Metal chemical composition analysis method based on deep learning

Abstract

The application provides a metal chemical composition analysis method based on deep learning, and belongs to the technical field of metal material detection and analysis. The application at least partially solves the problems that existing metal spectrum images have high dimension, noise and complex nonlinear characteristics, and traditional methods are difficult to model. The application comprises the following steps: collecting original spectrum data of metal samples with known chemical composition information, and preprocessing the original spectrum data to obtain preprocessed sample data and build a sample data training set; A metal chemical composition prediction model is built based on a convolutional neural network, and the model is trained using the sample data training set; The original spectrum data of the metal to be detected are collected, preprocessed and input into the metal chemical composition prediction model, and the content prediction value of each metal chemical composition in the metal to be detected is output. The application overcomes the problems of high dimension, noise and spectral line overlap, and improves the precision, automation degree and detection efficiency of metal chemical composition quantitative analysis.

Description

Technical Field

[0001] This invention belongs to the field of metal material detection and analysis technology, specifically relating to a deep learning-based method for analyzing the chemical composition of metals. Background Technology

[0002] Classical chemical analysis methods are among the earliest applied methods for detecting the chemical composition of metals, mainly including gravimetric analysis, titration analysis, and colorimetric analysis. These methods convert the analyte into a stable compound form through chemical reactions, and then achieve quantitative analysis through mass measurement, volume change, or color change. For example, gravimetric analysis separates the target element into a sparingly soluble compound through precipitation, which is then accurately weighed after washing, drying, or ignition to calculate its content. Titration analysis determines the element content based on the quantitative reaction between a standard solution and the analyte, using the volume of standard solution consumed at the reaction endpoint. Colorimetric analysis is based on the correlation between the intensity of solution color and substance concentration, achieving quantitative analysis by measuring absorbance using optical means. These methods have advantages such as mature theoretical foundations, high measurement accuracy, and good repeatability, playing a crucial role in standardized laboratory testing and benchmark analysis, and are often used as calibration bases for other analytical methods.

[0003] However, classical chemical analysis methods typically require complex sample pretreatment processes, including multiple steps such as dissolution, separation, and purification. These procedures are cumbersome and time-consuming, making them unsuitable for rapid detection. Furthermore, these methods are highly dependent on the experience and skill level of the operators, and operational errors between different personnel can affect the consistency of the results. In addition, many of these methods are destructive, requiring acid dissolution or chemical reactions of the sample, which can cause irreversible damage, limiting their application in scenarios involving valuable or non-reproducible samples. Moreover, these methods usually analyze single elements or small amounts of elements, making it difficult to achieve simultaneous rapid detection of multiple elements, resulting in low detection efficiency and difficulty in integrating with modern automated detection systems, thus limiting their application in industrial online detection and smart manufacturing.

[0004] Spectroscopic analysis methods are currently the most widely used techniques for metal composition detection, mainly including inductively coupled plasma mass spectrometry (ICP-MS), inductively coupled plasma optical emission spectroscopy (ICP-OES), X-ray fluorescence spectroscopy (XRF), and laser-induced breakdown spectroscopy (LIBS). ICP-MS uses high-temperature plasma to ionize the sample and separate ions based on their mass-charge ratio, achieving highly sensitive detection of trace and ultra-trace elements. ICP-OES detects the characteristic spectra emitted by excited-state atoms or ions, enabling rapid quantitative analysis of multiple elements. XRF uses X-rays to excite the sample to generate characteristic fluorescence signals, determining the element type and content based on fluorescence intensity; it is non-destructive and easy to operate. LIBS uses high-energy laser pulses to generate plasma on the sample surface and analyzes its emission spectrum, achieving rapid or even real-time element detection. These methods offer advantages such as high detection speed, simultaneous multi-element analysis, high automation, and applicability to various material types, and have been widely applied in metallurgy, materials engineering, and quality inspection.

