Method for establishing LIBS noise reduction model based on self-supervised learning

Abstract

The invention relates to the technical field of noise reduction model establishment methods, in particular to an LIBS noise reduction model establishment method based on self-supervised learning, which specifically comprises the following steps: step 1, determining a model architecture: based on a blind spot network architecture, adopting the blind spot network architecture, and based on self-supervised learning, adopting the blind spot network architecture; randomly shielding some data points in the original spectral data, and inferring a true value of a missing part according to other unshielded part data by using a blind spot convolution module, so as to realize noise reduction under the condition of no noise-free target data; 2, constructing a noise reduction model, wherein the noise reduction model comprises a mask module and a noise reduction coding module; no matter in a high-intensity wave band or a low-intensity wave band, the noise reduction model can effectively reduce noise, well reserve spectral detail features, accurately restore original spectral signals, ensure the accuracy of a detection result and avoid element misjudgment or missed judgment caused by distortion.

Description

Technical Field

[0001] The present invention relates to the technical field of methods for establishing a denoising model, and in particular to a method for establishing a LIBS denoising model based on self-supervised learning. Background Art

[0002] Laser induced breakdown spectroscopy (LIBS) is a branch technology based on atomic emission spectroscopy. It directly excites the sample surface through laser pulses to generate high-temperature plasma, and determines the elemental composition of the sample by analyzing the spectrum emitted by the plasma. This technology can directly analyze solid, liquid and gas samples, and is particularly suitable for rapid, on-site analysis, and has unique advantages in meeting the needs of rapid acquisition of elemental information.

[0003] Although LIBS technology has significant advantages in elemental analysis, its accuracy and practicality are severely restricted by noise problems. Spectral noise not only weakens sensitivity and accuracy, but is particularly prominent in the analysis of complex samples.

[0004] The main sources of noise include the instability of laser pulses, spontaneous radiation of plasma, background light interference, and changes in the surface state of the sample. These noises are intertwined, significantly reducing the quality of spectral data, especially for the quantitative analysis of low-concentration elements. The signal instability caused by laser source fluctuations and plasma dynamic changes will result in different intensities and shapes of each spectral excitation, increasing the difficulty of analysis. In addition, the interference of background light and scattered light often obscures the characteristic spectral lines of the target elements, especially in multi-element analysis, where spectral line overlap further exacerbates the interference. The inhomogeneity of the sample surface and changes in microstructure introduce additional noise, resulting in distortion of the spectral signal. These noise problems significantly limit the application of LIBS technology in high-precision measurement and real-time detection. Traditional noise reduction methods are difficult to fully and effectively deal with due to the complexity and diversity of noise sources, so the current LIBS industry is in urgent need of more advanced noise reduction solutions. Summary of the invention

[0005] In view of the technical problem that effective noise reduction cannot be performed in the prior art, the present invention provides a method for establishing a LIBS noise reduction model based on self-supervised learning.

[0006] The technical solution adopted by the present invention is: a method for establishing a LIBS denoising model based on self-supervised learning, which specifically includes the following steps: Step 1: Determine the model architecture: Based on the blind spot network architecture, the blind spot network architecture is adopted. The blind spot network architecture is based on self-supervised learning. Some data points are randomly blocked in the original spectral data. The blind spot convolution module is used to infer the true value of the missing part based on the remaining unblocked part of the data, so as to achieve noise reduction without noise-free target data. Step 2: construct a noise reduction model: the noise reduction model includes a mask module and a noise reduction encoding module; Step 3: Set the training process parameters; Step 4: Package and apply the model.

[0007] In one embodiment, in step 2, the specific method of constructing the noise reduction model is as follows: Mask module: At the beginning of each round of training, according to the preset mask ratio, randomly select the spectral data position to generate a Boolean mask, apply the mask to the input spectral data, set the spectral value of the True position in the mask to zero, generate some missing input data for training, and regenerate the mask in the next round of training; Denoising coding module: With 1D blind spot convolution as the core, the traditional convolution kernel is changed to a blind spot convolution kernel with a center of 0. The module consists of multiple 1D blind spot convolution layers, which is used to process one-dimensional spectral signals. The spectral data after mask processing is input, and the target output is the original spectrum. The local features of the spectral data are extracted through multi-layer blind spot convolution. The ReLU activation function is used to introduce nonlinearity after each convolution layer. The output layer restores the spectral signal through linear transformation and predicts the spectral value at the mask position. Design loss function: Use a custom loss function that combines mean square error and total variation regularization. ,in, Measure the error between the denoised spectral data and the original spectrum, Used to smooth the prediction signal, is the regularization coefficient.

