Asteroid material classification method, system and device based on multi-parameter constraint

Abstract

The invention discloses an asteroid material classification method, system and device based on multi-parameter constraint. The asteroid material classification method comprises the steps of S1, acquiring a spectrum, a characteristic parameter and a category label of an asteroid; s2, performing data preprocessing on the acquired spectrum, characteristic parameters and category labels, and constructing an asteroid material comprehensive classification data set; step S3, constructing a deep learning model AsteriNet and a random forest model RF; s4, training a deep learning model AsteriNet and a random forest model RF based on the asteroid material comprehensive classification data set to obtain respective optimal models; and S5, integrating the two optimal models to obtain an AsteriRF model, and carrying out asteroid classification based on the AsteriRF model. According to the method, the combined features of various parameters of the asteroid are introduced, and the asteroid material classification precision based on the short-wavelength range and the low-resolution spectrum can be remarkably improved. The AsteriNet model and the random forest model are integrated to obtain the AsteriRF model, so that the generalization performance of the model is further ensured.

Description

Technical Field

[0001] The present invention relates to the field of asteroid physical property inversion, and in particular to an asteroid material classification method, system and device based on multi-parameter constraints. Background Art

[0002] Asteroids are of great value to the study of the formation and evolution of the solar system and planets, as well as the origin of life on Earth. Asteroids of different materials are believed to have formed in different regions of the early solar system and may have experienced collisions and orbital migration during their long evolution. Asteroids of different materials contain different types and contents of mineral resources. Therefore, classifying asteroid materials is not only the basis for studying their physical properties, but also the key to supporting close-in exploration, space resource development, and planetary defense.

[0003] Asteroid material classification is mainly carried out using data such as spectrum, multi-color photometry, albedo and polarization. Among them, spectral type classification is a commonly used method. The longer the spectral range, the higher the separability of the asteroid type is generally. It is difficult to effectively distinguish many types, such as C and X complexes, relying solely on the visible light band or only the near-infrared band. The wavelength range, resolution and quality of the spectrum, as well as the performance of the classification method, all affect the reliability of the classification results. Most of the current asteroid spectral classification results are derived from ground-based telescope observation data. The Gaia space telescope released the GDR 3 data set in 2023, which contains visible light spectra of more than 50,000 asteroids. Compared with most ground-based observation data, Gaia's spectral range is shorter (0.37-1.03μm) and the spectral resolution is much lower, only 44nm. This is a considerable challenge for accurately determining the type of asteroid. Therefore, accurate asteroid material classification methods for ultra-low dispersion spectra are urgently needed.

[0004] At present, the manually designed asteroid spectral classification method usually uses principal component transformation, combined with the slope and subtle features of the spectrum to design the classification process. Whenever faced with a new spectral data set, this traditional manual interpretation method needs to design a new set of classification rules. In recent years, asteroid material classification methods based on machine learning have continued to develop, including the use of naive Bayes, support vector machines, random forests, and artificial neural networks (ANN) to achieve automatic classification of asteroids. Among them, ANN has powerful nonlinear fitting and data mining capabilities and fast reasoning efficiency, showing certain advantages in multiple methods.

[0005] However, existing methods are mainly based on spectral data only. In fact, there is a potential correlation between the material of an asteroid and its other physical properties, orbital characteristics, and photometric information. For example, some studies have found that the probability of an asteroid being C-type or S-type is significantly correlated with the semi-major axis of the orbit. Therefore, it is necessary to extract deep features that constrain the asteroid type from richer input information to improve the reliability of the classification results. In addition, the current artificial intelligence-based methods mainly use fully connected networks, which lack the introduction of advanced algorithms in the field of deep learning and task specificity, the network structure needs to be optimized, and the generalization accuracy of the classification algorithm needs to be improved.

