A granular intelligent identification method for lunar soil particle screening

Abstract

The present application relates to a kind of for lunar soil particle screening Particle Intelligent Recognition Method, comprising the following steps: obtaining the optical signal, image data and surface electric signal of the same lunar soil sample particle, constitute the multimodal feature set of each lunar soil sample particle;The multimodal feature set is standardized;Based on CNN+MLP+splicing fusion module structure, construct multimodal deep learning neural network model;Multimodal deep learning neural network model is supervised learning training, and the trained multimodal deep learning neural network model is obtained;The data after standardization processing are input into the trained multimodal deep learning neural network model, and particle classification is realized.The present application adopts intelligent recognition integrated deep learning model, utilizes intelligent algorithm to carry out the shape distinction, component identification, particle size sorting and other functions of particle, can realize automatic classification, clustering and attribute prediction, without artificial marking one by one, effectively improve work efficiency.

Description

Technical Field

[0001] This invention relates to the field of space exploration material analysis technology, and in particular to a smart particle identification method for screening lunar soil particles. Background Technology

[0002] As Earth's closest natural satellite, the Moon serves as a forward outpost and primary destination for human deep space exploration and development. Its surface is covered by lunar regolith, a loose granular medium formed by impacts and space weathering. Lunar regolith is rich in metals, non-metals, volatile components, and water ice resources, particularly He3, which is considered a significant candidate for future clean nuclear fusion energy, potentially providing sustained power to Earth's energy system for tens of thousands of years. Composed of solid particles ranging in size from submicron to submillimeter, lunar regolith exhibits high heterogeneity and complexity. Its fundamental physical properties, including mechanical, thermal, and electromagnetic properties, directly impact the feasibility and safety of lunar landing, drilling, sampling, and construction projects.

[0003] Because lunar soil samples exhibit a wide range of particle sizes, from 1µm fine dust to 1mm coarse sand, current particle analysis methods largely rely on electrostatic separators, scanning electron microscopy (SEM) images, and manual statistical analysis. These methods struggle to automate particle classification, severely impacting work efficiency. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for intelligent particle identification in lunar soil particle screening.

[0005] This invention is achieved through the following technical solution:

[0006] The present invention provides a particle intelligent identification method for screening lunar soil particles, comprising the following steps: acquiring the optical signal, image data and surface electrical signal of the same lunar soil sample particle to form a multimodal feature set for each lunar soil sample particle;

[0007] The multimodal feature set is standardized.

[0008] A multimodal deep learning neural network model is constructed based on a structure of convolutional neural network + multilayer perceptron + splicing and fusion module.

[0009] Supervised learning training is performed on the multimodal deep learning neural network model to obtain a well-trained multimodal deep learning neural network model;

[0010] The standardized data is input into a trained multimodal deep learning neural network model to achieve particle classification.

[0011] Optional, standardization methods include:

[0012] Image data is cropped and normalized.

[0013] The cropped and normalized image data is converted into a model input format of image tensor + numerical vector;

[0014] Filtering and baseline correction are performed on the optical signal data.

[0015] Preferably, the optical signal includes a forward scattering signal and a side scattering signal.

[0016] Optionally, the optical signal may also include a fluorescence signal or multiple fluorescence signals of different wavelengths.

[0017] Optionally, image data is input into a CNN, and the CNN outputs an image feature vector; other non-image data is input into an MLP, and the MLP outputs a signal feature vector; the splicing and fusion module performs feature fusion on the image feature vector and the signal feature vector, and sends it to the output layer, which is a classification task.

[0018] Optionally, the classification uses the cross-entropy loss function to output a probability distribution.

[0019] Optionally, the multimodal deep learning neural network model outputs particle category and confidence level.

[0020] Optionally, the output layer also performs a regression task, which uses linear activation to output one or more continuous variable values ​​to predict continuous physical property parameters of the particles.

[0021] Optionally, the output layer also performs clustering tasks, and can optionally output clustering results based on t-SNE / U-MAP dimensionality reduction and DBSCAN clustering results. Optional output means that the system can choose whether to enable or disable this output item according to actual application needs or user settings.

