A method and system for spectral transfer between different resolution LIBS devices

Abstract

The present application provides a spectrum transfer method and system between different resolution LIBS devices, comprising: obtaining first resolution spectrum data of a target sample obtained by a first LIBS device; obtaining second resolution spectrum data of the target sample based on a second LIBS device according to the obtained low resolution spectrum data and a trained spectrum correction model; wherein the spectrum correction model comprises: a residual dense network based on an attention mechanism, used for extracting shallow layer features and deep layer features of the first resolution spectrum data and fusing global features; and a learnable up-sampling, used for obtaining the second resolution spectrum data according to the predicted weight of the network and the global features. By constructing the spectrum correction model, the difference between the spectra obtained by different resolution instruments can be successfully corrected for laser-induced breakdown spectroscopy data transfer of different resolution instruments.

Description

Technical Field

[0001] This invention belongs to the field of laser-induced breakdown spectroscopy analysis technology, and particularly relates to a method and system for spectral transfer between LIBS devices of different resolutions. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Laser-induced breakdown spectroscopy (LIBS) has seen rapid development in civilian, military, and aerospace fields due to its long-range detection, high efficiency, and ability to modify sample surfaces (removing surface dust and probing internal sample points). LIBS technology has been used as a core payload to identify the composition of rocks and soil on Mars in the following missions: the ChemCam instrument on the Curiosity rover (launched in 2012), the SuperCam instrument on the Perseverance rover (launched in 2020), and the Zhurong lander on China's Mars 2020 probe (launched in 2020) with the MarsCoDe instrument.

[0004] However, in the practical application of LIBS technology, there is a problem: due to differences in instrument laser intensity, resolution, and experimental environment conditions, the quantitative analysis model established on the original spectrometer cannot be directly applied to another new instrument. The inability of the spectral model to be universal between different instruments and to adapt to changes in instruments and experimental conditions has become a major obstacle to the large-scale application of this technology. The solution to this problem is called calibration transfer in the field of spectroscopy. Calibration transfer can be generally divided into: (1) transfer of the model, i.e., model transfer learning, which involves training the model on the original instrument dataset and then retraining the model with some frozen structures using data from the new instrument to change some weights of the model structure; (2) transfer of data, which involves transferring the spectral data from the new instrument to the original instrument spectrum and then using the model of the original instrument for analysis.

[0005] Existing research methods mostly address the model transfer problem between spectra at the same resolution, but there is still a lack of in-depth research on the transfer of spectra between spectrometers with different resolutions. Furthermore, in spectral super-resolution within the LIBS field, the spectral resolutions of master and slave instruments differ, presenting challenges such as limited sample size and high dimensionality. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a method and system for spectral transfer between LIBS devices of different resolutions. By constructing a spectral correction model for the transfer of laser-induced breakdown spectral data from instruments of different resolutions, the differences between spectra obtained by instruments of different resolutions can be successfully corrected.

[0007] To achieve the above objectives, a first aspect of the present invention provides a method for spectral transfer between LIBS devices of different resolutions, comprising:

[0008] Acquire the first-resolution spectral data of the target sample obtained through the first LIBS device;

[0009] Based on the acquired low-resolution spectral data and the trained spectral correction model, the second-resolution spectral data of the target sample obtained based on the second LIBS device are obtained.

[0010] The spectral correction model consists of an attention-based residual dense network and a learnable upsampling layer. The attention-based residual dense network is used to extract global features from the fusion of shallow and deep features of the first-resolution spectral data. The learnable upsampling layer is used to obtain second-resolution spectral data by combining the global features with the weights predicted by the network.

[0011] A second aspect of the present invention provides a spectral transfer system between LIBS devices of different resolutions, comprising:

[0012] Acquisition module: used to acquire the first resolution spectral data of the target sample obtained by the first LIBS device;

[0013] Spectral data correction module: used to obtain the second resolution spectral data of the target sample based on the acquired low-resolution spectral data and the trained spectral correction model;

[0014] The spectral correction model comprises an attention-based residual dense network and a learnable upsampling layer. The attention-based residual dense network is used to extract global features from the fusion of shallow and deep features of the first-resolution spectral data. The learnable upsampling layer is used to obtain the second-resolution spectral data by combining the global features with the weights predicted by the network.

[0015] A third aspect of the present invention provides a computer device comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor communicates with the memory via the bus, and the machine-readable instructions, when executed by the processor, perform a spectral transfer method between LIBS devices of different resolutions.

[0016] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs a method for spectral transfer between LIBS devices of different resolutions.

