Liquid steel component online detection method based on LIBS technology and deep learning

Abstract

The invention discloses a molten steel component online detection method based on an LIBS technology and deep learning, and relates to the technical field of component online detection in the casting metallurgy process. The method comprises the following steps: acquiring one-dimensional spectral data of target molten steel through an LIBS-based light path excitation collection system, and preprocessing the one-dimensional spectral data to obtain a corresponding two-dimensional grey-scale map; inputting the two-dimensional grey-scale map into a trained molten steel component detection model, and outputting a multi-element content predicted value of the target molten steel; the molten steel component detection model is constructed based on an improved ResNet50 network. According to the invention, the light path excitation collection system is arranged based on the LIBS technology, and the light path excitation collection system is combined with the improved ResNet50 network, so that the detection efficiency and the detection precision of the molten steel components are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of online composition detection technology in the casting metallurgical process, and more specifically to an online steel composition detection method based on LIBS technology and deep learning. Background Technology

[0002] Rapid detection of steel composition is crucial for quality control in steelmaking processes. Traditional methods rely on taking liquid samples from molten steel with a sampling spoon, molding and cooling them into solids in a mold, then cutting and grinding them into test bars. Finally, the surface of the solid sample is excited by an electric arc or spark to generate elemental characteristic spectral feedback data. However, this process is time-consuming and cannot provide real-time feedback.

[0003] Laser-induced breakdown spectroscopy (LIBS) technology can generate plasma on the surface of molten steel by laser excitation, and analyze characteristic spectra to achieve qualitative and quantitative elemental detection. It offers advantages such as no need for pre-sample preparation, rapid online in-situ analysis, and simultaneous multi-element detection, making it suitable for on-site testing requirements in steel smelting. Most optical path designs use dichroic mirrors to separate the laser and signal light, utilizing coaxial transmission of the laser and signal light. However, this approach is severely affected by interference from multi-stage reflection paths, angular deviations, and limitations imposed by coating materials, all of which negatively impact the received spectral signal.

[0004] Therefore, how to improve the detection efficiency and accuracy of molten steel composition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, the present invention provides an online detection method for molten steel composition based on LIBS technology and deep learning, so as to at least solve some of the technical problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] This invention provides an online method for detecting the composition of molten steel based on LIBS technology and deep learning, comprising the following steps:

[0008] One-dimensional spectral data of the target molten steel is acquired through a LIBS-based optical path excitation and collection system, and the one-dimensional spectral data is preprocessed to obtain the corresponding two-dimensional grayscale image.

[0009] The two-dimensional grayscale image is input into the trained molten steel composition detection model, and the predicted value of the multi-element content of the target molten steel is output.

[0010] The steel composition detection model is built based on an improved ResNet50 network.

[0011] Furthermore, the LIBS-based optical path excitation and collection system includes a laser excitation optical path, a signal collection optical path, and a signal transmission and processing optical path;

[0012] In the laser excitation optical path, a laser, a focusing lens, and a reflector are arranged along the same optical axis. The laser emits a laser beam, which is focused by the focusing lens and reflected and refracted by the reflector placed at 45°, so that the laser beam is focused on the target molten steel, exciting the formation of plasma and emitting signal light.

[0013] In the signal collection optical path, a concave reflector and a convex reflector coaxially arranged with the concave reflector are provided; the concave reflector has a central hole; the signal light is reflected by the concave reflector to the convex reflector, then reflected and focused by the convex reflector, and coupled to the fiber optic receiver through the central hole;

[0014] In the signal transmission and processing optical path, the fiber optic receiver inputs the signal light containing the target molten steel composition information into the spectrometer to obtain one-dimensional spectral data, and the spectrometer preprocesses the one-dimensional spectral data to obtain a two-dimensional grayscale image corresponding to the target molten steel.

[0015] Furthermore, in the laser excitation optical path, the laser beam's focal point is positioned 0.5-1 mm below the surface of the target molten steel by means of the reflector.

[0016] Furthermore, the reflective coatings of the concave and convex mirrors cover a wide spectral range of 190-1100 nm.

[0017] Furthermore, the preprocessing includes:

[0018] Dynamic wavelength calibration is achieved by real-time correction of the spectrometer offset using Hg-Ar lamp standard spectral lines.