[0022] Clustering results can be used for group assignment, discovery of potential new particle types, and auxiliary verification of classification systems, possessing multi-level functional value. The clustering task is used to perform automatic grouping analysis of lunar soil sample particles based on their feature space structure in the absence of predefined labels or categories. Dimensionality reduction of high-dimensional fused features is achieved using t-SNE or UMAP algorithms, and then the DBSCAN algorithm is used to identify structurally consistent particle groups in the embedding space.

[0023] Compared with the prior art, this application has at least the following beneficial effects:

[0024] 1. This invention adopts an intelligent recognition integrated deep learning model and uses intelligent algorithms to carry out functions such as particle shape differentiation, component identification, and particle size sorting. It can not only realize automated classification, clustering and attribute prediction without manual labeling of each particle, but also solve the problem of insufficient component identification accuracy in the screening of lunar soil sample particles.

[0025] 2. This invention combines multiple channel parameters such as image, scattered light, fluorescence, particle size, and electrical properties to achieve multi-dimensional characterization at the single particle level, which can meet the analytical needs of lunar soil particles under diverse characterization backgrounds. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a schematic diagram of the structure of the multimodal deep learning neural network model in the embodiment;

[0028] Figure 2 This is a flowchart of the model training process in the embodiment;

[0029] Figure 3 This is a flowchart illustrating the extraction of deep representations using a three-layer perceptron in the embodiment.

[0030] Figure 4 This is a schematic diagram of the lunar soil particle screening device in the embodiment;

[0031] Figure 5 This is a schematic diagram of the structure of the multi-channel detector in the embodiment.

[0032] Reference numerals: 1-Microflow control system, 2-Flow chamber, 3-Particle detection system, 4-Particle identification system, 5-Particle screening system, 11-Compressed gas source, 12-Sheath fluid container, 13-Sample container, 14-Pressure valve, 21-Sample channel, 22-Sheath fluid tube, 31-Laser, 32-Detection system, 41-Computer, 42-Information storage, 51-Control module, 52-Collector. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0034] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other. It should also be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0035] In the description of this invention, it should be noted that the terms "front," "rear," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship that the product of this invention is usually placed in during use, or the orientation or positional relationship that is commonly understood by those skilled in the art. They are only used to facilitate the description of this invention and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0036] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0037] Example 1

[0038] The intelligent particle identification method for lunar soil particle screening provided in this embodiment includes the following steps:

[0039] Step 1: Based on the CNN + MLP + concatenation and fusion module structure, construct a multimodal deep learning neural network model, such as... Figure 1 As shown.

[0040] Step 2, model training, such as Figure 2 As shown, it includes:

[0041] 2.1 Supervised learning training was conducted based on the labeled standard particle sample set. The training data came from manually labeled lunar soil images and spectral datasets. The standard particle sample set mainly consisted of existing lunar soil sample images, compositional analysis results, and image and spectral samples constructed based on simulated lunar soil materials.

[0042] 2.2 The loss function is selected according to the task, such as classification cross-entropy, regression MSE or clustering loss;

[0043] 2.3. Apply data augmentation techniques (image rotation, noise perturbation, multimodal missing data simulation) to improve the robustness of the model;

[0044] 2.4 Adjust hyperparameters such as learning rate, batch size, and network depth to obtain the optimal initial model.

[0045] Step 3: Model optimization and verification.

[0046] 3.1 Evaluate model performance on independent validation sets and analyze metrics such as accuracy, recall, and F1 score;

[0047] 3.2. Locate the main sources of error through the confusion matrix (such as two types of particles being easily confused) and optimize accordingly;

[0048] 3.3 The model undergoes pruning, distillation, quantization, and other operations to compress the model size and improve inference speed;

[0049] 3.4. Incremental learning is supported. When the system encounters new types of particles during operation, it can be fine-tuned online to continuously optimize the model and obtain a well-trained multimodal deep learning neural network model.

[0050] In some embodiments, the optimization algorithm uses the Adam optimizer, with an initial learning rate of 0.001, a batch size of 32, and 100 training epochs.