[0006] Therefore, there is an urgent need to provide an asteroid material classification method based on multi-parameter constraints. Summary of the invention

[0007] In order to solve the above problems, the technical solution of the present invention provides an asteroid material classification method, system and device based on multi-parameter constraints, which can classify asteroid materials based on multiple asteroid parameters.

[0008] According to a first aspect of the technical solution of the present invention, there is provided an asteroid material classification method based on multi-parameter constraints, comprising:

[0009] Step S1, obtaining the spectrum, characteristic parameters and category label of the asteroid;

[0010] Step S2, performing data preprocessing on the acquired spectra, characteristic parameters and category labels to construct a comprehensive classification data set of asteroid materials;

[0011] Step S3, constructing a deep learning model AsterNet and a random forest model RF;

[0012] Step S4, training the deep learning model AsterNet and the random forest model RF respectively based on the asteroid material comprehensive classification dataset to obtain their respective optimal models;

[0013] Step S5: Integrate the two optimal models to obtain the AsterRF model, and classify asteroids based on the AsterRF model.

[0014] In the above scheme, the step S1 comprises:

[0015] S1.1. Obtain the spectrum of the asteroid by observing and processing it through ground-based or space telescopes;

[0016] S1.2. Acquire characteristic parameters of the asteroid through multiple databases, wherein the characteristic parameters include photometric parameters, physical parameters and orbital parameters;

[0017] S1.3. Collect and integrate existing asteroid classification results to obtain preliminary category labels for the asteroids.

[0018] In the above scheme, step S2 includes:

[0019] S2.1, normalize the spectral reflectance to 1 at 0.55 μm, and remove this channel, as well as the first channel and the last channel;

[0020] S2.2. Matching the plurality of databases based on the asteroid number to obtain all characteristic parameters of the corresponding asteroid;

[0021] S2.3. Based on the quality label or description of the characteristic parameter, remove objects whose characteristic parameter is null value, double / multiple asteroid parameters, and whose characteristic parameter reliability is lower than a threshold;

[0022] S2.4. Merging the existing type labels of the asteroids according to the required type, and using the merged type labels of the asteroids as the type labels of the asteroid material comprehensive classification dataset;

[0023] S2.5. Evaluate the contribution of a single characteristic parameter based on the information gain ratio, remove characteristic parameters with small contribution, measure the correlation between two parameters based on the Spearman correlation coefficient, and remove parameters with strong correlation with each other;

[0024] S2.6, expanding and standardizing the data of the spectrum and characteristic parameters;

[0025] S2.7. The expanded and standardized spectra and characteristic parameters are divided into multiple groups to constitute the comprehensive classification data set of asteroid materials.

[0026] In the above scheme, building the deep learning model AsterNet in step S3 includes:

[0027] The features of the spectral data and characteristic parameter data of the asteroid are extracted respectively through a dual-branch structure, and the fusion of the two branch features is completed through the self-attention mechanism to construct a deep learning model AsterNet of the asteroid material classification network based on multi-parameter constraints.

[0028] In the above scheme, constructing the random forest model RF in step S3 includes:

[0029] A random forest model RF containing 100 decision trees is constructed, wherein the input data of the random forest model RF is a one-dimensional vector concatenated with the asteroid spectrum and characteristic parameters.

[0030] In the above scheme, step S4 includes:

[0031] S4.1. Training and verifying the multiple groups of data in the asteroid material comprehensive classification dataset respectively, and using the average value of the verification results as the evaluation criteria for the deep learning model AsterNet and the random forest model RF;

[0032] S4.2. Use the random forest model RF to conduct three independent experiments with different initializations on all combinations of asteroid spectra and characteristic parameters. Based on the evaluation criteria of S4.1, obtain the top 100 parameter combinations with the highest verification accuracy, and take the highest verification accuracy as the random forest optimal model;

[0033] S4.3. The deep learning model AsterNet is subjected to three independent experiments with different initializations on the 100 parameter combinations selected in step S4.2, and the model with the highest verification accuracy is taken as the optimal deep learning model.