[0008] In one embodiment, in step 3, the specific method of setting the training process parameters is as follows: Network structure parameters: The denoising coding module adopts a double-layer blind spot convolution structure network. The first layer is set with 64 convolution kernels, and the second layer is set with 128 convolution kernels. The convolution kernel size is 3. The ReLU activation function is used to capture the local features in the sequence and introduce nonlinearity. Custom blind spot initializers and constraints are introduced in each layer. Training optimization parameters: Use Adam optimizer for optimization, set the initial learning rate to 0.001, the training batch size to 16, use a custom total variation regularization loss function, and set the regularization coefficient to 0.001; Training environment and cycles: Training was performed on an NVIDIA 1650 GPU with 15,000 training cycles.

[0009] In one embodiment, in step 4, the specific method of model packaging and application is as follows: After the denoising coding module is trained, it is packaged and saved. In actual applications, the user inputs the original noisy LIBS spectral data into the trained denoising coding module.

[0010] The beneficial effects of the present invention are as follows: compared with the prior art, the noise reduction model in the present invention has better noise reduction capabilities. The noise reduction model in the present invention can well retain the spectral detail characteristics while effectively reducing noise, accurately restore the original spectral signal, ensure the accuracy of the detection results, and avoid misjudgment or omission of elements caused by distortion, which is crucial for accurate element analysis. In addition, the noise reduction model in the present invention does not require a large amount of "clean data" without noise to participate in training, breaking through the limitations of traditional supervised learning methods and reducing the difficulty and cost of data acquisition. It has strong adaptive learning ability and is suitable for various situations with complex noise distribution or rich signal details. The entire noise reduction model does not require an explicit noise model and can be used normally when the noise distribution or level is unknown, expanding the application scenarios of LIBS technology, making its application in high-precision measurement and real-time detection possible, and promoting the further development and wide application of LIBS technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is the overall flow chart of the noise reduction model in the present invention; Figure 2 This is a comparison chart of the spectrum RSD before and after noise reduction in the present invention; Figure 3 It is a comparison chart of the denoising results of Savitzky-Golay filtering, moving average filtering, wavelet transform, median filtering, Gaussian filtering and the denoising model in the present invention. DETAILED DESCRIPTION

[0012] In the description of the present invention, it should be noted that the orientations or positional relationships indicated by terms such as “front”, “up”, “down”, “left”, “right”, “vertical” and “horizontal” are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0013] In order to solve the problems existing in the background technology, the present application proposes the following technical solution: a method for establishing a LIBS denoising model based on self-supervised learning, which specifically includes the following steps: Step 1: Determine the model architecture: Based on the blind spot network architecture, the blind spot network architecture is adopted. The blind spot network architecture is based on self-supervised learning. Some data points are randomly blocked in the original spectral data. The blind spot convolution module (the convolution kernel with the center point of 0) is used to infer the true value of the missing part based on the remaining unblocked data, so as to achieve noise reduction without noise-free target data; Step 2: construct a noise reduction model: the noise reduction model includes a mask module and a noise reduction encoding module; In step 2, the specific method of constructing the noise reduction model is as follows: Mask module: At the beginning of each round of training, according to the preset mask ratio (the default is 0.1, which can be dynamically adjusted according to the spectral noise level), the spectral data position is randomly selected to generate a Boolean mask, the mask is applied to the input spectral data, and the spectral value of the True position in the mask is set to zero. Part of the missing input data is generated for training, and the mask is regenerated in the next round of training to improve the robustness and generalization ability of the model.

[0014] Denoising coding module: With 1D blind spot convolution as the core, the traditional convolution kernel is changed to a blind spot convolution kernel with a center of 0. The module consists of multiple 1D blind spot convolution layers and is used to process one-dimensional spectral signals. The spectral data after mask processing is input, and the target output is the original spectrum. The local features of the spectral data are extracted through multi-layer blind spot convolution. The ReLU activation function is used after each convolution layer to introduce nonlinearity and enhance the model's expression ability. The output layer restores the spectral signal through linear transformation and predicts the spectral value at the mask position.