[0017] The above one or more technical solutions have the following beneficial effects:

[0018] In this invention, a spectral correction model is constructed for the transfer of laser-induced breakdown spectral data from instruments with different resolutions. The spectrum of the first-resolution LIBS instrument is transferred to the second-resolution spectral calibration, its resolution spectral information is reconstructed, and transferred to a second-resolution LIBS instrument with a complete and maintained calibration model, eliminating the expensive and time-consuming recalibration.

[0019] In this invention, the spectral correction model includes a residual dense network based on an attention mechanism and learnable upsampling. By incorporating the residual dense network based on an attention mechanism and learnable upsampling to reconstruct spectral data at the first resolution, the detailed information lost in the low-resolution spectrum can be reconstructed to the maximum extent, thereby improving the transfer effect between spectral data of different resolutions.

[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This refers to the spectral calibration transfer process in Embodiment 1 of the present invention;

[0023] Figure 2 This is a comparison of the spectra of the same sample from ChemCam and SuperCam in Embodiment 1 of the present invention;

[0024] Figure 3 This is a diagram of the residual dense network structure in Embodiment 1 of the present invention;

[0025] Figure 4(a) shows the effect of residual dense network prediction of Basalt (andesitic) sample SiⅠ 624.554nm without adding attention mechanism in Embodiment 1 of the present invention;

[0026] Figure 4(b) shows the effect of the residual dense network with added attention mechanism in Example 1 of the present invention on the prediction of Basalt (andesitic) sample SiⅠ 624.554nm transfer.

[0027] Figure 4(c) shows the predicted transfer effect of the spectral correction model on the SiⅠ sample at 624.554 nm in Embodiment 1 of the present invention.

[0028] Figure 5 This is a network structure diagram of the spectral correction model in Embodiment 1 of the present invention;

[0029] Figure 6 This is a schematic diagram of the position projection weight prediction of ChemCam to SuperCam in the wavelength range of 537.57nm to 852.77nm in Embodiment 1 of the present invention.

[0030] Figure 7(a) is a comparison of the SuperCam and ChemCam spectra of the Diopside sample in Example 1 of the present invention at 537.69 nm-852.77 nm.

[0031] Figure 7(b) is a comparison of the SuperCam spectrum and the spectrum after CNN transfer of the Diopside sample in Example 1 of the present invention in the range of 537.69 nm to 852.77 nm.

[0032] Figure 7(c) is a comparison of the SuperCam spectrum and the spectrum after RDN transfer of the Diopside sample in Example 1 of the present invention at 537.69 nm-852.77 nm.

[0033] Figure 7(d) is a comparison of the SuperCam spectrum and the spectrum after MetaRDN transfer of the Diopside sample in Example 1 of the present invention in the range of 537.69 nm to 852.77 nm.

[0034] Figure 8 This is a graph showing the predicted total integral intensity and the actual total integral intensity of the test set samples in Embodiment 1 of the present invention.

[0035] Figure 9 This is a performance comparison chart of CNN, RDN, and MetaRDN networks in Embodiment 1 of the present invention;

[0036] Figure 10 This is the SuperCam data enhancement process in Embodiment 1 of the present invention. Detailed Implementation

[0037] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0038] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0039] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0040] Example 1

[0041] like Figure 1 As shown, this embodiment discloses a method for spectral transfer between LIBS devices of different resolutions, including:

[0042] Acquire the first-resolution spectral data of the target sample obtained through the first LIBS device;

[0043] Based on the acquired low-resolution spectral data and the trained spectral correction model, the second-resolution spectral data of the target sample obtained based on the second LIBS device are obtained;

[0044] The spectral correction model consists of an attention-based residual dense network and a learnable upsampling layer. The attention-based residual dense network is used to extract global features from the fusion of shallow and deep features of the first-resolution spectral data. The learnable upsampling layer is used to obtain second-resolution spectral data by combining the global features with the weights predicted by the network.

[0045] This embodiment uses the laboratory simulated Mars environment dataset from two instruments, ChemCam (the slave instrument) and SuperCam (the master instrument), to illustrate the transmission of spectral data.

[0046] In this embodiment, the LIBS spectra obtained by the SuperCam team from 325 standard samples were used as the master instrument spectra, and the LIBS spectra obtained by the ChemCam team from 408 standard samples were used as the slave instrument spectra.

[0047] ChemCam Dataset: The physical forms of the samples were powder, glass, and ceramic. The powders were crushed and pressed into cakes, with a sample weight of approximately 3.5 g. Each prepared sample was then placed in a vacuum chamber filled with carbon dioxide to simulate the atmospheric pressure of Mars. A 1067 nm laser with a pulse energy of 14 mJ and a repetition rate of 3 Hz was used. The laser beam was projected onto a 1.56 m target through a Schmidt-Cassegrain telescope, and the returned plasma emission signal was collected. The laser created five different ablation pits on each sample, with each point ablated 50 times. LIBS spectra were obtained, resulting in a total of 250 spectra per sample.

