A deep-sea LIBS rare earth element quantitative detection method and system fusing spatial information

Abstract

The application provides a deep-sea LIBS rare earth element quantitative detection method and system fusing space information, and belongs to the element analysis detection technical field based on deep learning; first, a deep-sea LIBS sampling device is designed, and in-situ sampling of deep-sea sediments is realized through an ROV; then, according to a specific deep-sea area sampling strategy, multi-point sampling experiments are carried out to obtain a rare earth detection data set containing spectral characteristics, space position information and concentration labels; then, taking the sampling points as graph structure nodes, spectral data and space information are fused to construct a spectrum-position fused graph structure data; a rare earth element quantitative detection model is constructed, the internal law of the spatial distribution of rare earth elements is captured by using the information propagation mechanism in the graph structure, and the identification of the rare earth enrichment area is realized in combination with a regional aggregation method; the application realizes in-situ detection of rare earth elements and real-time identification of enrichment areas in a deep-sea environment, and provides intelligent technical support for deep-sea resource exploration.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of element analysis detection based on deep learning, and particularly relates to a deep-sea LIBS rare earth element quantitative detection method and system fusing spatial information. BACKGROUND

[0002] The complex deep-sea environment, high pressure, low visibility, and violent water disturbance bring great challenges to the in-situ detection and quantitative analysis of rare earth elements. The current seabed element detection methods mostly face problems such as poor in-situ performance, strong signal interference, low detection accuracy, and lack of spatial information, which are difficult to meet the efficient, accurate and automated needs of rare earth resource exploration. The existing rare earth element quantitative detection methods mainly include the following three kinds:

[0003] Laboratory post-processing analysis method: This method combines inductively coupled plasma mass spectrometry (ICP-MS) analysis method for rare earth element quantitative detection, which is one of the most widely used technologies in rare earth element detection. By collecting deep-sea sediment samples and performing acid dissolution, purification and other complex pretreatment processes in the laboratory, high-sensitivity quantitative analysis is performed using ICP-MS. Although it has high detection accuracy, this method is heavily dependent on sample recovery and laboratory conditions, and cannot achieve in-situ real-time detection. At the same time, the sample may be contaminated or its composition may change during transportation and processing, affecting the representativeness and reliability of the data. In addition, this method is costly and time-consuming, and is not suitable for large-scale, rapid seabed resource survey tasks.

[0004] Underwater laser-induced breakdown spectroscopy (LIBS) in-situ detection method: This method is tried to be applied to underwater element identification and sediment characterization due to its in-situ detection capability. Related technologies are mostly based on integrating LIBS probes into submersible platforms to achieve laser excitation and spectral collection of sediment surface. However, traditional underwater LIBS devices mostly directly excite the sediment surface, lacking sampling and enrichment capabilities for internal components, resulting in rare earth element signals being easily affected by environmental factors such as silt coverage and seawater disturbance, with low signal-to-noise ratio and poor stability. In addition, existing underwater LIBS systems are mostly single-point detection, without fusing spatial information, making it difficult to realize regional element distribution modeling and enrichment state identification.

[0005] Machine learning-based single-point modeling method: This method attempts to apply machine learning methods to the prediction of rare earth element concentrations from LIBS spectral data, using algorithms such as support vector regression (SVR) and random forests to model the collected single-point spectral data. This method improves the intelligence level of data processing to some extent, but most of the models are based on single-point spectral features, and do not consider the spatial correlation between sampling points, lacking the ability to express regional structural information. At the same time, this model is easily affected by the limitations of training samples, and has insufficient generalization ability, making it difficult to maintain prediction stability and accuracy in complex deep-sea geological backgrounds, and has not yet formed an integrated detection system that can be deployed and operated in deep-sea environments.

[0006] Therefore, the existing deep-sea rare earth element detection technology has different degrees of limitations in terms of in-situ nature, precision, anti-interference ability, and spatial information integration, making it difficult to meet the current needs of efficient, accurate, and automated deep-sea resource exploration. In this context, there is an urgent need to develop an integrated technology path that combines in-situ detection capability, high-sensitivity analysis performance, and regional modeling expression capability to break through the limitations of traditional methods and achieve rapid, stable, and intelligent identification of deep-sea rare earth elements, thereby providing strong technical support for the scientific evaluation and strategic development of seabed resources. SUMMARY

[0007] To solve the above problems, the present application provides an integrated technology path that combines in-situ detection capability, high-sensitivity analysis performance, and regional modeling expression capability to break through the limitations of traditional methods and achieve rapid, stable, and intelligent identification of deep-sea rare earth elements, thereby providing strong technical support for the scientific evaluation and strategic development of seabed resources.

[0008] The first aspect of the present application provides a deep-sea LIBS rare earth element quantitative detection method fusing spatial information, comprising the following processes:

[0009] S1, based on an integrated deep-sea LIBS sampling device and a regional sampling path and point distribution strategy, performing multi-point deep-sea sediment sampling in a target region to obtain rare earth element samples to be tested in the target region, and recording corresponding spatial position information, then detecting the spectral features of each sample to be tested by laser excitation;

[0010] S2, taking different sampling points as nodes Node in a graph structure, constructing the edge Edge structure of the graph combining the spectral features of the sampling points and the spatial distance, and encoding the spectral features of each sampling point and the spatial position information as node features, converting the sample data to be tested into graph structure data as standard input data for the rare earth element concentration detection model;

[0011] S3, input the graph structure data in S2 into the rare earth element concentration detection model based on the constructed and trained graph neural network (GNN), the model extracts the spatial correlation features of rare earth elements through the node feature propagation and aggregation mechanism of GNN, and outputs the rare earth element concentration of each sampling point;

[0012] S4, based on the rare earth element concentration of each sampling point in the target area, a regional aggregation algorithm is used to calculate the regional rare earth enrichment situation, and the enrichment grade is output.

[0013] Preferably, the specific structure and sampling process of the integrated deep-sea LIBS sampling device are as follows:

[0014] The device comprises a sample tank, a reaction chamber, an ROV mechanical arm, four sets of deep water motors and corresponding piston structures;

[0015] Among them, the fourth deep water motor is connected with the sampling piston through the sliding block, and is controlled and operated through the mechanical arm; the third deep water motor is connected with the acid liquid piston through the sliding block, the right chamber of the acid liquid piston is connected with the reaction chamber through the pipeline, and the acid inlet is arranged in the connecting chamber; the bottom of the reaction chamber is provided with a stirring and drainage integrated pump assembly; the reaction chamber is connected with the sample tank through the pipeline and the stop valve; wherein the first deep water motor is connected with the filtration and pressure reduction piston through the sliding block, the right chamber of the filtration and pressure reduction piston is connected with the sample tank through the pipeline, and the pressure sensor is arranged on the sample tank; wherein the second deep water motor is connected with the seawater piston through the sliding block, and the right chamber of the seawater piston is connected with the left chamber of the filtration and pressure reduction piston through the pipeline; the left chamber of the seawater piston is provided with a seawater inlet;