[0015] Design loss function: Use a custom loss function that combines mean square error (MSE) and total variation (TV) regularization. ,in, Weigh the error between the denoised spectral data and the original spectrum, guide the model to learn the main goal of the denoising task and ensure that the denoised data is not distorted. It is used to smooth the prediction signal, avoid overfitting, and reduce high-frequency noise by constraining the gradient of adjacent spectral points. is the regularization coefficient, balancing prediction accuracy and smoothness.

[0016] Step 3: Set the training process parameters; In step 3, the specific method of setting the training process parameters is as follows: Network structure parameters: The denoising coding module uses a double-layer blind spot convolutional structure network. The first layer has 64 convolution kernels and the second layer has 128 convolution kernels. The convolution kernel size is 3. The ReLU activation function is used to capture local features in the sequence and introduce nonlinearity. Custom blind spot initializers and constraints are introduced in each layer to ensure that the weight of the convolution kernel at a specific position is zero, realize the blind spot feature, and enhance the model's robustness to noise.

[0017] Training optimization parameters: The Adam optimizer is used for optimization, the initial learning rate is set to 0.001, and the training batch size is 16. A custom total variation regularization loss function is used, and the regularization coefficient is set to 0.001.

[0018] Training environment and cycles: Training was performed on an NVIDIA 1650 GPU with 15,000 training cycles.

[0019] Step 4: Model packaging and application;

[0020] In step 4, the specific methods for model packaging and application are as follows: After the noise reduction coding module training is completed, it is packaged and saved. In practical applications, users only need to input the original noisy LIBS spectral data into the trained noise reduction coding module to efficiently obtain the noise-reduced data, providing a more accurate and reliable data basis for subsequent element analysis and quantitative detection.

[0021] The overall explanation of the above technical solution is as follows: The present invention introduces self-supervised learning to overcome and solve the above-mentioned existing difficulties, and realizes the construction of an efficient LIBS spectral denoising model based on a blind spot network architecture.

[0022] Different from traditional supervised learning methods, the blind spot network architecture is trained on the input local mask data and then uses the original input data as the target data. It can achieve efficient denoising tasks even in the absence of noise-free target data.

[0023] The core idea of ​​this method is to randomly block certain data points in the original spectral data, and then use the blind spot convolution module (the convolution kernel with the center point at 0) to infer the true value of the missing part based on the remaining unblocked data. This denoising training strategy enables the blind spot network architecture to efficiently learn effective denoising features when data is scarce.

[0024] At the same time, the model only makes minimal assumptions about the noise statistics (data independence between spectral bands): (i) the signal data is not independent between bands; (ii) given the signal data, the noise is independent between bands; and the entire training process does not require any "clean data" to participate in training.

[0025] Since the entire model does not require an explicit noise model, it is applicable to various noise types and can be used when the noise distribution or level is unknown.

[0026] The noise reduction model established by the present invention includes the following modules: Mask module: This module is used to generate partially masked input spectral data.

[0027] That is, at the beginning of each round of training, the position in the spectral data is randomly selected for masking according to the preset mask ratio (the default is 0.1, and the threshold can be dynamically adjusted according to the spectral noise level. The higher the noise level, the higher the threshold level can be set accordingly) to generate a Boolean mask with the same length as the spectrum, where True indicates that the position is masked and False indicates that the original data is retained.

[0028] Subsequently, the generated Boolean mask is uniformly applied to all input spectral data, and the spectral values ​​at the positions corresponding to True are set to zero, thereby generating partially missing input data for training.

[0029] In the next round of training, a new mask is regenerated and the above process is repeated. By dynamically updating the mask, the model can be trained on different mask positions, thereby improving its robustness and generalization ability.

[0030] Noise reduction coding module: The core of this module is 1D blind spot convolution, which modifies the traditional convolution kernel into a blind spot convolution kernel with a center of 0.

[0031] This module consists of multiple 1D blind convolution layers and is designed for processing one-dimensional spectral signals. The input of the module is the spectral data after masking, and the target output is the original spectrum without masking. Then, local features in the spectral data are extracted through multiple layers of blind convolution.

[0032] The center weight of the blind spot convolution kernel is fixed to 0, ensuring that the center pixel does not participate in the calculation. This prevents the model from directly copying the input data while forcing the model to infer and predict the "pure data" of the center point based on the data before and after the center point.