[0051] Step 4: Using the trained multimodal deep learning neural network model, perform intelligent particle recognition, including:

[0052] 4.1 Acquire raw data of lunar soil sample particles, including multi-channel data such as optical signals, image data, and surface electrical signals, and synchronously bind the data of the same particle in different channels;

[0053] In some embodiments, the optical signals include forward scattering signals, side scattering signals, and multi-channel fluorescence signals.

[0054] 4.2 Standardize the raw data, including image cropping and normalization, signal filtering and baseline correction, and convert the image into an image tensor + numerical vector model input format to ensure consistent data input format.

[0055] Cropping aims to extract particle regions, highlight the characteristics of the particles themselves, reduce irrelevant information input, and improve recognition accuracy. Image normalization is used to unify the range of pixel values, avoid numerical shifts caused by varying illumination intensity, and improve stability. Signal filtering removes noise from scattering and fluorescence signals, focusing on effective information. Signal baseline correction corrects fluctuations in sensor zero-point and background signals, ensuring comparability of data collected in different batches or at different times, and improving data consistency. Image-to-tensor + numerical vector conversion unifies data from different modalities into a standard format that can be processed by neural networks.

[0056] 4.3 Input the standardized data into the trained multimodal deep learning neural network model, and output the category, confidence score, and group assignment. Specifically:

[0057] Image data is input into a convolutional neural network (CNN) to extract shape features;

[0058] Numerical features such as particle size, scattering intensity, and fluorescence intensity are input into a multilayer perceptron (MLP) for feature expansion.

[0059] Feature vectors from different sources are processed, aligned, and combined to generate a fused feature vector in a unified semantic space, which is used to jointly model feature vectors from different modalities.

[0060] Depending on the task requirements, set a classification head (such as identifying categories like vitreous / pyroxene / olivine), a regression head (predicting parameters such as elastic modulus), or a clustering head (to achieve unsupervised particle grouping).

[0061] Perform tasks such as classification (e.g., particle type), clustering (e.g., particle similarity grouping), and regression (e.g., physical property parameter estimation); output results such as particle category, confidence level, and group affiliation.

[0062] The process includes: classification, which determines the category of particles for efficient grouping and sorting applications; clustering, which automatically groups particles based on feature similarity when the category is unknown, discovering potential particle categories or structural patterns; regression, which predicts continuous physical properties of particles, such as elastic modulus and density; and outputting the category, confidence level, and assigned group, which provides the identification conclusion and confidence level for each particle, facilitating subsequent screening, retrieval, statistical analysis, and optimization.

[0063] In some embodiments, the input features are represented as:

[0064]

[0065] In the above formula, Represents the image tensor, which is a single-particle RGB capture of the boundary, texture, and morphological features of the particles acquired by a CMOS camera; Representative scattering vector: contains forward scattering (FSC) and side scattering (SSC) intensity values, used to represent particle size and structural roughness; The multi-channel fluorescence vector represents the material composition of the particle; Representing electrical characteristics: Input the surface electrical properties of the particles to identify the mineral type and dispersion characteristics of the particles.

[0066] In some embodiments, the classic convolutional network ResNet18 is used to extract deep features of the image, compressing the original image information into a low-dimensional representation, such as texture direction and edge sharpness, which is beneficial for subsequent classifier recognition.

[0067]

[0068] In the above formula, The input is a particle image. This demonstrates the use of a ResNet18 convolutional neural network for feature extraction from images. The output image feature vector is represented as a 512-dimensional vector; This indicates that the vector is a 512-dimensional vector in the real number space.

[0069] In some embodiments, all non-image data is stitched together using the following formula:

[0070]

[0071] In the above formula, It represents the characteristic vector of scattered light, including forward scattering (FSC), side scattering (SSC), etc. Represents the fluorescence intensity vector. Represents electrical characteristic values, This means concatenating the feature vectors of the three modalities along the feature dimension. The concatenated original numerical mode input vector The spliced ​​feature dimensions consist of 2 (scattering) + N (number of fluorescence channels) + 1 (electrical properties).

[0072] In some embodiments, such as Figure 3 As shown, a three-layer perceptron is used to extract deep representations. The first layer rapidly expands the information dimensionality, handles nonlinear relationships between features, and maps the original low-dimensional numerical features to a high-dimensional space, outputting a high-dimensional representation. The second layer extracts the interaction features between each feature dimension and stimulates high-order feature combinations through a nonlinear activation function. The third layer performs dimensionality reduction and compression on the feature representation, outputting a stable, low-dimensional, and compact numerical vector as a unified representation for fusion and alignment with image features.