[0016] Through the operation of the ROV mechanical arm, the sampling piston driven by the fourth deep water motor is inserted into the deep-sea sediment in the predetermined area, and the motor drives the piston to move backward to complete the collection of the sediment sample; then the piston is inserted into the reaction chamber, and is driven in reverse to push the sediment out into the reaction chamber;

[0017] The third deep water motor drives the acid liquid piston to inject a preset volume of quantitative acid liquid into the reaction chamber for sample pretreatment;

[0018] Start the stirring and drainage integrated pump to drive the impeller in the tank to rotate, so that the acid liquid and the sediment are fully mixed and reacted, and the rare earth elements in the sediment are fully dissolved into the acid liquid to realize the release of the components;

[0019] After the mixing reaction is completed, the first deep water motor drives the piston to extract the mixed sample treated in the reaction chamber into the LIBS detection special sample tank;

[0020] After the electric cut-off valve of the sample chamber is closed, the first deep water motor continues to draw, and due to the sealing of the system, the pressure in the chamber will gradually decrease, and the pressure in the sample chamber is monitored through the pressure sensor, and when the pressure in the sample chamber is reduced to the set value, the sample is in the target pressure and sealed state, and can be directly used for laser excitation experiment as the sample to be detected for LIBS detection.

[0021] Preferably, the construction process of the data set for training the rare earth element concentration detection model comprises:

[0022] S11, control the ROV mechanical arm to advance in a zigzag route parallel to the seafloor topography, and set sampling points equidistantly on the path to ensure uniform coverage of the entire target area in the horizontal and vertical directions; in addition, a cleaning reset operation is performed after each sampling, and then the next sampling point is reached;

[0023] S12, based on the sampling path, the integrated deep sea LIBS sampling device is used to obtain a rare earth enriched liquid sample at each sampling point, and laser excitation experiments are performed on the underwater ROV to obtain spectral data Sp1, Sp2,..., Sp N of N samples; at the same time, the spatial information St i of each sampling point is recorded, including longitude lg i , latitude lt i , and depth dp i , i.e. St i =[lg i , lt i , dp i ], and i∈[1,N]; finally, the spatial information St1, St2,..., St N of N samples is obtained;

[0024] S13, determine the rare earth category to be detected; then synchronously send the collected sample into the shipborne laboratory for concentration calibration experiment, and measure the rare earth element concentration values Rc1, Rc2,..., Rc N of N sampling points by inductively coupled plasma mass spectrometry, which are used as the concentration labels of the sample, wherein each rare earth element concentration value Rc i includes the concentration values of ten elements, and i∈[1,N];

[0025] S14, take the obtained multi-point spectral data Sp1, Sp2,..., Sp N and spatial position information St1, St2,..., St N as input data, take the rare earth element concentration values Rc1, Rc2,..., Rc N as target data, and integrate the input data and target data as a complete rare earth detection data set.

[0026] Preferably, the specific process of S2 is:

[0027] Graph node construction: each sampling point is regarded as a node Node in the graph, and there are N nodes in total. The LIBS spectral data and spatial position information corresponding to each node are jointly used as the feature representation of the node. The feature vector of each node is represented as X i , and i [1, N].

[0028] Graph edge construction: whether to establish an edge connection between each pair of nodes is determined based on the spatial distance and spectral similarity between the nodes. The three-dimensional spatial distance between each pair of sampling points is calculated and represented as Dis ij , j [1, N] and j≠i.

[0029] The spectral similarity Cos ij between the spectral data of the i th sampling point and the j th sampling point is calculated and represented as Cos ij ∈[0, 1] and Cos ij , the greater the spectral similarity between the two sampling points, and vice versa. Then, the overall similarity Sim ij between the i th sampling point and the j th sampling point is calculated.

[0030] Finally, the K sampling points with the highest similarity are selected to establish an edge Edge, and the feature vector containing spatial and spectral information is assigned to Edge:

[0031] E ij =[Dis ij , Sim ij ]

[0032] Where E ij is the Edge feature between node i and node j, and i, j [1, N].

[0033] Complete graph construction: based on the graph node construction and the graph edge construction, the data containing the spectral data Sp1, Sp2,..., Sp N and the spatial position information St1, St2,..., St N are represented as a complete graph G=(X, E), where X represents a set containing all N sampling point nodes, and each node is attached to a spectral-site fusion feature vector X i ; E represents a set of all edges, and each edge is attached to a corresponding Edge feature vector E ij .

[0034] Preferably, the rare earth element concentration detection model based on the graph neural network GNN includes an input feature encoding layer, a spatial perception module, and a rare earth element prediction module.

[0035] The input feature encoding layer is used to input the graph representation G=(X, E) containing the node feature vector X i and the Edge feature vector E ij , and encode the X i and E ij respectively through a multi-layer perception, and obtain the encoded node feature X and edge feature E

[0036] The spatial perception module sequentially inputs each node feature and edge feature output by the feature encoding layer into L spatial perception graph attention layers, and finally obtains the encoded feature of each node X wherein i∈[1, N], and the encoded feature is taken as the output feature of the spatial perception module; the spatial perception module is composed of L spatial perception graph attention layers; in each layer, the attention weight is calculated by combining the similarity between the node features and the spatial distance decay factor; through multi-layer spatial perception graph attention propagation, the model learns the context information of different neighborhood scales layer by layer;

[0037] The rare earth element prediction module is based on the obtained encoded feature of each node X Through a multi-layer perception, the node encoded feature is compressed, and then through a fully connected layer and a ReLU activation function, the rare earth element concentration prediction result of each sampling point is obtained

[0038] Preferably, the S4 specific process is:

[0039] The whole region to be detected is divided into R square grids of equal size, and each sampling point is divided into different local area grids;

[0040] For each divided region, the rare earth element concentration prediction values of all sampling points in the region are summarized, and the average concentration μ r and variance Δ r in the region are calculated, wherein μ r and Δ r represent the average concentration and variance of rare earth in the rth region, and r∈[1, R];

[0041] A multi-level discrimination rule is set, and threshold parameters are set, including a rare earth high enrichment mean parameter μ high , a high enrichment variance parameter Δ high , and a medium enrichment mean parameter μ mid ; these threshold parameters are used to determine whether the region is a rare earth enrichment area, and a rare earth enrichment grade label is assigned.