[0033] After each layer of blind spot convolution processing, the ReLU activation function introduces nonlinearity to enhance the model's expressiveness. Finally, the output layer restores the spectral signal through linear transformation, with the goal of predicting the spectral value at the mask position. Through this design, the model can effectively extract the local features of the spectral data while enhancing the robustness to noise. The specific structure is as follows: Loss function: This paper uses a custom loss function that combines mean square error (MSE) and total variation (TV) regularization, defined as follows: ,in, MSE stands for mean square error (MSE), which measures the error between the denoised spectral data and the original spectrum. It is used to guide the model to learn the main goal of the denoising task and ensure that the denoised data is not distorted. It stands for total variation (TV) regularization and is used to smooth the prediction signal to avoid overfitting.

[0034] By constraining the gradients of adjacent spectral points, the high-frequency noise of the model output is reduced while maintaining the smoothness of the signal. Represents the regularization coefficient, which controls the weight of the TV regularization term. The two parts of the loss are weighted by a regularization coefficient to balance the prediction accuracy and smoothness.

[0035] By minimizing this loss function, the model is able to more accurately restore the original spectral signal while maintaining the stability of the output.

[0036] The overall process of the present invention is shown as follows Figure 1 shown.

[0037] In the first stage of training, the mask scale is set to the default value (0.1).

[0038] The masked data is then used as input data of a noise reduction coding module. The noise reduction coding module constructed by the present invention is a double-layer blind spot convolution structure network.

[0039] The network extracts the features of the input sequence step by step through two convolutional layers. The first layer contains 64 convolution kernels and the second layer contains 128 convolution kernels. Both use a convolution kernel of size 3 and a ReLU activation function to capture local features in the sequence and introduce nonlinearity.

[0040] A custom blind spot initializer and constraints are introduced in each layer to ensure that the weights of the convolution kernel at specific positions are forced to zero, thereby realizing the blind spot feature, avoiding direct transmission of information at certain positions, and enhancing the model’s robustness to noise. In addition, the network uses a custom total variation regularization loss function, combining the data fitting error and the regularization term (regularization coefficient is 0.001).

[0041] The overall training process was optimized using the Adam optimizer, with an initial learning rate of 0.001 and a training batch size of 16. The final model was trained for 15,000 cycles on an NVIDIA 1650 GPU.

[0042] After the training of the noise reduction coding module is completed, the present invention encapsulates and saves it so as to be subsequently applied to the noise reduction prediction task of LIBS spectrum. During the prediction process, the user only needs to input the original noisy LIBS spectrum data into the trained noise reduction coding module to efficiently obtain the noise-reduced data.

[0043] This process is simple and efficient, without the need for additional complex operations or preprocessing steps, greatly improving the convenience and practicality of spectral noise reduction.

[0044] Through the noise reduction coding module of the present invention, the signal-to-noise ratio of LIBS spectral data is significantly improved, and the spectral features are clearer, thereby providing a more accurate and reliable data basis for subsequent elemental analysis and quantitative detection.

[0045] Among them, the LIBS data of six stainless steel samples were denoised by using the present invention and five existing traditional spectral denoising method models based on signal processing (Savitzky-Golay filtering, moving average filtering, wavelet transform, median filtering and Gaussian filtering), and then the denoising results of all methods were compared.

[0046] As an experimental example, the present invention uses 180 original noisy LIBS data of six stainless steel samples (30 spectral data of each sample) to train the model, and then uses the trained model to perform noise reduction on 300 original noisy LIBS data of six stainless steel samples (50 spectral data of each sample, and no overlap with the training data) to verify the performance of the present invention. For the comparison of the noise reduction results, the RSD index is used as the evaluation standard.

[0047] Relative Standard Deviation (RSD) is a statistical indicator used to measure the degree of dispersion and precision of data. Its calculation formula is as follows: ,in, is the standard deviation of the data, is the arithmetic mean of the data.

[0048] Figure 2 The figure is a comparison chart of the RSD index before and after noise reduction of the noisy spectral data according to the present invention. Figure 2As shown in the figure, the RSD of the six stainless steel spectral data was greatly reduced after the model denoising. Among them, the RSD of the JZG206B stainless steel data was reduced from 114.18% (before denoising) to 17.83% (after denoising), which fully proved the effectiveness and efficiency of the present invention.