[0048] The SuperCam dataset: In 2020, NASA launched the Perseverance rover on Mars, and the SuperCam instrument cluster on board was used for LIBS experiments and exploration. SuperCam's LIBS characteristics are very similar to ChemCam's, but SuperCam adds the ability to acquire spectrally resolved data in the 536–900 nm spectral range; that is, in this band, SuperCam acquires spectra with higher spectral resolution than ChemCam. The ultraviolet (UV) and violet-visible (VIO) spectral ranges on both SuperCam and ChemCam are covered by almost the same Czerny Turner spectrometer and unprocessed detectors. Unlike ChemCam, which uses a third Czerny Turner spectrometer, SuperCam uses a transmission spectrometer with gated intensifiers to cover the approximately 535–850 nm range, increasing the spectral resolution acquisition capability in this range. On SuperCam, plasma needs to be collected and analyzed between 245nm and 853nm. Apart from the two wavelength gaps between 340-385nm and 465-536nm imposed by the optical design, the full width at half maximum (FWHM) resolution below 500nm is better than 0.2nm, and the FWHM resolution above 500nm is better than 0.65nm.

[0049] like Figure 2 As shown, the laser characteristics determine the breakdown process generated by the LIBS plasma, such as the laser pulse length and wavelength, the properties of the sample itself, the sample's absorption of light at different laser wavelengths, the thermodynamic properties of the sample, and the pressure of the surrounding gas on the sample surface. Due to these differences, even samples with the same chemical composition will produce different spectra when acquired by SuperCam and ChemCam LIBS instruments, which hinders the accuracy and stability of quantitative analysis. In terms of spectral range, SuperCam and ChemCam share the following common spectral bands: 283.83 nm–341.36 nm (UV), 382.14–464.54 nm (VIO), and 537.57 nm–852.77 nm (VNIR).

[0050] This embodiment selected 275 identical samples, which can be broadly categorized into three types: silicate minerals, carbonate minerals, and sulfate minerals, totaling 1650 spectra. The Kennard-Stone algorithm was used to partition the sample dataset, with 260 samples in the training set and 15 samples in the test set. The spectra were obtained from the Planetary Data System website, and the three spectral regions recorded by the three independent detectors were normalized. Table 1 shows the corresponding content ranges (in wt%) of the main elements in the samples included in the training and test sets.

[0051] Table 1

[0052]

[0053] This embodiment provides a single-image super-resolution network, a Residual Dense Network (RDN) combined with a Convolutional Block Attention Module (CBAM) as a feature extraction module. The Residual Dense Network mainly consists of four parts: a shallow feature extraction layer, Residual Dense Blocks (RDBs), Dense Feature Fusion (DFF), and finally an upsampling layer (UPNet).

[0054] like Figure 3 and Figure 5 As shown, this embodiment uses the first three parts of a residual dense network: a shallow feature extraction layer including a first convolutional layer and a second convolutional layer for shallow feature extraction; multiple sequentially connected residual dense blocks (RDBs), each RDB having 3 convolutional layers; and dense feature fusion (DFF).

[0055] Low-resolution spectra are extracted using shallow features through the first two convolutional layers, with a CBAM attention layer following each convolutional layer to enhance focus on features of lower-intensity spectral lines. The output features from the CBAM attention layer after the second convolutional layer are input into the RDB module. The RDB module consists of three convolutional layers and a ReLU activation function, with a 1*1 bottleneck layer added at the end. The dense structure of the RDB and local residual connections are used to extract local features. Finally, the outputs of multiple RDB modules are concatenated, dimensionality is reduced using the bottleneck layer, and then further extracted into depth features through a convolutional layer, outputting the depth features. The depth features and the shallow features (i.e., the output of the attention module after the first convolutional layer) are residually concatenated through global residual learning to obtain the final output global feature F. DF .

[0056] The residual dense network fully utilizes all hierarchical features of the original low-resolution LR spectrum; the residual dense block (RDB) not only reads the state from the preceding RDB through the continuous memory (CM) mechanism, but also makes better use of all layers within it through local dense connections; then, accumulated features are adaptively preserved through local feature fusion (LFF). Dense feature fusion (DFF) adaptively fuses the hierarchical features of all RDBs in the LR space. Through global residual learning, shallow and deep features are combined to obtain global dense features from the original LR spectrum.

