Deep sea ore grade quantitative detection method based on laser light-sound fusion

Abstract

The invention belongs to the technical field of spectral quantitative analysis, and discloses a deep sea ore grade quantitative detection method based on laser light-sound fusion. The method comprises the following steps: synchronously acquiring an LIBS spectrum and a plasma sound wave signal, and respectively carrying out pretreatment and feature extraction on the spectrum signal and the sound wave signal; the inter-domain difference is eliminated through split-mode transfer learning correction, and then light-sound dual-mode fusion modeling is adopted, so that rapid and accurate quantitative detection of deep sea ore elements is realized. According to the method, a correction mechanism combining plasma acoustic wave signals and transfer learning is introduced into an LIBS detection system, so that a high-precision and high-robustness quantitative method for synchronous compensation of a physical mechanism and data distribution is realized; physical changes of excitation conditions are represented by acoustic information, so that signal fluctuation caused by a matrix effect is inhibited; by introducing a transfer learning strategy, the data distribution difference between a source domain and a target domain is aligned, and the quantitative precision and robustness of multiple elements of deep sea ore are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of spectral quantitative analysis technology, and particularly relates to a method for quantitative detection of deep-sea ore grade based on laser-optical-acoustic fusion. By introducing plasma acoustic wave signals as physical compensation for LIBS and combining transfer learning for data compensation, this method corrects the problem that the quantitative accuracy of LIBS in deep-sea ore detection is easily affected by sample position fluctuations and that matrix effects are severe, thereby improving the robustness and accuracy of multi-element grade analysis in complex environments. Background Technology

[0002] With the gradual depletion of terrestrial metal mineral resources, deep-sea mining has become a globally recognized emerging field. Deep-sea cobalt-rich crusts, as key seabed metal ores, are rich in critical metals such as Mn, Co, Cu, and Ti, possessing immense economic value and strategic significance. Grade determination of deep-sea ores, represented by cobalt-rich crusts, has become a crucial step in resource assessment and development decisions. Currently, chemical analysis techniques such as X-ray fluorescence (XRF), inductively coupled plasma mass spectrometry (ICP-MS), and inductively coupled plasma atomic emission spectrometry (ICP-AES) are used for elemental analysis of minerals such as cobalt-rich crusts and manganese nodules. However, these technologies have significant limitations. They require complex sample pretreatment in the laboratory, have long detection cycles, and rely on expensive equipment, making them unsuitable for the "rapid on-board detection" requirements in deep-sea mineral surveys. Therefore, there is an urgent need to develop a simple and efficient on-board quantitative detection technology.

[0003] Laser-induced breakdown spectroscopy (LIBS) utilizes high-energy pulsed lasers to ablate samples, generating transient plasmas, and then analyzes their emission spectra to achieve qualitative and quantitative elemental detection. This technique offers advantages such as rapid detection, simultaneous multi-element analysis, ease of operation, and minimal or no sample pretreatment, making it highly suitable for shipboard testing. However, in shipboard field testing, especially when dealing with complex and non-uniform deep-sea ore samples, LIBS signals face significant challenges. Firstly, the physical properties of different ore samples vary significantly; factors such as density and hardness directly alter the laser ablation efficiency. Secondly, samples with the same elemental concentration in different matrices will produce significantly different spectra, causing a shift in the relationship between spectrum and concentration—a phenomenon known as the matrix effect. It is worth noting that transient acoustic signals accompany plasma formation. These acoustic signals can reflect physical information such as the coupling efficiency between the laser and the sample, and are directly related to properties such as sample density and hardness. If the spectral signal and the acoustic signal are analyzed together, not only can the elemental characteristics be characterized by spectral information, but the physical changes of the excitation conditions can also be characterized by acoustic information, thereby achieving effective compensation for matrix effects and improving the stability and robustness of the quantitative model.

[0004] On the other hand, traditional LIBS modeling methods often assume that spectral data have the same distribution characteristics. However, fluctuations in experimental parameters, especially small changes in the relative distance between the laser focus and the sample surface, can significantly affect signal stability. In a shipboard environment, platform vibration and sample fixation deviations can make it difficult to maintain a consistent focus position precisely. This can cause differences in the intensity and feature distribution of photoacoustic data, which can shift the mapping relationship between photoacoustic features and elemental concentration, thus reducing the quantitative accuracy of the model. Although using data from ideal laboratory conditions as a training set can effectively align the relationship between photoacoustic features and elemental concentration, the differences between the data and field test data can make it difficult for laboratory-built models to be directly applied to field data.

[0005] While existing chemical analysis techniques such as X-ray fluorescence (XRF), inductively coupled plasma mass spectrometry (ICP-MS), and inductively coupled plasma atomic emission spectrometry (ICP-AES) offer high detection accuracy, they generally rely on complex sample pretreatment processes, resulting in long detection cycles and expensive equipment. These factors make them unsuitable as the best solution for rapid on-site detection in shipboard environments. Although existing laser-induced breakdown spectroscopy (LIBS) technology has potential for on-site detection, the stability and quantitative accuracy of LIBS signals are affected by multiple factors in the complex on-site environment. Fluctuations in experimental parameters can lead to changes in the plasma excitation state, causing instability in spectral signal intensity. Simultaneously, physical differences in density, hardness, and other properties between different ore samples can induce significant matrix effects, shifting the correlation between spectrum and elemental concentration. The combination of these two problems makes it difficult for traditional partial least squares (PLS) modeling methods to accurately predict elements in deep-sea ores, creating a technical bottleneck of insufficient prediction accuracy and poor stability in on-site quantitative models. Therefore, a method is needed that comprehensively considers how to utilize photoacoustic data under ideal laboratory conditions to achieve highly stable and accurate multi-element detection of deep-sea ores.

[0006] Based on the above analysis, the problems and shortcomings of the existing technology are as follows: (1) Existing high-precision chemical analysis technologies cannot meet the needs of rapid on-site detection: Although XRF, ICP-MS, ICP-AES and other technologies have high detection accuracy, they rely on complex and time-consuming laboratory sample pretreatment processes, and the equipment is expensive and bulky. This inherent characteristic determines that they are essentially unable to meet the urgent requirements of "shipborne, on-site, and rapid" detection in deep-sea mineral surveys, thus creating a gap in the application scenarios of the technology.

[0007] (2) A Chinese patent similar to this invention, entitled "A Method for Analyzing the Composition of Hot-Dip Galvanizing Liquid Based on Laser-Induced Breakdown Spectroscopy Combined with Machine Learning", application number: CN202410556349.7, application date: May 7, 2024; however, this solution is for materials with relatively simple elemental content and cannot solve the stability and accuracy bottlenecks faced by LIBS technology in shipboard testing: Although LIBS technology itself has the potential for on-site testing, its quantitative analysis stability and accuracy are seriously challenged in the complex shipboard environment, specifically due to two interrelated core issues: Significant matrix effect: Deep-sea ores (such as cobalt-rich crusts) have complex and non-uniform compositions. Differences in their physical properties, such as density and hardness, directly affect laser ablation efficiency, resulting in significantly different spectral signals from samples with the same elemental concentration, even when the matrix is ​​different. Traditional LIBS models struggle to overcome this shift in the spectral-concentration relationship caused by differences in physical properties.

[0008] Unavoidable fluctuations in experimental parameters: In a shipboard environment, factors such as platform vibration and sample fixation deviations cause the laser focusing position to be inconsistent. These fluctuations directly alter the excitation state of the plasma, causing random variations in the intensity and characteristics of the photoacoustic signal, thus disrupting the stability of the mapping relationship between photoacoustic characteristics and elemental concentration established in the laboratory.

[0009] (3) Existing modeling strategies have inherent limitations: Similar solutions to this invention, such as "A Laser-Induced Breakdown Spectroscopy Detection Device and Method Based on Acoustic Signal Assistance," only utilize the time-domain characteristics of acoustic signals. The dimensionality of spectral signals is much larger than that of acoustic signals, resulting in an imbalance of modal weights and a reduction in the quantitative value of acoustic signals. Furthermore, time-domain signals reflect the overall intensity of acoustic signals more and are more susceptible to fluctuations in experimental parameters, while changes in frequency-domain signals are directly related to changes in the physical properties of the sample, thus enabling more effective correction of matrix effects. Existing technologies rely solely on a single fusion strategy, making it difficult to achieve simultaneous high-precision quantification of multiple elements in complex samples such as deep-sea ores.

