activation functions used are Rectified Linear Unit (Relu) Sigmoid, Tanh, and Softmax. The output layer is one-dimensional, corresponding to the number of types of carbon steel.

Based on the spectral data of each representative sample shown in Figure 3, the aforementioned data reading, preprocessing, and database construction calculations are performed. While building the database, the spectral data of the sample to be identified is collated through a judgment calculation process to determine the classification. Figures 4 to 6 show schematic diagrams of the reading, learning, and judgment processes of the present invention. Specifically, deep learning algorithms used in image recognition are applied to the field of image recognition because they can extract features from input images and distinguish them by learning data. Here, spectral images are recognized, and the CNN structure goes through a series of processes from the input layer, convolutional layer, and pooling layer to the activation layer, fully connected layer, and output layer, with each layer gradually extracting more abstract features from the input image data. This process allows the deep learning model to learn patterns in complex images and extract important information for classification and recognition. Figure 7 shows a conceptual diagram of the CNN structure used in this invention. From left to right, the input layer (InputLayer∈R^16), two hidden layers (HiddenLayer∈R^12, HiddenLayer∈R^10), and output layer (OutputLayer∈R^1) are shown. The mechanism for image recognition and judgment is as follows: image data is first supplied to the input layer, and features are extracted through the convolutional layer and pooling layer. Next, a nonlinear transformation is performed by the activation layer, and these features are integrated through the fully connected layer. Finally, predictions for classification and recognition are made in the output layer. Through this series of processes, deep learning models gain the ability to recognize and judge images with high accuracy for specific tasks.

Table 1 shows the results of a comparison of the accuracy of numerical calculation AI and NN-AI using spectral data for each carbon steel. The accuracy of the numerical calculation AI was approximately 20%, while the accuracy of NN-AI was approximately four times higher at 84%. If we simply consider accuracy to be equal to the accuracy rate, the probability that an average person would correctly identify the five types of carbon steel is 20%, so the numerical calculation AI can be considered to be closer to AI modeled after an average person. On the other hand, this suggests that NN-AI is likely to be AI at the level of a professional or expert with experience in laser processing. Therefore, improvements to NN-AI and changes and trends in accuracy were examined, leading to the present invention.

Here, the accuracy of the concentration types and the accuracy of individual types were assessed. The accuracy rate is the number of answers judged correct by the user divided by the number of data points entered into the discrimination AI, and accuracy can be expressed as the number of correct answers divided by the data points used for testing. The data input to the training AI was set to n, the data used for training to 0.8*n, and the data used for testing to 0.2*n. Each training model was calculated using the data used for training (data used for training to 0.8*n). The accuracy of the models generated by each training calculation was then calculated using the value calculated using the data used for testing (data used for testing to 0.2*n) as the number of correct answers during testing divided by the data points used for testing. This model was a neural network model generated by the training AI through training. The coefficients n of 0.8 and 0.2 were determined based on empirical data in this field. To verify the accuracy of a model, data is typically divided into data used for training (training data) and data not used (validation data) in a ratio of 7:3 or 8:2. For example, we used "History of Artificial Intelligence and Deep Learning" by Sony (registered trademark). The website is https://www.comm.tcu.ac.jp/mds-center/Resources/SNCs/PDFs/07-0_WhatIsDeepLearning%C2%A9SNC.pdf.

The original NN-AI system saved each piece of judged spectral data in a database, modeled that database, and used it to judge carbon steel from the next piece of spectral data. As a result, when the AI had to build a model from the database and then make a judgment each time, it took more than a minute to judge 20 or more pieces of data. Therefore, as shown in Figure 8, by linking the CPU and GPU to constantly run database model generation as a separate process, the AI judgment process was shortened to less than 10 seconds.