[0042] The second aspect of the application provides a deep-sea LIBS rare earth element quantitative detection system integrating spatial information, comprising a water surface part and an underwater part;

[0043] The underwater part comprises a main cabin and an integrated deep-sea LIBS sampling device; the main cabin contains a laser control module, a spectrometer, a power supply module, a control module, and an in-cabin state monitoring feedback unit;

[0044] The LIBS optical probe is electrically connected to the spectrometer and comprises a laser head and an auxiliary sampling device, and the laser head is electrically connected to the laser control module; the integrated deep-sea LIBS sampling device is based on a regional sampling path and a point distribution strategy, and performs deep-sea sediment sampling at multiple points in a target region to obtain a sample to be measured, and performs excitation experiments through the laser head to obtain spectral characteristics through the auxiliary sampling device;

[0045] The water surface part comprises a control computer, wherein a rare earth element concentration detection model based on a graph neural network (GNN) that is constructed and trained, a spectral-position fused graph structure data modeling module, and a regional aggregation algorithm module are integrated in the control computer;

[0046] The spectral-position fused graph structure data modeling module takes different sampling points as nodes in a graph structure, constructs an edge structure of the graph in combination with spectral characteristics and spatial distances of the sampling points, and encodes the spectral characteristics and spatial position information of each sampling point as node features, and converts sample data to be measured into graph structure data as standard input data of the rare earth element concentration detection model;

[0047] The rare earth element concentration detection model extracts spatial correlation features of rare earth elements through a node feature propagation and aggregation mechanism of GNN, and outputs rare earth element concentrations of each sampling point;

[0048] The regional aggregation algorithm module calculates a regional rare earth enrichment situation using a regional aggregation algorithm based on rare earth element concentrations of each sampling point in a target region, and outputs an enrichment grade.

[0049] Preferably, the specific structure of the integrated deep-sea LIBS sampling device is as follows:

[0050] The device comprises a sample cabin, a reaction chamber, an ROV mechanical arm, four groups of deep water motors, and corresponding piston structures;

[0051] The fourth deep water motor is connected with the sampling piston through a sliding block and is controlled and operated through a mechanical arm; the third deep water motor is connected with the acid liquid piston through a sliding block, the right chamber of the acid liquid piston is connected with the reaction chamber through a pipeline, and an acid adding port is arranged in the connecting chamber; a stirring and drainage integrated pump assembly is arranged at the bottom of the reaction chamber; the reaction chamber is connected with the sample cabin through a pipeline and a stop valve; the first deep water motor is connected with the filtration and decompression piston through a sliding block, the right chamber of the filtration and decompression piston is connected with the sample cabin through a pipeline, and a pressure sensor is arranged on the sample cabin; the second deep water motor is connected with the seawater piston through a sliding block, and the left chamber of the seawater piston is connected with the left chamber of the filtration and decompression piston through a pipeline; the left chamber of the seawater piston is provided with a seawater inlet.

[0052] Preferably, the rare earth element concentration detection model based on the graph neural network GNN comprises an input feature encoding layer, a spatial perception module and a rare earth element prediction module.

[0053] The input feature encoding layer is used for taking a graph representation G=(X, E) containing a node feature vector X i and an Edge feature vector E ij as input, performing feature encoding on X i and E ij respectively through a multilayer perceptron, and obtaining encoded node features X and edge features E

[0054] The spatial perception module sequentially inputs each node feature and edge feature output by the feature encoding layer into L spatial perception graph attention layers, and finally obtains the encoded feature of each node X wherein i∈[1,N], and the encoded feature is taken as the output feature of the spatial perception module; the spatial perception module is composed of L spatial perception graph attention layers; in each layer, the attention weight is calculated by combining the similarity between the node features and the spatial distance decay factor; through multilayer spatial perception graph attention propagation, the model learns the context information of different neighborhood scales layer by layer;

[0055] The rare earth element prediction module obtains the encoded feature X of each node based on the encoded feature, compresses the node encoded feature through a multilayer perceptron, and obtains the rare earth element concentration prediction result of each sampling point through a fully connected layer and a ReLU activation function

[0056] Compared with the prior art, the application has the following innovations:

[0057] (1) Integrated deep-sea laser-induced breakdown spectroscopy (LIBS) sampling device: An integrated deep-sea LIBS device with sediment sampling, reaction chamber acid dissolution, particle filtration and electric field enrichment functions is proposed, which can realize in-situ extraction and signal enhancement of rare earth elements;

[0058] (2) Spectrum-position fusion graph structure data modeling method: A "spectrum-position fusion" graph modeling method for deep-sea rare earth element detection is constructed, which encodes the spectral features and spatial position information as graph node features, and constructs an edge structure considering spatial distance and spectral similarity, to express the distribution relationship of rare earth elements in complex geological environment;

[0059] (3) Rare earth element quantitative detection model and regional aggregation method based on GNN: A graph neural network detection model combined with spectrum-position graph input is constructed, which extracts the context information between sampling points through a spatial perception graph attention layer, and introduces a regional aggregation mechanism, finally realizes the accurate detection and identification of rare earth elements at the sampling point and regional level.

[0060] The beneficial effects brought by the innovation points of the present application include:

[0061] Realize in-situ quantitative detection in deep-sea environment: The integrated deep-sea LIBS sampling and analysis device has high pressure corrosion resistance and automatic sampling function, combined with the constructed graph neural network model, can complete sample collection, signal collection, concentration analysis and enrichment area identification on site at the seabed, breaking through the limitations of traditional laboratory analysis process, realizing integrated sampling detection and identification process;

[0062] Enhance the identification ability and regional adaptability of the enrichment area: The model introduces the fusion expression of spatial proximity and spectral characteristics during modeling, which can not only accurately detect single point samples, but also intelligently identify the rare earth enrichment trend in a specific area, with excellent regional adaptability and identification ability;

[0063] Improve the efficiency and intelligent level of resource exploration: The present application realizes high integration and intelligent deployment at the hardware and algorithm level, significantly improves the efficiency and automation level of deep-sea rare earth resource exploration, and provides technical support for subsequent resource evaluation and mining decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 The overall implementation logic flowchart of the detection method of the present application.

[0065] Figure 2 The structure diagram of the integrated deep-sea LIBS sampling device of the present application.

[0066] Figure 3 The structure diagram of the rare earth element concentration detection model based on GNN of the present application.

[0067] Figure 4 The rare earth element quantitative detection system architecture diagram of the present application.

[0068] Figure 5 The rare earth element average concentration detection heat map in the embodiment of the present application.

[0069] Figure 6 The sampling point level rare earth element concentration detection precision comparison result graph in the embodiment of the present application.

[0070] Figure 7 The regional level rare earth element concentration detection precision comparison result graph in the embodiment of the present application. DETAILED DESCRIPTION

[0071] The overall implementation logic of the present application is as shown in the figure. Figure 1 The present application proposes a deep-sea LIBS rare earth element quantitative detection method and system fusing spatial information. Its main process is as follows: first, design and build an integrated deep-sea LIBS sampling device, realize in-situ sampling of deep-sea sediments through a remotely operated vehicle (ROV), and integrate functions such as extraction, acid dissolution, filtration and enrichment, effectively improving the excitation and detection signal quality of rare earth elements; then, according to a specific deep-sea area sampling strategy, carry out multi-point sampling experiments, and obtain a rare earth detection data set containing spectral features, spatial position information and concentration labels; then, take the sampling point as a graph structure node, fuse the spectral data and spatial information to construct a spectrum-position fused graph structure data; then, construct a rare earth element quantitative detection model based on a graph neural network (GNN), use the information propagation mechanism in the graph structure to capture the internal law of the spatial distribution of rare earth elements, and combine a regional aggregation method to realize the identification of rare earth enrichment areas; finally, encapsulate and deploy the sampling device and the detection model to realize in-situ detection of rare earth elements and real-time identification of enrichment areas in deep-sea environments, and provide intelligent and regional technical support for deep-sea resource exploration.

