Mineral oxygen element analysis method based on multi-mode LIBS

Abstract

The invention discloses a multi-mode LIBS mineral oxygen element analysis method. The method comprises the following steps: firstly, preprocessing an LIBS spectrum, and then fusing the spectrum and a mineral microscopic image by using a deep learning model to carry out feature extraction and oxygen analysis. The core of the method is model training: shielding a spectrum, pairing the spectrum with a corresponding mineral image (positive samples and negative samples), and generating a fused spectrum-image representation by using mask reconstruction and comparative learning. The element content is predicted based on this representation, and the model is optimized according to the oxygen prediction results. According to the method, accurate, robust and interpretable quantitative analysis of the oxygen element in the complex mineral matrix can be realized.

Description

Technical Field

[0001] This application relates to the technical field of LIBS mineral oxygen element analysis, and particularly to a multi-modal LIBS-based mineral oxygen element analysis method. Background Art

[0002] The efficient and accurate analysis of mineral oxygen elements is a core requirement in geological exploration and materials science. Traditional methods such as EPMA and XRF have high precision, but the processes are cumbersome and costly. Although the LIBS technology is fast and non-destructive, it faces bottlenecks such as weak oxygen signals, strong matrix effects, poor generalization for small samples, and the black-box nature of the model.

[0003] Existing technologies mostly focus on single modalities, lacking the ability of multi-source information fusion and self-supervised learning, and are difficult to handle complex mineral scenarios. Summary of the Invention

[0004] The multi-modal LIBS-based mineral oxygen element analysis method provided by this application can achieve accurate, robust, and interpretable quantitative analysis of oxygen elements in complex mineral matrices.

[0005] In a first aspect, this application provides a multi-modal LIBS-based mineral oxygen element analysis method. The multi-modal LIBS-based mineral oxygen element analysis method includes: performing logarithmic normalization and discrete wavelet transform on the original LIBS spectral signal to obtain a target LIBS spectral signal; using a deep learning model to extract features from the target LIBS spectral signal to obtain target features, and performing oxygen element analysis on the target features to obtain an oxygen element analysis result; wherein, the deep learning model is trained in the following manner: obtaining a training data set; wherein, the training data set includes training LIBS spectral signals and mineral microscopic images; performing spectral segment masking on the training LIBS spectral signals to obtain masked LIBS spectral signals, and pairing the mineral microscopic images and training LIBS spectral signals of the same mineral as positive samples, and the rest as negative samples; inputting the masked LIBS spectral signals, positive samples, and negative samples into an expert module in the deep learning model for mask reconstruction and contrast learning to construct a joint semantic space of the spectral signal and the mineral microscopic image to obtain a spectral-image representation; wherein, each expert module is composed of a variational encoder; performing element prediction according to the spectral-image representation to obtain an element predicted content; the element predicted content includes an oxygen element predicted content; adjusting the weight of the oxygen element branch according to the oxygen element predicted content to complete the training of the deep learning model.

[0006] Among them, inputting the masked LIBS spectral signal into the expert module in the deep learning model for mask reconstruction includes: learning the context information of the masked LIBS spectral signal through the expert module to reconstruct the masked LIBS spectral signal, so that the deep learning model can capture the context dependence relationship of the oxygen element spectral lines.

[0007] Among them, the multi-modal LIBS mineral oxygen element analysis method further includes: performing spatio-temporal modeling on the original LIBS spectral signal to suppress the random noise of single breakdown and weaken the stable background signal; and performing region adaptive spectral peak enhancement on the original LIBS spectral signal.

[0008] Among them, the multi-modal LIBS mineral oxygen element analysis method further includes: calculating the spectral cosine similarity and image structure similarity of samples within a batch, constructing an undirected graph with samples as nodes and similarity as edge weights; passing the undirected graph through a two-layer graph convolutional network to aggregate the features of similar samples.

[0009] Among them, the multi-modal LIBS mineral oxygen element analysis method further includes: in the spectral branch, based on the improved gradient-weighted class activation mapping technology, extracting the three-order gradient information of the target output layer relative to the end convolutional feature map through the backpropagation mechanism, constructing the spatial attention weight distribution of the oxygen element fingerprint spectrum line, and generating a spectral heat map; in the image branch, extracting the microstructural region related to the oxygen content, superimposing it on the original mineral microimage to verify the spectral-morphology synergy effect.

[0010] Among them, in the spectral branch, based on the improved gradient-weighted class activation mapping technology, extracting the three-order gradient information of the target output layer relative to the end convolutional feature map through the backpropagation mechanism, and constructing the spatial attention weight distribution of the oxygen element fingerprint spectrum line, including: using the oxygen content regression value as the supervision signal, calculating the first-order partial derivative matrix of the output layer prediction value with respect to the final convolutional feature tensor to characterize the linear response intensity of each feature channel in the spatial position; further solving the second-order partial derivative and the third-order partial derivative to quantify the non-linear contribution degree of the feature map; strengthening the response of the region in the feature map that has a strong non-linear correlation with the oxygen atom spectrum line through high-order derivative constraints; performing spatial weight weighted summation on all feature channels; restoring the heat map to the original spectral scale through bilinear interpolation, and extracting the local maximum region on the continuous wavelength coordinate.