[0010] Traditional transfer learning methods either perform only feature transfer or only instance transfer, and most rely solely on spectral signals. They have never designed modal processing or two-layer progressive structures for dual-modal signals, requiring specialized design to address the differences between LIBS photoacoustic signals. Similar solutions, such as "A Quantitative Analysis Method for Heavy Metals in Soil Powder Based on Transfer Learning using LIBS," employ only a single-layer instance transfer strategy, failing to correct feature space shifts and analyzing only spectral modes. Therefore, they cannot fundamentally achieve effective correction of matrix effects. Summary of the Invention

[0011] To address the problems of existing LIBS technology in shipborne quantitative detection of deep-sea ores, this invention discloses a method for quantitative detection of deep-sea ore grade based on laser-optical-acoustic fusion. This method achieves simultaneous compensation from both physical mechanisms and data distribution perspectives, thereby improving the quantitative accuracy and robustness of the model in complex environments. It provides an efficient and stable technical approach for shipborne on-site detection of multiple elements such as Mn, Co, Cu, and Ti in deep-sea ores. The technical solution is as follows: This invention is implemented as follows: a quantitative detection method for deep-sea ore grade based on laser-optical-acoustic fusion. This method constructs a quantitative framework driven by both physical mechanism compensation and data distribution compensation. The initial shock wave after laser breakdown is extracted and subjected to FFT as acoustic modal information. The optical and acoustic modal information is then processed through a progressive two-layer structure of feature transfer and instance transfer, and a feature-level early fusion model is trained to achieve simultaneous, high-precision quantitative analysis of multiple elements in deep-sea ores. Specifically, the method includes the following steps: S1, Signal Acquisition: Under ideal conditions in the source domain laboratory and fluctuations in the sample position in the target domain, a pulsed laser and a signal acquisition device are synchronously triggered to acquire time-consistent laser-induced breakdown spectral signals and plasma acoustic signals, and to construct source domain datasets and target domain datasets. S2, Signal preprocessing and feature extraction: The laser-induced breakdown spectral signal is preprocessed and the spectral mode feature vector is extracted. The plasma acoustic wave signal is preprocessed and the acoustic mode feature vector is extracted. A multimodal feature database containing element concentration labels of spectral signal and acoustic wave signal is obtained. S3, Modal Transfer Learning Correction: At the feature level, a subset of common features that are strongly correlated with and stable with element concentration in the source and target domains are selected to achieve feature space alignment; at the instance level, based on the weighted fusion similarity of spectral and acoustic modes, the sample that best matches the distribution of the target domain is selected from the source domain, and the training set with the target domain is used to construct an enhanced training set. S4, Optical-Acoustic Dual-Modal Fusion Modeling: A feature-level early fusion strategy is used to directly splice optical-acoustic features on the enhanced training set. A quantitative model is established through partial least squares regression. Within the preset number of principal components, cross-validation is used to select the optimal model parameters. S5, Model Validation and Output: Use the final model to predict the elemental concentration of the target domain test set and calculate the evaluation index to output the quantitative detection results.

[0012] In step S1, acquiring time-consistent laser-induced breakdown spectral signals and plasma acoustic signals includes: The cobalt-rich crust pellet sample was penetrated using a pulsed laser, and the spectral signal was acquired using a spectrometer. Microphone collects sound wave signals Two datasets were constructed, collecting photoacoustic signals under ideal laboratory conditions in the source domain and under sample position fluctuation conditions in the target domain, respectively. For each environment, both signals were synchronously triggered by a signal trigger to ensure that the data was collected under the same laser pulse, resulting in a temporally consistent source domain dataset. and target domain dataset ,in, This refers to the sample number under ideal laboratory conditions. This refers to the sample number under conditions of sample position fluctuation.

[0013] In step S2, the laser-induced breakdown spectral signal is preprocessed and spectral mode feature vectors are extracted, including: For spectral signals Baseline removal, smoothing filtering, integral area normalization, and inter-point averaging preprocessing are performed. Outlier signals are removed using cosine similarity to obtain a stable spectral intensity curve. The preprocessed spectral signal is then analyzed. Feature extraction is performed to obtain spectral features. ,in, For the first The intensity of spectral lines at each wavelength point; all spectral features constitute the spectral mode eigenvector. .

[0014] Furthermore, the plasma acoustic signal is preprocessed, including: For sound wave signals Perform time window interception and capture the value after laser breakdown. Effective segment acoustic signal Perform a Fast Fourier Transform on the intercepted acoustic signal to obtain its frequency domain characteristics. ; In the formula, , For the first One frequency component, For the corresponding complex amplitude; The modulus value is then relatively normalized and converted to a decibel scale: ; The acoustic modal eigenvectors are obtained as follows: ; The multimodal feature database containing spectral and acoustic signals was obtained as follows: ; ; In the formula, and These are spectral and acoustic characteristic modes, respectively. For element concentration labels.

[0015] In step S3, the modal transfer learning correction includes: introducing a transfer learning algorithm based on fine-tuning strategy and adopting a two-layer strategy that combines feature transfer and instance transfer; At the feature level, standardization and feature selection are performed on spectral and acoustic features respectively; operations are performed independently on spectral and acoustic modes within the source domain feature set. and target domain feature set The SelectKBest algorithm was used to filter the elements with the highest correlation to their concentrations. Features, and based on the intersection of the two domain filtering results As a subset of common features and These are the filtering features for the source and target domains, respectively; if Then, according to the sum of correlation coefficients Supplementing common features, among which, and The first The correlation coefficients between each feature and its concentration in the source and target domains are obtained; finally, the transferred feature matrix is ​​obtained. and To achieve feature space alignment; In the instance transfer layer, the transferred features are modally standardized to obtain... and Then, the weighted cosine similarity is calculated. To measure the degree of matching between the source domain samples and the target domain distribution. These are the spectral mode weighting parameters. and These are the spectral and acoustic modal cosine similarities, respectively; within the candidate proportion set Search for the optimal migration ratio ,in, For migration ratio, To select the most similar The augmented training set is constructed by merging the source domain samples and the target domain training set. The coefficient of determination on the validation set This represents the total number of samples in the source domain.

[0016] In step S4, the optical-acoustic dual-modal fusion modeling includes: Enhanced training set after transfer correction in the optical-acoustic fusion stage The modeling strategy employed is feature-level early fusion; feature-level early fusion involves directly splicing spectral features at the input level. Acoustic characteristics Construct joint features We directly train a partial least squares regression model to perform overall modeling of multimodal features.

[0017] In step S4, during early feature-level fusion, the input layer incorporates common spectral features. Harmony and common characteristics Direct splicing to form a fusion feature As model input; through the PLS regression layer, the principal components of PLS ​​are represented in linear combination form: , For the first One principal component direction coefficient; within the preset number of principal components, each candidate number is trained, and the 5-fold cross-validation selection principle is adopted to select the number of principal components with the minimum RMSE of the validation set as the standard.

[0018] In step S5, model validation and elemental quantification output include: The model performance was evaluated using 5-fold cross-validation and test sample validation, and the coefficients of determination for the training and test sets were calculated. Given the root mean square error (RMSE) and the relative standard deviation (RSD), we have: ; ; ; In the formula, For the first The true concentration of each sample To predict concentration, This is the average concentration. The total number of samples, The standard deviation of the prediction results; The output data, after transfer correction, is evaluated using the performance metrics of the feature-level early fusion model. The system automatically selects the optimal parameter strategy as the final quantitative model, achieving model-level ensemble. Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, this invention proposes a method based on laser-optical-acoustic fusion combined with transfer learning. It first simultaneously acquires the spectral and acoustic signals of laser-induced plasma, and then accurately extracts the frequency domain features of the effective time period of the acoustic signal using Fast Fourier Transform. Simultaneously, this invention introduces a transfer learning algorithm based on a fine-tuning strategy to transfer the optical-acoustic signal data under ideal laboratory conditions to optical-acoustic data acquired under conditions of sample position fluctuation, thereby helping to build a better regression model. Then, this invention uses a feature-level early fusion strategy to deeply fuse the acoustic signal features and spectral features, achieving a high-precision, robust quantitative method that simultaneously compensates for both physical mechanisms and data distribution.