[0057] In neural networks, network capacity and complexity are directly proportional; storing more information increases network complexity, leading to a significant increase in network parameters. Therefore, attention mechanisms are used to improve representational capabilities: focusing on important features and suppressing unnecessary ones.

[0058] Convolutional Block Attention (CBAM) modules are added after the first and second convolutional layers used for shallow feature extraction. CBAM is a simple yet effective attention module for feedforward convolutional neural networks, consisting of spatial attention and channel attention. Given an intermediate feature map, the module infers attention maps sequentially along two independent dimensions (channel and spatial), and then multiplies the attention maps with the input feature map for adaptive feature refinement.

[0059] The inputs to the channel attention layers pass through parallel max pooling and average pooling layers. The outputs from the max pooling and average pooling layers are then summed after passing through a multilayer perceptron. Finally, the summed results are passed through an activation function to obtain the channel attention weights.

[0060] The spatial attention module consists of a max pooling layer, an average pooling layer, and a 7×7 convolutional layer connected in sequence.

[0061] Channel attention allows a CNN to learn weights at each channel to better capture useful features in the input. This is achieved by performing global average pooling and global max pooling on each channel, followed by passing the input through a multilayer perceptron to compute an importance score for each channel. Spatial attention allows a CNN to learn weights at each spatial location to better capture spatial features in the input. This is achieved by performing two convolution operations on the input: the first convolution produces channel feature maps, and the second convolution produces spatial feature maps. Then, a multilayer perceptron is used to compute an importance score for each spatial location to produce the final weight matrix.

[0062] Combining channel attention and spatial attention creates the CBAM attention mechanism. CBAM can be integrated into existing CNN architectures to improve their performance and accuracy.

[0063] Attention mechanisms can be summarized as follows:

[0064]

[0065]

[0066] F is the output of the convolutional layer in the shallow feature extraction layer. As input, CBAM sequentially infers a 1D channel attention map M. c A 2D spatial attention map M sF″ is the final output feature map.

[0067] Because CBAM is a lightweight, general-purpose module, it can be seamlessly integrated into any CNN architecture with negligible overhead, and can be trained end-to-end with the base CNN.

[0068] Adding the CBAM attention layer increases the attention given to spectral lines with lower intensity, resulting in better transmission performance. Figures 4(a)-4(c) The transfer effect of Basalt (andesitic) samples in the 607-643 nm range was shown. The RDN with added CBAM achieved better correction for the less intense peaks in a certain spectrum.

[0069] When faced with the problem of reconstructing information from low-resolution spectra and transferring it to high-resolution spectra, the prior knowledge of high-dimensional LIBS spectra is often extremely complex, and its degradation is often unknown. Furthermore, directly learning the mapping from low-dimensional space to high-dimensional space will increase the difficulty.

[0070] In super-resolution problems, upsampling is involved, which in turn involves the upsampling method and its placement within the network. For example, Dong et al. first employed a pre-upsampling SR framework. Specifically, they upsampled LR images to coarse HR images of the desired size using traditional methods (e.g., bicubic interpolation), and then applied deep CNNs to these images to reconstruct high-quality details. Since the most difficult upsampling operation was performed, the CNN only needed to refine the coarse images, significantly reducing the learning difficulty. However, upsampling the input data to the target data size before feeding it into the model increases computation, leading to longer training times. Furthermore, traditional interpolation methods for upsampling only improve resolution based on their own signal without providing more information. Instead, they often introduce side effects such as computational complexity, noise amplification, and spectral peak shift. To improve computational efficiency and fully utilize deep learning techniques to automatically improve resolution, researchers suggest performing most of the computation in a low-dimensional space, replacing predefined upsampling with end-to-end learnable layers integrated at the end of the model.

[0071] The model in this embodiment adopts a post-upsampling structure, adding an end-to-end learnable upsampling module suitable for spectral data at the end of the model.

[0072] The fourth part of RDN uses sub-pixel convolutional layers for upsampling, which is an image-specific upsampling method and performs upsampling at integer multiples. This is unsuitable for upsampling non-integer multiples from low to high resolution using one-dimensional spectral features. Hu et al. proposed a meta-upsampling method that addresses arbitrary scaling factors. This module dynamically predicts the weights of the magnification filters based on the input scaling factor and then uses these weights to generate HR images of arbitrary sizes. For a low-resolution image, Meta-SR can magnify it by any factor using only one model.

[0073] This embodiment employs bicubic interpolation to extract the features F from the RDN. DF Extending to the dimension of high-resolution spectra, we obtain FLR(I′).

[0074] For spectral data, this embodiment simplifies the weight prediction process. This process uses a convolutional layer and a ReLU activation function to project the input position and predict its weight features.