[0011] Among them, in the image branch, extracting the microstructural region related to the oxygen content, superimposing it on the original image to verify the spectral-morphology synergy effect, including: adopting a cascaded dilated convolution architecture to construct a multi-scale context-aware model of the mineral microimage, and strengthening the response intensity of the oxygen diffusion sensitive region through a cross-layer attention mechanism; among them, the oxygen diffusion sensitive region includes the grain boundary oxidation zone and the pore oxidation ring; performing channel normalization processing on the extracted oxidation sensitive feature map: applying a sub-pixel convolutional layer to upsample the heat map to the original image resolution, and performing pixel-level superposition with the original mineral microimage through the α blending formula; establishing a spectral-image bimodal response space mapping model to verify the spectral-morphology synergy effect.

[0012] Among them, the multi-modal LIBS mineral oxygen element analysis method further includes: importance scoring based on an expert module, pruning redundant channels, and compressing the deep learning model.

[0013] Among them, the multi-modal LIBS mineral oxygen element analysis method further includes: adopting 8-bit integer quantization to simulate quantization errors during the training of the deep learning model.

[0014] Among them, the multi-modal LIBS mineral oxygen element analysis method further includes: dynamically selecting the number of activated expert modules according to the signal-to-noise ratio and the number of interference peaks of the original LIBS spectral signal.

[0015] The beneficial effects of this application are as follows: Different from the prior art, the multi-modal LIBS mineral oxygen element analysis method provided in this application includes: performing logarithmic normalization and discrete wavelet transform on the original LIBS spectral signal to obtain the target LIBS spectral signal; using a deep learning model to extract features from the target LIBS spectral signal to obtain target features, and performing oxygen element analysis on the target features to obtain the oxygen element analysis result; among them, the deep learning model is trained in the following manner: obtaining a training data set; where the training data set includes training LIBS spectral signals and mineral microscopic images; performing spectral segment masking on the training LIBS spectral signals to obtain masked LIBS spectral signals, and pairing the mineral microscopic images and training LIBS spectral signals of the same mineral as positive samples, and the rest as negative samples; inputting the masked LIBS spectral signals, positive samples, and negative samples into the expert module in the deep learning model for mask reconstruction and contrast learning to construct a joint semantic space of the spectral signal and the mineral microscopic image to obtain a spectrum-image representation; where each expert module is composed of a variational encoder; performing element prediction according to the spectrum-image representation to obtain the predicted element content; the predicted element content includes the predicted oxygen element content; adjusting the weight of the oxygen element branch according to the predicted oxygen element content to complete the training of the deep learning model. Through the above method, through multi-modal fusion and self-supervised enhancement, accurate, robust, and interpretable quantitative analysis of oxygen elements in complex mineral matrices is achieved. Description of the Drawings

[0016] In order to more clearly illustrate the technical solutions in the embodiments of this application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of this application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts. Among them: Figure 1 It is a schematic flowchart of an embodiment of the multi-modal LIBS mineral oxygen element analysis method provided by this application; Figure 2 is a schematic flowchart of an embodiment of a method for training a deep learning model provided by this application; Figure 3 is a schematic structural diagram of an embodiment of a multi-modal LIBS mineral oxygen element analysis system provided by this application; Figure 4 is a schematic structural diagram of an embodiment of a computer-readable storage medium provided by this application. Detailed implementation manners

[0017] Next, the technical solutions in the embodiments of this application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of this application. It can be understood that the specific embodiments described herein are only used to explain this application, rather than limiting this application. Additionally, it should be noted that for the sake of description, only parts related to this application rather than all structures are shown in the accompanying drawings. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of this application.

[0018] Referring to "

[0019] " Figure 1 ," Figure 1 is a schematic flowchart of an embodiment of a multi-modal LIBS mineral oxygen element analysis method provided by this application. The multi-modal LIBS mineral oxygen element analysis method includes: Step 11: Perform logarithmic normalization and discrete wavelet transform on the original LIBS spectral signal to obtain a target LIBS spectral signal.

[0020] In some embodiments, the oxygen element spectral signal is often weak, and there is a significant dynamic range difference from the strong signals of other elements. Traditional normalization methods (such as maximum-minimum normalization) will cause weak signals to be compressed or submerged in noise, thus affecting the prediction accuracy. This application innovatively adopts the logarithmic normalization method to effectively highlight the characteristics of weak signals and improve the signal expressiveness of oxygen elements.

[0021] The logarithmic normalization formula is as follows:

[0022] Where: is the intensity of the original spectral signal; is the maximum value of the spectral signal in the sample; is the normalized spectral signal.

[0023] Logarithmic normalization has the following advantages: Weak signal enhancement: Logarithmic transformation can stretch the amplitude of weak signals in the spectrum while compressing the dynamic range of strong signals, making the characteristic spectral lines of oxygen elements more prominent.

[0024] Improved robustness: Reduces the influence of detector response range differences and laser energy fluctuations on spectral data, optimizing the input quality of deep learning models.

[0025] For example, the characteristic spectral lines of oxygen elements in the LIBS spectrum of mineral samples are usually weak and may be masked by background noise in unnormalized data. After logarithmic normalization, the signal of the oxygen element spectral lines is significantly enhanced, and its characteristic information is more easily captured by the convolutional neural network.

[0026] Furthermore, in order to improve the signal-to-noise ratio of the oxygen element spectral lines, this application uses the discrete wavelet transform method to denoise the spectral signal. The discrete wavelet transform can effectively separate the low-frequency background signal and the high-frequency characteristic signal, enhancing the expressiveness of the characteristic spectral lines of oxygen elements.

[0027] The specific steps are as follows: Perform wavelet decomposition on the spectral signal to decompose it into components of different frequencies; Remove background noise from the low-frequency components and retain the high-frequency characteristics; Perform wavelet reconstruction on the denoised components to obtain the enhanced spectral signal.