[0005] However, spectral analysis methods still have certain limitations. For example, methods such as ICP-MS and ICP-OES typically rely on large, precision instruments, resulting in complex system structures, high purchase and maintenance costs, and the need for auxiliary systems such as gas and power supplies, making them difficult to apply in field or mobile scenarios. Furthermore, existing spectral analysis methods generally rely on manual experience or traditional statistical methods to extract and model spectral features during data processing, such as selecting characteristic spectral lines, performing peak fitting, or establishing linear regression models to predict elemental content. However, in multi-element coexistence and complex material systems, spectral signals often suffer from problems such as spectral line overlap, matrix interference, and nonlinear responses, making it difficult to effectively characterize the complex relationship between spectral data and elemental content, thus limiting analytical accuracy and model generalization ability. Therefore, how to improve the automated analysis capability and quantitative analysis accuracy of complex spectral data while fully utilizing existing spectral detection methods has become a critical issue that urgently needs to be addressed.

[0006] X-ray analysis methods mainly include techniques such as scanning electron microscopy-energy dispersive spectroscopy (SEM-EDS). The basic principle is to bombard the sample surface with a high-energy electron beam, causing inner-shell electron transitions in the sample's atoms and releasing characteristic X-rays. By analyzing the energy and intensity of these characteristic signals, the elemental composition and spatial distribution can be detected. This type of method not only provides information on elemental types and abundances but can also be combined with microscopic imaging techniques to obtain the microscopic morphological characteristics of materials, enabling comprehensive analysis of structure and composition. It is widely used in fields such as material failure analysis, micro-area compositional studies, and interface analysis. Furthermore, this type of method has high spatial resolution at the micrometer and even nanometer scales, enabling detailed analysis of local areas, which is of great significance in studying the relationship between material microstructure and properties.

[0007] However, these methods typically have a small detection range, primarily suitable for localized micro-area analysis, and are difficult to rapidly assess the composition of large-sized samples or bulk materials. Furthermore, their equipment is expensive, the system structure is complex, they have strict requirements for experimental environments (such as vacuum conditions), the operation process is relatively complex, and they rely heavily on specialized personnel. In addition, the interaction between the electron beam and the sample during quantitative analysis is complex and easily affected by factors such as sample morphology, surface condition, and matrix effects, thus impacting detection accuracy. Moreover, these methods have relatively low detection efficiency, making them unsuitable for the demands of large-scale, high-efficiency testing in industrial settings, and also unsuitable for online real-time detection scenarios, thus limiting their application in practical industrial settings.

[0008] While classical chemical analysis methods offer high precision, their complex and time-consuming processes, coupled with their destructive nature, make them unsuitable for rapid detection. Spectroscopic analysis methods (such as X-ray fluorescence spectroscopy), while boasting fast detection speeds and strong multi-element analysis capabilities, often rely on manual experience for characteristic spectral line selection and modeling during data processing. In complex systems, these methods are susceptible to factors such as spectral line overlap, matrix effects, and noise interference, resulting in insufficient precision and stability in quantitative analysis.

[0009] In quantitative analysis of metal composition based on metal spectral images acquired by CCD spectrometers, existing literature generally agrees that spectral analysis methods have significant limitations in complex spectral modeling. For example, related studies indicate that the relationship between spectral images and elemental content often exhibits complex nonlinear relationships, which traditional linear regression or shallow models struggle to accurately describe, thus limiting analytical precision. Furthermore, due to the high dimensionality and band correlation of spectral image data, local features and multi-scale information exist between different bands. Traditional methods rely on manually selecting feature lines, making it difficult to fully utilize the full spectral information. In addition, spectral image signals are often accompanied by noise and fluctuations, and traditional methods have limited capabilities in feature extraction and noise suppression.

[0010] Therefore, we propose a deep learning-based method for analyzing the chemical composition of metals. Summary of the Invention

[0011] The present invention aims to solve at least one of the technical problems existing in the prior art by providing a deep learning-based method for metal chemical composition analysis.

[0012] This invention provides a deep learning-based method for analyzing the chemical composition of metals, comprising the following steps: S1: acquiring raw spectral data of a metal sample with known chemical composition information, preprocessing the raw spectral data to obtain preprocessed sample data, and constructing a sample data training set based on the preprocessed sample data; S2: constructing a metal chemical composition prediction model based on a convolutional neural network, and training the metal chemical composition prediction model using the sample data training set; S3: acquiring raw spectral data of the metal to be tested, preprocessing the raw spectral data of the metal to be tested, inputting it into the metal chemical composition prediction model, and outputting the predicted content values ​​of each metal chemical component in the metal to be tested.

[0013] Further, in step S1, the original spectral data of the metal sample is preprocessed, specifically including: background subtraction, smoothing filtering, normalization and standardization of the original spectral data of the metal sample.