[0019] Adaptive noise reduction is achieved by combining wavelet transform and moving average filtering;

[0020] The one-dimensional spectral data after dynamic wavelength calibration and adaptive noise reduction is converted into a two-dimensional grayscale image.

[0021] Furthermore, the improved ResNet50 network includes: an input layer, an initial convolutional layer, a spectral feature enhancement module, a multi-scale residual block, a spectral pyramid pooling layer, and a fully connected layer;

[0022] The two-dimensional grayscale image is input into the trained molten steel composition detection model through the input layer;

[0023] The initial feature map is obtained by extracting features from the two-dimensional grayscale image through the initial convolutional layer; the initial convolutional layer uses a 1×7 convolutional kernel.

[0024] The initial feature map is enhanced by the spectral feature enhancement module to obtain spectral enhanced features;

[0025] The spectral enhancement features are input to the multi-scale residual block after passing through a 3×3 convolutional layer and a max pooling layer. After processing by the multi-scale residual block, multi-scale residual features are generated.

[0026] After processing the multi-scale residual features through the spectral pyramid pooling layer, the output is then passed through a fully connected layer.

[0027] Furthermore, the spectral feature enhancement module includes two branches and an SE attention layer; wherein:

[0028] The first branch uses a 1×7 convolution to extract the global spectral features of the initial feature map;

[0029] The second branch uses a 7×1 dilated convolution to extract local spectral features from the initial feature map;

[0030] The global spectral features and the local spectral features are concatenated and then fused through the SE attention layer to obtain spectral enhancement features.

[0031] Furthermore, the multi-scale residual block comprises three consecutive 3x3 convolutional layers, wherein the first 3x3 convolutional layer has 2 input channels and a kernel size of 3.

[0032] Furthermore, the spectral pyramid pooling layer includes three parallel pooling layers, specifically a 1×1 global average pooling layer, a 3×3 max pooling layer, and a 6×6 max pooling layer;

[0033] The 1×1 global average pooling layer is used to capture the overall intensity distribution of the spectrum;

[0034] The 3×3 max pooling layer is used to extract local spectral peak features of adjacent overlapping peaks;

[0035] The 6×6 max-pooling layer is used to focus on a wide-bandgap region;

[0036] The outputs of the 1×1 global average pooling layer, the 3×3 max pooling layer, and the 6×6 max pooling layer are concatenated and then input into the fully connected layer.

[0037] Furthermore, the fully connected layer comprises multiple parallel sub-networks, each outputting a predicted content value for one element; and the loss function of the fully connected layer is expressed as:

[0038]

[0039] Where L represents the loss function; N represents the total number of elements; i represents the element index; ωi represents the weight of the i-th element; and yi represents the true value of the i-th element. This represents the predicted value of the i-th element.

[0040] As can be seen from the above technical solution, compared with the prior art, the present invention discloses an online detection method for molten steel composition based on LIBS technology and deep learning, which has the following beneficial effects:

[0041] This invention establishes an optical path excitation and collection system based on LIBS technology, and combines the optical path excitation and collection system with an improved ResNet50 network, which effectively improves the detection efficiency and accuracy of molten steel composition.

[0042] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0044] Figure 1 This is a schematic diagram of the online steel composition detection method based on LIBS technology and deep learning, provided for an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram of the structure of a LIBS-based optical path excitation and collection system provided in an embodiment of the present invention.

[0046] Figure 3 This is a schematic diagram of the signal acquisition optical path provided in an embodiment of the present invention.

[0047] Figure 4 A schematic diagram of the improved ResNet50 network structure provided in an embodiment of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] This invention discloses an online method for detecting the composition of molten steel based on LIBS technology and deep learning. (See also...) Figure 1 As shown, the specific steps include the following:

[0050] One-dimensional spectral data of the target molten steel was acquired using a LIBS-based optical path excitation and collection system, and the one-dimensional spectral data was preprocessed to obtain the corresponding two-dimensional grayscale image.

[0051] The two-dimensional grayscale image is input into the trained molten steel composition detection model, which outputs the predicted values ​​of the multi-element content of the target molten steel. The molten steel composition detection model is built based on the improved ResNet50 network.