[0019] Secondly, this invention verifies that in quantitative tasks involving Mn, Co, Cu, and Ti, the frequency domain characteristics of acoustic signals provide more crucial information for improving quantitative performance. Compared to the time domain characteristics (peak-to-peak value) of acoustic signals, they possess more feature dimensions, which is reflected in the quantitative results, particularly in the training set determination coefficient. The values ​​increased from 0.844, 0.975, 0.968, and 0.763 to 0.914, 0.983, 0.981, and 0.849. The effective period of the acoustic signal is dominated by the initial shock wave, which propagates and directly drives the microphone diaphragm to vibrate. The acoustic signal acquired at this time has not been affected by sample vibration coupling or environmental reflection interference. Therefore, its signal characteristics are determined solely by the laser pulse parameters and the inherent physicochemical properties of the sample, representing the core stage reflecting the essence of the sample. Subsequent acoustic signals are mainly affected by environmental interference and have no analytical value. Therefore, extracting the effective time window (0.24-0.34 ms) further improves the model's predictive performance compared to the frequency domain characteristics of the acoustic signal across the entire time domain, increasing the prediction set determination coefficient. The relative standard deviation (RSDp) of the prediction set decreased from 6.978%, 4.765%, 8.377%, and 13.335% to 5.980%, 4.621%, 6.517%, and 12.965%, respectively, as the RSDs increased from 0.913, 0.978, 0.978, and 0.839 to 0.935%, 0.980%, 0.980%, 0.980%, 0.987%, and 0.853.

[0020] Third, this invention also verifies that compared to a two-layer transfer learning model that only analyzes spectral modes, the optical-acoustic fusion-based modal transfer learning strategy of this invention can effectively improve the stability of the transfer effect. This is because the multimodal complementarity of spectral and acoustic signals improves cross-domain stability and is more suitable for shipboard on-site inspection scenarios. Furthermore, the modal weighted transfer method can adaptively balance the importance of the two modes, optimize instance selection, and reduce the transfer risk of a single mode. Reflected in the quantitative results, the training set determination coefficient… The values ​​improved from 0.966, 0.989, 0.992, and 0.961 to 0.979, 0.990, 0.993, and 0.982, respectively; the root mean square error (RMSEP) of the prediction set decreased from 0.805 wt.%, 0.018 wt.%, 0.006 wt.%, and 0.109 wt.% to 0.777 wt.%, 0.015 wt.%, 0.005 wt.%, and 0.091 wt.%, respectively; and the relative standard deviation (RSDp) of the prediction set decreased from 5.714%, 5.437%, 6.667%, and 11.858% to 5.309%, 4.495%, 6.236%, and 9.931%, respectively.

[0021] Fourth, this invention achieves a high-precision and robust quantitative method by introducing a correction mechanism combining plasma acoustic signals and transfer learning into the LIBS detection system, thus simultaneously compensating for both physical mechanisms and data distribution. This invention utilizes acoustic information to characterize the physical changes in excitation conditions, thereby suppressing signal fluctuations caused by matrix effects. Simultaneously, by introducing a transfer learning strategy, this invention can align the data distribution differences between the source domain (ideal laboratory conditions) and the target domain (complex field environment), enabling the model to better align the relationship between photoacoustic features and element concentrations while aligning the feature relationships between the training and test sets. This invention can significantly improve the quantitative accuracy and robustness of multi-element analysis in deep-sea ores. Compared with traditional PLS modeling methods, this invention achieves higher training set determination coefficients in the quantitative analysis of Mn, Co, Cu, and Ti compared to traditional PLS modeling methods. The values ​​were improved from 0.948, 0.981, 0.984, and 0.932 to 0.979, 0.990, 0.994, and 0.982, respectively; the root mean square error (RMSEP) of the prediction set decreased from 0.975 wt.%, 0.019 wt.%, 0.007 wt.%, and 0.134 wt.% to 0.777 wt.%, 0.015 wt.%, 0.005 wt.%, and 0.091 wt.%, respectively; and the relative standard deviation (RSDp) of the prediction set decreased from 6.856%, 5.670%, 8.087%, and 14.567% to 5.309%, 4.495%, 6.115%, and 9.931%, respectively. The model stability was effectively improved, fully verifying the practicality of the proposed method in deep-sea ore shipborne field detection.

[0022] Fifth, the optical-acoustic migration fusion method proposed in this invention can obtain stable elemental quantitative results under shipboard operating conditions. Shipboard LIBS systems are often affected by factors such as platform vibration and sample surface undulations, resulting in significant sample position fluctuations. This causes obvious prediction biases when using models built in traditional laboratories on board. By combining acoustic features and a two-layer migration mechanism, this invention can significantly improve the accuracy of the model in the shipboard environment compared to before calibration, thereby reducing the need for repeated on-site measurements and increasing the effective data output per voyage. The results can be used as an independent data processing module in shipboard LIBS devices for marine surveys, mineral exploration, and digital geology platforms. It has clear industrial application scenarios and high engineering value, and can further promote the standardization and commercialization of shipboard real-time detection equipment. Simultaneously, it can also promote the further development of LIBS technology in the field of in-situ detection.

[0023] Sixth, existing LIBS systems primarily rely on spectral single-modal methods, lacking research on incorporating acoustic signals into model construction, and even more so, lacking systematic methods for achieving cross-scene model transfer under conditions of sample position fluctuation. Traditional methods often depend on spectral normalization, peak intensity ratios, or machine learning modeling under fixed laboratory conditions, lacking effective handling of common issues such as focal length shift and matrix effects. This invention uses acoustic frequency domain features generated by the same laser action as compensation information, and designs a two-layer transfer strategy of feature mapping and instance selection, enabling the laboratory model to adapt to different fluctuation conditions, forming a complete optical-acoustic multimodal fusion and cross-scene transfer system. No mature solution for this technical approach has been found in existing publicly available research.

[0024] LIBS technology has long faced two engineering challenges: firstly, fluctuations in experimental parameters, particularly small changes in the relative distance between the laser focus and the sample surface, can significantly affect signal stability; secondly, matrix effects can cause shifts in the spectrum-concentration relationship. This invention characterizes the physical changes in excitation conditions using acoustic information, thereby effectively compensating for matrix effects, improving the stability and robustness of the quantitative model. Furthermore, it employs a two-layer processing approach of feature transfer and instance transfer to correct the data, ensuring stable predictive performance even under highly perturbed sample conditions. Experimental data show that after processing with this invention, the R², RMSEP, and RSDp of elements such as Mn, Co, Cu, and Ti are significantly improved compared to the unprocessed values, experimentally validating a feasible solution to the aforementioned problems.

[0025] Seventh, traditional views generally hold that acoustic signals are unsuitable for chemical composition analysis. Therefore, in LIBS systems, acoustic signals are usually considered an incidental phenomenon of the laser breakdown process and not used for quantitative modeling. This invention demonstrates that acoustic frequency domain features can effectively improve the quantitative results under highly perturbed sample conditions. Furthermore, traditional LIBS modeling methods often assume that spectral data have the same distribution characteristics. However, factors such as sample height perturbation can lead to differences in spectral data intensity and feature distribution, causing a shift in the mapping relationship between spectral features and elemental concentrations, thus reducing the quantitative accuracy of the model. Although using data from ideal laboratory conditions as the training set can effectively align the relationship between spectral features and elemental concentrations, the difference between the data and actual test data makes it difficult to directly apply the laboratory-built model to the test data. This invention constructs a unified transfer space through optical-acoustic dual-modal features and introduces an instance selection method based on weighted cosine similarity, utilizing optical-acoustic data from ideal laboratory conditions to achieve highly stable and accurate multi-element detection. Attached Figure Description