[0034] In the above scheme, step S5 includes:

[0035] S5.1. Input the spectrum and parameters of the asteroid to be classified, and obtain the random forest optimal model and the deep learning optimal model corresponding to the input parameter combination;

[0036] S5.2, performing the standardized preprocessing in steps S2.1, S2.2, and S2.6 on the spectrum and parameters of the asteroid to be classified;

[0037] S5.3. Use the random forest optimal model and the deep learning optimal model to respectively identify the asteroid to be classified, perform weighted average on the predicted values ​​of the two, and obtain the prediction result through softmax activation, which is the output of the integrated model AsterRF, thereby obtaining the classification result of the asteroid to be classified.

[0038] In the above scheme, step S2.5 includes:

[0039] The characteristic parameters with information gain ratio lower than 0.01 and the characteristic parameters with Spearman correlation coefficient higher than 0.9 were removed.

[0040] According to a second aspect of the technical solution of the present invention, there is provided an asteroid material classification system based on multi-parameter constraints, the system being used to implement any of the above-described asteroid material classification methods based on multi-parameter constraints, the system comprising:

[0041] Acquisition module, which obtains the spectrum, characteristic parameters and category labels of asteroids;

[0042] A preprocessing module preprocesses the acquired data of the spectrum, characteristic parameters and category labels to construct a comprehensive classification data set of asteroid materials;

[0043] Model building module, building deep learning model AsterNet and random forest model RF;

[0044] A training module, which trains the deep learning model AsterNet and the random forest model RF based on the asteroid material comprehensive classification dataset to obtain their respective optimal models;

[0045] The integration module integrates the two optimal models to obtain the AsterRF model, and classifies asteroids based on the AsterRF model.

[0046] According to a third aspect of the technical solution of the present invention, there is provided an electronic device, the electronic device comprising:

[0047] A memory storing executable instructions;

[0048] A processor, wherein the processor runs the executable instructions in the memory to implement any of the above-mentioned asteroid material classification methods based on multi-parameter constraints.

[0049] Beneficial effects of the present invention:

[0050] The present invention discloses a method, system and device for asteroid material classification based on multi-parameter constraints, which introduces the combined features of multiple parameters of asteroids and can significantly improve the accuracy of asteroid material classification based on short-wavelength range and low-resolution spectra. Among them, the network structure that uses dual channels to process spectra and characteristic parameters respectively and fuses features through a self-attention mechanism is an effective method for fusing multi-source information of asteroids. The integration of AsterNet and random forest further ensures the generalization performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0052] Figure 1 A flowchart of the asteroid material classification method based on multi-parameter constraints disclosed in the present invention;

[0053] Figure 2 The present invention discloses a flow chart of constructing a deep learning model AsterNet and a random forest model RF for the asteroid material classification method based on multi-parameter constraints.

[0054] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0055] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0056] The terms "first", "second", etc. in the specification and claims of the present disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein, for example.

[0057] In addition, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus that includes a series of steps or elements is not necessarily limited to those steps or elements explicitly listed, but may include other steps or elements not explicitly listed or inherent to such process, method, product, or apparatus.

[0058] Multiple includes two or more.

[0059] It should be understood that the term "and / or" used in this disclosure is only a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone.

[0060] like Figure 1 and Figure 2 As shown, the first aspect of the technical solution of the present invention provides an asteroid material classification method based on multi-parameter constraints, including:

[0061] Step S1, obtaining the spectrum, characteristic parameters and category label of the asteroid;

[0062] Step S2: preprocess the acquired spectra, characteristic parameters and category labels to construct a comprehensive classification data set of asteroid materials;

[0063] Step S3, constructing a deep learning model AsterNet and a random forest model RF;

[0064] Step S4: Based on the asteroid material comprehensive classification dataset, the deep learning model AsterNet and the random forest model RF are trained respectively to obtain their respective optimal models;

[0065] Step S5: Integrate the two optimal models to obtain the AsterRF model, and classify asteroids based on the AsterRF model.