[0073]

[0074] In the above formula, Output feature vector; , , These are the weight matrices of a three-layer fully connected network, used for the linear transformations of the first, second, and third layers, respectively. It is a non-linear activation function; , , These are the bias vectors corresponding to each layer; Original input feature vector (such as numerical modal features such as particle size, scattered light intensity, fluorescence intensity, etc.).

[0075] A multi-branch fusion neural network is employed. The image branch uses a pre-trained convolutional neural network to extract image features, outputting a 1×512 image feature vector. The signal branch uses scattering, fluorescence, and electrical features to concatenate a 1×(2+N+1) vector (where N is the number of fluorescence channels), and extracts signal features through a three-layer fully connected network (64, 128, and 256 nodes, respectively). The feature fusion layer concatenates the image features and signal features to form a 1×768 feature vector, which is then fed into the output layer after passing through a fusion layer with 256 nodes and ReLU activation.

[0076]

[0077]

[0078] In the above formula, This is the concatenated fused feature vector. This is a feature concatenation operation, which concatenates two vectors along their feature dimensions. For image feature vectors, For numerical modal eigenvectors, This is the final feature vector output after fusion. It is a non-linear activation function. This is the weight matrix. This is the bias vector.

[0079] The output layer is for classification tasks: In classification tasks, the output layer uses the Softmax activation function, and the dimension of the output vector is equal to the number of target particle categories. Each dimension corresponds to the predicted probability of a particle type, and the maximum value corresponds to the final classification result predicted by the model. This maximum probability value can also be used as a confidence index for particle recognition.

[0080] As a preferred option, the output layer also performs a regression task: in the regression task, the output layer uses a linear activation function to output one or more continuous numerical variables to predict the physical properties of the particles, such as particle size estimates, roughness index, and mean fluorescence intensity.

[0081] Optionally, the output layer also performs a clustering task: In this task, the model performs non-linear dimensionality reduction on the intermediate feature representations using t-SNE or UMAP methods, mapping high-dimensional features to a two-dimensional or three-dimensional embedding space, and further performs density clustering analysis on the dimensionality reduction results based on the DBSCAN algorithm. Finally, it outputs a clustering label for each particle, which can be used for unsupervised particle grouping or exploration of unknown particle types. This clustering result is an optional output item, mainly used to assist in classification system validation, sorting strategy optimization, and visualization interpretation.

[0082] Preferably, the classification uses cross-entropy loss, outputs a probability distribution, and determines which mineral category it belongs to. The cross-entropy loss function is then used for optimization.

[0083]

[0084]

[0085] In the above formula, This represents the probability distribution belonging to each class; The activation function is used to convert the output of a neural network into a probabilistic form. This is the classification weight matrix; This is the feature vector after multimodal fusion; For classification bias vector, The loss function value for the classification task. This is the one-hot encoding of the real label. If the real category is the k-th category, then this position is 1, and the rest are 0. To predict the probability of the model belonging to the k-th class, This represents the total number of categories.

[0086] Preferably, the regression uses continuous variables for prediction, such as particle size and roughness value, and is optimized using mean square error.

[0087]

[0088]

[0089] In the above formula, For continuous values ​​predicted by the model, such as particle size estimates, modulus and other physical property predictions; This is the weight matrix for the regression task; This is the final feature vector output after fusion; This is the bias vector for the regression task. For regression loss function, The regression value predicted by the model. This represents the actual measurement results.

[0090] In some embodiments, weighted fusion can be used to introduce fusion weights for each modality. These weights can be: fixedly designed, such as empirical values ​​or hyperparameters; learned and automatically adjusted through neural networks or attention mechanisms; or dynamically generated, such as adaptively adjusted based on modality quality / missing conditions.

[0091] This embodiment employs a multimodal deep learning neural network architecture to fuse and analyze the multidimensional features of each particle, such as image, optical scattering, fluorescence intensity, and electrical properties, to achieve high-precision classification and attribute recognition.