[0026] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure; Figure 1 This is a flowchart of a method for quantitative detection of deep-sea ore grade based on laser-optical-acoustic fusion provided in an embodiment of the present invention; Figure 2 This is a roadmap of a method for quantitative detection of deep-sea ore grade based on laser-optical-acoustic fusion provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the LIBS optical-acoustic signal acquisition system provided in an embodiment of the present invention; Figure 4 These are typical spectra of the samples provided in this embodiment of the invention, collected at nine different altitudes. Figure 5 These are time-domain diagrams of typical acoustic signals collected at nine different altitudes using samples provided in this embodiment of the invention. Figure 6 This is a frequency domain diagram of the effective phase of the acoustic signal after FFT transformation provided in the embodiments of the present invention; Figure 7 This is a schematic diagram of the transfer learning method provided in an embodiment of the present invention; Figure 8 This is a PCA dimensionality reduction diagram showing the alignment effect of the spectral feature distributions of the source and target domains before and after transfer learning, provided by an embodiment of the present invention; where (a) is the inter-domain distance before transfer, which is 126.9529, and (b) is the inter-domain distance after transfer, which is 54.7451. Figure 9 This is a PLS model calibration curve of Mn, Co, Cu, and Ti elements in the source domain provided in this embodiment of the invention; Figure 10 This is a PLS model calibration curve of Mn, Co, Cu, and Ti elements in the target domain provided in this embodiment of the invention; Figure 11 This is a calibration curve of Mn, Co, Cu, and Ti elements after processing by a combination of optical-acoustic fusion and transfer learning methods provided in an embodiment of the present invention. Detailed Implementation

[0027] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0028] LIBS (Liquid-Induced Bioassay) is an emission spectroscopic analysis technique that uses high-energy laser pulses to irradiate the surface (solid) or interior (liquid or gas) of a sample, generating a transient high-temperature plasma. When the focused laser pulse reaches a threshold energy density, the sample surface rapidly ablates, forming plasma. During plasma cooling, excited-state ions and atoms transition to lower energy levels or the ground state, emitting photons with specific frequencies. The characteristic spectra of these emissions reflect the elemental composition of the sample. Simultaneously, the plasma expansion process excites acoustic signals with energy information. The effective period of the acoustic signal is dominated by the initial shock wave, which propagates and directly drives the microphone diaphragm to vibrate. The acoustic signal acquired at this stage is not affected by sample vibration coupling or environmental reflections; therefore, its signal characteristics are determined solely by the laser pulse parameters and the inherent physicochemical properties of the sample, representing the core stage reflecting the sample's essence.

[0029] This invention leverages the characteristics of plasma acoustic signals to overcome the limitations of solely relying on spectral intensity. By performing feature-level early fusion of spectral and acoustic signals, the model can simultaneously learn signal changes resulting from variations in sample chemical composition and physical information, thereby improving the model's interpretability and robustness. Transfer learning treats data collected at a fixed distance (source domain) and data collected at varying distances (target domain) as two domains with different data distributions. This invention achieves the transfer of knowledge from the source domain to the target domain through feature transfer (screening for common cross-domain features) and instance transfer (screening for source domain samples with similar distributions). It utilizes experimental data under ideal conditions to improve model performance, ensuring high-accuracy prediction performance even under complex conditions.

[0030] To further ensure that the optical-acoustic fusion transfer method of this invention is applicable to different shipboard environments, different matrix materials, and different surface morphology conditions, in specific implementation, this invention adopts a processing sequence of feature transfer first, followed by instance transfer. This is because there is a significant shift in the spectral feature distributions of the source and target domains. Directly mixing and training would cause the model to be interfered with by irrelevant features, making it unable to characterize the elemental concentration of the sample. The core of feature transfer is to first screen highly correlated features from the source and target domains separately, and then merge to select common features. These common features are the core features that simultaneously correlate the source and target domains, filtering out a large number of noisy features and achieving a unified alignment of the feature spaces of the source and target domains. If instance transfer is performed directly, the model will be affected by sample position fluctuations due to the inconsistency of the feature space, and will not be able to effectively learn the correlation rules of elemental concentration.

[0031] During instance migration, this invention employs weighted cosine similarity to analyze spectral and acoustic modes. This is because the two types of features differ significantly in scale and variation. Directly concatenating them or relying solely on single-mode analysis often results in one mode having an excessively high weight in distance calculations, leading to selection errors in the screening results. Weighted cosine similarity can simultaneously consider the directional consistency of feature vectors and adjust the contribution of the two modes to similarity calculations through weighting, allowing optical-acoustic information to participate in the sample matching process in a more balanced manner.

[0032] To ensure the stability of the above instance migration under different environments, this invention modally weights... Migration ratio set The parameters define the preferred range. In practical applications, Typically limited to the range of 0.6 to 0.8, a screening structure is formed with spectral modes as the primary component and acoustic modes as a secondary component, ensuring that both contribute effectively and preventing the weakening of certain types of information due to weight imbalance. The instance migration ratio set in this invention... The migration ratio can be set between 0.2 and 0.7. A threshold that is too low (<0.2) will result in insufficient source domain samples, while a threshold that is too high (>0.7) will introduce a large number of samples that are significantly different from the target domain, which carries the risk of negative migration.

[0033] Through the above technical approach, this invention has formed a stable modeling mechanism applicable to shipboard conditions in terms of feature alignment, sample selection, and dual-modal fusion, enabling the model to maintain high adaptability under different sample position fluctuations and different matrix material conditions.

[0034] The innovations of this invention are as follows: This invention proposes a dual-modal optical-acoustic cross-domain transfer modeling system for complex field environments, achieving consistent correction of spectral and acoustic feature distributions under different acquisition scenarios. This invention is the first to combine optical-acoustic dual-modal information with hierarchical transfer learning, constructing a correction system that combines feature transfer and instance transfer. This enables high-quality data from laboratory conditions to be effectively transferred to target domain scenarios with sample position fluctuations, solving the problem of traditional LIBS's difficulty in quantification under dynamic field environments. This invention constructs an optical-acoustic fusion quantitative modeling mechanism for deep-sea ores, utilizing the complementary characteristics of spectral information and plasma acoustic response to achieve stable grade detection under the influence of sample position fluctuations and matrix effects. This invention uses spectral features to characterize elemental chemical information and combines acoustic frequency domain features to reflect physical information, constructing a unified quantitative model through early fusion of dual-modal features. The acoustic mode is more stable under conditions of positional changes and matrix differences, providing physical compensation for the spectral mode, ensuring good predictive consistency of the final model under different altitudes and matrix conditions. This fusion mechanism overcomes the limitations of traditional single-modal methods in terms of robustness in complex scenarios, providing an engineering-ready solution for rapid on-site evaluation of deep-sea mineral resources.

[0035] Example 1, as Figure 1 As shown, the method for quantitative detection of deep-sea ore grade based on laser-optical-acoustic fusion provided in this embodiment of the invention includes the following steps: S1, Signal Acquisition: Under ideal conditions in the source domain laboratory and fluctuations in the sample position in the target domain, a pulsed laser and a signal acquisition device are synchronously triggered to acquire time-consistent laser-induced breakdown spectral signals and plasma acoustic signals, and to construct source domain datasets and target domain datasets. LIBS spectra and plasma acoustic signals were acquired. A pulsed laser was used to break down the cobalt-rich crust pellet samples (39 samples of different concentrations) to be tested, and the spectral signals were acquired using a spectrometer. Simultaneously, a microphone is used to collect sound wave signals. Two datasets were constructed, requiring the acquisition of photoacoustic signals from two environments: a source domain (ideal laboratory conditions) and a target domain (sample position fluctuation conditions). For each environment, both signals were synchronously triggered using a signal trigger to ensure acquisition under the same laser pulse, resulting in a time-consistent source domain dataset. and target domain dataset ,in, This indicates the sample number under ideal laboratory conditions. This indicates the sample number under conditions of sample position fluctuation.

[0036] This embodiment sets up two types of environments: a source domain (fixed sample height) and a target domain (nine height variations) to simulate continuous sample position shifts under shipboard conditions. By constructing nine height conditions in the target domain, this embodiment can systematically cover typical position ranges that may occur in on-site detection, making subsequent migration correction practically applicable. Simultaneously, a signal trigger connects the laser, spectrometer, and data acquisition card; both spectral and acoustic signals are acquired using synchronous triggering to ensure a strict correspondence between photoacoustic modes.