[0075] Position projection: R(i) is the relative offset of the high-resolution channel relative to the low-resolution channel. The input spectral data is the feature F of the low-resolution spectrum extracted by the residual dense network. DF Each data point is indexed as i, and its position is projected to find i′ on the low-resolution spectrum. Data point i is determined by the characteristics of i′ in the low-resolution spectrum. The following projection operator is used to map these two data points:

[0076]

[0077] Where T is the transformation function. It's a floor function, r is the scaling factor.

[0078] Weight prediction: After projection, the corresponding weights need to be calculated. R(i) is used as the input to the weight convolutional layer, representing the relative offset of the high-resolution channel relative to the low-resolution channel, as shown in the following formula:

[0079]

[0080] The network predicts the convolutional kernel weights and concatenates the weight features (the weights assigned to hr feature points by lr feature points) with FLR(I′) through channels. The weight features refer to the output of R(i) after passing through the convolutional layer and ReLU. FLR(I′) is the feature F of the low-resolution spectrum. DF The FLR(I′) obtained is obtained by bicubic interpolation. The bicubic interpolated FLR(I′) is concatenated with the weighted feature channels to obtain FLR′(I′).

[0081] Figure 6This diagram illustrates the combination of position projection and weight prediction. In the diagram, 'w' represents the weight features obtained above, which are assigned to FLR(I') during channel concatenation. The convolutional layer then combines the two features to complete the mapping.

[0082] Feature mapping: A 1×3 convolutional layer and a fully connected layer further extract high-frequency features such as feature spectral peaks from the fused feature FLR′(I′). Finally, a fully connected layer and a ReLU activation function are used to complete the mapping from low resolution to high resolution.

[0083] Figure 4(c) shows the spectrum of the Basalt (andesitic) sample transmitted by MetaRDN. Compared with the RDN feature extraction network that only incorporates CBAM, MetaRDN not only solves the noise fluctuation problem of the global spectrum, but also achieves better correction effect for the smaller intensity spectral peaks of SiⅠ.

[0084] To address the spectral super-resolution problem in the LIBS field, where the spectral resolutions of master and slave instruments differ, and the challenges of limited samples and high dimensionality, this embodiment employs an RDN neural network based on the CBAM attention mechanism combined with a learnable upsampling layer (MetaRDN) to transfer spectra at different resolutions, predicting SuperCam spectra from ChemCam spectra.

[0085] Spectral transfer results for CNN, RDN, and MetaRDN:

[0086] Figure 7(a) shows the SuperCam and ChemCam spectra of the diopside sample in the 537.69nm-852.77nm range. First, a CNN was used to calibrate and transfer the SuperCam and ChemCam standard datasets, as shown in Figure 7(b). Since the CNN network has few layers and cannot effectively extract features, the example used RDN and MetaRDN to perform the same experiments, obtaining the calibrated and transferred results for the diopside sample as shown in Figure 7(b). Figures 7(c)-7(d) As shown in Figure 7(c), the inset plot illustrates the resolution improvement from a single peak at 589.158 nm for NaⅠ to a double peak at 589.158 nm and 589.755 nm for NaⅠ after calibration transfer.

[0087] To compare the correction effects and explain the intensity variations between the compared spectra at each wavelength point, the concept of intensity error is used, such as... Figure 8 The figure shows the relationship between the total integrated intensity and the actual integrated intensity of the ChemCam spectra of the test set samples after transfer via MetaRDN. In this study, the intensity difference is equal to the sum of the intensity differences between the predicted and actual data at each wavelength point compared to the sum of the intensities of the actual spectra. The formula is:

[0088]

[0089] For the same sample spectral datasets from ChemCam and SuperCam, since both the master and slave instrument spectra are divided into three segments—violet-visible, ultraviolet, and visible-near-infrared—we also divided the spectra into these three segments for transmission. This involved three experiments, transmitting the corresponding master and slave instrument spectral segments, and discarding wavelengths that did not overlap between the master and slave instruments. The wavelength ranges of the three overlapping spectral segments are listed in the header of Table 2. Table 2 shows the comparison of the transmitted spectra with the original spectra in terms of intensity error and Pearson correlation coefficient between the CNN, RDN, and MetaRDN algorithms when dealing with the same sample spectral datasets from ChemCam and SuperCam. The Pearson correlation coefficient finds the ratio between the covariance and standard deviation of two objects. Mathematically, it can be described as follows:

[0090]

[0091] The results listed in Table 2 clearly show that the prediction error using the MetaRDN algorithm is significantly lower than that of the RDN and CNN algorithms, and according to the quantitative analysis results below, the performance of the RDN algorithm is better than that of CNN.