[0014] Furthermore, in step S2, the metal chemical composition prediction model adopts a one-dimensional convolutional neural network structure, including an input layer, a convolutional feature extraction module, a feature fusion module, and an output layer.

[0015] Furthermore, the convolutional feature extraction module includes a convolutional layer, a normalization layer, an activation layer, and an attention mechanism module.

[0016] Furthermore, the convolutional layer includes a first one-dimensional convolutional layer and a second one-dimensional convolutional layer; the first one-dimensional convolutional layer has a kernel size of 3 and a kernel count of 32; the second one-dimensional convolutional layer has a kernel size of 3 and a kernel count of 64.

[0017] Furthermore, the attention mechanism module employs both the Self-Attention mechanism and the ECA mechanism.

[0018] Further, in step S2, the metal chemical composition prediction model is trained using the sample data training set. Specifically, this includes: inputting the preprocessed sample data into the metal chemical composition prediction model and obtaining the corresponding true content of chemical components as a supervision signal; constructing a regression model and setting differentiated weight parameters for different chemical components to balance the learning differences between major and minor chemical components; and training the regression model using the preprocessed sample data and the supervision signal to obtain the metal chemical composition prediction model for predicting the content of metal chemical components.

[0019] Furthermore, the regression model is trained using the preprocessed sample data and the supervision signal. The training parameters include: learning rate, batch size, and number of training epochs; the learning rate is 0.001, the batch size is 32, and the number of training epochs is 150.

[0020] Furthermore, the regression model is trained using the preprocessed sample data and the supervision signal. During the model training process, the regression model is optimized using a loss function, wherein the loss function is a mean squared error loss function or a weighted loss function.

[0021] Furthermore, the regression model is trained using the preprocessed sample data and the supervision signal. The specific steps also include: introducing a validation set to evaluate the model performance during the model training process to obtain evaluation results, and then fine-tuning the parameters based on the evaluation results.

[0022] The beneficial effects of this invention are as follows: This invention achieves end-to-end automated modeling of metal spectral data by constructing a deep learning model based on a one-dimensional convolutional neural network. This avoids the reliance on manual selection of feature lines and experience-based modeling in traditional methods, significantly reducing the impact of human factors on the analysis results and improving the automation and consistency of the analysis process. Addressing the high dimensionality, strong noise, and complex nonlinear relationships in laser-induced breakdown spectral data, this invention utilizes a convolutional neural network to extract multi-layer features from the spectral data. This effectively captures local correlation features and deep nonlinear relationships between different wavelengths, significantly improving the accuracy of quantitative analysis of metal chemical composition, especially in complex compositional systems and cases of overlapping spectral lines. Furthermore, this invention achieves fully automated processing from spectral data input to elemental content output, completing spectral feature identification and quantitative composition analysis without manual intervention. It can directly convert the acquired spectral data into corresponding metal chemical composition results, achieving fully automated identification and rapid analysis, significantly improving detection efficiency and practical application value. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the steps of a deep learning-based metal chemical composition analysis method according to an embodiment of the present invention. Figure 2 This is a structural diagram of the convolutional neural network in the deep learning-based metal chemical composition analysis method according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the Self-Attention mechanism of the neural convolutional network in the deep learning-based metal chemical composition analysis method according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the ECA mechanism of the neural convolutional network diagram of the deep learning-based metal chemical composition analysis method according to an embodiment of the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] Please refer to the following: Figure 1 The specific embodiments of the present invention provide a deep learning-based method for analyzing the chemical composition of metals, comprising the following steps: Step S1: Acquiring raw spectral data of metal samples with known chemical composition information, and preprocessing the raw spectral data of the metal samples to obtain preprocessed sample data, and constructing a sample data training set based on the preprocessed sample data; Step S2: Constructing a metal chemical composition prediction model based on a convolutional neural network, and training the metal chemical composition prediction model using the sample data training set; Step S3: Acquiring raw spectral data of the metal to be tested, preprocessing the raw spectral data of the metal to be tested, inputting it into the metal chemical composition prediction model, and outputting the predicted content values ​​of each metal chemical component in the metal to be tested.

[0026] In this embodiment, in step S1, the original spectral data of the metal sample is preprocessed. The steps specifically include: background subtraction, smoothing filtering, normalization and standardization of the original spectral data of the metal sample.