[0037] S2, Signal preprocessing and feature extraction: The laser-induced breakdown spectral signal is preprocessed and the spectral mode feature vector is extracted. The plasma acoustic wave signal is preprocessed and the acoustic mode feature vector is extracted. A multimodal feature database containing element concentration labels of spectral signal and acoustic wave signal is obtained. Extraction of spectral and acoustic signals, preprocessing and feature extraction, and analysis of spectral signals. Baseline removal, smoothing filtering, integral area normalization, and inter-point averaging preprocessing are performed, and outlier signals are removed using cosine similarity to obtain a stable spectral intensity curve. These operations are necessary because deep-sea mineral LIBS signals are significantly affected by matrix effects, and sample position fluctuations cause nonlinear shifts in the overall spectral morphology. These processing steps significantly reduce the overall spectral intensity differences at different altitudes, thus alleviating the distribution differences caused by altitude drift. In quantitative results, taking Mn as an example, the coefficient of determination of the preprocessed spectral model is significantly higher than that of the training set before preprocessing. The accuracy improved from 0.724 to 0.979 (a 35.22% increase); the prediction root mean square error (RMSEP) decreased from 1.434 wt.% to 0.943 wt.% (a 34.24% decrease). This applies to the preprocessed spectral signal. Feature extraction is performed to obtain spectral features. ,in, Indicates the first The intensity of spectral lines at each wavelength point. All spectral features constitute the spectral mode eigenvector. .

[0038] For sound wave signals Perform time window interception and capture the value after laser breakdown. - effective segment acoustic signal in milliseconds In this example, a window of 0.24-0.34ms is selected as the effective window. The selected acoustic signal is then subjected to a Fast Fourier Transform (FFT) to obtain its frequency domain characteristics. , ,in, Indicates the first One frequency component, To correspond to the complex amplitude, its magnitude is taken, relatively normalized, and converted to a decibel scale: Acoustic modal features were obtained. Ultimately, a multimodal feature database containing spectral and acoustic signals is obtained. and ,in, and Representing the spectral and acoustic characteristic modes, respectively. For element concentration labels.

[0039] S3, Modal Transfer Learning Correction: At the feature level, a subset of common features that are strongly correlated with and stable with element concentration in the source and target domains are selected to achieve feature space alignment; at the instance level, based on the weighted fusion similarity of spectral and acoustic modes, the sample that best matches the distribution of the target domain is selected from the source domain, and the training set with the target domain is used to construct an enhanced training set. To address the inconsistency in optical-acoustic feature distribution between the source domain (ideal laboratory conditions) and the target domain (sample position fluctuation conditions), this embodiment employs a fine-tuning-based transfer learning algorithm, combining feature transfer and instance transfer in a two-layer strategy. This is because sample position fluctuations cause only some features in the spectral feature dimension to remain stable between the two domains. By filtering common features through feature transfer, stable variables that maintain consistency across altitude conditions can be preferentially retained, making source domain knowledge transferable; however, feature transfer alone cannot solve the sample distribution offset problem. This embodiment further introduces instance transfer after feature transfer, using bimodal weighted cosine similarity to filter source domain samples. Samples more closely related to the physical conditions of the target domain can be added to the enhanced training set, thereby enabling the model to have better data coverage at different altitudes. First, standardization and feature selection are performed on spectral and acoustic features separately at the feature level. Operations are performed independently on spectral and acoustic modes. Taking spectral features as an example, in the source domain feature set... and target domain feature set The SelectKBest algorithm was used to filter the elements with the highest correlation to their concentrations. Features, and based on the intersection of the two domain filtering results As a subset of common features and These represent the filtering features of the source and target domains, respectively. If... Then, according to the sum of correlation coefficients Supplementing common features, among which, and They represent the first The correlation coefficients between each feature and its concentration in the source and target domains are obtained; finally, the transferred feature matrix is ​​obtained. and This achieves feature space alignment. In the instance transfer layer, the transferred features are first subjected to modality-specific normalization to obtain... and Then, the weighted cosine similarity is calculated. To measure the degree of matching between the source domain samples and the target domain distribution. These are the spectral mode weighting parameters. and These are the spectral and acoustic modal cosine similarities, respectively. Further, in the candidate proportion set... Search for the optimal migration ratio ,in, Indicates the migration ratio. This indicates the highest similarity. The augmented training set is constructed by merging the source domain samples and the target domain training set. The coefficient of determination on the validation set This represents the total number of samples in the source domain.

[0040] To address the inconsistency in optical-acoustic feature distribution between the source domain (ideal laboratory conditions) and the target domain (sample position fluctuation conditions), a transfer learning algorithm based on a fine-tuning strategy is introduced, and a two-layer strategy combining feature transfer and instance transfer is adopted.

[0041] Step 3.1: Feature Transfer Layer – Achieving cross-domain feature space alignment; This layer filters common features that are strongly correlated with element concentration and have stable distribution in both the source and target domains, eliminates domain-specific noise features, achieves unification of the feature spaces of the two domains, and eliminates feature distribution shifts caused by differences in the detection environment.

[0042] Step 3.1.1: Modal Feature Filtering Based on SelectKBest In the source domain feature set and target domain feature set The correlation of modal features was calculated separately. The Pearson coefficient was used. Combining the SelectKBest algorithm to filter the top elements with the highest correlation to element concentration One feature, among which and They represent the first The first sample and the first One characteristic, Indicates the first The sample at the th The values ​​that can be taken on each feature Indicates the first The known elemental concentrations of a sample and These represent the average values ​​of the feature and concentration, respectively. The source region spectral screening features are obtained. Target domain spectral screening features Source domain acoustic screening features Target domain acoustic screening features .

[0043] Step 3.1.2: Extraction of the intersection of cross-domain common features; The top results filtered by both the source domain and the target domain Find the intersection of the feature sets to obtain and To prevent the number of intersections from being too small, a supplementary strategy was designed. Then according to the sum of correlation coefficients Subsidies are provided based on public characteristics, among which, and They represent the first The correlation coefficients of the features with concentration in the source and target domains.

[0044] Step 3.1.3: Modal standardization after common feature extraction; After extracting the common stable features of the two domains, the optical and acoustic features are standardized separately. ,in, Represents the original values ​​of the same feature. This represents the average of all samples for that feature. This represents the standard deviation of all samples for this feature. This is to ensure that spectral and acoustic modes are comparable between the source and target domains, so that subsequent similarity calculations, instance selection, and fusion modeling are not affected by differences in feature dimensions.

[0045] Step 3.2: Instance Transfer Layer – Implements similar sample selection and enhances training set construction; This layer performs further cross-domain correction on the source and target domain data after feature transfer at the sample level to construct an enhanced training set.

[0046] Step 3.2.1: Calculate the modal similarity; Calculate the cosine similarity between the source and target domain training sets for the transferred spectral and acoustic modes, respectively: , , and For source domain spectral and acoustic signal data, and For target domain spectral and acoustic signal data.

[0047] Step 3.2.2: Dual-modal weighted similarity fusion; The cosine similarity between the two modules is weighted according to a set parameter. Perform weighted fusion. This is used to measure the degree of matching between the source domain samples and the target domain distribution.

[0048] Step 3.2.3: Source domain sample screening and migration ratio search based on weighted similarity; The source domain samples are sorted from highest to lowest fusion similarity, and then further refined in the candidate proportion set. Search for the optimal migration ratio ,in, This indicates the highest similarity. The augmented training set is constructed by merging the source domain samples and the target domain training set. The coefficient of determination on the validation set.

[0049] S4, Optical-Acoustic Dual-Modal Fusion Modeling: A feature-level early fusion strategy is used to directly splice optical-acoustic features on the enhanced training set. A quantitative model is established through partial least squares regression. Within the preset number of principal components, cross-validation is used to select the optimal model parameters. Enhanced training set after transfer correction in the optical-acoustic fusion stage The modeling strategy employed is feature-level early fusion; feature-level early fusion involves directly splicing spectral features at the input level. Acoustic characteristics Construct joint features We directly train a partial least squares regression model to perform overall modeling of multimodal features.

[0050] In early feature-level fusion, the input layer incorporates common spectral features. Harmony and common characteristics Direct splicing to form a fusion feature As model input; through the PLS regression layer, the principal components of PLS ​​are represented in linear combination form: , For the first The principal component direction coefficients are calculated. Within a preset range of principal component numbers, the number of candidate components is trained using a 5-fold cross-validation selection principle, with the minimum RMSE (Real-Timestamp) of the validation set used as the criterion for selecting the principal component number. In this embodiment, feature-level early fusion is employed because by splicing common spectral features and acoustic frequency domain features, the complementary relationship between the two modes can be directly learned at the input layer. Spectra reflect the optical response of elemental content, while acoustic waves reflect physical information. The two are affected differently by sample position fluctuations; therefore, their joint modeling can effectively offset the instability factors of a single mode. Experiments show that compared to using only spectra, the dual-modal fusion model has a higher coefficient of determination and lower prediction error.