[0072] The application will be further described below in conjunction with specific embodiments.

[0073] The implementation process of the deep-sea LIBS rare earth element quantitative detection method fusing spatial information in the embodiment is as follows:

[0074] S1, design an integrated deep-sea laser-induced breakdown spectroscopy (LIBS) sampling device; realize in-situ collection of deep-sea sediment samples through ROV operation, wherein the sampling device has sediment extraction, reaction tank acid injection dissolution, particle filtration and electric field enrichment functions, and is used to extract rare earth element samples to be tested;

[0075] S2, design deep-sea area sampling strategy and build dataset; based on S1 integrated deep-sea LIBS sampling device, design area sampling path and point distribution strategy, obtain multi-point rare earth element samples covering the target area, and obtain spectral data through LIBS optical probe excitation experiment, and based on the sampling path and point distribution strategy, obtain and record the corresponding spatial position information (including latitude, longitude and depth), and obtain the rare earth element concentration label through laboratory calibration, thereby building a rare earth detection dataset containing spectral features, spatial information and concentration label;

[0076] S3, graph structure data modeling of spectrum-site fusion; different sampling points are used as nodes (Node) in the graph structure, the edge (Edge) structure of the graph is constructed by combining the spectral features and spatial distance of the sampling points, and the spectral features and spatial position information of each sampling point are encoded as node features, so as to convert the dataset of S2 into graph structure data;

[0077] S4, build a rare earth element concentration detection model based on graph neural network (GNN); use the graph structure data described in S3 to extract the spatial correlation features of rare earth elements through the node feature propagation and aggregation mechanism of GNN, realize the quantitative detection of the rare earth element concentration of each sampling point, and realize the identification and detection of the regional rare earth enrichment trend by combining the regional aggregation method;

[0078] S5, system deployment and application; integrate, package and deploy the rare earth element concentration detection model to realize in-situ quantitative detection and enrichment area identification of rare earth elements in deep-sea environment, and output the sampling point level and regional level rare earth concentration prediction results in real time.

[0079] I. Specific implementation of integrated deep-sea LIBS sampling device

[0080] The composition of deep-sea sediments is complex and the environment is extreme. In order to exclude the influence of other components in the sediments on the LIBS signal of rare earth elements, and enhance the excitation and collection efficiency of the LIBS signal of rare earth elements, the present application designs an integrated deep-sea LIBS sampling device, which can filter out other particulate matters in the sediments, and can enrich rare earth elements through electric field, so as to finally realize the collection of rare earth elements in the sediments and the enhancement of the LIBS signal. The sampling device is shown as Figure 2

[0081] The device includes a sample cabin, a reaction chamber, a ROV mechanical arm, four sets of deep water motors and their corresponding piston structures.

[0082] ​The fourth deep water motor is connected with the sampling piston through a sliding block and is controlled and operated through a mechanical arm; the third deep water motor is connected with the acid liquid piston through a sliding block, the right chamber of the acid liquid piston is connected with the reaction chamber through a pipeline, and an acid adding port is arranged in the connecting chamber; a stirring and drainage integrated pump assembly is arranged at the bottom of the reaction chamber; the reaction chamber is connected with the sample cabin through a pipeline and a stop valve; the first deep water motor is connected with the filtration and decompression piston through a sliding block, the right chamber of the filtration and decompression piston is connected with the sample cabin through a pipeline, and a pressure sensor is arranged on the sample cabin; the second deep water motor is connected with the seawater piston through a sliding block, and the left chamber of the seawater piston is connected with the left chamber of the filtration and decompression piston through a pipeline; the left chamber of the seawater piston is provided with a seawater inlet.

[0083] The sampling process: the purpose of this process is to complete the extraction, acid dissolution, stirring and concentration pretreatment of deep sea sediments, and finally obtain a rare earth element sample that can be used for LIBS detection, which specifically includes the following steps:

[0084] (1) sediment sampling: through the operation of the ROV mechanical arm, the sampling piston driven by the fourth deep water motor is inserted into the deep sea sediment in the predetermined area, the motor drives the piston to move backward to complete the collection of the sediment sample; then the piston is inserted into the reaction chamber and is driven in reverse to push the sediment out into the reaction chamber;

[0085] (2) automatic sealing of the reaction cabin and acid injection: after pulling out the piston, the spring cover on the reaction chamber automatically closes to ensure that the cabin environment is airtight; the third deep water motor drives the acid liquid piston to inject a preset volume of quantitative acid liquid into the reaction chamber for sample pretreatment;

[0086] (3) sample dissolution and mixing: start the stirring and drainage integrated pump to drive the impeller in the cabin to rotate at high speed, so that the acid liquid and the sediment are fully mixed and reacted, and the rare earth elements in the sediment are fully dissolved into the acid liquid to realize the release of the components;

[0087] (4) sample transfer: after the mixing reaction is completed, the first deep water motor drives the piston to extract the mixed sample treated in the reaction chamber into the sample cabin special for LIBS detection;

[0088] (5) sample cabin pressure adjustment: after closing the sample cabin electric stop valve, the deep water motor continues to extract, and due to the sealing of the system, the pressure in the cabin will gradually decrease, and the pressure of the sample cabin is monitored through the pressure sensor; when the pressure of the sample cabin decreases to a set value (the present application sets it to 11 MPa), the sample is ready;

[0089] (6) obtaining the sample to be detected: at this time, the sample is in a target pressure and sealed state, as a sample to be detected S for LIBS detection, and can be directly used for laser excitation experiment.

[0090] Device cleaning and circulation reset process: the purpose of this process is to complete the cleaning and reset of the sample chamber and the pipeline in the system, to prepare for the next sampling cycle, so as to ensure the long-term stability of the system, which includes the following steps:

[0091] (1) seawater flushing preparation after sampling is completed: after detection is completed, the second deep water motor drives the seawater piston to draw a predetermined volume of clean seawater; then open the stop valve and continue to push the piston to the flushing interface position;

[0092] (2) sample chamber flushing: the first deep water motor continues to drive the seawater piston to push all the seawater in the piston out, completing the flushing of the sample chamber to prevent residual contamination from affecting the next detection;

[0093] (3) reaction chamber cleaning: simultaneously start the reverse function of the stirring and drainage integrated pump to discharge the residual liquid in the reaction chamber and the sample outside the chamber, completing the cleaning of the chamber;

[0094] (4) device reset: all motors are homed and ready for the next sampling task.