[0092] Example 2

[0093] This embodiment discloses a lunar soil particle screening device, which is based on an improved flow cytometer. For example... Figure 4 As shown, the lunar soil particle screening device specifically includes a microflow rate control system 1, a flow chamber 2, a particle detection system 3, a particle identification system 4, and a particle screening system 5. Both the particle detection system 3 and the particle screening system 5 are connected to the particle identification system 4.

[0094] The microfluidic control system 1 includes a compressed air source 11, a sheath fluid container 12, and a sample container 13. The compressed air source 11 is connected to the sheath fluid container 12 and the sample container 13 via air vents, and a pressure valve 14 is installed on the air vents. The sample container 13 is used to store lunar soil sample solution.

[0095] The flow chamber 2 includes a sample channel 21 and a sheath fluid tube 22. The sheath fluid tube 22 is concentrically located outside the sample channel 21. The sample container 13 is connected to the rear end of the sample channel 21 through a pipe, and the sheath fluid container 12 is connected to the rear end of the sheath fluid tube 22 through a pipe. The sample channel 21 has a nozzle at the front end. A single lunar soil sample particle suspension is ejected from the nozzle at the front end of the sample channel 21 under the action of liquid flow pressure. The sheath fluid flows from all sides of the sheath fluid tube 22 to the nozzle, surrounds the outer periphery of the lunar soil sample particle, and is ejected from the nozzle.

[0096] Optionally, the particle detection system 3 includes a laser 31 for irradiating and exciting lunar soil sample particles and a detection system 32. The laser 31 is used to irradiate and excite lunar soil sample particles to generate one or more optical response signals, the type of which depends on the particle type. The detection system 32 includes a photodetector for acquiring the fluorescence generated by the lunar soil sample particles after photoexcitation, an image acquisition device for acquiring image signals of the lunar soil sample particles, and a microelectrode measuring device for measuring the surface electrical properties of the lunar soil sample particles.

[0097] Because some particles possess natural fluorescence response capabilities due to their physical composition, in another embodiment, the particle detection system 3 also includes a fluorescence detector for acquiring fluorescence signals; the fluorescence detector is also a type of photodetector. It is worth noting that there can be one or more fluorescence detectors. Figure 5 As shown, when multiple fluorescence detectors are used, the combination structure consists of multiple detectors and a spectrometer, with each channel corresponding to a different center wavelength (e.g., 530nm, 605nm, 690nm). Some lunar soil sample particles are excited to emit multi-band fluorescence signals under laser irradiation, which are then filtered through different filters and fed into the photodetectors of each channel. The light intensity of each particle at different wavelengths is recorded simultaneously for subsequent component analysis and particle classification.

[0098] The particle recognition system 4 includes a computer 41 and an information storage device 42.

[0099] The particle screening system 5 includes a control module 51 and a collector 52. The control module 51 is connected to a computer 41. The control module 51 includes a deflector, an air pore control device, or an optical sorter, etc. Among them, the deflector can be an electromagnetic deflector.

[0100] The working principle of the lunar soil particle sorting device is as follows: Lunar soil samples are mixed with deionized water to prepare a lunar soil solution. Compressed gas is introduced through a pressure valve to control the solution flow rate. The solution flows into the sample channel in the sorting system and passes through the laser sorting system in a single-particle arrangement. Multimodal data, including forward / lateral scattering, fluorescence excitation, and electrical properties, are collected from the particles. Intelligent algorithms are used to perform particle shape differentiation, component identification, and particle size sorting. The sorting information is fed back to the sorting control module, which then sorts the collected particles and places them into different particle collection tubes to complete the sorting and identification process. This device solves the core problems of low efficiency, high destructiveness, and insufficient component identification accuracy in lunar soil sample particle sorting, providing a flow cytometry system suitable for large-sized, irregular mineral particles.

[0101] Optionally, in some embodiments, for lunar soil with a wide range of particle sizes, a large-aperture nozzle with an aperture of 50–500 μm is used to support the passage of large-sized particles, prevent clogging, adapt to the high hardness and multi-faceted characteristics of lunar soil, and combine with a high-viscosity sheath fluid formulation to achieve stable flow of 0.1–1000 μm particles, thus solving the clogging problem.