[0051] S5, Model Validation and Output: Use the final model to predict the elemental concentration of the target domain test set and calculate the evaluation index to output the quantitative detection results.

[0052] The model performance was evaluated using 5-fold cross-validation and test sample validation, and the coefficient of determination was calculated. Root mean square error (RMSE) and relative standard deviation (RSD). , , ,in, For the first The true concentration of each sample To predict concentration, This is the average concentration. The total number of samples, This represents the standard deviation of the prediction results. The final output, after transfer correction, is evaluated using the performance metrics of the early fusion model. The system automatically selects the optimal parameter strategy as the final quantitative model, achieving model-level fusion.

[0053] This embodiment fully demonstrates the practical application value of the present invention under simulated shipboard conditions, proving that the technology can achieve rapid, accurate and stable quantitative detection of deep-sea ore grade in complex environments.

[0054] Example 2: In terms of acoustic signal feature extraction, in addition to using FFT to extract frequency domain features, wavelet packet transform can also be used to extract joint time and frequency domain features.

[0055] Example 3: Regarding transfer learning algorithms, in addition to transfer methods based on feature selection and similarity calculation, deep transfer learning methods such as Transfer Component Analysis (TCA) and Joint Distribution Adaptation (JDA) can also be used to achieve domain adaptation.

[0056] Example 4: In the multimodal fusion stage, in addition to PLS regression, other regression models such as support vector regression (SVR), artificial neural network (ANN), and convolutional neural network (CNN) can also be used as the fused predictor.

[0057] Example 5: The deep-sea ore grade quantitative detection system based on laser-optical-acoustic fusion provided in this embodiment of the invention includes: The signal acquisition device includes a 1064nm Nd:YAG pulsed laser, an electrically driven three-dimensional displacement platform, a spectral signal receiving system, and an acoustic acquisition module. It is used to break down the sample under source and target domain conditions and acquire LIBS spectral signals and plasma acoustic signals. The electrically driven three-dimensional displacement platform is used to change the relative height of the sample to the laser focus. The spectral signal receiving system is an Avantes four-channel spectrometer. The acoustic acquisition module includes a microphone (cardioid, frequency response 20Hz-20kHz) and a high-speed data acquisition card with a sampling rate of 500kHz. The synchronization triggering device is an ASG8100, which is communicatively connected to the signal acquisition device. It is used to synchronize the timing of the pulsed laser, spectrometer and data acquisition card to ensure that the timing of the acquired optical-acoustic signals is consistent. A data processing and control device, communicatively connected to a signal acquisition device, and configured for: (a) Preprocessing and feature extraction of spectral and acoustic signals to form a multimodal feature database; (b) Perform submodal transfer learning corrections to construct an enhanced training set; (c) Perform optical-acoustic dual-modal fusion modeling and select the final quantitative model; (d) Output the quantitative detection results of element concentration.

[0058] To further demonstrate the positive effects of the above embodiments, the present invention conducts the following experiments based on the above technical solutions.

[0059] The implementation process of the method used in this invention is as follows: Figure 2 As shown, LIBS spectra and plasma acoustic signals were obtained from the LIBS experimental system. After preprocessing the spectral data and extracting the acoustic frequency domain features, feature transfer and instance transfer were performed on them based on a fine-tuned transfer strategy. Then, the acoustic features and spectral features were fused at the feature level in the early stage, and a quantitative model was established through partial least squares regression. The above steps were performed on the four elements Mn, Co, Cu, and Ti respectively to establish independent quantitative models. To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings.

[0060] Experimental Environment and Materials: Thirty-nine manually pressed cobalt-rich crust standard samples (the concentrations of Mn, Co, Cu, Ti, etc. are shown in Table 1) were used for model training and validation. The samples were then processed according to... Figure 3The diagram illustrates the construction of a LIBS photoacoustic signal acquisition system. The core components include: a 1064nm Nd:YAG pulsed laser (pulse energy 40mJ, repetition rate 1Hz); an Avantes four-channel spectrometer; a microphone (cardioid, frequency response 20Hz-20kHz) paired with a high-speed data acquisition card (sampling rate 500kHz) to acquire acoustic signals; an ASG8100 signal trigger; and a precision motorized three-dimensional stage for accurately controlling the sample position and simulating the relative distance (height) change between the laser focus and the sample surface. Data processing, transfer learning, and modeling algorithms are implemented in a Python environment. The following sections will further illustrate this process in detail with accompanying figures and examples, demonstrating the feasibility and superiority of the proposed correction method.

[0061] Table 1 Concentration of 39 groups of cobalt-rich crust pellet samples

[0062] The specific steps are as follows: Step 1: Acquisition of LIBS spectrum and plasma acoustic wave signals.

[0063] Thirty-nine cobalt-rich crust samples were sequentially placed on a three-dimensional displacement stage. For source domain data acquisition, the laser focus was precisely fixed at the sample surface (defined as "0" height), and laser penetration was performed on each sample. The penetration point position was changed every 10 triggers. Each sample was simultaneously triggered with 1000 sets of spectral and acoustic signals via an ASG8100. The data acquired under these conditions served as a reference dataset under stable conditions. For target domain data acquisition, to simulate sample position fluctuations in a shipboard environment, the sample height was systematically changed using the displacement stage. At nine different heights (-1mm, -0.75mm, -0.5mm, -0.25mm, 0mm, +0.25mm, +0.5mm, +0.75mm, +1mm), 900 sets of photoacoustic signals were also acquired for each sample. Figure 4 , Figure 5 The images show spectral and acoustic signal graphs at different altitudes. The data collected under these conditions simulates complex scenarios involving positional fluctuations and signal drift in real-world applications. Based on sample concentration and experimental design, the dataset is divided into training and prediction sets, as detailed in Table 2.

[0064] Table 2. Division of source domain, target domain, and training and prediction sets of samples involved in transfer learning.

[0065] * Acoustic signal data were collected at 9 different altitudes. Step 2: Preprocessing and feature extraction of spectral and acoustic signals.

[0066] In the preprocessing stage, baseline correction and Savitzky-Golay smoothing are performed on each raw spectrum to remove noise. The smoothed spectra are then normalized by integral area, unifying the integral intensity of all spectra to the same scale. Multiple measurements of each sample under the same conditions are averaged (averaging in groups of 10), and the cosine similarity to the median reference spectrum is calculated. Outliers with the lowest similarity (top 20%) are removed, resulting in stable and clean spectral data, and a spectral feature vector is constructed.

[0067] like Figure 5 As shown, when preprocessing the acoustic signal, its time-domain characteristics are analyzed, and the time window from 0.24ms to 0.34ms is determined as the effective information segment. The acoustic signal data within this window is extracted and subjected to a Fast Fourier Transform (FFT) to convert the signal to the frequency domain, as shown... Figure 6 As shown. The amplitude spectrum is converted to a decibel (dB) scale to enhance the dynamic range and comparability of the features. The top 26 maximum amplitude intensity values ​​are extracted from the frequency domain signal to form the frequency domain feature vector for each acoustic observation.

[0068] Step 3: Modal transfer learning correction.

[0069] The specific methods and processes of using transfer learning are as follows: Figure 7 As shown, the spectral and acoustic features of the training sets in the source and target domains are processed independently. Using the SelectKBest algorithm, the correlation between each feature and the concentration of the target element (e.g., Mn) is calculated, and the most relevant feature subsets are selected from both domains. The intersection of the selected features from both domains is taken as the cross-domain stable "common feature subset." This process achieves alignment between the source and target domains in the feature space. Afterwards, based on the common feature subset, the spectral and acoustic features are standardized separately. The weighted cosine similarity (weights: 0.6 for spectral and 0.4 for acoustic) between each sample in the source domain and the sample in the target domain in the spectral and acoustic feature spaces is calculated. The performance on the target domain validation set (R0) is then evaluated. 2 The optimization process involves determining the optimal instance transfer ratio (e.g., 40%) from candidate ratios [0.2, 0.3, 0.4, 0.5, 0.6, 0.7], and then selecting a subset of samples from the source domain that are most similar to the target domain distribution. These source domain samples, selected through feature transfer and instance transfer, are then merged with the target domain training set to form an enhanced fine-tuning training set. This strategy effectively utilizes knowledge from stable laboratory data (source domain) to assist and improve the performance of models built in changing environments (target domain).