[0092] Table 2 compares the relative errors between the results obtained from the quantitative spectral analysis after the CNN, RDN, and MetaRDN test sets are transferred and the results obtained from the SuperCam spectral prediction.

[0093] Table 2:

[0094]

[0095]

[0096] Quantitative analysis results of the transferred spectra: Quantitative analysis was performed using a CNN regression model, built with the Inception V2 architecture, which allows for increasing network depth and width. This model was trained using SuperCam spectral data from 260 samples in the training set and tested using 15 samples in the test set. Simultaneously, LIBS spectral data from 15 transferred ChemCam samples were used as the test set to determine the effectiveness of the post-transfer quantitative analysis. Figures 7(a)-7(d) As shown, after transferring the test set data using MetaRDN, the performance of the transferred spectra in quantitative analysis is close to that of the SuperCam spectra. Table 3 presents the quantitative analysis results and relative root mean square errors of the spectra transferred from the instrument to the main instrument, as shown in the following formulas:

[0097]

[0098] The root mean square error relative errors are: Si: 7.59%, Ti: 15.69%, Al: 20.59%, Fe: 31.67%, Mg: 6.56%, Ca: 9.93%, Na: 6.88%, K: 15.47%.

[0099] Table 3 Comparison of the root mean square error of predictions on the CNN quantitative analysis model between the SuperCam test set and the ChemCam test set transmitted by MetaRDN;

[0100] Table 3

[0101] Testing sets RMSEP_SC RMSEP_CT Relative_rmse <![CDATA[SiO2]]> 3.8906 4.1859 7.59％ <![CDATA[TiO2]]> 0.4577 0.5295 15.69％ <![CDATA[Al2O3]]> 2.5535 3.0792 20.59％ <![CDATA[FeO T ]]> 2.4312 3.2012 31.67％ MgO 1.5328 1.6334 6.56％ CaO 1.1002 0.9909 9.93％ <![CDATA[Na2O]]> 1.1192 1.1962 6.88％ <![CDATA[K2O]]> 1.2762 1.4736 15.47％

[0102] Furthermore, this embodiment also uses the quantitative analysis model to perform performance testing and comparison on the spectra transferred by the three calibration transfer models, and the results are as follows: Figure 9 As shown in Table 2, the evaluation metric used for comparison is the relative root mean square error. Although the propagation performance of CNN on the SuperCam and ChemCam datasets is similar to that of RDN, the root mean square error of the quantitative results is unsatisfactory. MetaRDN, with the addition of CBAM and a learnable upsampling layer, achieved excellent propagation performance and quantitative analysis results. Except for the prediction root mean square error of element Al, which is 20.59% higher than the original value on the SuperCam test set, and Fe, which is 31.67% higher, the prediction root mean square error of other elements is close to the original value on the SuperCam test set.

[0103] SuperCam Data Augmentation: The data augmentation process is as follows Figure 10 As shown, the spectral data of 260 ChemCam samples used to train the transfer model are first used to predict the transferred data X using the trained MetaRDN model. CT Then X CT With X SC Combined, they form the enhanced data X DA Using X DARetraining the CNN quantitative model can expand the size and diversity of the training dataset, improve the model's generalization ability and robustness, reduce overfitting, and thus effectively improve the performance of the convolutional neural network. Table 4 shows the quantitative prediction results of the SuperCam test set after data augmentation for model retraining. The bolded negative values ​​in the last column represent the relative error reduction in the root mean square error of the element content prediction after data augmentation compared to the original values ​​in the SuperCam test set. In the results, after the ChemCam data calibration transfer, the method of expanding the SuperCam training data had a negative effect on Si and Al, resulting in a decrease in their prediction accuracy. The negative effect on Si was more obvious. In Table 1, the SiO2 content ranged from 0.02 to 97.71 wt% in the entire dataset, and the content varied greatly. Compared with other elements, the data augmentation on the training set was not diversified enough compared to the test set. Data augmentation was only performed on the training set, while the test set was not augmented accordingly, resulting in an inconsistent distribution between the training and test sets, which affected the model's generalization ability and prediction results. The Fe content ranges from 0.04 to 64.85 wt%, with an average of 7.05 wt% and a median of 6.08 wt%. Although the content data has a wide distribution range, the data is concentrated in a small content range with small content variations. This means that even though the calibration transfer effect for Fe is not as good as that for Si, the data enhancement can still improve the quantitative analysis results.

[0104] Table 4 compares the root mean square error of prediction on the original CNN quantitative analysis model with the root mean square error of prediction on the quantitative analysis model after data augmentation and retraining.