[0052] Next, the LIBS-based optical path excitation and collection system and the improved ResNet50 network mentioned in the above embodiments of the present invention will be described respectively.

[0053] 1. LIBS-based optical excitation and collection system:

[0054] This LIBS-based optical path excitation and collection system includes a laser excitation optical path, a signal collection optical path, and a signal transmission and processing optical path; see [link / reference] Figure 2 and Figure 3 As shown, specifically:

[0055] (1) Laser excitation optical path:

[0056] In the laser excitation optical path, a laser, a focusing lens, and a reflector are arranged along the same optical axis. The laser is a near-infrared pulsed laser source with an output wavelength of 1064nm. The emitted laser beam is focused by the parallel focusing lens and reflected and refracted by the reflector placed at 45°, so that the laser beam is focused on the target molten steel in the medium frequency furnace, exciting the formation of plasma and emitting signal light.

[0057] Among them, the reflector is placed in the same area as the signal acquisition, minimizing the mirror size while meeting the focusing requirements of the receiving optical path, and avoiding obstruction of the plasma radiation light; at the same time, by precisely controlling the laser beam to position the laser focus at a depth of 0.5-1mm below the surface of the target molten steel, a stable output with an effective spot diameter of 3-5mm is achieved.

[0058] (2) Signal collection optical path:

[0059] In the signal collection optical path, this embodiment of the invention innovatively configures a coaxial dual-mirror group consisting of a concave mirror and a convex mirror, that is, the primary mirror and the secondary mirror are coaxially arranged; the primary mirror is a concave mirror (concave mirror Φ150mm, central aperture 40mm), and the concave mirror is provided with a central hole; the secondary mirror is a convex mirror (Φ86mm). The curvature and spacing parameters of the two are optimized through optical simulation to form a cooperative folding optical path structure; the signal light emitted by the plasma is reflected by the concave mirror to the convex mirror, and then reflected and focused by the convex mirror, and coupled to the fiber optic receiver through the central hole, realizing efficient transmission and precise convergence of the optical path;

[0060] It should be noted here that, in Figure 2 The concave mirror in the image is a single unit, except that it has a central hole in the center. Figure 2 The "gap" in the concave mirror is actually to indicate the central hole, so as to explain that the convex mirror reflects the signal light and couples it to the fiber optic receiver through the central hole.

[0061] The reflective coatings of the concave and convex mirrors cover a wide spectral range of 190-1100nm, effectively expanding the spectral response capability; the central hole design of the primary mirror maximizes the utilization rate of the reflective surface while ensuring unobstructed light transmission.

[0062] This invention abandons the traditional dichroic mirror and adopts a coaxial dual-mirror group (concave primary mirror + convex secondary mirror). By optimizing the curvature matching relationship and spatial arrangement angle of the mirrors through simulation model, the optical path deflection loss can be significantly reduced. This allows the plasma radiation signal to be converged to the fiber optic receiver with a transmission efficiency of >90%, and the signal-to-noise ratio is improved to 48dB, greatly improving the signal-to-noise ratio and stability of spectral signal acquisition. For example, with a curvature radius of 677.12mm for the concave mirror and 897.76mm for the convex mirror, and a spacing of 87.259mm, simulation using Zemax software shows that the optical path transmission efficiency is consistently above 90%. The signal-to-noise ratio (SNR) is compared with the intensity of a characteristic peak (Fe 438.35nm) and the background noise (multiple measurements). The secondary reflection by the concave and convex mirrors increases the intensity of the characteristic peak by 40% (from 2800mV to 3990mV), with a maximum SNR of 48dB.

[0063] (3) Optical path for signal transmission and processing:

[0064] In the optical path of signal transmission and processing, the fiber optic receiver inputs the signal light containing the composition information of the target molten steel into the spectrometer to obtain one-dimensional spectral data, and the spectrometer preprocesses the one-dimensional spectral data to obtain a two-dimensional grayscale image corresponding to the target molten steel.