[0070] Step 4: Optical-acoustic dual-modal fusion modeling.

[0071] An early feature-level fusion strategy was employed to splice the optical-acoustic signals. Early fusion directly concatenates the spectral and acoustic feature vectors, corrected through transfer learning, to form a high-dimensional fusion feature matrix. After standardization and feature selection, this matrix was input into a partial least squares regression (PLS) model for training. Five-fold cross-validation was used to determine the optimal number of principal components (range 1-30) for PLS.

[0072] Step 5: Model validation and elemental quantitative output.

[0073] Perform steps 1-4 above on each of the four elements Mn, Co, Cu, and Ti to establish independent quantitative models.

[0074] The model performance is evaluated based on the calibration curves of the target domain test set (see Table 3). The final system output is the quantitative prediction of the grade of each element in the unknown cobalt-rich crust sample (prediction set).

[0075] Table 3 Evaluation parameters of calibration curves for different methods

[0076] Implementation Results: This invention, through comparative analysis of acoustic signal feature selection, verifies that in quantitative tasks involving Mn, Co, Cu, and Ti, the frequency domain features of acoustic signals provide more crucial information for improving quantitative performance. Compared to the time domain features (peak-to-peak value), they possess more feature dimensions, which is reflected in the quantitative results, particularly in the training set determination coefficient. The values ​​increased from 0.844, 0.975, 0.968, and 0.763 to 0.914, 0.983, 0.981, and 0.849. The effective period of the acoustic signal is dominated by the initial shock wave, which propagates and directly drives the microphone diaphragm to vibrate. The acoustic signal acquired at this time has not been affected by sample vibration coupling or environmental reflection interference. Therefore, its signal characteristics are determined solely by the laser pulse parameters and the inherent physicochemical properties of the sample, representing the core stage reflecting the essence of the sample. Subsequent acoustic signals are mainly affected by environmental interference and have no analytical value. Therefore, extracting the effective time window (0.24-0.34 ms) further improves the model's predictive performance compared to the frequency domain characteristics of the acoustic signal across the entire time domain, increasing the prediction set determination coefficient. The relative standard deviation (RSDp) of the prediction set decreased from 6.978%, 4.765%, 8.377%, and 13.335% to 5.980%, 4.621%, 6.517%, and 12.965%, respectively, as the RSDs increased from 0.913, 0.978, 0.978, and 0.839 to 0.935%, 0.980%, 0.980%, 0.980%, 0.987%, and 0.853.

[0077] Table 3 provides a clear overview of the quantitative performance of each method. The comparison methods are: PLS quantitative analysis of source domain data, PLS quantitative analysis of target domain data, quantitative analysis of target domain prediction set data directly transferred from the model trained on source domain data, quantitative analysis using only the optical-acoustic fusion method, quantitative analysis using only the transfer learning method, and quantitative analysis of optical-acoustic fusion transfer.

[0078] In the quantitative analysis of Mn, Co, Cu, and Ti, the performance of the photoacoustic fusion migration method of this invention and the direct PLS quantitative method on the target domain test set are compared. The results are shown in Table 3. (Training set determination coefficient) The accuracy improved from 0.948, 0.981, 0.984, and 0.932 to 0.979, 0.990, 0.993, and 0.982, respectively; the root mean square error (RMSEP) of the prediction set decreased from 0.975 wt.%, 0.019 wt.%, 0.007 wt.%, and 0.134 wt.% to 0.777 wt.%, 0.015 wt.%, 0.005 wt.%, and 0.091 wt.%, respectively; and the relative standard deviation (RSDp) of the prediction set decreased from 6.856%, 5.670%, 8.087%, and 14.567% to 5.309%, 4.495%, 6.115%, and 9.931%, respectively, demonstrating higher prediction accuracy and better resistance to signal drift. Taking the quantitative analysis of cobalt (Co) as an example, compared with the "direct migration" method that simply applies a stability model to changing conditions, the performance (R) of this invention is significantly better. 2 =0.980) is far superior to the latter (R 2 =0.559), which fully demonstrates the effectiveness of the modal feature transfer and instance transfer framework of the present invention.

[0079] This invention also verifies that, compared to models built using only the photoacoustic fusion method, the photoacoustic fusion transfer model of this invention can effectively improve quantitative accuracy. This is because, although introducing plasma acoustic signals can compensate for matrix effects and improve prediction accuracy, fluctuations in sample position can cause differences in the intensity and feature distribution of photoacoustic data. These differences can shift the mapping relationship between photoacoustic features and elemental concentrations. Reflected in the quantitative results, the coefficient of determination of the training set... The values ​​were improved from 0.882, 0.983, 0.975, and 0.782 to 0.979, 0.990, 0.993, and 0.982, respectively; the root mean square error (RMSEP) of the prediction set decreased from 0.824 wt.%, 0.016 wt.%, 0.006 wt.%, and 0.119 wt.% to 0.777 wt.%, 0.015 wt.%, 0.005 wt.%, and 0.091 wt.%, respectively; and the relative standard deviation (RSDp) of the prediction set decreased from 5.980%, 4.621%, 6.517%, and 12.965% to 5.309%, 4.495%, 6.236%, and 9.931%. Compared to two-layer transfer learning models that analyze only spectral modes, the modal-specific transfer learning strategy based on optical-acoustic fusion in this invention effectively improves the stability of transfer results. This is achieved through multimodal complementarity of spectral and acoustic signals, enhancing cross-domain stability and making it more suitable for shipboard on-site inspection scenarios. Furthermore, the modal-weighted transfer method adaptively balances the importance of the two modes, optimizes instance selection, and reduces the transfer risk of a single mode. This is reflected in the quantitative results, such as the training set determination coefficient. The values ​​improved from 0.966, 0.989, 0.992, and 0.961 to 0.979, 0.990, 0.993, and 0.982, respectively; the root mean square error (RMSEP) of the prediction set decreased from 0.805 wt.%, 0.018 wt.%, 0.006 wt.%, and 0.109 wt.% to 0.777 wt.%, 0.015 wt.%, 0.005 wt.%, and 0.091 wt.%, respectively; and the relative standard deviation (RSDp) of the prediction set decreased from 5.714%, 5.437%, 6.667%, and 11.858% to 5.309%, 4.495%, 6.236%, and 9.931%.

[0080] like Figure 8 As shown in Figures (a) and (b), the PCA dimensionality reduction visualization demonstrates that after the transfer learning processing of this invention, the feature distributions of the source and target domains achieve better alignment, explaining the improved model robustness from a data distribution perspective. The calibration curves of each element (Mn, Co, Cu, Ti) after processing by the method of this invention are shown below. Figure 11 Compared to the source domain ( Figure 9 ) and target domain ( Figure 10 The calibration curves obtained by direct processing all showed good linearity and prediction accuracy, verifying the universality and effectiveness of the present invention in multi-element quantitative analysis.

[0081] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A deep-sea ore grade quantitative detection method based on laser photo-acoustic fusion, characterized in that, The method constructs a quantitative framework driven by physical mechanism compensation and data distribution compensation, intercepts the initial shock wave segment after laser breakdown for FFT as acoustic modal information, and realizes the synchronous high-precision quantification of deep-sea ore multi-elements through the progressive double-layer structure of light-acoustic modal through feature migration and instance migration, retraining of the feature-level early fusion model, and the specific steps include the following steps. S1, signal acquisition: under the ideal conditions in the source domain laboratory and the sample position fluctuation conditions in the target domain, the pulsed laser and the signal acquisition device are triggered synchronously to obtain time-consistent laser-induced breakdown spectroscopy signals and plasma acoustic wave signals, and source domain datasets and target domain datasets are constructed; S2, signal preprocessing and feature extraction: the laser-induced breakdown spectroscopy signals are preprocessed and the spectral modal feature vectors are extracted, the plasma acoustic wave signals are preprocessed and the acoustic modal feature vectors are extracted, and a multi-modal feature database containing the spectral signal and the acoustic signal element concentration label is obtained; S3, modal transfer learning correction: in the feature level, the common feature subset strongly related to the element concentration and stable in the source domain and the target domain is screened to realize the alignment of the feature space; in the instance level, based on the fusion similarity of the spectral and acoustic modal, the part of samples in the source domain most matched with the target domain distribution is screened, and the enhanced training set is constructed by combining with the target domain training set; S4, light-acoustic dual-modal fusion modeling: the feature-level early fusion strategy is adopted to directly splice the light-acoustic features on the enhanced training set, and the quantitative model is established by partial least squares regression, and the optimal model parameters are selected by cross-validation in the preset principal component number range; S5, model verification and output: the final model is used to predict the element concentration of the target domain test set, and the evaluation indexes are calculated to output the quantitative detection results.