Figure 9 shows the accuracy over time (number of days) for the judgment accuracy when judging each carbon steel type from spectral data (individual judgment) and the judgment (concentration judgment) divided into low, medium, and high carbon steel concentrations (low concentration: S15C and S20C, medium concentration: S35C, high concentration: S45C and S55C). Initially, each judgment showed an accuracy of over 80%, but for the individual judgment (Figure 9), the accuracy dropped sharply to below 60% on the eighth day, and for the low, medium, and high concentration judgment (Figure 10), the accuracy also showed a slight drop on the fourth day. This change was due to a significant change in the fixed position of the spectrometer's incident fiber, which caused a change in the spectral data and led the NN-AI to recognize it as different data from previous data. Therefore, additional data was acquired and the database was rebuilt with the incident fiber angle fixed, and the accuracy of the individual judgment recovered to the 80% range. If the measurement conditions (sample position, laser position, and device settings) are fixed, for example, data can be measured for approximately six hours or more, and the AI can be trained using the large amount of data obtained, resulting in an accuracy rate of over 80% if the accuracy is good. In this case, it was necessary to ensure stability during long-term measurements by preventing wear inside the laser. Furthermore, even if accuracy and accuracy rates decrease for a certain period of time due to fluctuations in the positional relationship between the laser head and sample, or the positional relationship with the incident fiber, experience shows that accuracy tends to recover within 50 data points at the earliest, and around 100 data points on average, and the accuracy rate also recovers.

Figures 11 and 12 show the correlation between the accuracy rate and precision. The graphs are plotted for each number of judgment cycles. The number of judgment cycles refers to the number of repeated calculations performed to minimize prediction error. The X-axis represents the number of judgment cycles, the Y1-axis represents the accuracy rate, and the Y2-axis represents accuracy. For individual judgments, accuracy remained stable at 86-88%, while the accuracy rate initially remained below 50% and increased to 80% by the fourth cycle. In other words, if the accuracy rate is 80%, it can be assumed that there will be almost no judgment errors if 6-9 spectral data points are acquired during measurement, as with commercially available portable LIBS devices, and judgments are made by majority vote. On the other hand, for concentration-specific judgments, even when accuracy exceeds 90%, the accuracy rate fluctuates irregularly, ranging from 67-85%. This is due to the frequent misjudgments of S15C and S35C. Ultimately, the accuracy rate remained stable at 86% thanks to the "supervised learning" function, demonstrating that concentration-specific discrimination is an AI with fewer errors than individual judgments. The preferred wavelength range for analysis is 350 to 550 nm, and it is particularly preferable to set 400 to 410 nm as the AI judgment standard, as this can increase the accuracy rate of judgment.

Using a steel plate cutting laser and a small spectrometer, it became possible to measure the emission spectrum of each carbon steel and use NN-AI to identify the carbon steel. Although the emission spectrum was weaker in intensity than measurements using existing nanosecond lasers, it was possible to obtain spectra showing the elemental peaks of iron, chromium, and manganese in carbon steel without using a vacuum spectrometer, argon gas, or nitrogen gas. The accuracy of the developed NN-AI tended to vary depending on the hardware environment, but by improving the AI program and data acquisition, it was possible to process the AI judgment within 10 seconds and achieve an accuracy rate of over 80%.

Claims (2)

In the LIBS measurement of carbon steel, a near-infrared multimode fiber laser light source attached to a cutting laser head is used, and the laser pulse width and oscillation signal are controlled by a pulse controller. A vacuum spectrometer, argon gas, or nitrogen gas is not used, and the laser pulse width is set to 10 to 300 microseconds in the wavelength range of 200 nm to 600 nm. In the elemental analysis of carbon steel, the emission spectrum of the carbon element is not used, but the peaks of the emission spectra of the elements iron, chromium, and manganese are used to perform judgment calculation processing and distinguish the type of carbon steel.
A method for distinguishing carbon steel, characterized in that a spectrometer trigger signal is generated from a laser oscillation signal and sent to a pulse generator, which is then sent to the spectrometer, the spectrometer acquiring data synchronized with the laser oscillation and spectroscopy in synchronization, reading and transferring data based on the spectral data of a typical sample, performing calculations to reconstruct a learning database related to the elements of carbon steel using the results of elemental analysis of the carbon steel, and while constructing the database, performing a judgment calculation process to compare the spectral data of the sample to be judged, and using AI using a convolutional neural network to judge the type of carbon steel, training the peak heights included in the wavelength range of actual spectral measurement data as learning data, thereby distinguishing between types of carbon steel S15C to S55C. 2. The method for distinguishing carbon steel according to claim 1, wherein the determination calculation process for collating the spectral data of the sample to be determined is performed in parallel, with the CPU processing data transmission and integration between the three processes of spectral data acquisition, database construction calculation, and material determination calculation, and the GPU processing the database construction and material determination calculations, so that database model generation is always performed as a separate process by cooperation between the CPU and GPU, and the type of carbon steel is determined by using AI using a convolutional neural network to train the peak heights included in the wavelength range of actual spectral measurement data as learning data, and distinguishing the type of carbon steel from S15C to S55C.