[0065] The preprocessing process includes: after receiving the one-dimensional spectral data, the spectrometer generates one-dimensional intensity-wavelength data (2048 channels); then, the spectrometer offset is corrected in real time using the Hg-Ar lamp standard spectral lines to achieve dynamic wavelength calibration; next, wavelet transform (db4 basis function) and moving average filtering are combined for adaptive noise reduction; finally, the one-dimensional spectral data (spectral intensity moment) after dynamic wavelength calibration and adaptive noise reduction is reconstructed into a two-dimensional grayscale image of 256×1024 pixels; during the training of the molten steel composition detection model, a large number of two-dimensional grayscale images can be acquired to form a spectral dataset, which can be used as input for training the molten steel composition detection model;

[0066] Specifically, the spectrometer offset is corrected in real time using Hg-Ar lamp standard spectral lines. This involves: nonlinear interpolation correction using standard spectral lines such as Hg 253.65nm and Ar 763.51nm, with a wavelength error < ±0.05nm, for wavelength calibration; and linear mapping of the original ADU values ​​(0-65535) to the [0,1] interval to achieve intensity normalization, using the following formula:

[0067]

[0068] Where Inorm(x,y) represents the normalized image value; Iraw(x,y) represents the original image value, that is, the value of the image before processing; Imin represents the global minimum image value; Imax represents the global maximum image value; where Imax and Imin take the global extreme values ​​of the entire dataset; this calculation can map the original image value to the interval [0,1];

[0069] The input dimension of the above two-dimensional grayscale image is set to 256 (height) × 1024 (width) × 1 (channel) to match the number of channels of the spectrometer. The pixel values ​​are mapped to the [0,1] interval. The formula Inorm=(I-Imin) / (Imax-Imin) is used for spectral normalization, and finally the network input layer is optimized.

[0070] Based on the above, this embodiment of the invention reconstructs one-dimensional spectral data into a 256×1024 pixel grayscale image while preserving the original resolution; and optimizes the input quality through dynamic wavelength calibration (error < ±0.05nm) and adaptive noise reduction.

[0071] 2. Improve the ResNet50 network:

[0072] The improved ResNet50 network includes: an input layer, an initial convolutional layer, a spectral feature enhancement module, multi-scale residual blocks, a spectral pyramid pooling layer, and a fully connected layer; specifically:

[0073] (1) Input the two-dimensional grayscale image into the trained molten steel composition detection model through the input layer;

[0074] (2) The initial feature map is obtained by extracting features from the two-dimensional grayscale image through the initial convolutional layer; the initial convolutional layer cancels the first 7×7 convolution of the traditional ResNet network and uses a 1×7 horizontal convolutional kernel instead.

[0075] (3) The initial feature map is enhanced by a spectral feature enhancement module to obtain spectral enhancement features. The spectral feature enhancement module includes two branches and an SE attention layer. The first branch uses a 1×7 convolution (64 channels, ReLU activation) to extract global spectral features in the wavelength direction of the initial feature map. The second branch uses a 7×1 dilated convolution (dilation = 3, 64 channels) to capture local intensity correlations in the initial feature map and obtain local spectral features. The global spectral features and local spectral features are concatenated and then fused by dynamic weighting of the channels through the SE attention layer (compression ratio = 16) to obtain spectral enhancement features.

[0076] In this embodiment of the invention, by fusing the global features and local intensity correlations of the wavelength direction through the above two branches, the discrimination of overlapping spectral peaks is improved by 40% compared with the traditional ResNet. Taking Fe 438.35nm and Cr 438.42nm as an example, with a peak spacing of 0.07nm, the discrimination (PDI) of the overlapping spectral peaks of LIBS by the traditional 7×7 convolutional CNN is in the range of 0.3-0.4, and the minimum resolvable peak spacing is 0.15nm, which is difficult to distinguish. The PDI of trace elements such as S and P is even smaller. By using a 1×7 horizontal convolutional kernel, the minimum resolvable peak spacing can be reduced to 0.09nm.

[0077] (4) The spectral enhancement features are input into the multi-scale residual block after passing through a 3×3 convolutional layer and a max pooling layer. After processing by the multi-scale residual block, multi-scale residual features are generated.

[0078] The multi-scale residual block (MS-ResBlock) contains three consecutive 3x3 convolutional layers. The first 3x3 convolutional layer has 2 input channels and a kernel size of 3. After the input image passes through the convolution operation of the three 3x3 convolutional layers, spatial attention weights are calculated and generated. Then, after passing through residual connections, the features are reconstructed and output.