[0028] After the above background noise processing, the characteristic spectral lines of oxygen elements are clearer in the spectrogram, which helps the subsequent feature extraction of the model. At the same time, background removal also helps to maintain the consistency of the LIBS spectrum.

[0029] Step 12: Use a deep learning model to extract features from the target LIBS spectral signal to obtain target features, and perform oxygen element analysis on the target features to obtain the oxygen element analysis result.

[0030] In some embodiments, refer to Figure 2 , the deep learning model is trained in the following manner: Step 21: Obtain a training dataset; among them, the training dataset includes training LIBS spectral signals and mineral microscopic images.

[0031] In some embodiments, to improve the generalization ability and robustness of the model, this application introduces the following enhancement techniques into the training dataset: Random noise addition: Simulate the noise under actual detection conditions to generate more diverse training data.

[0032] Spectral shift and scaling: Randomly adjust the position and intensity of spectral peaks to simulate the impact of equipment errors or laser power changes on the data.

[0033] Step 22: Perform spectral segment masking on the training LIBS spectral signal to obtain the masked LIBS spectral signal, and pair the mineral microscopic image and the training LIBS spectral signal of the same mineral as positive samples, with the rest as negative samples.

[0034] In some embodiments, spectral segment masking is performed on the training LIBS spectral signal to facilitate subsequent spectral mask reconstruction. Spectral mask reconstruction mainly involves randomly masking 30% of the spectral segments of the LIBS spectrum (such as near the oxygen characteristic peak at 777 nm), learning context information through an encoder-decoder structure, reconstructing the complete spectrum, and forcing the model to capture the context dependencies of the oxygen element spectral lines.

[0035] In some embodiments, image-spectral contrast alignment is performed on the mineral microscopic image and the training LIBS spectral signal. For example, pair the microscopic image and the LIBS spectrum of the same mineral as positive samples, and use different mineral samples as negative samples. By minimizing the contrast loss, the semantic distance between modalities is reduced, and the correlation between spectral features and mineral microstructures is established.

[0036] Step 23: Input the masked LIBS spectral signal, positive samples, and negative samples into the expert module in the deep learning model for mask reconstruction and contrast learning, construct the joint semantic space of the spectral signal and the mineral microscopic image, and obtain the spectral-image representation; each expert module is composed of a variational encoder.

[0037] In some embodiments, input the masked LIBS spectral signal into the expert module in the deep learning model for mask reconstruction in the following way: learn the context information of the masked LIBS spectral signal through the expert module, reconstruct the masked LIBS spectral signal, so that the deep learning model can capture the context dependencies of the oxygen element spectral lines.

[0038] In some embodiments, the deep learning model may include a variational mixture of experts network. The variational mixture of experts network is composed of several expert modules. Each expert is composed of a variational encoder. After inputting spectral or image features, it generates latent variables that follow a Gaussian distribution, and constrains them to be close to the standard normal distribution through the KL divergence to filter noise and retain discriminative features.

[0039] The variational mixture of experts network adopts sparse dynamic routing. Based on the complexity of the input spectrum (such as the background noise level), the sparse dynamic routing module calculates the expert activation weights, and only selects the weighted fusion of the outputs of 2-3 most relevant experts to achieve a balance between computational efficiency and feature expressiveness.

[0040] Step 24: Perform elemental prediction based on the spectrum-image representation to obtain the predicted elemental content; the predicted elemental content includes the predicted oxygen elemental content.

[0041] Step 25: Adjust the weight of the oxygen element branch according to the predicted oxygen elemental content to complete the training of the deep learning model.

[0042] In some embodiments, due to the low signal intensity of oxygen element detection and the susceptibility of its characteristic spectral lines to interference, the traditional loss function cannot effectively focus on the performance of the oxygen element branch during the optimization process. Therefore, this application designs a dynamic weighted loss function (Dynamic Weighted Loss Function), which ensures that the model preferentially optimizes the prediction accuracy of oxygen element during training by assigning dynamically adjusted weights to the oxygen element branch.

[0043] 1. Loss function formula: The loss function of this application adopts the form of weighted mean squared error (WMSE): Where:

[0044] represents the actual elemental content; represents the model prediction value; represents the dynamic weight, which is set to a higher value for oxygen element signals; N represents the total number of samples.

[0045] 2. Dynamic weight update strategy: During the training process, the dynamic weight will be adjusted in real time according to the following two points: a. Error feedback mechanism: After each round of training, the weight of the oxygen element branch is dynamically adjusted according to the oxygen element prediction error‍. The greater the error, the higher the weight.

[0046] Where:

[0047] M represents the number of oxygen element samples; represents the dynamic adjustment factor (for example, taking values from 5 to 2.0, adjusted according to the training effect).

[0048] represents the predicted value of oxygen element content, represents the actual value of oxygen element content.

[0049] b. Branch importance mechanism: Different initial weights are set for the oxygen element and other element branches (for example, the initial weight value of the oxygen element is 0, and that of other elements is 1.0), and they are dynamically adjusted according to the convergence of branch training.

[0050] The advantages of dynamic weighting are as follows: Prioritize optimizing oxygen element prediction: The dynamic weights ensure that the model focuses on fitting the weak signals of oxygen elements during the optimization process.

[0051] Adaptive adjustment: The weights are updated in real time according to the error, enabling the model to converge faster and be more stable.

[0052] Global cooperation: Even if a higher weight is assigned to the oxygen element branch, the performance of other branches will be improved through the sharing of full-spectrum information.