[0102] Optionally, the flow chamber can be a ceramic-coated flow chamber, which helps reduce lunar soil abrasion.

[0103] Optionally, in some embodiments, the laser 31 integrates dual laser beams (405 nm + 785 nm), and the laser 31 excites characteristic fluorescence of minerals (such as silicate fluorescence peaks). A highly sensitive photodetector array receives forward scattering (FSC) and side scattering (SSC) signals generated by the particles under laser irradiation, capturing physical information related to particle size and other factors.

[0104] Optionally, in some embodiments, a silicon carbide ceramic coating is used with a sapphire detection window to reduce lunar soil wear and improve equipment lifespan.

[0105] Based on the lunar soil particle screening device, this embodiment also discloses a method of using the lunar soil particle screening device, including the following steps:

[0106] Step 1: Sample Preparation

[0107] Preliminary drying and foreign matter removal are performed using real or simulated lunar soil samples to ensure that particles do not clump together. The sample is then added to deionized water, and ultrasonic treatment and surfactants are used to disperse the particles, preventing agglomeration. Sample concentration is controlled to ensure that only a single particle passes through the detection zone with each injection, preventing overlap. Finally, the prepared lunar soil sample solution is stored in sample container 13.

[0108] Step 2, Sample delivery

[0109] A suspension of lunar soil sample particles is injected into sample channel 21 using compressed gas, while simultaneously injecting sheath fluid, either deionized water or a low-conductivity buffer solution. This ensures that the lunar soil sample particles are positioned at the center of the channel and pass through sequentially as individual particles, enveloped by the sheath fluid. Meanwhile, the microfluidic system 1 controls the flow rate within a typical range of (1–100 μL / min), forming a stable and continuous flow of single particles. The system maintains flow stability through a precision pressure or pump control platform to prevent air bubbles or blockage.

[0110] Step 3, Sample Testing

[0111] When lunar soil sample particles pass through the detection window of particle detection system 3, laser 31 emits a focused beam. Forward scattering (FSC) and side scattering (SSC) signals are collected by a photodetector to reflect the particle's size, structure, and roughness. If the sample contains fluorescently labeled or naturally photoluminescent particles, the fluorescence channel simultaneously collects fluorescence intensities at different wavelengths. The multi-channel detector records the fluorescence information for each wavelength, reflecting the particle's compositional characteristics. Simultaneously, an image acquisition device is activated to capture two-dimensional morphological images of the particles as they pass through the detection point, acquiring grayscale and RGB images to capture contours, textures, and edge features. A microelectrode measuring instrument measures the particle surface electrical properties, obtaining zeta potential data to identify the mineral type and dispersion stability of the particles. Finally, a timestamp system synchronously binds the image, optical signal, and electrical data of each particle to construct a multimodal feature set.

[0112] It is worth noting that the binding here refers to associating the feature data of the same particle collected from different sensor channels such as image, scattering, fluorescence, and electrical properties with the particle's ID number and saving it as a particle feature dataset.

[0113] The particle detection system can collect and record all the raw sensor data as each particle passes by, and associate these multi-channel data into a complete particle feature record by particle number.

[0114] Step 4: Data Processing and Storage

[0115] The particle identification system 4 processes the detection data and then identifies the particles in the lunar soil sample. The particle identification system employs the intelligent particle identification method described in Example 1. A multimodal deep learning neural network model receives particle features from multiple sensor channels, including images, scattered light, fluorescence signals, and electrical parameters. It fuses multimodal information to complete tasks such as classification, regression, and clustering of lunar soil sample particles.

[0116] The particle identification system 4 records all detection data, identification results, and physical sorting paths of each particle as structured entries and stores them in the database, supporting subsequent retrieval, tracing, and analysis.

[0117] Optionally, the particle identification system 4 displays particle size distribution, particle type ratio, and characteristic parameter statistics on a visual interface, and can filter specific types of particles for analysis or recycling according to task requirements.