[0066] Step S1 specifically includes the following steps:

[0067] S1.1. Obtain the spectrum of asteroids by observing and processing them through ground-based or space telescopes;

[0068] The spectrum is a one-dimensional spectrum, including at least: wavelength channels and corresponding relative reflectances, and ideally also a quality label for each reflectance value. The data format is .txt or .csv.

[0069] S1.2. Obtain the characteristic parameters of asteroids through multiple databases, including photometric parameters, physical parameters and orbital parameters;

[0070] Since the availability of various characteristic parameters of different asteroids is different, and they may be the integration results of multiple databases. Therefore, it is necessary to obtain various characteristic parameters from multiple databases. Specifically, the photometric parameters can be derived from the LCDB database of Planetary DataSystem, including absolute magnitude (H), phase slope parameter (G) and maximum light curve amplitude (AmpMax). The available sources of physical parameters are the same as above, including rotation period, effective diameter and albedo. Orbital parameters come from Minor Planet Center (MPC Asteroid Orbit Database), including semi-major axis, eccentricity, orbital inclination, perigee / aphelion distance, orbit type and other parameters.

[0071] S1.3. Collect existing asteroid classification results, integrate them, and obtain preliminary category labels for asteroids.

[0072] Among them, the existing asteroid classification results include integration from published literature or public databases. There is a priority during integration. Preferably, the label given based on the visible light near-infrared spectrum has a higher priority than the label obtained using only the visible light or near-infrared spectrum, and takes precedence over the label given by multi-color photometry; the latest published labels are used first. By setting the priority, asteroids can use more reliable labels to ensure the accuracy of classification. For example, in this embodiment, the label based on the visible light near-infrared spectrum is given a higher priority than the label obtained using only the visible light or near-infrared spectrum. If the visible light near-infrared spectrum gives Class C and the visible light spectrum alone gives Class S, Class C is used as the preliminary category label because the spectral range is longer and the judgment is more accurate.

[0073] Step S2 includes:

[0074] S2.1, normalize the spectral reflectance to 1 at 0.55 μm, and remove this channel, as well as the first channel and the last channel;

[0075] Among them, since all spectra at 0.55μm are normalized to 1, there is no effective information to distinguish different types. In order to reduce the amount of unnecessary data, this channel is removed. In order to ensure the quality of the spectra in the dataset, for data with spectral quality labels, it is necessary to remove the channels in the low-quality set. The channels in the low-quality set are usually the first channel and the last channel, so they need to be removed. If there are still low-quality channels in the remaining channels, they need to be removed together.

[0076] S2.2. Match multiple databases based on asteroid numbers to obtain all characteristic parameters of the corresponding asteroids;

[0077] Since the asteroid database has its own asteroid number, which is the unique identifier of the asteroid, multiple databases can be matched according to the asteroid number to obtain all the characteristic parameters of the corresponding asteroid. In this embodiment, the LCDB database and the MPC asteroid orbit database are matched to obtain the characteristic parameters of the corresponding asteroid.

[0078] S2.3. Based on the quality labels or descriptions of characteristic parameters, remove objects with null values ​​for characteristic parameters, double / multiple asteroid parameters, and characteristic parameter reliability below the threshold;

[0079] Among them, the objects whose characteristic parameter reliability is lower than the threshold include objects whose rotation period results are unreliable due to reasons such as rolling.

[0080] S2.4. Merge the type labels of existing asteroids according to the required type;

[0081] For different application goals such as close-up detection, the degree of subdivision of types varies. For low-resolution spectra in the optical band, the separability between some categories, especially subcategories, is low.