[0053] Furthermore, the deep learning model adopts a pre-training - fine-tuning transfer strategy. For example: Unsupervised pre-training: On a large-scale unlabeled mineral dataset, the network is pre-trained through masked reconstruction and contrast learning to learn the general spectral - image representation.

[0054] Few-shot fine-tuning: Freeze the encoder and only fine-tune the classification head, and use a small amount of labeled data (such as 50 samples) to adapt to the downstream oxygen content regression task.

[0055] In some embodiments, spatio-temporal dynamic compensation enhancement and region-adaptive spectral peak enhancement can be performed on the LIBS spectral signal.

[0056] For example, for spatio-temporal dynamic compensation enhancement, a multi-pulse sequence spatio-temporal modeling temporal convolutional network (TCN) can be used to perform causal convolution on the spectral sequences of multiple LIBS breakdowns of the same sample along the time axis, extract breakdown stability features (such as plasma temperature fluctuations), and fuse the results of multiple measurements with weights to suppress the random noise of a single breakdown. And use a residual attention mechanism to calculate the residual between the current spectrum and the historical average spectrum, and strengthen the difference region (such as oxygen peak fluctuations) through a channel attention module to weaken the stable background signal.

[0057] For example, for region-adaptive spectral peak enhancement, a neighborhood focusing window technique can be used. For example, within the neighborhood of the oxygen characteristic peak (such as 777nm ± 5nm), adaptive Gaussian filtering is used to smooth the noise, and at the same time, non-local means are used to enhance the peak intensity. And a matrix interference suppression technique, for the strong peak regions of interfering elements such as iron and titanium (such as Fe II 275nm), a band-stop filter is designed to suppress the interference components in the frequency domain. That is, in this application, multi-pulse sequence spatio-temporal modeling can be performed on the original LIBS spectral signal to suppress the random noise of a single breakdown and weaken the stable background signal; and region-adaptive spectral peak enhancement can be performed on the original LIBS spectral signal.

[0058] The above deep learning model has an interpretable multi-modal fusion decision-making, which can be specifically reflected in graph structure semantic propagation and dual-path interpretability analysis.

[0059] In some embodiments, graph structure semantic propagation may include dynamic sample graph construction and graph convolutional feature aggregation.

[0060] Dynamic sample graph construction is mainly reflected in calculating the spectral cosine similarity and image structure similarity (SSIM) of samples within a batch, and constructing an undirected graph with samples as nodes and similarity as edge weights.

[0061] Graph convolutional feature aggregation is mainly reflected in aggregating the features of similar samples through a two-layer graph convolutional network (GCN) to enhance the representation robustness under the condition of small samples (such as the migration of rare mineral euxenite to similar matrix samples). That is, the multi-modal LIBS mineral oxygen element analysis method of the present application further includes: calculating the spectral cosine similarity and image structure similarity of samples within a batch, constructing an undirected graph with samples as nodes and similarity as edge weights; passing the undirected graph through a two-layer graph convolutional network to aggregate the features of similar samples.

[0062] In some embodiments, dual-path interpretability analysis may include spectral heatmap generation and mineral structure association visualization.

[0063] Spectral heatmap generation is mainly reflected in that in the spectral CNN branch, based on the improved gradient-weighted class activation mapping technology, the three-step gradient information of the target output layer relative to the end convolutional feature map is extracted through the backpropagation mechanism, and the spatial attention weight distribution of the oxygen element fingerprint spectral line is constructed. That is, the multi-modal LIBS mineral oxygen element analysis method of the present application further includes: in the spectral branch, based on the improved gradient-weighted class activation mapping technology, the three-step gradient information of the target output layer relative to the end convolutional feature map is extracted through the backpropagation mechanism, the spatial attention weight distribution of the oxygen element fingerprint spectral line is constructed, and the spectral heatmap is generated.

[0064] Specifically, it may involve gradient propagation calculation, weight dynamic fusion, and heatmap generation and spectral line positioning.

[0065] Among them, gradient propagation calculation is mainly reflected in using the oxygen content regression value as the supervision signal to calculate the first-order partial derivative matrix of the output layer prediction value with respect to the final convolutional feature tensor to represent the linear response intensity of each feature channel at the spatial position; further solving the second-order partial derivative and the third-order partial derivative to quantify the non-linear contribution degree of the feature map.

[0066] Among them, weight dynamic fusion is mainly reflected in strengthening the regional response in the feature map that has a strong non-linear association with the oxygen atom spectral line (such as the 776.5nm and 844.6nm triplets in the near-infrared region) through high-order derivative constraints.

[0067] Among them, the generation of the heat map and the spectral line positioning are mainly reflected in the implementation of spatial weight weighted summation on the full feature channels; the low-resolution heat map is restored to the original spectral scale through bilinear interpolation, and the local maximum region is extracted on the continuous wavelength coordinate to achieve the accurate decoupling of the oxygen characteristic peak under the interference of the iron matrix.

[0068] The visualization of the mineral structure association is mainly reflected in the image branch, where the microscopic structure regions strongly correlated with the oxygen content (such as the grain boundary oxidation zone) are extracted and superimposed on the original image to verify the spectral-morphology synergistic effect. Specifically, it can involve the extraction of oxidation characteristic regions, the fusion and spatial alignment of heat maps, and the cross-modal collaborative verification. That is, the multi-modal LIBS mineral oxygen element analysis method of the present application further includes: in the image branch, extracting the microscopic structure regions related to the oxygen content and superimposing them on the original mineral microscopic image to verify the spectral-morphology synergistic effect.