[0118] Step 5: Particle sorting

[0119] The output of the multimodal deep learning neural network model is linked in real time with the particle sorting control module 51. The classification results are mapped into digital control signals to activate electromagnetic deflection, airflow control or optical sorting devices, and the target particles are introduced into different collectors 52 to achieve real-time physical separation.

[0120] This invention constructs a multi-parameter fusion model to jointly analyze forward scattering (FSC, particle size), side scattering (SSC, surface roughness), and characteristic fluorescence intensity (FL, component identification). A three-dimensional classification threshold is then constructed, which involves creating a three-dimensional feature space based on these three optical parameters and classifying or discriminating particle categories within this space. In this three-dimensional space, different particle categories exhibit different distributions along the FSC-SSC-FL coordinate axes. By setting multiple decision planes, the model determines which category a particle belongs to or which sorting channel it enters.

[0121] By training a convolutional neural network (CNN) to identify scattered light signal patterns, the system distinguishes particle shapes (e.g., angular vs. spherical) and compositional categories (e.g., olivine, pyroxene). During the sorting execution phase, the system evaluates the brittleness characteristics of particles based on the particle identification results and automatically selects either a flexible sorting mode or a standard sorting mode. The flexible sorting mode, by reducing the electric field deflection voltage or decreasing the airflow impact force, is suitable for structurally fragile or highly important particles, ensuring particle integrity and recovery quality during the sorting process.

[0122] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent particle identification in lunar soil particle screening, characterized in that, A lunar soil particle screening device was adopted, which includes a flow chamber having a sample channel, a nozzle disposed at the front end of the sample channel, and a sheath fluid tube concentrically disposed outside the sample channel; a microflow rate control system for delivering sheath fluid to the sheath fluid tube of the flow chamber and delivering lunar soil sample solution to the sample channel of the flow chamber; and a particle detection system for detecting lunar soil sample particles and transmitting the detection data to the particle identification system. Includes the following steps: Lunar soil sample particle suspension is injected into the sample channel by compressed gas, and sheath fluid is injected simultaneously so that the lunar soil sample particles are placed in the center of the channel under the sheath fluid, and the individual particles pass through in sequence. When lunar soil sample particles pass through the detection window of the particle detection system, the particle detection system acquires the optical signal, image data and surface electrical signal of the same lunar soil sample particle, forming a multimodal feature set for each lunar soil sample particle. The multimodal feature set is standardized. A multimodal deep learning neural network model is constructed based on a structure of convolutional neural network + multilayer perceptron + splicing and fusion module. Supervised learning training is performed on the multimodal deep learning neural network model to obtain a well-trained multimodal deep learning neural network model; The standardized data is input into a trained multimodal deep learning neural network model to achieve particle recognition.

2. The particle intelligent recognition method according to claim 1, characterized in that, The standardization process includes: Image data is cropped and normalized. The cropped and normalized image data is converted into a model input format of image tensor + numerical vector; Filtering and baseline correction are performed on the optical signal data.

3. The particle intelligent recognition method according to claim 1 or 2, characterized in that, Includes the following steps: The optical signals include forward scattering signals and side scattering signals.

4. The particle intelligent recognition method according to claim 3, characterized in that, The optical signal also includes a fluorescence signal or multiple fluorescence signals in different wavelength bands.

5. The particle intelligent recognition method according to claim 3, characterized in that, Image data is input into a CNN, and the CNN outputs image feature vectors. Other non-image data are input into the MLP, and the MLP outputs a signal feature vector. The splicing and fusion module performs feature fusion on image feature vectors and signal feature vectors, and sends them to the output layer, which is a classification task.

6. The particle intelligent recognition method according to claim 5, characterized in that, The classification uses the cross-entropy loss function and outputs a probability distribution.

7. The particle intelligent recognition method according to claim 1 or 5, characterized in that, The multimodal deep learning neural network model outputs particle category and confidence level.

8. The particle intelligent recognition method according to claim 6, characterized in that, The output layer also performs a regression task, which uses linear activation to output one or more continuous variable values ​​to predict continuous physical property parameters of particles.

9. The particle intelligent recognition method according to claim 6, characterized in that, The output layer also performs clustering tasks, and can optionally output clustering results based on t-SNE / U-MAP dimensionality reduction and DBSCAN clustering results.

Images