[0027] The process involves acquiring two-dimensional spectral images of the metal sample with known chemical composition information using a high-precision spectral acquisition device, preferably a CCD-based spectrometer. This spectrometer can accurately record the light intensity distribution information of the sample at different wavelengths, thereby effectively characterizing the elemental composition features of the sample. During data acquisition, repeated sampling at multiple locations and in multiple directions can be performed to obtain high-quality and diverse datasets, thereby improving the model's generalization ability and robustness to different samples. Furthermore, the acquisition parameters of the spectral images include, but are not limited to, wavelength range, spectral resolution, and CCD sampling accuracy; for example, the wavelength range can be set to 200–900 nm, and the spectral resolution can be set to 0.05–0.2 nm to ensure accurate capture of the characteristic spectral lines of key elements.

[0028] Specifically, the metal samples with known chemical composition information cover steel or alloy materials with different proportions to ensure the diversity and representativeness of the dataset. During the data acquisition phase, a CCD-based spectrometer is used to acquire spectral images of the samples. Specifically, the spectrometer guides the light signal reflected or emitted from the sample surface through a slit into a grating, which, after dispersion, projects it onto a high-precision CCD sensor according to wavelength. This CCD sensor records the light intensity information corresponding to each wavelength through its two-dimensional pixel array, thereby generating a high-resolution spectral image. Furthermore, during the acquisition process, parameters such as the spectrometer's exposure time, integration times, and grating angle can be adjusted to optimize signal intensity and spectral resolution. Simultaneously, by scanning the sample surface at multiple points, more spatially representative spectral data is obtained. To improve data stability and reliability, it is preferable to perform multiple repeated acquisitions for each sample and perform averaging processing, ultimately constructing a stable and high-quality training image dataset.

[0029] Specifically, since LIBS spectra typically contain continuous background radiation and random noise, the original spectral data of the metal samples with known chemical composition information needs to be systematically processed. First, background subtraction is performed on the original spectra to eliminate interference from the continuous spectrum on the characteristic spectral lines. Second, a smoothing filter method (Savitzky-Golay filtering) is used to denoise the spectra, reducing the impact of random noise on model training. Subsequently, the spectral intensity is normalized to ensure a uniform scale across different samples, thereby improving the model's adaptability to different data distributions. Furthermore, the data is standardized (e.g., zero-mean unit variance standardization) to accelerate the convergence process of the metal chemical composition prediction model and significantly improve the stability of the training process. In this embodiment, the processed spectral data is uniformly converted into a fixed-length one-dimensional vector to obtain the preprocessed sample data, which is then used as the input to the metal chemical composition prediction model.

[0030] Please refer to the following: Figure 2 In this embodiment, in step S2, the metal chemical composition prediction model adopts a one-dimensional convolutional neural network structure, including an input layer, a convolutional feature extraction module, a feature fusion module, and an output layer.

[0031] Specifically, the metal chemical composition prediction model employs a convolutional neural network (CNN) or its improved structure. By performing convolution operations on one-dimensional spectral data, it can effectively extract local correlation features and global spectral features between different bands. Compared to the problem of information loss caused by manually selecting feature lines in traditional methods, CNN can automatically extract features from the full spectrum data end-to-end, thus preserving more complete spectral information. Furthermore, to improve the model's ability to represent different elements and achieve simultaneous prediction of multi-element content, the model can adopt a multi-branch structure or a multi-task learning mechanism. For example, by constructing a multi-branch convolutional neural network, a shared backbone network can be combined with independent feature extraction branches, which can significantly improve the accuracy of multi-element regression analysis while achieving effective feature interaction.

[0032] In this embodiment, the convolutional feature extraction module includes a convolutional layer, a normalization layer, an activation layer, and an attention mechanism module.

[0033] In this embodiment, the convolutional layer includes a first one-dimensional convolutional layer and a second one-dimensional convolutional layer; the first one-dimensional convolutional layer has a kernel size of 3 and a kernel count of 32; the second one-dimensional convolutional layer has a kernel size of 3 and a kernel count of 64.