[0069] Among them, the extraction of oxidation characteristic regions is mainly reflected in the adoption of a cascaded dilated convolution architecture to construct a multi-scale context-aware model of the microscopic image, and the response intensity of oxygen diffusion sensitive regions such as the grain boundary oxidation zone and the pore oxidation ring is enhanced through a cross-layer attention mechanism.

[0070] Among them, ii. The fusion and spatial alignment of heat maps are mainly reflected in the channel normalization processing of the extracted oxidation-sensitive feature maps: applying a sub-pixel convolutional layer to upsample the heat map to the resolution of the original image, and realizing the pixel-level superposition with the in-situ microscopic image through the Alpha Blending Formula (AlphaBlending Formula is an image superposition technology that fuses two images by weight through the transparency parameter (α value) and is often used for the superposition display of the heat map and the original image).

[0071] Among them, the cross-modal collaborative verification is mainly reflected in the establishment of a spectral-image bimodal response space mapping model.

[0072] In the present application, the deep learning model can be deployed as a lightweight adaptive inference engine. For example, the lightweight adaptive inference engine has dynamic computing resource allocation and hardware-level accelerated deployment.

[0073] The dynamic computing resource allocation includes complexity-aware routing and hierarchical model pruning.

[0074] The complexity-aware routing is mainly reflected in dynamically selecting the number of activated experts according to the signal-to-noise ratio (SNR) and the number of interference peaks of the input spectrum (low noise → 1 expert, high noise → 3 experts) to balance the accuracy and latency. That is, the multi-modal LIBS mineral oxygen element analysis method of the present application further includes: dynamically selecting the number of activated expert modules according to the signal-to-noise ratio and the number of interference peaks of the original LIBS spectral signal.

[0075] Hierarchical model pruning is mainly reflected in the importance scoring based on expert modules (such as the L1 norm of weights), pruning redundant channels, and compressing the model size to 30% of the original size. That is, the multi-modal LIBS mineral oxygen element analysis method of this application also includes: importance scoring based on expert modules, pruning redundant channels, and compressing the deep learning model.

[0076] Hardware-level acceleration deployment includes FPGA pipeline optimization and quantization-aware training.

[0077] FPGA pipeline optimization is mainly reflected in mapping modules such as spectral preprocessing (wavelet denoising), expert routing, and feature fusion into a hardware parallel pipeline, with the single-inference time ≤ 5ms.

[0078] Quantization-aware training is mainly reflected in adopting 8-bit integer quantization, simulating quantization errors during training, and ensuring that the accuracy loss after deployment < 0.5%. That is, the multi-modal LIBS mineral oxygen element analysis method of this application also includes: adopting 8-bit integer quantization and simulating quantization errors during the training of the deep learning model.

[0079] The above technical solutions of this application have the following technical effects.

[0080] First, the technical effects of the multi-modal self-supervised pre-training framework are as follows: Spectral-image cross-modal representation learning can construct a joint semantic space of spectra and microscopic images through masked reconstruction and contrast learning tasks, achieving deep alignment of features between modalities. This mechanism enables the model to effectively mine the implicit correlations of unlabeled data under small-sample conditions, breaking through the high dependence of traditional single-modal models on labeled data.

[0081] Enhanced noise robustness is mainly reflected in the collaborative processing of logarithmic normalization and wavelet denoising, significantly improving the signal-to-noise ratio of weak oxygen signals while retaining the physical interpretability of spectra; the variational expert network filters high-frequency noise through probability encoding, strengthening the anti-interference ability against unstable factors such as plasma fluctuations and matrix interference.

[0082] Dynamic sparse computing optimization is mainly reflected in the expert routing mechanism adaptively selecting activation paths according to input complexity, reducing redundant computing consumption while maintaining model capacity, especially suitable for processing long-tail distribution scenarios with strong heterogeneity among mineral samples.

[0083] Second, the technical effects of spatio-temporal dynamic compensation enhancement are as follows: Temporal stability optimization is mainly reflected in the time convolutional network (TCN) modeling the time-varying characteristics of multi-pulse breakdown sequences, suppressing random noise in single measurements through weighted fusion, and simultaneously capturing the plasma evolution law to improve the temporal consistency of oxygen signals.

[0084] The selective enhancement of spectral peaks is mainly reflected in the synergistic effect of the neighborhood focusing window and non-local mean filtering, which realizes the improvement of the local signal-to-noise ratio in the oxygen characteristic peak region. The band-stop filter effectively suppresses the strong spectral peaks of interfering elements such as iron and titanium, forming a directional enhancement of the target signal.

[0085] III. The technical effects of interpretable multi-modal fusion decision-making are as follows: The breakthrough in spectral localization accuracy is mainly reflected in that the improved gradient-weighted class activation mapping technology can accurately lock the weak response spectral lines of oxygen elements through three-step gradient constraints, and can still achieve spectral peak decoupling under complex matrix interference, which is significantly better than the traditional linear response analysis method.

[0086] The cross-modal physical correlation verification is mainly reflected in the spatial alignment mechanism between the microstructure thermal map and the spectral feature weights, revealing the internal correlation between the oxygen element enrichment area (such as the grain boundary oxidation zone) and the intensity of the characteristic spectral line, and providing visual evidence for the mechanism research of the mineral oxidation process.