2. The method according to claim 1, wherein, In step S1, time-consistent laser-induced breakdown spectroscopy signals and plasma acoustic wave signals are obtained, including: The pulse laser is used to break the cobalt-rich crust tablet sample to be tested, and a spectrum signal is collected by using a spectrometer , and a microphone collects a sound wave signal Two types of data sets are constructed, and the photo-acoustic signals under ideal conditions in the source domain laboratory and under sample position fluctuation conditions in the target domain are collected; each environment synchronously triggers two signals through a signal trigger to ensure that the data sets are collected under the same laser pulse, and the source domain data set is consistent in time sequence and the target domain data set , wherein, is the sample serial number under the ideal conditions in the laboratory, is the sample serial number under the sample position fluctuation conditions. 3.The deep-sea ore grade quantitative detection method based on laser photo-acoustic fusion according to claim 1, characterized in that, In step S2, the laser-induced breakdown spectroscopy signals are preprocessed and the spectral modal feature vectors are extracted, including: Spectral signal Baseline removal, smoothing filtering, integral area normalization and point-to-point average preprocessing are performed to obtain stable spectral intensity curves by removing abnormal signals through cosine similarity; the preprocessed spectral signal Feature extraction is performed to obtain spectral features , wherein, is the intensity of the spectrum line of the th wavelength point; all spectral features constitute a spectral modal feature vector .

4. The method according to claim 3, wherein, The plasma acoustic wave signals are preprocessed, including: on the acoustic wave signal time windowing, taking the effective segment of the acoustic wave signal after laser breakdown ; fast Fourier transform on the windowed acoustic wave signal to obtain frequency domain features:​ wherein , is the th frequency component, is the corresponding complex amplitude; The modulus value is normalized and converted to decibel scale: The acoustic modal feature vector is: The multi-modal feature database containing the spectral signal and the acoustic signal is: wherein and are spectral and acoustic feature modalities, respectively, is an elemental concentration label. 5.The deep-sea ore grade quantitative detection method based on laser photo-acoustic fusion according to claim 1, characterized in that, In step S3, the modal transfer learning correction includes: introducing a transfer learning algorithm based on fine-tuning strategy, and adopting a double-layer strategy combining feature transfer and instance transfer; At the feature level, standardization and feature selection are performed on spectral and acoustic features respectively; operations are performed independently on spectral and acoustic modes within the source domain feature set. and target domain feature set The SelectKBest algorithm was used to filter the elements with the highest correlation to their concentrations. Features, and based on the intersection of the two domain filtering results As a subset of common features and These are the filtering features for the source and target domains, respectively; if Then, according to the sum of correlation coefficients Supplementing common features, among which, and The first The correlation coefficients between each feature and its concentration in the source and target domains are obtained; finally, the transferred feature matrix is ​​obtained. and To achieve feature space alignment; In the instance migration layer, the migrated features are modality-standardized to obtain and , and then the weighted cosine similarity is calculated to measure the matching degree of the source domain samples and the target domain distribution, is the weight parameter of the spectral modality, and are the cosine similarities of the spectral and acoustic modalities, respectively; the optimal migration ratio is searched in the candidate ratio set , wherein, is the migration ratio, is the enhanced training set constructed by combining the target domain training set with the source domain samples with the highest similarity is the determination coefficient on the validation set, is the total number of source domain samples. 6.The deep-sea ore grade quantitative detection method based on laser photo-acoustic fusion according to claim 5, characterized in that, In the feature transfer layer, the common features related to the element concentration and stable in the source domain and the target domain are screened, the domain-specific noise features are removed, the cross-domain feature space is aligned, the feature distribution deviation caused by the detection environment difference is eliminated, and the specific steps include: SelectKBest-based modality-specific feature selection: modality-specific feature correlation calculation is performed on the source domain feature set and the target domain feature set respectively; the top features with the highest correlation with the element concentration are selected using the Pearson coefficient combined with the SelectKBest algorithm to obtain the source domain spectral screening features , target domain spectral screening features , source domain acoustic screening features and target domain acoustic screening features ; The expression of the Pearson coefficient is: wherein and are the first sample and the first feature, is the value of the first sample on the first feature, is the known elemental concentration of the first sample, and are the average values of the feature and the concentration, respectively; Intersection extraction of cross-domain public features: intersection of the feature sets filtered by the source domain and the target domain respectively ;​​ Design a complementary strategy, if , then the sum of the correlation coefficient Subsidy public feature dimension, where, And The first The correlation coefficient of the feature in the source domain and the target domain with the concentration; After extracting the common stable features of the source domain and the target domain, the light-acoustic features are standardized respectively, that is: wherein, is the original value of the feature, is the mean value of all samples of the feature, is the standard deviation of all samples of the feature.

7. The method according to claim 5, wherein the method is characterized by, In the instance transfer layer, the source domain and the target domain data after feature transfer are corrected at the sample level to realize similar sample screening and enhanced training set construction, and the specific steps include: Modality similarity calculation: the cosine similarity between the source domain and the target domain training sets is calculated for the migrated optical and acoustic modalities, respectively, that is: wherein and are source domain spectral and acoustic signal data, and are target domain spectral and acoustic signal data; Dual-modal weighted similarity fusion: cosine similarity of two modules according to the set weight parameter Weighted fusion is performed to measure the matching degree of the source domain sample and the target domain distribution; Source domain sample screening and migration ratio search based on weighted similarity: arrange the source domain samples in descending order of fusion similarity, and search for the optimal migration ratio in the candidate ratio set ​ In the formula, The highest similarity is The enhanced training set is constructed by merging the source domain sample with the highest similarity and the target domain training set The determination coefficient on the validation set, N is the total number of source domain samples. 8.The deep-sea ore grade quantitative detection method based on laser photo-acoustic fusion according to claim 1, characterized in that, In step S4, the photo-acoustic dual-modality fusion modeling includes: Light-acoustic fusion stage in migration correction after enhanced training set The modeling strategy of feature level early fusion is adopted above; feature level early fusion is directly splicing spectral features at the input level and acoustic features , construct joint features , directly train the partial least squares regression model, and model the overall multi-modal features. 9.The deep-sea ore grade quantitative detection method based on laser photo-acoustic fusion according to claim 1, characterized in that, In step S4, in the early fusion of features, the input layer will form the fusion features by directly splicing the spectral common features and acoustic common features , as model inputs; through the PLS regression layer, the principal components of PLS are expressed in the form of linear combination: , is the direction coefficient of the th principal component; the candidate number is trained within the preset range of the number of principal components, and the principle of 5-fold cross-validation is adopted to select the number of principal components according to the standard of minimum RMSE of the validation set. 10.The deep-sea ore grade quantitative detection method based on laser photo-acoustic fusion according to claim 1, characterized in that, In step S5, model verification and element quantitative output includes: Model performance was evaluated using 5-fold cross-validation and test sample validation to calculate the coefficient of determination (R2), root mean square error (RMSE), and relative standard deviation (RSD) for the training and test sets. , root mean square error (RMSE), and relative standard deviation (RSD) for the training and test sets. wherein is the true concentration of the nth sample, is the true concentration of the nth sample, is the predicted concentration of the nth sample, is the average concentration of the samples, is the total number of samples, is the standard deviation of the predicted results; The performance index of the output data after migration correction through the feature level early fusion model, the system automatically selects the optimal parameter strategy as the final quantitative model, and realizes model level integration.

Images