A Method for Establishing a LIBS Denoising Model Based on Self-Supervised Learning

Abstract

The present invention relates to the technical field of noise reduction model establishment methods, specifically: a method for establishing a LIBS noise reduction 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 occluded in the original spectral data, and the blind spot convolution module is used to infer the true values of the missing parts according to the remaining unoccluded data, so as to achieve noise reduction in the case of no noise-free target data; Step 2, construct the noise reduction model: And the noise reduction model includes a masking module and a noise reduction encoding module. The noise reduction model in the present invention can well retain the spectral detail features while effectively reducing noise in both high-intensity bands and low-intensity bands, accurately restore the original spectral signal, ensure the accuracy of the detection results, and avoid misjudgment or missed judgment of elements caused by distortion.

Description

Technical Field

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

[0002] Laser-induced breakdown spectroscopy (LIBS), as a branch technology based on atomic emission spectroscopy, directly excites the surface of a sample through a laser pulse to generate a 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, having unique advantages for the need to quickly obtain elemental information.

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

[0004] The main sources of noise include the instability of the laser pulse, the spontaneous emission of the 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 having the most obvious impact on 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 for each spectral excitation, increasing the difficulty of analysis. In addition, the interference of background light and scattered light often masks the characteristic spectral lines of the target element. Especially in multi-element analysis, spectral line overlap further exacerbates the interference. The non-uniformity and microscopic structure changes of the sample surface introduce additional noise, resulting in spectral signal distortion. These noise problems significantly limit the application of the LIBS technology in high-precision measurement and real-time detection. Due to the complexity and diversity of the noise sources, traditional noise reduction methods are difficult to fully and effectively handle, so there is an urgent need for more advanced noise reduction solutions in the current LIBS industry. Summary of the Invention

[0005] Aiming at the technical problem that effective noise reduction cannot be carried out 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 noise reduction model based on self-supervised learning, specifically including the following steps:

[0007] Step 1, determining 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 occluded in the original spectral data, and the blind spot convolution module is used to infer the true values of the missing parts according to the remaining unoccluded part of the data, so as to achieve noise reduction in the case of no target data without noise.

[0008] Step 2, construct a noise reduction model: The noise reduction model includes a masking module and a noise reduction encoding module;

[0009] Step 3, set the training process parameters;

[0010] Step 4, perform model encapsulation and application.

[0011] In one embodiment, in Step 2, the specific method for constructing the noise reduction model is as follows;

[0012] Masking module: At the beginning of each round of training, according to the preset masking ratio, randomly select the spectral data positions to generate a boolean mask, apply the mask to the input spectral data, set the spectral values at the positions corresponding to True in the mask to zero, generate partially missing input data for training, and regenerate the mask in the next round of training;

[0013] Noise reduction encoding module: With 1D blind spot convolution as the core, change the traditional convolution kernel to a blind spot convolution kernel with a center of 0. The module consists of multiple 1D blind spot convolution layers, which are used to process one-dimensional spectral signals. Input the spectral data after masking processing, and the target output is the original spectrum. Extract the local features of the spectral data through multiple layers of blind spot convolution. Use the ReLU activation function to introduce non-linearity after each convolution. The output layer restores the spectral signal through a linear transformation and predicts the spectral values at the masked positions;

[0014] Design the loss function: Adopt a custom loss function that combines mean square error and total variation regularization, , where, Measure the error between the denoised spectral data and the original spectrum, Used to smooth the predicted signal, is the regularization coefficient.

[0015] In one embodiment, in Step 3, the specific method for setting the training process parameters is as follows:

[0016] Network structure parameters: The noise reduction encoding module adopts a two-layer blind spot convolution structure network. Set 64 convolution kernels in the first layer and 128 convolution kernels in the second layer. The size of the convolution kernels is 3 for both layers. Use the ReLU activation function to capture the local features in the sequence and introduce non-linearity. Introduce a custom blind spot initializer and constraint conditions for each layer;

[0017] Training optimization parameters: Use the Adam optimizer for optimization. Set the initial learning rate to 0.001, the training batch size to 16, adopt a custom total variation regularization loss function, and set the regularization coefficient to 0.001;

[0018] Training environment and cycle: Train on an NVIDIA 1650 GPU, and the training cycle is 15000.

[0019] In one of the embodiments, in step four, the specific method for model encapsulation and application is as follows:

[0020] After the training of the noise reduction coding module is completed, it is encapsulated and saved. In actual application, the user inputs the original noisy LIBS spectral data into the trained noise reduction coding module.