[0087] The graph structure semantic propagation ability is mainly reflected in the dynamic sample graph construction and graph convolution aggregation mechanism, enabling the model to transfer the implicit knowledge of similar matrix samples, enhancing the representation robustness of rare minerals and low-content samples, and solving the strong assumption limitation of the traditional method on the independent and identically distributed samples.

[0088] IV. The technical effects of the dynamic weighted loss function are as follows: The weak signal priority learning mechanism is mainly reflected in dynamically adjusting the oxygen element branch weight through error feedback, forcing the model to preferentially fit the subtle changes in the low signal-to-noise ratio region, and overcoming the defect of insufficient gradient update of the traditional mean square error loss for weak signals.

[0089] The multi-task collaborative optimization is mainly reflected in strengthening the prediction of oxygen elements while constraining the representation space of other elements through full-spectrum information sharing, avoiding overfitting of a single target, and achieving an overall improvement in performance.

[0090] V. The technical effects of the lightweight adaptive inference engine are as follows: The balance between computational efficiency and accuracy is mainly reflected in the complexity-aware routing and hierarchical pruning technologies, which significantly compress the model parameters while ensuring the analysis accuracy of oxygen elements, and adapt to the resource limitations of edge computing devices.

[0091] The hardware-level deployment optimization is mainly reflected in the coordination of the FPGA pipeline design and quantization-aware training, realizing end-to-end acceleration from algorithms to hardware, meeting the strict requirements of field in-situ detection for real-time performance, and maintaining industrial-level stability.

[0092] Furthermore, the comprehensive technical advantages are as follows: This solution systematically addresses the four major bottlenecks of traditional LIBS technology in oxygen element analysis through the deep integration of multi-modal self-supervised learning, dynamic compensation enhancement, and interpretability design: Small sample generalization: Cross-modal pre-training and graph semantic propagation achieve knowledge transfer; Weak signal detection: An end-to-end optimization link of logarithmic transformation - dynamic weighted loss; Complex matrix interference: A collaborative anti-interference mechanism of spectral peak directional enhancement and band-stop filtering; Model black-boxing: Dual-channel interpretability analysis validates physical consistency.

[0093] This technology system provides a complete solution for the rapid, accurate, and interpretable analysis of mineral oxygen elements, and is particularly suitable for industrial scenarios such as geological exploration and metallurgical detection that have strict requirements for robustness and interpretability.

[0094] Furthermore, there are interactions among the above technical effects, specifically as follows: 1. Spatiotemporal complementarity between self-supervised pre-training and dynamic compensation.

[0095] Positive cycle: Self-supervised pre-training (spectral mask reconstruction + image contrast learning) generates robust multi-modal representations, providing input features with less noise and clearer physical meaning for the spatiotemporal dynamic compensation module (TCN time series modeling, neighborhood focusing window); while the temporally stabilized spectra and spectral peak enhanced data output by the compensation module feed back to the pre-training process, improving the learning efficiency of self-supervised tasks.

[0096] Case: During the pre-training stage, mask reconstruction forces the model to learn the context relevance of oxygen spectral lines (such as the correlation between the 777nm and 844.6nm triplets). The spatiotemporal compensation module then uses this knowledge to accurately locate interference regions (such as Fe II 275nm) during the inference stage, dynamically adjusting the parameters of the band-stop filter to form a "learning - application - feedback" closed loop.

[0097] 2. Bidirectional optimization between interpretability analysis and dynamic loss function.

[0098] Gradient flow collaboration: The dynamic weighted loss function (high weight focusing on oxygen elements) guides the model to preferentially optimize oxygen-related feature channels, enhancing the heatmap localization accuracy of the interpretability analysis module; conversely, the oxygen-sensitive regions revealed by the heatmap (such as grain boundary oxidation zones) can provide spatial attention constraints for the loss function, dynamically adjusting the weights of different regions to avoid weak signal dilution caused by global average weighting.

[0099] Effect superposition: In the iron matrix interference scenario, the dynamic weighting of the loss function suppresses the gradient update of the non-oxygen region, while the interpretable heatmap synchronously verifies whether this suppression accurately points to the interference spectral line. The two form a "target-oriented - process verification" dual verification mechanism, which improves the positioning accuracy by more than 30% compared with a single module.

[0100] 3. Joint synergy of graph semantic propagation and lightweight inference.

[0101] Knowledge transfer acceleration: The graph convolutional network (GCN) aggregates the features of similar minerals through sample similarity to generate enhanced representations, enabling the lightweight inference engine (dynamic routing + pruning) to still compensate for the information loss of individual samples using group knowledge while reducing the computational load, especially improving the generalization ability of small-sample / rare minerals.

[0102] Resource-accuracy balance: The lightweight module dynamically selects the number of activated experts according to the input complexity (such as signal-to-noise ratio), and the cross-sample prior knowledge provided by graph propagation (such as the spectral line patterns of similar matrix minerals) can assist the routing module in making quick decisions, reducing the trial-and-error cost of expert selection, and achieving the optimal ratio of "computational efficiency - knowledge density".

[0103] 4. End-to-end collaboration of multi-modal fusion and hardware acceleration.

[0104] Feature-hardware joint optimization: The multi-modal fusion decision module (spectral heatmap + image overlay) generates high-dimensional semantic features, and the FPGA pipeline and 8-bit quantization technology of the lightweight engine perform hardware-level adaptation for the data distribution of these features (such as fixed-point number precision allocation), avoiding the accuracy loss caused by the algorithm-hardware disconnection in traditional solutions.