[0095] Therefore, the sampling process of the integrated deep-sea LIBS sampling device can obtain the sample S to be detected, and then the next sampling process is prepared through the cleaning and reset preparation process. By using the integrated deep-sea LIBS sampling device, the sampling process is first performed, which can effectively obtain the rare earth element sample to be detected; then, the device cleaning and circulation reset process is performed to thoroughly flush and reset the sample chamber and pipeline system, preparing for the next sampling process.

[0096] II. Design of deep-sea area sampling strategy and construction of rare earth detection data set

[0097] First, based on the integrated deep-sea LIBS sampling device, an area sampling strategy suitable for deep-sea environment is developed; combined with the sampling process and the device cleaning and circulation reset process, a multi-point deep-sea sediment sampling experiment is carried out to obtain multi-point LIBS spectral data and corresponding spatial position information (including latitude, longitude and depth) of the target area; then the rare earth element concentration label is obtained through laboratory calibration to construct the rare earth detection data set, which includes:

[0098] S2-1, control the ROV manipulator to advance along a Z-shaped route parallel to the seafloor topography, and set sampling points at equal intervals on the path to ensure uniform coverage of the entire target area in the horizontal and vertical directions; in addition, perform cleaning and reset operation after each sampling, and then go to the next sampling point;

[0099] S2-2, based on the ROV sampling path of S2-1, using the LIBS sampling device to perform a sampling process and a device cleaning and recycling reset process, to obtain a rare earth enriched liquid sample at each sampling point, and to obtain spectral data Sp1, Sp2,..., Sp of N samples through laser excitation experiments under water ROV N ; at the same time, record the spatial information St of each sampling point i , including longitude lg i , latitude lt i , and depth dp i , i.e. St i = [lg i , lt i , dp i ], and i ∈ [1, N]; finally, obtain the spatial information St1, St2,..., St N of N samples;

[0100] S2-3, determine the rare earth elements to be detected, including lanthanum (La), cerium (Ce), praseodymium (Pr), neodymium (Nd), samarium (Sm), europium (Eu), gadolinium (Gd), terbium (Tb), dysprosium (Dy), and holmium (Ho) for a total of ten kinds; then, synchronously send the collected samples into a shipborne laboratory for concentration calibration experiments, and use inductively coupled plasma mass spectrometry (ICP-MS) to measure the rare earth element concentration values Rc1, Rc2,..., Rc N of N sampling points, which are used as the concentration labels of the samples, wherein each rare earth element concentration value Rc i includes the concentration values of ten elements, and i ∈ [1, N];

[0101] S2-4, construct a data set, use the multi-point spectral data Sp1, Sp2,..., Sp N and spatial position information St1, St2,..., St N obtained in the above steps as input data, use the rare earth element concentration values Rc1, Rc2,..., Rc N as target data, and integrate the input data and the target data as a complete rare earth detection data set;

[0102] S2-5, repeat steps S2-1 to S2-5, so as to obtain M groups of rare earth detection data sets from different areas.

[0103] III. Spectral-Position Fusion Graph Structure Data Modeling Process

[0104] Based on the constructed rare earth detection data set, in order to capture the distribution characteristics of rare earth elements in space and mine the local correlation between spectral data, the present application carries out spectrum-site fusion graph structure data modeling; Specifically, each sampling point is regarded as a node (Node) in the graph structure, and the spatial distance and spectral feature similarity between nodes are used to construct edge (Edge) connection relationship, and the node and edge are respectively given multi-dimensional feature representation, realizing the graph structure data modeling of multi-source information fusion; The specific process includes:

[0105] Graph node construction: each sampling point is regarded as a node (Node) in the graph structure and has N nodes, and the LIBS spectral data and spatial position information corresponding to each node are used as the feature representation of the node; Each node feature vector can be expressed as:

[0106] X i =[Sp i ,St i ]=[Sp i ,lg i ,lt i ,dp i ]

[0107] Wherein, X i represents the node feature vector of the i-th sampling point, Sp i , St i respectively the spectral data and spatial position information of the i-th sampling point, lg i , lt i , dp i respectively the longitude, latitude and depth position of the i-th sampling point, and i∈[1,N];

[0108] Graph edge construction: whether to establish edge connection is determined based on the spatial distance and spectral similarity between each pair of nodes; The three-dimensional spatial distance between each pair of sampling points is calculated as:

[0109]

[0110] Wherein, Dis ij represents the spatial distance between the i-th sampling point and the k-th sampling point, lg j , lt j , dp j respectively the longitude, latitude and depth position of the j-th sampling point, k∈[1,N] and j≠i;

[0111] The spectral similarity between the spectra is calculated as:

[0112]

[0113] Wherein, Cos ijCos represents the spectral similarity between the i-th sampling point and the k-th sampling point. ij ∈[0,1] and Cos ij A larger value indicates a higher spectral similarity between the two sampling points, and vice versa; Sp j Let represent the spectral data of the j-th sampling point, j∈[1,N] and j≠i; ||*|| is the modulo operation;

[0114] Then, calculate the overall similarity Sim between the i-th sampling point and the k-th sampling point. ij :

[0115]

[0116] Where α1 and α2 are weight coefficients; for the i-th sampling point, the overall similarity between this sampling point and other sampling points is calculated, Sim i1 Sim i2 Sim iN Sim i1 and Sim iN Let represent the overall similarity between the i-th sampling point and the 1st and Nth sampling points, respectively;

[0117] Finally, the K sampling points with the highest similarity are selected to establish an edge connection, and the edge is assigned a feature vector containing spatial and spectral information:

[0118] E ij =[Dis ij Sim ij ]

[0119] Among them, E ij Let i be the edge feature between node i and node j, where i, k∈[1, N].

[0120] Complete graph construction: Based on graph node construction and graph edge construction, the dataset contains spectral data Sp1, Sp2, ..., Sp... N and spatial location information St1, St2, ..., St N The complete graph representation is G = (X, E), where X represents the set containing all N sampling point nodes, and each node is accompanied by a spectral-position fusion feature vector X. i E represents the set of all edges, and each edge is accompanied by a corresponding Edge feature vector E. ij Finally, the graph representation G = (X, E) is used as the input to the rare earth element quantitative detection model based on graph neural networks, and is used to predict the rare earth element concentration values ​​Rc1, Rc2, ..., Rc at each sampling point. N .