[0105] Table 4:

[0106]

[0107] In this embodiment, a MetaRDN calibration transfer model was constructed for transferring laser-induced breakdown spectral data from instruments with different resolutions. Calibration results using a common standard dataset from ChemCam and SuperCam demonstrate that this method can successfully correct differences between spectra obtained from instruments with different resolutions. The model is effective in calibrating the spectra themselves and, by inputting the corrected spectra into the quantitative analysis model, yields results close to those of the primary instrument. After using the transferred ChemCam training set spectra for training set data augmentation in the SuperCam quantitative model, except for Al2O3 and CaO errors which showed no significant change and SiO2 prediction error which increased, the root mean square errors of prediction for other major elements were significantly improved. This ensures that the method proposed in this embodiment can transfer the spectrum of a low-resolution instrument to a high-resolution spectral calibration, reconstruct its high-resolution spectral information, and transfer it to a high-resolution instrument with a well-maintained calibration model. This eliminates the costly and time-consuming recalibration and, in turn, can be used for retraining the high-resolution instrument model, improving its predictive performance.

[0108] In practical applications of laser-induced breakdown spectroscopy (LIBS), quantitative analysis models established on one spectrometer cannot be directly applied to another new instrument due to differences in laser intensity, resolution, and experimental conditions. Low-resolution spectrometers reduce the signal-to-noise ratio, blur spectral and characteristic peak details, leading to the synthesis of multiple peaks into a single envelope peak. The problem of transferring spectra at different resolutions is similar to the super-resolution problem in image processing. With the rapid development of deep learning technology, deep learning-based super-resolution models have been actively explored and have achieved state-of-the-art performance on various super-resolution benchmarks. In this work, a residual dense network from the field of image super-resolution is used, combined with a convolutional attention mechanism module and an upsampling layer of arbitrary scale, to transfer the same standard samples from the ChemCam carried by Curiosity and the SuperCam carried by Perseverance. This transfer algorithm has excellent transfer performance; the Pearson correlation coefficient between the transferred spectrum and the SuperCam spectrum is 0.9901. Quantitative analysis and evaluation of the transferred test set sample spectra were performed using a convolutional neural network. The relative errors of the root mean square error (RMSE) of the ChemCam transferred spectra compared to the RMSE of the SuperCam test set were: Si: 7.59%, Ti: 15.69%, Al: 20.59%, Fe: 31.67%, Mg: 6.56%, Ca: 9.93%, Na: 6.88%, K: 15.47%. The relative errors of the major elements' transferred spectra from ChemCam to SuperCam were relatively small. After transferring the ChemCam training set used for the transfer model to expand the SuperCam quantitative model training set, the major elements that showed significant improvements in the quantitative analysis performance of SuperCam test samples were: Ti - 44.66% reduction in predicted RMSE, Fe - 36.13% reduction, Mg - 54.64% reduction, Na - 69.44% reduction, and K - 44.25% reduction.

[0109] This embodiment constructs a MetaRDN calibration transfer model for transferring laser-induced breakdown spectral data from instruments with different resolutions. Calibration results using a common standard dataset from ChemCam and SuperCam demonstrate that this method can successfully correct differences between spectra obtained from instruments with different resolutions. The model is effective in calibrating the spectra themselves, and when the corrected spectra are input into the quantitative analysis model, results close to those obtained from the primary instrument are obtained. After using the transferred ChemCam training set spectra for training set data augmentation in the SuperCam quantitative model, except for Al2O3 and CaO errors which show no significant change and SiO2 prediction error which increases, the root mean square errors of prediction for other major elements are significantly improved. This ensures that the method proposed in this study can transfer the spectrum of a low-resolution instrument to a high-resolution spectral calibration, reconstruct its high-resolution spectral information, and transfer it to a high-resolution instrument with a well-maintained calibration model. This eliminates the costly and time-consuming recalibration and, in turn, can be used for retraining the high-resolution instrument model, improving the model's predictive performance.

[0110] Example 2

[0111] The purpose of this embodiment is to provide a spectral transfer system between LIBS devices of different resolutions, including:

[0112] Acquisition module: used to acquire the first resolution spectral data of the target sample obtained by the first LIBS device;

[0113] Spectral data correction module: used to obtain the second resolution spectral data of the target sample based on the acquired low-resolution spectral data and the trained spectral correction model;

[0114] The spectral correction model consists of an attention-based residual dense network and a learnable upsampling layer. The attention-based residual dense network is used to extract global features from the fusion of shallow and deep features of the first-resolution spectral data. The learnable upsampling layer is used to obtain second-resolution spectral data by combining the global features with the weights predicted by the network.