[0105] Case: The sub-pixel convolutional upsampling of the microscopic image branch is deeply coupled with the parallel computing unit of the FPGA, achieving high-resolution overlay of the heatmap while maintaining a latency of about 5ms, accelerating 12 times compared with CPU deployment and reducing power consumption by 83%.

[0106] 5. Emergent properties of system-level gain.

[0107] The above interactions give rise to system-level advantages that cannot be achieved by traditional technologies: Strong generalization with small samples: Self-supervised pre-training provides general representations, graph propagation injects domain knowledge, and dynamic loss focuses on key signals. The three form a progressive learning from "general to specific". Noise immunity and interpretability unification: Spatiotemporal compensation suppresses physical noise, and the interpretability module eliminates model cognitive noise, purifying the information flow from the data domain and the model domain respectively. Industrial Deployment Feasibility: The collaboration between the lightweight engine and multi-modal decision-making enables the direct migration of laboratory-level algorithms to portable LIBS devices, breaking through the "algorithm-product" transformation gap.

[0108] Conclusion: Engineering significance of non-linear coupling. Through non-linear coupling between modules (such as cross-feedback, joint optimization, hardware-algorithm collaboration), this technical system transforms the linear gain of a single module into an exponential improvement in system-level performance. For example: Self-supervised pre-training (improving the basic accuracy by 20%) + dynamic compensation (improving by 15%) → The actual improvement after collaboration reaches 45%; Interpretability analysis (reducing the positioning error by 25%) + dynamic loss (reducing the error by 30%) → The comprehensive error is reduced by 60% after collaboration.

[0109] This synergistic effect stems from the deep integration of the entire "data-model-hardware" link, marking the entry of LIBS oxygen element analysis from "single-point improvement" into the new paradigm of "system intelligence".

[0110] See Figure 3 , Figure 3 FIG. is a schematic structural diagram of an embodiment of a multi-modal LIBS mineral oxygen element analysis system provided by this application. The multi-modal LIBS mineral oxygen element analysis system 30 includes a processor 31 and a memory 32 coupled to the processor 31; Among them, the memory 32 is used to store computer programs, and the processor 31 is used to execute the computer programs to implement the following methods: Perform logarithmic normalization and discrete wavelet transform on the original LIBS spectral signal to obtain the target LIBS spectral signal; use a deep learning model to extract features from the target LIBS spectral signal to obtain target features, and perform oxygen element analysis on the target features to obtain the oxygen element analysis result; among them, the deep learning model is trained in the following manner: Obtain a training data set; among them, the training data set includes training LIBS spectral signals and mineral microscopic images; perform spectral band masking on the training LIBS spectral signals to obtain masked LIBS spectral signals, and pair the mineral microscopic images and training LIBS spectral signals of the same mineral as positive samples, and the rest as negative samples; input the masked LIBS spectral signals, positive samples, and negative samples into the expert module in the deep learning model for mask reconstruction and contrast learning to construct a joint semantic space of the spectral signal and the mineral microscopic image to obtain a spectral-image representation; each expert module is composed of a variational encoder; perform element prediction based on the spectral-image representation to obtain the predicted element content; the predicted element content includes the predicted oxygen element content; adjust the weight of the oxygen element branch according to the predicted oxygen element content to complete the training of the deep learning model.

[0111] In some embodiments, the processor 31 is further configured to execute a computer program to implement the method of any of the above embodiments.

[0112] Referring to Figure 4 , Figure 4 FIG. is a schematic structural diagram of an embodiment of a computer-readable storage medium provided by the present application. The computer-readable storage medium 40 is used to store a computer program 41, and when the computer program 41 is executed by a processor, it is used to implement the following method: Perform logarithmic normalization and discrete wavelet transform on the original LIBS spectral signal to obtain a target LIBS spectral signal; use a deep learning model to extract features from the target LIBS spectral signal to obtain target features, and perform oxygen element analysis on the target features to obtain an oxygen element analysis result; wherein, the deep learning model is trained in the following manner: Obtain a training data set; wherein, the training data set includes training LIBS spectral signals and mineral microscopic images; perform spectral band masking on the training LIBS spectral signals to obtain masked LIBS spectral signals, and pair positive samples of the mineral microscopic images and training LIBS spectral signals of the same mineral, and the rest as negative samples; input the masked LIBS spectral signals, positive samples, and negative samples into the expert module in the deep learning model for mask reconstruction and contrast learning to construct a joint semantic space of the spectral signal and the mineral microscopic image to obtain a spectral-image representation; wherein, each expert module is composed of a variational encoder; perform element prediction based on the spectral-image representation to obtain an element predicted content; the element predicted content includes an oxygen element predicted content; adjust the weight of the oxygen element branch according to the oxygen element predicted content to complete the training of the deep learning model.

[0113] In some embodiments, when the computer program 41 is executed by a processor, it is further configured to implement the method of any of the above embodiments.

[0114] In several implementation manners provided by the present application, it should be understood that the disclosed method and device can be implemented in other ways. For example, the device implementation manner described above is only illustrative. For example, the division of the modules or units is only a logical function division, and there may be other division manners in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0115] If the integrated units in the above-mentioned other embodiments are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods described in various embodiments of this application. The foregoing storage medium includes: various media that can store program codes such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs.

[0116] The above are only the embodiments of this application, and do not limit the patent scope of this application accordingly. Any equivalent structural or equivalent process transformation made by using the content of the specification and drawings of this application, or directly or indirectly applied in other related technical fields, shall be similarly included in the patent protection scope of this application.