[0121] Four, the rare earth element concentration detection model based on GNN

[0122] In order to fully tap the correlation of rare earth element concentration in spatial distribution, and improve the quantitative identification ability of LIBS spectrum data, the application constructs a rare earth element concentration detection model based on GNN; the model takes the constructed spectrum-site fusion graph structure data as input, uses the feature propagation and aggregation mechanism of graph neural network to learn the correlation between nodes in the global graph structure, realizes accurate prediction of the rare earth element concentration of each sampling point, and realizes identification and detection of the rare earth enrichment trend in the region through regional level aggregation algorithm; the rare earth element concentration detection model based on GNN is as shown in Figure 3 The specific steps include the following:

[0123] Input feature encoding layer: take the graph representation G=(X, E) containing node feature vector X i and edge feature vector E ij as input, encode X i and E ij respectively through multilayer perceptron, and get encoded node feature and edge feature

[0124] Spatial perception module: this module models the spatial correlation of the encoded node feature and edge feature , and realizes spatial structure constraint modeling; this module is composed of L spatial perception graph attention layers; in each layer, the attention weight is calculated by combining the similarity between node features and the spatial distance decay factor, so that the nodes with high similarity and adjacent nodes can obtain higher weight, and the interference of distant noise nodes on the representation of the center node is inhibited, thereby improving the accuracy of local spatial modeling; finally, through multi-layer spatial perception graph attention propagation, the model can learn the context information of different neighborhood scales layer by layer, and realize in-depth mining of the distribution rule of rare earth elements; the calculation of the spatial perception graph attention layer is specifically represented as:

[0125]

[0126] Wherein, and respectively represent the output and input node features of the i th node in the l th spatial perception graph attention layer, and l∈[1, L]; N(i) represents the neighbor node set of the i th node, ReL∪(*) is the ReLU activation function, and ω ij represents the spatial attention weight between the i th node and the j th node; is the basic attention weight between the i th node and the j th node, and βij is the spatial decay factor, and β ij The calculation is expressed as:

[0127]

[0128] where, and respectively represent the feature representation of node i and node j at the l-th layer, when l = 0, it is the output feature of the feature encoding layer and Dis ij is the spatial distance between the i-th sampling point and the j-th sampling point, and σ is the spatial decay factor parameter;

[0129] Therefore, the feature encoding layer output of each node feature and edge feature is input into L spatial perception graph attention layers in turn, and finally the encoding feature of each node is obtained where i ∈ [1, N], and it is taken as the output feature of the spatial perception module.

[0130] Rare earth element prediction module: based on the obtained encoding feature of each node Through a multi-layer perception machine, the node encoding feature is compressed, and then through a fully connected layer and a ReLU activation function, the rare earth element concentration prediction result of each sampling point is obtained

[0131] Regional aggregation algorithm: based on the rare earth element concentration prediction result of each sampling point In order to realize the identification of regional rare earth element distribution situation, and accurately judge and divide the regional enrichment degree, the regional aggregation algorithm is designed in combination with spatial information, which specifically includes the following steps:

[0132] (1) The whole region to be detected is divided into R square grids of equal size, and each sampling point is divided into different local area grids;

[0133] (2) For each delimited region, the rare earth element concentration prediction value of all sampling points in the region is summarized, and the average concentration μ r and variance Δ r of the region are calculated r and Δ r respectively represent the average concentration and variance of the rare earth of the r-th region, and r ∈ [1, R];

[0134] (3) Set multi-level discrimination rules and threshold parameters, including rare earth high enrichment mean parameter μ high , high enrichment variance parameter Δ high , and medium enrichment mean parameter μ midThe threshold parameters are used to determine whether the region is a rare earth enrichment area, and a rare earth enrichment grade label is assigned, as shown in Table 1:

[0135] Table 1 Multi-level discrimination rule of rare earth enrichment area

[0136] Rare earth enrichment grade Discrimination condition High enrichment area μ r > μ high and Δ r < Δ high ]]> Medium enrichment area μ mid <μ r <μ high or μ r >μ high , Δ r >]]> Low enrichment area μ r <μ mid ]]>

[0137] V. Quantitative detection system based on rare earth element concentration detection model

[0138] GNN-based rare earth element concentration detection model training: based on the constructed rare earth detection dataset and the spectral-site fusion graph structure data modeling method, the standard model training dataset is obtained, and the model is trained. In the training process, the mean square error (MSE) is used to calculate the loss function of element concentration prediction; the Adam optimization algorithm is used for parameter updating in model training, and finally the trained rare earth element concentration detection model is obtained.

[0139] The detection system is as shown in Figure 4 The detection system includes a water surface part and an underwater part.

[0140] The underwater part includes a main cabin and an integrated deep-sea LIBS sampling device; the main cabin contains a laser control module, a spectrometer, a power supply module, a control module, and an in-cabin state monitoring feedback unit.

[0141] The LIBS optical probe is electrically connected to the spectrometer, including a laser head and an auxiliary sampling device, and the laser head is electrically connected to the laser control module; the integrated deep-sea LIBS sampling device is based on the regional sampling path and the point distribution strategy, and performs multi-point deep-sea sediment sampling in the target area to obtain the sample to be tested, and through the auxiliary sampling device, the laser head is used to perform excitation experiment to obtain the spectral characteristics.

[0142] The water surface part includes a control computer, which integrates a modularly packaged rare earth element concentration detection model based on graph neural network GNN, a spectral-site fusion graph structure data modeling module, and a regional aggregation algorithm module, which are constructed and trained;

[0143] The spectral-site fusion graph structure data modeling module takes different sampling points as nodes Node in the graph structure, constructs the edge Edge structure of the graph combining the spectral characteristics and spatial distance of the sampling points, and encodes the spectral characteristics and spatial position information of each sampling point as node features, and converts the sample data to be tested into graph structure data as standard input data of the rare earth element concentration detection model;

[0144] The rare earth element concentration detection model extracts the spatial correlation features of rare earth elements through the node feature propagation and aggregation mechanism of GNN, and outputs the rare earth element concentration of each sampling point;

[0145] a regional aggregation algorithm module, which uses a regional aggregation algorithm based on the rare earth element concentration of each sampling point in the target region to calculate the rare earth enrichment state at the regional level and output the enrichment grade.

[0146] The staff selects the target region and further deployment according to the obtained rare earth element concentration distribution and enrichment grade results, combined with geological background information and actual exploration needs.

[0147] Six, experimental results

[0148] In order to verify the effectiveness of the deep-sea LIBS rare earth element quantitative detection method proposed in the present application in the prediction of rare earth element concentration, the present application carries out experimental test, and obtains the average concentration distribution of ten rare earth elements (La, Ce, Pr, Nd, Sm, Eu, Gd, Tb, Dy, Ho) in different regions, as shown in Figure 5 The heat map analysis result shows that the method of the present application can effectively realize the regional level detection of the spatial distribution of rare earth elements, and further verifies the feasibility of the method in the quantitative analysis of deep-sea rare earth elements.