[0115] Example 3

[0116] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method.

[0117] Example 4

[0118] The purpose of this embodiment is to provide a computer-readable storage medium.

[0119] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the above method.

[0120] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0121] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0122] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for spectral transfer between LIBS devices of different resolutions, characterized in that, include: Acquire the first-resolution spectral data of the target sample obtained through the first LIBS device; The first resolution spectral data is low resolution spectral data; Based on the acquired low-resolution spectral data and the trained spectral correction model, the second-resolution spectral data of the target sample obtained using the second LIBS device is obtained; the second-resolution spectral data is high-resolution spectral data. The spectral correction model consists of an attention-based residual dense network and a learnable upsampling layer. The attention-based residual dense network is used to extract global features from the fusion of shallow and deep features of the first-resolution spectral data. The learnable upsampling layer is used to obtain second-resolution spectral data by combining the global features with the weights predicted by the network. Learnable upsampling is used to obtain second-resolution spectral data by combining the global features with the weights predicted by the network, specifically as follows: The global features obtained by the attention-based residual dense network are upsampled using bicubic interpolation. The global features obtained by the attention-based residual dense network are mapped to one-dimensional spectral data and processed by weighted convolutional layers and activation functions to obtain weighted features. The weighted convolution and the result of upsampling are concatenated by channels; The channel stitching results are sequentially passed through a convolutional layer, a fully connected layer, and an activation function to obtain the second-resolution spectral data.

2. The spectral transfer method between LIBS devices of different resolutions as described in claim 1, characterized in that, The attention-based residual dense network includes a shallow feature extraction layer that incorporates the attention mechanism, multiple residual dense blocks, and dense feature fusion. The shallow feature extraction layer of the fusion attention mechanism is used to extract shallow features from the first resolution spectral data. The residual dense block is used to extract local features based on the shallow features. The dense feature fusion is used to fuse the local features extracted from multiple residual dense blocks and the shallow features to obtain global features.

3. The spectral transfer method between LIBS devices of different resolutions as described in claim 2, characterized in that, The shallow feature extraction layer of the fusion attention mechanism is used to extract shallow features from the first resolution spectral data, specifically including: The output of the first resolution spectral data after passing through the first convolutional layer is used as the input of the first convolutional block attention module; The output of the attention module of the first convolutional block is used as the input of the second convolutional layer; The output of the second convolutional layer is used as the input to the second convolutional block attention module, which outputs the shallow features of the first resolution spectral data.

4. The spectral transfer method between LIBS devices of different resolutions as described in claim 2, characterized in that, The local features extracted from the multiple residual dense blocks are fused through a fully connected layer. The fused feature result is then passed through a convolutional layer and combined with the shallow features through global residual learning to obtain global features.

5. The spectral transfer method between LIBS devices of different resolutions as described in claim 1, characterized in that, Also includes: The second-resolution spectral data were analyzed using the quantitative analysis model of the second LIBS device to obtain the analysis results.

6. The spectral transfer method between LIBS devices of different resolutions as described in claim 1, characterized in that, The first LIBS device shown is ChemCam on the Mars rover; the LIBS device is SuperCam on the Mars rover.

7. A spectral transfer system between LIBS devices of different resolutions, characterized in that, include: Acquisition module: used to acquire the first resolution spectral data of the target sample obtained by the first LIBS device; the first resolution spectral data is low resolution spectral data; Spectral data correction module: used to obtain second-resolution spectral data of the target sample based on the acquired low-resolution spectral data and the trained spectral correction model; the second-resolution spectral data is high-resolution spectral data; The spectral correction model consists of an attention-based residual dense network and a learnable upsampling layer. The attention-based residual dense network is used to extract global features from the fusion of shallow and deep features of the first-resolution spectral data. The learnable upsampling layer is used to obtain second-resolution spectral data by combining the global features with the weights predicted by the network. Learnable upsampling is used to obtain second-resolution spectral data by combining the global features with the weights predicted by the network, specifically as follows: The global features obtained by the attention-based residual dense network are upsampled using bicubic interpolation. The global features obtained by the attention-based residual dense network are mapped to one-dimensional spectral data and processed by weighted convolutional layers and activation functions to obtain weighted features. The weighted convolution and the result of upsampling are concatenated by channels; The channel stitching results are sequentially passed through a convolutional layer, a fully connected layer, and an activation function to obtain the second-resolution spectral data.

8. A computer device, characterized in that, include: The computer device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform a spectral transfer method between LIBS devices of different resolutions as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs a spectral transfer method between LIBS devices of different resolutions as described in any one of claims 1 to 6.

Images