[0034] Specifically, through layer-by-layer convolutional operations, the metal chemical composition prediction model can achieve deep extraction of local spectral features and progressively enhance its feature representation capabilities. Given the significant local correlation of spectral data in the wavelength dimension, this network structure can effectively capture subtle feature changes and correlation information between adjacent bands. Each convolutional layer is followed by a cascaded batch normalization (BN) layer and an Exponential Linear Unit (ELU) activation function. The BN layer is used to standardize intermediate features, effectively mitigating internal covariate shift issues, thereby improving the stability of the model training process and accelerating convergence. Furthermore, the ELU activation function is introduced, overcoming the information loss defect of the traditional ReLU activation function in the negative value region by performing an exponential nonlinear transformation in the negative value region; this design significantly enhances the model's ability to represent weak spectral signals, especially suitable for scenarios with strong noise interference and low-intensity spectral lines commonly found in LIBS data.

[0035] Furthermore, to enhance the model's robustness, a Dropout layer is selectively introduced after some convolutional or fully connected layers, with the deactivation ratio preferably set to 0.3. During model training, the Dropout layer reduces the co-fitness between neuron nodes by randomly deactivating some neurons with a preset probability. This mechanism can significantly suppress overfitting of the model to the training data, thereby greatly improving the model's generalization ability and prediction accuracy when facing unknown samples.

[0036] In this embodiment, the attention mechanism module employs both Self-Attention and ECA mechanisms. By introducing the attention mechanism, the model gains the ability to adaptively select key features across the entire spectrum, thereby significantly enhancing the regression prediction accuracy for complex multi-element systems.

[0037] Please refer to the following: Figure 3 Specifically, the Self-Attention mechanism can capture the global dependencies between different spectral bands, automatically assigning higher weights to relevant spectral bands, thereby strengthening the model's focus on key element features. Input features are first processed through a weight matrix. , and Key vectors, query vectors, and value vectors are generated by mapping respectively. Then, the query vectors and key vectors are multiplied by a dot product to calculate the relevance score, and the scores are normalized by the Softmax function to obtain the attention weights. Finally, the value vectors are weighted and summed using these weights to aggregate contextual information and output fused features.

[0038] Please refer to the following: Figure 4 Specifically, the ECA attention uses adaptive one-dimensional convolution to weight and integrate features from each channel, improving the response to important spectral bands and enhancing the model's ability to capture local information. First, the input feature map... (dimension is) After global pooling, the spatial dimensions are compressed, resulting in a shape of... A global descriptor is generated; this descriptor learns the non-linear dependencies between channels through a sub-network (usually a multilayer perceptron MLP) containing activation functions; finally, it outputs a... The vector is used as the channel weight to weight and label each channel of the original feature map, thereby highlighting important features and suppressing irrelevant information.

[0039] In this embodiment, in step S2, the metal chemical composition prediction model is trained using the sample data training set. The steps specifically include: inputting the preprocessed sample data into the metal chemical composition prediction model and obtaining the corresponding true content of chemical components as a supervision signal; constructing a regression model and setting differentiated weight parameters for different chemical components to balance the learning differences between major chemical components and trace chemical components; and training the regression model using the preprocessed sample data and the supervision signal to obtain the metal chemical composition prediction model for predicting the content of metal chemical components.

[0040] Specifically, by setting different weight parameters for different chemical components, the learning differences between major and minor chemical components can be effectively balanced, thereby improving the prediction accuracy of minor chemical components.

[0041] In this embodiment, the regression model is trained using the preprocessed sample data and the supervision signal. The training parameters include: learning rate, batch size, and number of training epochs; the learning rate is 0.001, the batch size is 32, and the number of training epochs is 150.

[0042] In this embodiment, the regression model is trained using the preprocessed sample data and the supervision signal. During the model training process, the regression model is optimized using a loss function; wherein the loss function is a mean squared error loss function or a weighted loss function.

[0043] In this embodiment, the regression model is trained using the preprocessed sample data and the supervision signal. The specific steps also include: introducing a validation set to evaluate the model performance during the model training process to obtain evaluation results, and then performing parameter tuning based on the evaluation results.

[0044] Specifically, during model training, network parameters are iteratively optimized by introducing a loss function. Preferably, the loss function is the mean squared error (MSE) loss or a weighted loss function. To address the sample imbalance problem caused by significant differences in the content of different elements in the dataset, the weighted loss function assigns appropriate weight coefficients to high-concentration or low-concentration samples, thereby effectively overcoming the bias phenomenon in the model's learning process and improving overall prediction accuracy.