[0079] (5) Multi-scale feature fusion is performed on multi-scale residual features through spectral pyramid pooling layer (SPP) to improve spectral resolution, and then output through fully connected layer;

[0080] The spectral pyramid pooling layer comprises three parallel pooling layers: a 1×1 global average pooling layer, a 3×3 max pooling layer, and a 6×6 max pooling layer. The 1×1 global average pooling layer captures the overall intensity distribution of the spectrum; the 3×3 max pooling layer extracts local spectral peak features of adjacent overlapping peaks; and the 6×6 max pooling layer focuses on broad spectral regions. The outputs of the 1×1 global average pooling layer, the 3×3 max pooling layer, and the 6×6 max pooling layer are concatenated into a 512-dimensional feature vector, fusing global statistics and local details, and then input to the fully connected layer. The fully connected layer outputs predicted values ​​of the elemental content of the target molten steel through a classifier.

[0081] The fully connected layer described above adopts a parallel regression head design, which contains multiple parallel sub-networks, each of which outputs a predicted value of the content of a single element; the input dimension is 512 up to the output dimension of 1; this parallel regression head design supports the simultaneous detection of major elements and trace impurities, with a dynamic range spanning four orders of magnitude;

[0082] Its loss function uses the weighted Huber Loss, expressed as:

[0083]

[0084] Where L is the loss function, used to measure the difference between the model's predicted value and the true value; N is the total number of elements, representing the total number of samples in the dataset; i is the element index, used to traverse each element from 1 to N; ωi is the weight of the i-th element, used to adjust the importance of different samples in the loss calculation; yi is the true value of the i-th element, also called the target value. The predicted value of the i-th element is the output value given by the steel composition detection model for the i-th element.

[0085] (6) Output post-processing:

[0086] The data output post-processing adopts an online self-calibration mechanism. By automatically performing dark current correction and wavelength reference calibration environmental compensation every 2 hours, when a new element spectral peak is detected, an online fine-tuning mode is triggered, which only updates the parameters of the last two fully connected layers to achieve incremental learning. This optimizes the results based on physical constraints and the incremental learning mechanism, thus achieving closed-loop feedback control.

[0087] In summary, this invention provides an online steel composition detection method based on LIBS technology and deep learning. Through co-designed coaxial dual-mirror optical path optimization and anti-interference deep learning algorithms, it overcomes existing technological limitations, shortening the detection cycle to 120ms and improving efficiency by over 90% compared to traditional methods, thus meeting the requirements for rapid furnace-front control. For example, using coaxial dual mirrors with a 600μm wide-spectrum UV-resistant optical fiber for spectral signal acquisition is more than 110ms faster than conventional LIBS equipment. Through data preprocessing and an improved ResNet50 network model, the inference time is controlled within 190ms. Traditional spectrometer testing, from solid-state sampling, cooling and solidification to spectral analysis, takes approximately 7 minutes, making a 90% or more efficiency improvement over traditional methods feasible. However, shortening the detection cycle to 120ms is difficult to achieve and can be omitted here.

[0088] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0089] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for online detection of molten steel composition based on LIBS technology and deep learning, characterized in that, Includes the following steps: One-dimensional spectral data of the target molten steel is acquired through a LIBS-based optical path excitation and collection system, and the one-dimensional spectral data is preprocessed to obtain the corresponding two-dimensional grayscale image. The two-dimensional grayscale image is input into the trained molten steel composition detection model, and the predicted value of the multi-element content of the target molten steel is output. The steel composition detection model is built based on an improved ResNet50 network.