[0021] 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 strength. Whether in high-intensity bands or low-intensity bands, the noise reduction model in the present invention can effectively reduce noise while well retaining the spectral detail features, 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 elemental 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 the training, breaks through the limitations of traditional supervised learning methods, and reduces the difficulty and cost of data acquisition. It has strong adaptive learning ability and is applicable to various situations with complex noise distributions 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, expanding the application scenarios of LIBS technology and making its application in high-precision measurement and real-time detection possible, promoting the further development and wide application of LIBS technology. Description of the Drawings

[0022] Figure 1 is the overall flowchart of the noise reduction model in the present invention;

[0023] Figure 2 is the comparison chart of the spectral RSD before and after noise reduction in the present invention;

[0024] Figure 3 is the comparison chart of the noise reduction results of Savitzky-Golay filtering, moving average filtering, wavelet transform, median filtering, Gaussian filtering and the noise reduction model in the present invention. Detailed Embodiments

[0025] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "front", "upper", "lower", "left", "right", "vertical", "horizontal", etc. is based on the orientation or positional relationship shown in the drawings, and is 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 should not be construed as a limitation of the present invention.

[0026] 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 noise reduction model based on self-supervised learning, specifically including the following steps:

[0027] Step 1, determine the model architecture: Based on the blind spot network architecture, adopt the blind spot network architecture. The blind spot network architecture is based on self-supervised learning. Randomly occlude some data points in the original spectral data, and use the blind spot convolution module (a convolution kernel with a center of 0) to infer the true values of the missing parts according to the remaining unoccluded data, so as to achieve noise reduction in the case of no target data containing noise.

[0028] Step 2, construct a noise reduction model: And the noise reduction model includes a masking module and a noise reduction encoding module.

[0029] In Step 2, the specific method for constructing the noise reduction model is as follows:

[0030] Masking module: At the beginning of each round of training, according to the preset masking ratio (default is 0.1, which can be dynamically adjusted according to the spectral noise level), randomly select the spectral data positions to generate a boolean mask, apply the mask to the input spectral data, set the spectral values at the positions corresponding to True in the mask to zero, and generate partially missing input data for training. Regenerate the mask in the next round of training to improve the robustness and generalization ability of the model.

[0031] Noise reduction encoding module: With 1D blind spot convolution as the core, change the traditional convolution kernel 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. Input the spectral data processed by the mask, and the target output is the original spectrum. Extract the local features of the spectral data through multiple layers of blind spot convolution. Use the ReLU activation function after each convolution to introduce non-linearity and enhance the model's expression ability. The output layer restores the spectral signal through a linear transformation and predicts the spectral values at the masked positions.

[0032] Design the loss function: Adopt a custom loss function that combines mean square error (MSE) and total variation (TV) regularization. , where Measure 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 predicted signal, avoid overfitting, and reduce high-frequency noise by constraining the gradients of adjacent spectral points. is the regularization coefficient, which balances the prediction accuracy and smoothness.

[0033] Step 3, set the parameters of the training process;

[0034] In Step 3, the specific method for setting the parameters of the training process is as follows:

[0035] Network structure parameters: The noise reduction 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 size of each convolution kernel is 3, and the ReLU activation function is used for both layers to capture local features in the sequence and introduce non - linearity. A custom blind - spot initializer and constraint conditions are introduced for each layer to ensure that the weights of the convolution kernels at specific positions are zero, realizing the blind - spot feature and enhancing the robustness of the model to noise.

[0036] 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 adopted, and the regularization coefficient is set to 0.001.

[0037] Training environment and epochs: Training is carried out on an NVIDIA 1650 GPU, and the number of training epochs is 15000.

[0038] Step 4: Perform model encapsulation and application;

[0039] In Step 4, the specific method for performing model encapsulation and application is as follows:

[0040] After completing the training of the noise reduction coding module, it is encapsulated and saved. In actual applications, users only need to input the original noisy LIBS spectral data into the trained noise reduction coding module to efficiently obtain the denoised data, providing a more accurate and reliable data basis for subsequent element analysis and quantitative detection.

[0041] The overall explanation of the above technical solution is as follows:

[0042] 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 noise reduction model based on the blind - spot network architecture.

[0043] Different from traditional supervised learning methods, the blind - spot network architecture trains on local masked data of the input, and then uses the original input data as the target data. Without noise - free target data, it can also achieve efficient noise reduction tasks.