[0045] In summary, the embodiments disclosed herein have at least the following technical effects: This invention achieves end-to-end automated modeling of metal spectral data by constructing a deep learning model based on a one-dimensional convolutional neural network. This avoids the reliance on manual selection of feature lines and experience-based modeling in traditional methods, significantly reducing the impact of human factors on the analysis results and improving the automation and consistency of the analysis process. Addressing the high dimensionality, strong noise, and complex nonlinear relationships in laser-induced breakdown spectral data, this invention utilizes a convolutional neural network to extract multi-layer features from the spectral data. This effectively captures local correlation features and deep nonlinear relationships between different wavelengths, significantly improving the accuracy of quantitative analysis of metal chemical composition, especially in complex compositional systems and cases of overlapping spectral lines. Furthermore, this invention achieves fully automated processing from spectral data input to elemental content output, completing spectral feature identification and quantitative composition analysis without manual intervention. It can directly convert the acquired spectral data into corresponding metal chemical composition results, achieving fully automated identification and rapid analysis, significantly improving detection efficiency and practical application value.

[0046] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A deep learning-based method for analyzing the chemical composition of metals, characterized in that, Includes the following steps: S1: Collect raw spectral data of metal samples with known chemical composition information, preprocess the raw spectral data of the metal samples to obtain preprocessed sample data, and construct a sample data training set based on the preprocessed sample data. S2: Construct a metal chemical composition prediction model based on a convolutional neural network, and train the metal chemical composition prediction model using the sample data training set; S3: Collect the original spectral data of the metal to be tested, preprocess the original spectral data of the metal to be tested, input it into the metal chemical composition prediction model, and output the predicted content of each metal chemical component in the metal to be tested.

2. The deep learning-based metal chemical composition analysis method according to claim 1, characterized in that, In step S1, the raw spectral data of the metal sample is preprocessed, specifically including the following steps: The original spectral data of the metal sample were subjected to background subtraction, smoothing filtering, normalization, and standardization.

3. The deep learning-based metal chemical composition analysis method according to claim 1, characterized in that, In step S2, the metal chemical composition prediction model adopts a one-dimensional convolutional neural network structure, including an input layer, a convolutional feature extraction module, a feature fusion module, and an output layer.

4. The deep learning-based metal chemical composition analysis method according to claim 3, characterized in that, The convolutional feature extraction module includes a convolutional layer, a normalization layer, an activation layer, and an attention mechanism module.

5. The deep learning-based metal chemical composition analysis method according to claim 4, characterized in that, The convolutional layer includes a first one-dimensional convolutional layer and a second one-dimensional convolutional layer; The kernel size of the first one-dimensional convolutional layer is 3, and the number of kernels is 32; The second one-dimensional convolutional layer has a kernel size of 3 and a kernel count of 64.

6. The deep learning-based metal chemical composition analysis method according to claim 4, characterized in that, The attention mechanism module employs the Self-Attention mechanism and the ECA mechanism.

7. The deep learning-based metal chemical composition analysis method according to claim 1, characterized in that, In step S2, the metal chemical composition prediction model is trained using the sample data training set. The specific steps include: The preprocessed sample data is input into the metal chemical composition prediction model, and the corresponding true content of chemical composition is obtained as a supervision signal. A regression model was constructed, and differentiated weight parameters were set for different chemical components to balance the learning differences between major and minor chemical components. The regression model is trained using the preprocessed sample data and the supervision signal to obtain the metal chemical composition prediction model for predicting the content of metal chemical components.

8. The deep learning-based metal chemical composition analysis method according to claim 7, characterized in that, The regression model is trained using the preprocessed sample data and the supervision signal. The training parameters include: learning rate, batch size, and number of training rounds. The learning rate is 0.001, the batch size is 32, and the number of training rounds is 150.

9. The deep learning-based metal chemical composition analysis method according to claim 7, characterized in that, The regression model is trained using the preprocessed sample data and the supervision signal. During the model training process, the regression model is optimized using a loss function; wherein the loss function is a mean squared error loss function or a weighted loss function.

10. The deep learning-based metal chemical composition analysis method according to claim 7, characterized in that, The regression model is trained using the preprocessed sample data and the supervision signal. The specific steps also include: introducing a validation set to evaluate the model performance during the model training process to obtain evaluation results, and then fine-tuning the parameters based on the evaluation results.

Images