2. The online steel composition detection method based on LIBS technology and deep learning according to claim 1, characterized in that, The LIBS-based optical path excitation and collection system includes a laser excitation optical path, a signal collection optical path, and a signal transmission and processing optical path. In the laser excitation optical path, a laser, a focusing lens, and a reflector are arranged along the same optical axis. The laser emits a laser beam, which is focused by the focusing lens and reflected and refracted by the reflector placed at 45°, so that the laser beam is focused on the target molten steel, exciting the formation of plasma and emitting signal light. In the signal collection optical path, a concave reflector and a convex reflector coaxially arranged with the concave reflector are provided; the concave reflector has a central hole; the signal light is reflected by the concave reflector to the convex reflector, then reflected and focused by the convex reflector, and coupled to the fiber optic receiver through the central hole; In the signal transmission and processing optical path, the fiber optic receiver inputs the signal light containing the target molten steel composition information into the spectrometer to obtain one-dimensional spectral data, and the spectrometer preprocesses the one-dimensional spectral data to obtain a two-dimensional grayscale image corresponding to the target molten steel.

3. The online steel composition detection method based on LIBS technology and deep learning according to claim 2, characterized in that, In the laser excitation optical path, the laser beam is refracted by the reflector to position the laser focus at a depth of 0.5-1 mm below the surface of the target molten steel.

4. The online steel composition detection method based on LIBS technology and deep learning according to claim 2, characterized in that, The reflective coatings of the concave and convex mirrors cover a wide spectral range of 190-1100 nm.

5. A method for online detection of molten steel composition based on LIBS technology and deep learning according to any one of claims 1-4, characterized in that, The preprocessing includes: Dynamic wavelength calibration is achieved by real-time correction of the spectrometer offset using Hg-Ar lamp standard spectral lines. Adaptive noise reduction is achieved by combining wavelet transform and moving average filtering; The one-dimensional spectral data after dynamic wavelength calibration and adaptive noise reduction is converted into a two-dimensional grayscale image.

6. The online steel composition detection method based on LIBS technology and deep learning according to claim 1, characterized in that, The improved ResNet50 network includes: an input layer, an initial convolutional layer, a spectral feature enhancement module, a multi-scale residual block, a spectral pyramid pooling layer, and a fully connected layer; The two-dimensional grayscale image is input into the trained molten steel composition detection model through the input layer; The initial feature map is obtained by extracting features from the two-dimensional grayscale image through the initial convolutional layer; the initial convolutional layer uses a 1×7 convolutional kernel. The initial feature map is enhanced by the spectral feature enhancement module to obtain spectral enhanced features; The spectral enhancement features are input to the multi-scale residual block after passing through a 3×3 convolutional layer and a max pooling layer. After processing by the multi-scale residual block, multi-scale residual features are generated. After processing the multi-scale residual features through the spectral pyramid pooling layer, the output is then passed through a fully connected layer.

7. The online steel composition detection method based on LIBS technology and deep learning according to claim 6, characterized in that, The spectral feature enhancement module comprises two branches and an SE attention layer; wherein: The first branch uses a 1×7 convolution to extract the global spectral features of the initial feature map; The second branch uses a 7×1 dilated convolution to extract local spectral features from the initial feature map; The global spectral features and the local spectral features are concatenated and then fused through the SE attention layer to obtain spectral enhancement features.

8. The online steel composition detection method based on LIBS technology and deep learning according to claim 6, characterized in that, The multi-scale residual block contains three consecutive 3x3 convolutional layers, wherein the first 3x3 convolutional layer has 2 input channels and a kernel size of 3.

9. The online steel composition detection method based on LIBS technology and deep learning according to claim 6, characterized in that, The spectral pyramid pooling layer includes three parallel pooling layers, specifically a 1×1 global average pooling layer, a 3×3 max pooling layer, and a 6×6 max pooling layer. The 1×1 global average pooling layer is used to capture the overall intensity distribution of the spectrum; The 3×3 max pooling layer is used to extract local spectral peak features of adjacent overlapping peaks; The 6×6 max-pooling layer is used to focus on a wide-bandgap region; The outputs of the 1×1 global average pooling layer, the 3×3 max pooling layer, and the 6×6 max pooling layer are concatenated and then input into the fully connected layer.

10. The online steel composition detection method based on LIBS technology and deep learning according to claim 6, characterized in that, The fully connected layer comprises multiple parallel sub-networks, each outputting a predicted content value for one element; and the loss function of the fully connected layer is expressed as: Where L represents the loss function; N represents the total number of elements; i represents the element index; ωi represents the weight of the i-th element; and yi represents the true value of the i-th element. This represents the predicted value of the i-th element.

Images