[0082] Therefore, based on the Bus-DeMeo classification, the asteroid materials are classified and merged, and the subclasses are merged into their equivalent major categories, such as Xc, Xk and Xe merged into the X category; categories with small sample numbers are merged into major categories with similar spectral types or similar mineralogical interpretations, such as O, R and K merged into S category. After the merger, there are 6 types of asteroids: A-type, representing asteroid types with obvious olivine characteristics in the spectrum. C-type (including B and C subclasses), representing carbonate-rich asteroid groups. D-type (including T-type), showing a red slope in the spectrum, may be composed of organic-rich primitive materials, mainly distributed in the outer region of the solar system. S-type (including K, L, O, Q, R and S subclasses), representing a major group of asteroids dominated by silicates. V-type, with strong 1μm and 2μm absorption characteristics, is mainly associated with Vesta and is rich in pyroxene. X-type (including X-type subclasses), including metal-rich asteroids.

[0083] S2.5. Evaluate the contribution of a single characteristic parameter based on the information gain ratio, remove characteristic parameters with small contribution, measure the correlation between two parameters based on the Spearman correlation coefficient, and remove parameters with strong correlation with each other;

[0084] Since not all matched parameters are of great significance to the classification of asteroid materials, removing characteristic parameters with small contributions and characteristic parameters that are strongly correlated with each other and have highly redundant information can reduce the data dimension and reduce the complexity of the experiment. The specific steps are:

[0085] The information gain ratio (IGR) is used to evaluate the contribution of individual characteristic parameters, rank the importance of input characteristic parameters, and help identify the effectiveness of each thermal parameter in distinguishing asteroid types. The expression equation is as follows:

[0086]

[0087] IG(C,P)=H(C)-H(C|P) (2)

[0088]

[0089] Among them, H(C) is the information entropy of the target variable C (i.e., asteroid type), |C i | is the number of samples of category i, |C| is the number of all samples, and H(C|P) is the conditional entropy of the target variable C given the characteristic parameter P. IG(C,P) is the information gain. H(P) measures the amount of information in the characteristic parameter P. IGR(C,P) is the information gain ratio.

[0090] The Spearman correlation coefficient is used to measure the correlation between two characteristic parameters. The value range of the correlation coefficient r is between (-1, 1). When r = 0, it means that the two variables are independent; the larger the absolute value of r, the stronger the correlation. The formula is:

[0091]

[0092] Among them, R(X i ) and R(Y i ) are the ranks of variables X and Y of the i-th sample, and are the rank means of variables X and Y respectively, and n is the total number of samples.

[0093] According to the results obtained from formulas (1) to (4), characteristic parameters with information gain ratios lower than 0.01 and Spearman correlation coefficients higher than 0.9, such as the longitude of the ascending node, the argument of perihelion, the angle of mean anomaly, the mean motion, the semi-diameter and the synodic period, are removed. The remaining parameters are retained as available input parameters for subsequent model construction.

[0094] S2.6, expanding and standardizing the data of the spectrum and characteristic parameters;

[0095] For spectral data, the mean and standard deviation of each wavelength are statistically classified, and a Gaussian distribution centered on the mean is used to generate simulated spectra according to the standard deviation. For absolute magnitude (H), phase slope parameter (G), and maximum light curve amplitude (AmpMax), random Gaussian noise with the median of the statistical error as the boundary is added to their own values ​​for expansion. For rotation period and effective diameter, the error tends to increase as the estimated value increases, and the boundary of the Gaussian noise is set to the median of the error as a percentage of the estimated value. Orbital parameters are generally more reliable, and the noise boundary is set to 0.001. The orbit type remains unchanged. Finally, Z-score standardization is applied to each characteristic parameter. Z-score is used to convert the values ​​in the data set into the form of a standard normal distribution. Data of different scales can be compared and analyzed on the same scale.

[0096] S2.7. The expanded and standardized spectra and characteristic parameters are divided into multiple groups to form a comprehensive asteroid material classification data set.