[0049] Then, based on the other five traditional spectral denoising methods (Savitzky-Golay filtering, moving average filtering, wavelet transform, median filtering and Gaussian filtering), 300 original noisy LIBS data (consistent with the data used to verify the effect of the invention model above) were also denoised. All the results are shown in Table 1, which systematically evaluates the denoising performance of traditional denoising methods (including Savitzky-Golay filtering, moving average filtering, wavelet transform, median filtering and Gaussian filtering) and the invention on one-dimensional spectral data.

[0050] Table 1 RSD of six stainless steel data after different noise reduction methods;

[0051] The above analysis is as follows: Comparison of noise reduction capabilities: It can be seen intuitively from the RSD change results of the six stainless steel data after noise reduction by each noise reduction model in Table 1 that the noise reduction model in the present invention shows a significant noise reduction effect on all six stainless steel sample data, and its noise reduction performance is significantly better than other traditional noise reduction methods.

[0052] For example, in the JZG204B sample, the RSD value of the denoising model in the present invention after denoising is 19.79%, while that of the Gaussian filter is 74.1%, an increase of about 73%. In the JZG205 sample, the RSD value of the denoising model in the present invention is 15.86%, while the RSD values ​​of other methods range from 26.87% (wavelet transform) to 84.25% (original data), a significant increase.

[0053] From the overall mean, the average RSD value of the six stainless steel test set data after denoising by the denoising model in the present invention is 18.29%, which is significantly lower than other methods and is about 42% higher than the optimal wavelet transform in traditional technology. This fully demonstrates that the denoising effect of BSDN in different samples is highly consistent.

[0054] Figure 3 It is a comparison chart of the denoising results of Savitzky-Golay filtering, moving average filtering, wavelet transform, median filtering, Gaussian filtering and the denoising model of the present invention; Fidelity comparison: Figure 3The full-band visualization display of the spectral denoising results of each denoising model in (a) shows that, compared with the denoising model in the present invention and the wavelet transform denoising model, the other four models all have serious distortion in the high-intensity band.

[0055] For example, Figure 3 In the red box in (a), the spectral lines after denoising by Savitzky-Golay filtering, moving average filtering, median filtering and Gaussian filtering are significantly lower than the original spectral lines (gray background spectral lines).

[0056] Furthermore, the noise reduction model in the present invention is compared with the wavelet transform which performs well in the high intensity band. Figure 3 By comparing the spectral lines in the red marked box in the local magnification of (b), it can be clearly seen that the wavelet transform has distortion caused by transition smoothing in the low-intensity band.

[0057] For example, Figure 3 In the fourth red box in (b), the original spectrum has four peaks, but after wavelet transformation, it only becomes two peaks, which is seriously distorted. However, BSDN well preserves these four peaks while reducing noise. The above results fully prove that the noise reduction model in the present invention has excellent noise reduction fidelity in both high-intensity bands and low-intensity bands.

[0058] In contrast, the noise reduction model in the present invention significantly improves the noise reduction effect while retaining the spectral detail characteristics through its adaptive learning ability, and is particularly suitable for situations where the noise distribution is complex or the signal details are rich. The results show that the noise reduction model in the present invention is superior to other traditional spectral noise reduction technologies in terms of both the fidelity and efficiency of the noise reduction results.

[0059] In terms of noise reduction capability, the present invention performed noise reduction processing on the LIBS data of six stainless steel samples and compared them with five traditional spectral noise reduction methods. The results showed that the noise reduction model in the present invention exhibited strong noise reduction capabilities on all sample data.

[0060] Taking the JZG204B sample as an example, the RSD before denoising was as high as 88.46%, which was reduced to 29.32% after denoising by this model, while Gaussian filtering could only reduce it to 74.10%, and the improvement of this model was about 73%. From the overall mean, the average RSD value of the six stainless steel test set data after denoising by this model was 19.83%, which was significantly lower than the best wavelet transform (31.49%) in the traditional method, and was improved by about 42% compared with the wavelet transform. This fully proves the high consistency of its denoising effect in different samples, can effectively deal with complex and diverse noises, greatly improves the quality of spectral data, and provides a solid and reliable data foundation for subsequent element analysis and quantitative detection.