[0044] The core idea of this method is: By randomly occluding some data points in the original spectral data, and then using the blind - spot convolution module (convolution kernel with a center point of 0) to infer the true values of the missing parts based on the remaining un - occluded data. This noise reduction training strategy enables the blind - spot network architecture to efficiently learn effective noise reduction features in the case of scarce data.

[0045] At the same time, the model 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 the training.

[0046] 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.

[0047] The noise reduction model established by the present invention includes the following modules:

[0048] Mask module: This module is used to generate partially masked input spectral data.

[0049] That is, at the beginning of each round of training, first randomly select positions in the spectral data for masking according to a preset masking ratio (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), and generate a boolean mask (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.

[0050] Subsequently, apply the generated boolean mask uniformly to all input spectral data, set the spectral values at the positions corresponding to True to zero, so as to generate partially missing input data for training.

[0051] In the next round of training, generate a new mask and repeat the above process. By dynamically updating the mask in this way, the model can be trained at different mask positions, thereby improving its robustness and generalization ability.

[0052] Noise reduction encoding module: The core of this module is 1D BlindSpot Convolution, that is, modifying the traditional convolution kernel into a BlindSpot convolution kernel with a center of 0.

[0053] This module consists of multiple 1D BlindSpot convolution layers and is designed specifically for processing one-dimensional spectral signals. The input of the module is the spectral data after masking processing, and the target output is the original spectrum without masking processing. Subsequently, local features in the spectral data are extracted through multiple layers of BlindSpot convolution.

[0054] The central weight of the BlindSpot convolution kernel is fixed at 0, ensuring that the central pixel does not participate in the calculation, so as to avoid the model directly copying the input data while forcing the model to infer and predict the "clean data" of the central point based on the data before and after the central point.

[0055] After the blind-spot convolution processing for each layer, the ReLU activation function is introduced to introduce non-linearity and enhance the model's expressive power. Finally, the output layer restores the spectral signal through a linear transformation, aiming to predict the spectral values at the masked positions. 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:

[0056] Loss function: The present invention uses a custom loss function that combines mean squared error (MSE) and total variation (TV) regularization, which is defined as follows:

[0057] , where represents the mean squared error (MSE), which measures the error between the denoised spectral data and the original spectrum, and is used to guide the model to learn the main objective of the denoising task and ensure that the data after denoising is not distorted. represents the total variation (TV) regularization, which is used to smooth the predicted signal and avoid overfitting.

[0058] By constraining the gradients of adjacent spectral points, the high-frequency noise in 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.

[0059] By minimizing this loss function, the model can more accurately restore the original spectral signal while maintaining the stability of the output.

[0060] The overall process of the present invention is shown as Figure 1 shown.

[0061] In the first stage of training, the masking ratio is first set to the default value (0.1).

[0062] Then, the masked data is used as the input data for the denoising encoding module. The denoising encoding module constructed in the present invention is a double-layer blind-spot convolution structure network.

[0063] This network gradually extracts the features of the input sequence through two convolutional layers. The first layer contains 64 convolutional kernels, and the second layer contains 128 convolutional kernels. Convolutional kernels of size 3 and the ReLU activation function are used to capture the local features in the sequence and introduce non-linearity.

[0064] For each layer, a custom blind-spot initializer and constraint conditions are introduced to ensure that the weights of the convolutional kernels at specific positions are forced to be zero, thereby achieving the blind-spot characteristic, avoiding directly transmitting information at certain positions, and enhancing the robustness of the model to noise. In addition, the network uses a custom total variation regularization loss function, combining the data fitting error and the regularization term (the regularization coefficient is 0.001).

[0065] The overall training process is optimized using the Adam optimizer, with an initial learning rate of 0.001, a training batch size of 16, and the final model is trained for 15,000 epochs on an NVIDIA 1650 GPU.

[0066] After the training of the noise reduction coding module of the present invention is completed, it is encapsulated and saved for subsequent application to the noise reduction prediction task of LIBS spectra. During the prediction process, users only need to input the original noisy LIBS spectral data into the trained noise reduction coding module to efficiently obtain the denoised data.

[0067] This process is simple and efficient, without additional complex operations or preprocessing steps, greatly improving the convenience and practicality of spectral noise reduction.

[0068] 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, thus providing a more accurate and reliable data basis for subsequent elemental analysis and quantitative detection.

[0069] Among them, by using the present invention and five existing traditional spectral noise reduction method models based on signal processing (Savitzky-Golay filtering, moving average filtering, wavelet transform, median filtering, and Gaussian filtering) to denoise the LIBS data of six stainless steel samples, and then comparing the denoising results of all methods.