Claims

1. A method for mineral oxygen element analysis based on multimodal LIBS, characterized in that: The multimodal LIBS mineral oxygen element analysis method includes: The original LIBS spectral signal is logarithmically normalized and discrete wavelet transformed to obtain the target LIBS spectral signal; Using a deep learning model to extract features of the target LIBS spectral signal to obtain target features, and performing oxygen element analysis on the target features to obtain oxygen element analysis results; The deep learning model is trained in the following way: Acquire a training data set; wherein the training data set includes training LIBS spectral signals and mineral microscopic images; The training LIBS spectrum signal is masked to obtain the masked LIBS spectrum signal, and the mineral microscopic image of the same mineral and the training LIBS spectrum signal are paired as positive samples, and the rest are used as negative samples; The masked LIBS spectral signal, positive samples, and negative samples are input into the expert module in the deep learning model for mask reconstruction and contrast learning, and the joint semantic space of the spectral signal and the mineral microscopic image is constructed to obtain the spectrum-image representation; wherein each expert module is composed of a variational encoder; Perform element prediction based on the spectrum-image representation to obtain element predicted content; the element predicted content includes oxygen element predicted content; The weight of the oxygen element branch is adjusted according to the predicted oxygen element content to complete the training of the deep learning model.

2. The multimodal LIBS mineral oxygen element analysis method according to claim 1, characterized in that: The step of inputting the masked LIBS spectral signal into an expert module in a deep learning model for mask reconstruction comprises: The context information of the masked LIBS spectral signal is learned through an expert module, and the masked LIBS spectral signal is reconstructed, so that the deep learning model captures the context dependency of the oxygen element spectral line.

3. The method for mineral oxygen element analysis based on multimodal LIBS according to claim 1, characterized in that: The multimodal LIBS mineral oxygen element analysis method further includes: Perform spatiotemporal modeling of multi-pulse sequences on the original LIBS spectral signal to suppress the random noise of single breakdown and weaken the stable background signal; And regional adaptive spectral peak enhancement of the original LIBS spectral signal.

4. The method for mineral oxygen element analysis based on multimodal LIBS according to claim 1, characterized in that: The multimodal LIBS mineral oxygen element analysis method further includes: Calculate the spectral cosine similarity and image structure similarity of samples in a batch, and construct an undirected graph with nodes as samples and edge weights as similarities; The undirected graph is passed through a two-layer graph convolutional network to aggregate the features of similar samples.

5. The method for mineral oxygen element analysis based on multimodal LIBS according to claim 1, characterized in that: The multimodal LIBS mineral oxygen element analysis method further includes: In the spectral branch, based on the improved gradient weighted class activation mapping technology, the third-order gradient information of the target output layer relative to the terminal convolution feature map is extracted through the back-propagation mechanism, the spatial attention weight distribution of the oxygen element fingerprint spectrum is constructed, and the spectral heat map is generated; In the image branch, the microstructural areas related to oxygen content are extracted and superimposed on the original mineral microscopic image to verify the spectral-morphological synergistic effect.

6. The method for mineral oxygen element analysis based on multimodal LIBS according to claim 5, characterized in that: In the spectral branch, based on the improved gradient weighted class activation mapping technology, the third-order gradient information of the target output layer relative to the terminal convolution feature map is extracted through the back propagation mechanism to construct the spatial attention weight distribution of the oxygen element fingerprint spectrum, including: Taking the oxygen content regression value as the supervision signal, the first-order partial derivative matrix of the output layer prediction value to the final convolution feature tensor is calculated to characterize the linear response strength of each feature channel in the spatial position; the second-order partial derivative and the third-order partial derivative are further solved to quantify the nonlinear contribution of the feature map; By using high-order derivative constraints, the regional response in the characteristic map that has strong nonlinear correlation with the oxygen atomic spectral lines is strengthened; The spatial weighted summation is performed on all feature channels; the heat map is restored to the original spectral scale through bilinear interpolation, and the local maximum area is extracted on the continuous wavelength coordinates.

7. The method for mineral oxygen element analysis based on multimodal LIBS according to claim 5, characterized in that: In the image branch, the microstructure area related to the oxygen content is extracted and superimposed on the original image to verify the spectrum-morphology synergistic effect, including: A cascaded dilated convolutional architecture is used to build a multi-scale context-aware model of mineral microscopic images, and the response intensity of oxygen diffusion-sensitive areas is enhanced through a cross-layer attention mechanism; wherein the oxygen diffusion-sensitive areas include grain boundary oxidation zones and pore oxidation rings; The extracted oxidation-sensitive feature map is subjected to channel normalization: the sub-pixel convolution layer is applied to upsample the thermal map to the original image resolution, and the alpha blending formula is used to perform pixel-level superposition with the original mineral microscopic image; A spectrum-image dual-modal response space mapping model was established to verify the spectrum-morphology synergistic effect.

8. The method for mineral oxygen element analysis based on multimodal LIBS according to claim 1, characterized in that: The multimodal LIBS mineral oxygen element analysis method further includes: Importance scoring based on expert modules, pruning redundant channels, and compressing deep learning models.

9. The method for mineral oxygen element analysis based on multimodal LIBS according to claim 1, characterized in that: The multimodal LIBS mineral oxygen element analysis method further includes: 8-bit integer quantization is adopted to simulate quantization error in the training of the deep learning model.

10. The method for mineral oxygen element analysis based on multimodal LIBS according to claim 1, characterized in that: The multimodal LIBS mineral oxygen element analysis method further includes: The number of activated expert modules is dynamically selected according to the signal-to-noise ratio and the number of interference peaks of the original LIBS spectral signal.

Images