[0149] In addition, the present application is compared and analyzed with three existing mainstream algorithms for predicting rare earth element concentration, including support vector machine (SVM), random forest (RF) and gradient boosting regression tree (GBRT). And taking the mean absolute percentage error (MAPE) as the evaluation index, the prediction accuracy of the model for rare earth element concentration at the sampling point level and the regional level is compared; specifically, a plurality of sampling points are selected for the region to be detected, and the rare earth element concentration prediction value of each sampling point is obtained, the average of all sampling points is taken and compared with the real element concentration to calculate the MAPE to obtain the final detection accuracy; similarly, for the evaluation of the prediction accuracy of the rare earth element concentration at the regional level, the whole region is first divided into grids, and the average of the sampling point rare earth detection concentration in the region is taken, and the MAPE is calculated to obtain the rare earth concentration detection accuracy in the region; the experimental results are shown in Figure 6 and Figure 7

[0150] ​Since the method of the present application can effectively alleviate the interference caused by spectral overlap and complex soil composition through integrated deep-sea LIBS sampling device for rare earth sampling, thereby improving the discrimination degree and quantitative ability of rare earth element identification; in addition, the method introduces a spatial decay attention factor based on GNN, which can fully utilize the spatial correlation of rare earth element distribution in the geological background; by jointly modeling the spectral features of adjacent sampling points and the spatial information of the target point, the discrimination ability of the model to small differences is finally improved. Therefore, compared with existing methods (support vector machine SVM, random forest RF and gradient boosting regression tree GBRT), the method of the present application shows higher prediction accuracy in the detection of ten rare earth elements, and embodies stronger feature extraction ability and fine modeling level.

[0151] In addition, the method of the present application enhances the robustness of the model to local anomalies by introducing a graph modeling method based on spatial information, thereby avoiding the influence of single-point data fluctuations on the overall prediction results, at the same time, the propagation mechanism of GNN on graph data can deeply capture the potential relationship between different sampling points, thereby enhancing the generalization ability of the model between different geological regions, significantly reducing the performance fluctuation in different element detection applications. The experimental results further verify the stability of the method of the present application in the detection of ten rare earth elements, thereby ensuring that it still maintains high-precision, low-fluctuation and reliable prediction performance under complex geological conditions.

[0152] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make various modifications and changes to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0153] Although the specific embodiments of the present application have been described above, they are not intended to limit the scope of protection of the present application, and those skilled in the art should understand that various modifications or changes made to the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A method for deep-sea LIBS rare earth element quantitative detection fusing spatial information, characterized in that, The process comprises the following steps: S1. Based on the integrated deep-sea LIBS sampling device and the regional sampling path and point distribution strategy, multi-point deep-sea sediment sampling is performed in the target region to obtain rare earth element samples to be tested in the target region, and the corresponding spatial position information is recorded, and then the spectral characteristics of each sample to be tested are detected by laser excitation; S2. Different sampling points are taken as nodes Node in a graph structure, the edge Edge structure of the graph is constructed in combination with the spectral characteristics and spatial distance of the sampling points, and the spectral characteristics and spatial position information of each sampling point are encoded as node characteristics, and the sample data to be tested is converted into graph structure data as standard input data of the rare earth element concentration detection model; the specific process is as follows: Figure node construction: each sampling point is regarded as a node Node in the graph and there are nodes in total, and the LIBS spectral data and spatial position information corresponding to each node are taken as the feature representation of the node; the feature vector of each node is represented as , and ; Graph construction: the edge connection is determined by the spatial distance and spectral similarity between each pair of nodes; the three-dimensional spatial distance between each pair of sampling points is calculated as , and ; Computing spectral similarity between spectra , representing the spectral similarity between the first sample point and the first sample point, and the greater the value of the spectral similarity between the two sample points, and conversely, the lower the spectral similarity; thereafter, computing the overall similarity between the first sample point and the first sample point; Finally, the K sampling points with the highest similarity are selected to establish an edge Edge connection, and a feature vector containing spatial and spectral information is assigned to the Edge: ； wherein, is an Edge feature between a node and a node and ; Complete graph construction: Based on graph node construction and graph edge construction, data containing spectral data and spatial location information is characterized as a complete graph representation , wherein represents a set containing all sampling point nodes, and each node is attached to a spectrum-site fused feature vector ; represents a set of all edges, and each edge is attached to a corresponding Edge feature vector ; S3. The graph structure data in S2 is input into the rare earth element concentration detection model based on the graph neural network GNN which is constructed and trained, the model extracts the spatial correlation characteristics of the rare earth elements through the node feature propagation and aggregation mechanism of the GNN, and outputs the rare earth element concentration of each sampling point; S4. Based on the rare earth element concentration of each sampling point in the target region, a regional aggregation algorithm is used to calculate the regional rare earth enrichment situation, and the enrichment grade is output.

2. The method for deep-sea LIBS rare earth element quantitative detection fusing spatial information according to claim 1, characterized in that, The specific structure and sampling process of the integrated deep-sea LIBS sampling device are as follows: The device comprises a sample tank, a reaction chamber, a ROV mechanical arm, four deep water motors and corresponding piston structures; Among them, the fourth deep water motor is connected with the sampling piston through a sliding block and is controlled and operated through the mechanical arm; the third deep water motor is connected with the acid liquid piston through a sliding block, the right chamber of the acid liquid piston is connected with the reaction chamber through a pipeline, and an acid inlet is arranged in the connecting chamber; a stirring and drainage integrated pump assembly is arranged at the bottom of the reaction chamber; the reaction chamber is connected with the sample tank through a pipeline and a stop valve; the first deep water motor is connected with the filtration and pressure reduction piston through a sliding block, the right chamber of the filtration and pressure reduction piston is connected with the sample tank through a pipeline, and a pressure sensor is arranged on the sample tank; the second deep water motor is connected with the seawater piston through a sliding block, and the left chamber of the seawater piston is connected with the left chamber of the filtration and pressure reduction piston through a pipeline; the left chamber of the seawater piston is provided with a seawater inlet; Through the operation of the ROV mechanical arm, the sampling piston driven by the fourth deep water motor is inserted into the deep-sea sediment in the predetermined area, the motor drives the piston to move backward to complete the collection of the sediment sample; then the piston is inserted into the reaction chamber and is driven in reverse to push the sediment out into the reaction chamber; The third deep water motor drives the acid liquid piston to inject a preset volume of quantitative acid liquid into the reaction chamber for sample pretreatment; Start the stirring and drainage integrated pump to drive the impeller in the tank to rotate, so that the acid liquid and the sediment are fully mixed and reacted, and the rare earth elements in the sediment are fully dissolved into the acid liquid to realize the release of the components; After the mixing reaction is completed, the first deep water motor drives the piston to extract the mixed sample in the reaction chamber to the LIBS detection special sample tank. After the electric cut-off valve of the sample chamber is closed, the first deep water motor continues to draw, and due to the system sealing, the pressure in the chamber will gradually decrease, and the pressure in the sample chamber is monitored through the pressure sensor, and when the pressure in the sample chamber is reduced to the set value, the sample is in the target pressure and sealed state, and can be directly used as a sample for laser excitation experiment.