[0070] As an experimental example, the present invention uses 180 original noisy LIBS data of six stainless steel samples (30 spectral data for each sample) for model training, and then uses the trained model to denoise 300 original noisy LIBS data of six stainless steel samples (50 spectral data for each sample and no overlap with the training data) to verify the performance of the present invention. For the comparison of the denoising results, the RSD index is used as the evaluation standard.

[0071] The relative standard deviation (RSD, Relative Standard Deviation) is a statistical index used to measure the degree of data dispersion and precision. Its calculation formula is as follows: , where is the standard deviation of the data, is the arithmetic mean of the data.

[0072] Figure 2 This is the comparison chart of the RSD index before and after denoising the noisy spectral data by the present invention. As Figure 2As shown, the relative standard deviations (RSDs) of the spectral data of six stainless steels were significantly reduced after noise reduction by the model. Among them, the RSD of the stainless steel data of category JZG206B decreased from 114.18% (before noise reduction) to 17.83% (after noise reduction), fully demonstrating the effectiveness and high efficiency of the present invention.

[0073] Then, based on the other 5 traditional spectral noise reduction method models (Savitzky-Golay filtering, moving average filtering, wavelet transform, median filtering, and Gaussian filtering), the same 300 original noisy LIBS data (the same as the data used when verifying the effect of the invention model above) were also processed for noise reduction. All the results are shown in Table 1, and the noise reduction performance of the traditional noise reduction methods (including Savitzky-Golay filtering, moving average filtering, wavelet transform, median filtering, and Gaussian filtering) and the present invention on one-dimensional spectral data was systematically evaluated.

[0074] Table 1 RSDs of six stainless steel data after being processed by different noise reduction methods;

[0075]

[0076] The above analysis is as follows:

[0077] Comparison of noise reduction capabilities: From the RSD change results of each noise reduction model for the six stainless steel data in Table 1, it can be intuitively seen that the noise reduction model in the present invention shows significant noise reduction effects on all six stainless steel sample data, and its noise reduction ability is significantly better than other traditional noise reduction methods.

[0078] For example, in the JZG204B sample, the RSD value after noise reduction by the noise reduction model in the present invention is 19.79%, while that of Gaussian filtering is 74.1%, an improvement of about 73%. In the JZG205 sample, the RSD value of the noise reduction 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), with a significant improvement.

[0079] From the overall average value, the average RSD value of the six stainless steel test set data after noise reduction by the noise reduction model in the present invention is 18.29%, significantly lower than that of other methods, and it is about 42% higher than the optimal wavelet transform in the traditional technology, which fully shows that the noise reduction effect of BSDN in different samples has a high degree of consistency.

[0080] Figure 3 It is a comparison chart of the noise reduction results of Savitzky-Golay filtering, moving average filtering, wavelet transform, median filtering, Gaussian filtering, and the noise reduction model in the present invention;

[0081] Comparison of fidelity: From Figure 3It can be found from the full-band visualization of the spectral noise reduction results of each noise reduction model in (a) that, compared with the two noise reduction models of the noise reduction model and wavelet transform in the present invention, there are serious distortion phenomena in the high-intensity bands of the other four models.

[0082] For example, in Figure 3 in the red marked box of (a), the spectral lines after Savitzky-Golay filtering, moving average filtering, median filtering and Gaussian filtering are significantly lower than the original spectral line (gray background spectral line).

[0083] Furthermore, by comparing the noise reduction model in the present invention with the wavelet transform that performs well in the high-intensity band, from Figure 3 the comparison of the spectral lines in the locally enlarged red marked box of (b), it can be clearly found that there is a distortion phenomenon caused by over-smoothing in the low-intensity band of the wavelet transform.

[0084] For example, in Figure 3 in the 4th red marked box of (b), there are 4 spectral peaks in the original spectral line, but only 2 peaks remain after wavelet transform, showing serious distortion, while BSDN well preserves these 4 peaks while reducing noise. The above results fully prove that the noise reduction model in the present invention has excellent noise reduction and fidelity both in the high-intensity band and the low-intensity band.

[0085] In contrast, through its adaptive learning ability, the noise reduction model in the present invention significantly improves the noise reduction effect while retaining the spectral detail features, especially suitable for situations with complex noise distribution or rich signal details. The results show that, whether in terms of the fidelity or efficiency of the noise reduction results, the noise reduction model in the present invention is superior to other traditional spectral noise reduction techniques.