[0097] Building the deep learning model AsterNet in step S3 includes:

[0098] The features of the asteroid's spectral data and characteristic parameter data are extracted separately through a dual-branch structure, and the fusion of the two branch features is completed through the self-attention mechanism to construct a deep learning model AsterNet for asteroid material classification network based on multi-parameter constraints.

[0099] One of the branches in the dual-branch structure extracts local features of the spectrum through one-dimensional convolution, and uses the bottleneck layer of identity mapping to promote efficient feature extraction and prevent gradient vanishing. Subsequently, adaptive average pooling is used to reduce the dimension of the features. At the same time, the other branch processes the asteroid characteristic parameters organized in a certain order through a fully connected layer. Subsequently, the features of the two branches are fused through a multi-head self-attention module. The fused features pass through the subsequent fully connected layer and are activated using the Softmax function to complete the classification. Among them, the activation function of the feature extraction layer is LeakyReLU.

[0100] Constructing the random forest model RF in step S3 includes:

[0101] A random forest model RF containing 100 decision trees is constructed. The input data of the random forest model RF is a one-dimensional vector composed of the asteroid spectrum and characteristic parameters.

[0102] Step S4 includes:

[0103] S4.1. Train and validate multiple sets of data in the asteroid material comprehensive classification dataset separately, and use the average value of the validation results as the evaluation criteria for the deep learning model AsterNet and the random forest model RF;

[0104] In this embodiment, 10-fold cross validation is preferably used to evaluate model performance. The expanded data set is randomly divided into 10 parts with equal sample sizes, and the experiment is repeated 10 times, each time 9 parts are taken as training sets and the remaining 1 part is taken as validation set. The overall effect of the deep learning model AsterNet and the random forest model RF is evaluated by the average value of the 10 validation results.

[0105] S4.2. Use the random forest model RF to conduct three independent experiments with different initializations on all combinations of asteroid spectra and characteristic parameters. Based on the evaluation criteria of S4.1, the top 100 parameter combinations with the highest verification accuracy are obtained, and the model with the highest verification accuracy is selected as the optimal model.

[0106] Since different input characteristic parameters affect the classification performance of the model, and there are 13 parameters used in this embodiment, the number of parameter combinations reaches 8191. It is very time-consuming and unnecessary to perform all parameter ablation experiments on the AsterNet model. Therefore, based on the data set generated by S2, a fast-training random forest model is used to perform three independent experiments with different initializations on all parameter combinations to obtain the top 100 parameter combinations with the highest verification accuracy. The preliminary screening of the parameter combinations is conducive to improving the training speed and efficiency of the AsterNet model.

[0107] S4.3. The AsterNet model is subjected to three independent experiments with different initializations on the 100 parameter combinations selected in step S4.2, and the model with the highest verification accuracy is taken as the optimal model.

[0108] Specifically, the training was performed for 80 rounds with a learning rate of 1×10 -4 , the batch size is 128, the dropout is 0.3, the loss function is cross entropy, and the optimizer is Adam.

[0109] Since the asteroid characteristic parameters required for the model to achieve the highest accuracy may not be available at the same time, based on the above 100 parameter combinations, the available parameter scenarios are divided according to the difficulty of parameter acquisition, and the models are trained and evaluated respectively to obtain parameter combinations suitable for the following 7 situations. The classification reliability gradually increases from the front to the back:

[0110] 1. Only spectral data: Applicable to newly discovered asteroids, in which case the input to the model is only the spectrum.

[0111] 2. After orbit determination is completed, orbit parameters can be obtained: the recommended input parameter at this time is orbit type.

[0112] 3. Obtain absolute magnitude and orbital parameters: The recommended input parameter combination at this time is absolute magnitude, orbit type, eccentricity, orbit inclination and aphelion distance.

[0113] 4. Obtain the light curve of the asteroid's complete rotation period: The recommended input parameter combination at this time is absolute magnitude, maximum light curve amplitude, rotation period, orbit type and aphelion distance.