3. The method for deep-sea LIBS rare earth element quantitative detection fusing spatial information according to claim 1, characterized in that, The construction process of the data set for training the rare earth element concentration detection model includes: S11, control the ROV mechanical arm to advance along a Z-shaped route parallel to the seafloor topography, and set sampling points at equal intervals on the path, ensuring uniform coverage of the entire target area in the horizontal and vertical directions; in addition, a cleaning and resetting operation is performed after each sampling, and then the next sampling point is reached; S12, based on the sampling path, uses an integrated deep-sea LIBS sampling device to obtain rare earth enriched liquid samples at each sampling point, and obtains the total rare earth enriched liquid samples through laser excitation experiments conducted by an underwater ROV. Spectral data of each sample Simultaneously, spatial information of each sampling point is recorded. Including longitude ,latitude ,depth ,Right now ,and ; ultimately obtained Spatial information of each sample ; S13, determining the rare earth category to be detected; then synchronously sending the collected sample into the shipborne laboratory to conduct concentration calibration experiment, and using inductively coupled plasma mass spectrometry to measure the rare earth element concentration values of the sampling points , taking the rare earth element concentration values as the concentration label of the sample, wherein each rare earth element concentration value includes the concentration values of ten elements, and ;​ S14, integrating the acquired multi-point spectral data and spatial position information as input data, rare earth element concentration values as target data, and integrating the input data and the target data as a complete set of rare earth detection data 4. The deep-sea LIBS rare earth element quantitative detection method fusing spatial information according to claim 1, characterized in that: The rare earth element concentration detection model based on the graph neural network GNN includes an input feature encoding layer, a spatial perception module, and a rare earth element prediction module; The input feature encoding layer is used to represent the graph as input with node feature vectors and edge feature vectors The multi-layer perception is used to encode the and features respectively, and the encoded node features and edge features are obtained.​ The spatial perception module sequentially inputs each node feature and edge feature output by the feature encoding layer into L spatial perception graph attention layers, and finally obtains the encoding feature of each node wherein and takes it as the output feature of the spatial perception module; the spatial perception module is composed of L spatial perception graph attention layers; in each layer, the attention weight is calculated by combining the similarity between node features and a spatial distance decay factor; through multi-layer spatial perception graph attention propagation, the model learns the context information of different neighborhood scales layer by layer; The rare earth element prediction module is based on the obtained each node coding feature , through a multi-layer perception to node coding feature for feature compression, and then through a full connection layer and a ReLU activation function to obtain the rare earth element concentration prediction result of each sampling point .

5. The method for deep-sea LIBS rare earth element quantitative detection fusing spatial information according to claim 1, characterized in that: The specific process of S4 is: The whole region to be detected is divided into equal-sized square grids, and each sampling point is divided into different local region grids; For each delineated region, the rare earth element concentration predictions for all the sampling points within the region are summed and the average concentration within the region is calculated and variance , and denote the rare earth average concentration and variance, respectively, for the th region, and ; A multi-level discrimination rule is set, and threshold parameters are set, including a high enrichment mean parameter of rare earth , a high enrichment variance parameter of rare earth , and a medium enrichment mean parameter of rare earth ; the threshold parameters are used to determine whether the region is a rare earth enrichment area, and a rare earth enrichment grade label is given.

6. A deep-sea LIBS rare earth element quantitative detection system fusing spatial information, applied to the deep-sea LIBS rare earth element quantitative detection method fusing spatial information according to any one of claims 1 to 5, characterized in that: The water surface part and the underwater part are included. The underwater part includes a main body chamber and an integrated deep-sea LIBS sampling device; the main body chamber contains a laser control module, a spectrometer, a power supply module, a control module, and an in-chamber state monitoring feedback unit; The LIBS optical probe is electrically connected to the spectrometer and includes a laser head and an auxiliary sampling device; the laser head is electrically connected to the laser control module; the integrated deep-sea LIBS sampling device performs multi-point deep-sea sediment sampling in the target area based on the regional sampling path and point distribution strategy, obtains the sample to be tested, and performs excitation experiments through the laser head to obtain spectral characteristics through the auxiliary sampling device; The water surface part includes a control computer, which integrates the modularly packaged rare earth element concentration detection model based on the graph neural network GNN, the spectral-point fused graph structure data modeling module, and the regional aggregation algorithm module; The spectral-point fused graph structure data modeling module uses different sampling points as nodes Node in the graph structure, constructs the edge Edge structure of the graph by combining the spectral characteristics and spatial distances of the sampling points, and encodes the spectral characteristics and spatial position information of each sampling point as node features, and converts the sample data to be tested into graph structure data as standard input data of the rare earth element concentration detection model; The rare earth element concentration detection model extracts the spatial correlation characteristics of rare earth elements through the node feature propagation and aggregation mechanism of GNN, and outputs the rare earth element concentration of each sampling point; The regional aggregation algorithm module calculates the regional rare earth enrichment situation based on the rare earth element concentration of each sampling point in the target area using the regional aggregation algorithm, and outputs the enrichment grade.

7. The deep-sea LIBS rare earth element quantitative detection system integrating spatial information of claim 6, characterized in that: The specific structure of the integrated deep-sea LIBS sampling device is: The device includes a sample chamber, a reaction chamber, an ROV mechanical arm, four deep water motors, and corresponding piston structures; The fourth deep water motor is connected with the sampling piston through a sliding block and is controlled and operated through a mechanical arm; the third deep water motor is connected with the acid liquid piston through a sliding block, the right chamber of the acid liquid piston is connected with the reaction chamber through a pipeline, and an acid adding port is arranged in the connecting chamber; a stirring and drainage integrated pump assembly is arranged at the bottom of the reaction chamber; the reaction chamber is connected with the sample cabin through a pipeline and a stop valve; the first deep water motor is connected with the filtration and decompression piston through a sliding block, the right chamber of the filtration and decompression piston is connected with the sample cabin through a pipeline, and a pressure sensor is arranged on the sample cabin; the second deep water motor is connected with the seawater piston through a sliding block, and the left chamber of the seawater piston is connected with the left chamber of the filtration and decompression piston through a pipeline; the left chamber of the seawater piston is provided with a seawater inlet.

8. The deep-sea LIBS rare earth element quantitative detection system integrating spatial information of claim 6, characterized in that: The rare earth element concentration detection model based on the graph neural network GNN comprises an input feature encoding layer, a spatial perception module and a rare earth element prediction module. The input feature encoding layer is used to represent the graph as input with node feature vectors and edge feature vectors The multi-layer perception is used to encode the and respectively, and the encoded node features and edge features are obtained.​ The spatial perception module sequentially inputs each node feature and edge feature output by the feature encoding layer into L spatial perception graph attention layers, and finally obtains the encoding feature of each node wherein and takes it as the output feature of the spatial perception module; the spatial perception module is composed of L spatial perception graph attention layers; in each layer, the attention weight is calculated by combining the similarity between node features and a spatial distance decay factor; through multi-layer spatial perception graph attention propagation, the model learns the context information of different neighborhood scales layer by layer; The rare earth element prediction module is based on the obtained encoding features of each node , the node encoding features are compressed by a multi-layer perception, and then a fully connected layer and a ReLU activation function are used to obtain the rare earth element concentration prediction result of each sampling point .

Images