[0086] In terms of noise reduction ability, by performing noise reduction processing on the LIBS data of six stainless steel samples in the present invention and comparing with 5 traditional spectral noise reduction methods, the results show that the noise reduction model in the present invention demonstrates strong noise reduction strength in all sample data.

[0087] Taking the JZG204B sample as an example, the RSD was as high as 88.46% before noise reduction, and dropped to 29.32% after noise reduction by this model, while Gaussian filtering could only reduce it to 74.10%. The improvement amplitude of this model is about 73%. From the overall mean value, the average RSD value of the six stainless steel test set data after noise reduction by this model is 19.83%, significantly lower than that of the wavelet transform (31.49%), which performs best among the traditional methods, and is about 42% higher than the wavelet transform. This fully proves the high consistency of its noise reduction effect in different samples, can effectively cope with complex and diverse noises, greatly improves the quality of spectral data, and provides a solid and reliable data basis for subsequent element analysis and quantitative detection.

[0088] In terms of fidelity, the noise reduction model in the present invention has obvious advantages. Compared with traditional noise reduction methods, there are serious distortion phenomena in the high-intensity band for the other four models. For example, the spectral lines after noise reduction by Savitzky-Golay filtering, moving average filtering, median filtering, and Gaussian filtering are significantly lower than the original spectral lines. Compared with wavelet transform, which performs better in the high-intensity band, wavelet transform has distortion caused by over-smoothing in the low-intensity band. For example, in a specific area, there are 4 spectral peaks in the original spectral line, but only 2 spectral peaks remain after wavelet transform. However, the noise reduction model in the present invention can effectively reduce noise while well retaining the spectral detail features in both high-intensity and low-intensity bands, 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.

[0089] 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, breaks through the limitations of traditional supervised learning methods, and reduces the difficulty and cost of data acquisition. It has strong adaptive learning ability and is applicable to various situations with complex noise distributions 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, greatly 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.

[0090] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for establishing a LIBS noise reduction model based on self-supervised learning, characterized in that Specifically, it 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. Randomly occlude some data points in the original spectral data, and use the blind spot convolution module to infer the true values of the missing parts according to the remaining unoccluded data, so as to achieve noise reduction in the case of no noise-free target data; Step 2, construct the noise reduction model: And the noise reduction model includes a masking module and a noise reduction encoding module; Step 3, set the training process parameters: Step 4, perform model encapsulation and application; In Step 2, the specific method for constructing the noise reduction model is as follows; Masking module: At the beginning of each round of training, according to the preset masking ratio, randomly select the spectral data positions to generate a boolean mask, apply the mask to the input spectral data, set the spectral values at the positions corresponding to True in the mask to zero, and generate partially missing input data for training. Generate a new mask in the next round of training; Noise reduction encoding module: With 1D blind spot convolution as the core, change the traditional convolution kernel to a blind spot convolution kernel with 0 in the center. The module consists of multiple 1D blind spot convolution layers and is used to process one-dimensional spectral signals. Input the spectral data processed by the mask, and the target output is the original spectrum. Extract the local features of the spectral data through multiple layers of blind spot convolution. Use the ReLU activation function to introduce non-linearity after each convolution. The output layer restores the spectral signal through a linear transformation and predicts the spectral values at the masked positions; Design loss function: A custom loss function that combines mean squared error and total variation regularization is adopted. , where measures the error between the denoised spectral data and the original spectrum. is used to smooth the predicted signal. is the regularization coefficient.

2. The method for establishing a LIBS noise reduction model based on self-supervised learning according to claim 1, wherein In Step 3, the specific method for setting the training process parameters is as follows: Network structure parameters: The noise reduction encoding module adopts a double-layer blind spot convolution structure network. Set 64 convolution kernels in the first layer and 128 convolution kernels in the second layer. The size of the convolution kernels is 3 for both layers. Use the ReLU activation function to capture the local features in the sequence and introduce non-linearity. A custom blind spot initializer and constraint conditions are introduced for each layer; Training optimization parameters: Use the Adam optimizer for optimization. Set the initial learning rate to 0.001, the training batch size to 16, and adopt a custom total variation regularization loss function. Set the regularization coefficient to 0.001; Training environment and epochs: Train on an NVIDIA 1650 GPU for 15000 epochs.

3. The method for establishing a LIBS noise reduction model based on self-supervised learning according to claim 1, characterized in that In Step 4, the specific method for performing model encapsulation and application is as follows: After completing the training of the noise reduction encoding module, encapsulate and save it. In actual application, the user inputs the original noisy LIBS spectral data into the trained noise reduction encoding module.

Images