[0114] 5. Albedo parameters are known: The recommended input parameter combination is absolute magnitude, albedo, orbit type and eccentricity. The source of albedo is calculated based on absolute magnitude and diameter inverted from infrared or optical observation data, or inferred based on orbital groups.

[0115] 6. The phase slope parameter G is known: The recommended input parameter combination at this time is absolute magnitude, G, orbit type, orbit inclination and eccentricity.

[0116] 7. Both albedo and phase slope parameters G are available: At this time, the recommended input parameter combination is absolute magnitude, G, albedo, orbit type and aphelion distance.

[0117] Step S5 includes:

[0118] S5.1. Input the spectrum and parameters of the asteroid to be classified, and obtain the optimal AsterNet model and RF model corresponding to the input parameter combination;

[0119] Among them, a parameter available scenario is selected according to the input parameter combination, and the optimal AsterNet model and RF model corresponding to the scenario are selected.

[0120] S5.2, performing standardized preprocessing on the spectrum and parameters of the asteroid to be classified as in steps S2.1, S2.2, and S2.6;

[0121] S5.3. Use the optimal AsterNet model and RF model to identify the asteroid to be classified respectively, perform weighted average of the prediction values ​​of the two, and obtain the prediction result through softmax activation, which is the output of the integrated model AsterRF, thereby obtaining the classification result of the asteroid to be classified;

[0122]

[0123] Among them, OA AN and OA RF are the 10-fold cross-validation average accuracy of AsterNet and RF, P AN and P RF are the predicted values ​​of the two for the classified objects.

[0124] A second aspect of the technical solution of the present invention provides an asteroid material classification system based on multi-parameter constraints, which is used to implement the asteroid material classification method based on multi-parameter constraints as described above. The system includes:

[0125] Acquisition module, which obtains the spectrum, characteristic parameters and category labels of asteroids;

[0126] The preprocessing module preprocesses the acquired spectra, characteristic parameters and category label data to construct a comprehensive classification dataset of asteroid materials;

[0127] Model building module, building deep learning model AsterNet and random forest model RF;

[0128] The training module trains the deep learning model AsterNet and the random forest model RF based on the asteroid material comprehensive classification dataset to obtain their respective optimal models;

[0129] The integration module integrates the two optimal models to obtain the AsterRF model, and classifies asteroids based on the AsterRF model.

[0130] A third aspect of the technical solution of the present invention provides an electronic device, the electronic device comprising:

[0131] A memory storing executable instructions;

[0132] A processor runs executable instructions in the memory to implement the asteroid material classification method based on multi-parameter constraints as described above.

[0133] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0134] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0135] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0136] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A method for classifying asteroid materials based on multi-parameter constraints, characterized in that: include: Step S1, obtaining the spectrum, characteristic parameters and category label of the asteroid; Step S2, performing data preprocessing on the acquired spectra, characteristic parameters and category labels to construct a comprehensive classification data set of asteroid materials; Step S3, constructing a deep learning model AsterNet and a random forest model RF; Step S4, training the deep learning model AsterNet and the random forest model RF respectively based on the asteroid material comprehensive classification dataset to obtain their respective optimal models; Step S5: Integrate the two optimal models to obtain the AsterRF model, and classify asteroids based on the AsterRF model.

2. The asteroid material classification method based on multi-parameter constraints according to claim 1, characterized in that: The step S1 comprises: S1.

1. Obtain the spectrum of the asteroid by observing and processing it through ground-based or space telescopes; S1.

2. Acquire characteristic parameters of the asteroid through multiple databases, wherein the characteristic parameters include photometric parameters, physical parameters and orbital parameters; S1.

3. Collect and integrate existing asteroid classification results to obtain preliminary category labels for the asteroids.