[0061] In terms of fidelity, the denoising model in the present invention has obvious advantages. Compared with traditional denoising methods, the other four models have serious distortion in the high-intensity band. For example, the spectrum lines after denoising by Savitzky-Golay filtering, moving average filtering, median filtering and Gaussian filtering are significantly lower than the original spectrum lines. Compared with the wavelet transform that performs well in the high-intensity band, the wavelet transform has distortion caused by transition smoothing in the low-intensity band. For example, in a specific area, the original spectrum line has 4 peaks, which only become 2 peaks after wavelet transform. The denoising model in the present invention, whether in the high-intensity band or the low-intensity band, can effectively reduce noise while retaining the spectral detail characteristics well, accurately restore the original spectral signal, ensure the accuracy of the detection results, and avoid misjudgment or omission of elements due to distortion, which is crucial for accurate element analysis.

[0062] In addition, the noise reduction model in the present invention does not require a large amount of "clean data" without noise to participate in training, breaking through the limitations of traditional supervised learning methods and reducing the difficulty and cost of data acquisition. It has strong adaptive learning ability and is suitable for various situations with complex noise distribution or rich signal details. Moreover, the entire noise reduction model does not require an explicit noise model and can be used normally when the noise distribution or level is unknown, which greatly expands the application scenarios of LIBS technology, making its application in high-precision measurement and real-time detection possible, and promoting the further development and widespread application of LIBS technology.

[0063] While the embodiments of the present invention have been shown and described, it will be apparent to those skilled in the art that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for establishing a LIBS denoising model based on self-supervised learning, characterized in that: The specific steps include: Step 1: Determine the model architecture: Based on the blind spot network architecture, the blind spot network architecture is adopted. The blind spot network architecture is based on self-supervised learning. Some data points are randomly blocked in the original spectral data. The blind spot convolution module is used to infer the true value of the missing part based on the remaining unblocked part of the data, so as to achieve noise reduction without noise-free target data. Step 2: construct a noise reduction model: the noise reduction model includes a mask module and a noise reduction encoding module; Step 3: Set the training process parameters; Step 4: Package and apply the model.

2. The method for establishing a LIBS denoising model based on self-supervised learning according to claim 1, characterized in that: In step 2, the specific method of constructing the noise reduction model is as follows; Mask module: At the beginning of each round of training, according to the preset mask ratio, randomly select the spectral data position to generate a Boolean mask, apply the mask to the input spectral data, set the spectral value of the True position in the mask to zero, generate some missing input data for training, and regenerate the mask in the next round of training; Denoising coding module: With 1D blind spot convolution as the core, the traditional convolution kernel is changed to a blind spot convolution kernel with a center of 0. The module consists of multiple 1D blind spot convolution layers, which is used to process one-dimensional spectral signals. The spectral data after mask processing is input, and the target output is the original spectrum. The local features of the spectral data are extracted through multi-layer blind spot convolution. The ReLU activation function is used to introduce nonlinearity after each convolution layer. The output layer restores the spectral signal through linear transformation and predicts the spectral value at the mask position. Design loss function: Use a custom loss function that combines mean square error and total variation regularization. ,in, Measure the error between the denoised spectral data and the original spectrum, Used to smooth the prediction signal, is the regularization coefficient.

3. The method for establishing a LIBS denoising model based on self-supervised learning according to claim 1, characterized in that: In step 3, the specific method of setting the training process parameters is as follows: Network structure parameters: The denoising coding module adopts a double-layer blind spot convolution structure network. The first layer is set with 64 convolution kernels, and the second layer is set with 128 convolution kernels. The convolution kernel size is 3. The ReLU activation function is used to capture the local features in the sequence and introduce nonlinearity. Custom blind spot initializers and constraints are introduced in each layer. Training optimization parameters: Use Adam optimizer for optimization, set the initial learning rate to 0.001, the training batch size to 16, use a custom total variation regularization loss function, and set the regularization coefficient to 0.001; Training environment and cycles: Training was performed on an NVIDIA 1650 GPU with 15,000 training cycles.

4. The method for establishing a LIBS denoising model based on self-supervised learning according to claim 1, characterized in that: In step 4, the specific methods for model packaging and application are as follows: After the denoising coding module is trained, it is packaged and saved. In actual applications, the user inputs the original noisy LIBS spectral data into the trained denoising coding module.

Images