3. The asteroid material classification method based on multi-parameter constraints according to claim 2 is characterized in that: The step S2 comprises: S2.1, normalize the spectral reflectance to 1 at 0.55 μm, and remove this channel, as well as the first channel and the last channel; S2.

2. Matching the plurality of databases based on the asteroid number to obtain all characteristic parameters of the corresponding asteroid; S2.

3. Based on the quality label or description of the characteristic parameter, remove objects whose characteristic parameter is null value, double / multiple asteroid parameters, and whose characteristic parameter reliability is lower than a threshold; S2.

4. Merging the existing type labels of the asteroids according to the required type, and using the merged type labels of the asteroids as the type labels of the asteroid material comprehensive classification dataset; S2.

5. Evaluate the contribution of a single characteristic parameter based on the information gain ratio, remove characteristic parameters with small contribution, measure the correlation between two parameters based on the Spearman correlation coefficient, and remove parameters with strong correlation with each other; S2.6, expanding and standardizing the data of the spectrum and characteristic parameters; S2.

7. The expanded and standardized spectra and characteristic parameters are divided into multiple groups to constitute the comprehensive classification data set of asteroid materials.

4. The asteroid material classification method based on multi-parameter constraints according to claim 1, characterized in that: Building the deep learning model AsterNet in step S3 includes: The features of the spectral data and characteristic parameter data of the asteroid are extracted respectively through a dual-branch structure, and the fusion of the two branch features is completed through the self-attention mechanism to construct a deep learning model AsterNet of the asteroid material classification network based on multi-parameter constraints.

5. The asteroid material classification method based on multi-parameter constraints according to claim 3 is characterized in that: Constructing the random forest model RF in step S3 includes: A random forest model RF containing 100 decision trees is constructed, wherein the input data of the random forest model RF is a one-dimensional vector concatenated with the asteroid spectrum and characteristic parameters.

6. The asteroid material classification method based on multi-parameter constraints according to claim 5, characterized in that: Step S4 includes: S4.

1. Train and verify the multiple groups of data in the asteroid material comprehensive classification dataset respectively, and use the average value of the verification results as the evaluation criteria for the deep learning model AsterNet and the random forest model RF; S4.

2. Use the random forest model RF to conduct three independent experiments with different initializations on all combinations of asteroid spectra and characteristic parameters. Based on the evaluation criteria of S4.1, obtain the top 100 parameter combinations with the highest verification accuracy, and take the highest verification accuracy as the random forest optimal model; S4.

3. The deep learning model AsterNet is subjected to three independent experiments with different initializations on the 100 parameter combinations selected in step S4.2, and the highest verification accuracy is taken as the optimal deep learning model.

7. The asteroid material classification method based on multi-parameter constraints according to claim 6, characterized in that: Step S5 includes: S5.

1. Input the spectrum and parameters of the asteroid to be classified, and obtain the random forest optimal model and the deep learning optimal model corresponding to the input parameter combination; S5.2, performing the standardized preprocessing in steps S2.1, S2.2, and S2.6 on the spectrum and parameters of the asteroid to be classified; S5.

3. Use the random forest optimal model and the deep learning optimal model to respectively identify the asteroid to be classified, perform weighted average on the predicted values ​​of the two, and obtain the prediction result through softmax activation, which is the output of the integrated model AsterRF, thereby obtaining the classification result of the asteroid to be classified.

8. The asteroid material classification method based on multi-parameter constraints according to claim 3 is characterized in that: Step S2.5 includes: The characteristic parameters with information gain ratio lower than 0.01 and the characteristic parameters with Spearman correlation coefficient higher than 0.9 were removed.

9. An asteroid material classification system based on multi-parameter constraints, characterized in that: The system is used to implement the asteroid material classification method based on multi-parameter constraints as described in any one of claims 1 to 8, and the system includes: Acquisition module, which obtains the spectrum, characteristic parameters and category labels of asteroids; A preprocessing module preprocesses the acquired data of the spectrum, characteristic parameters and category labels to construct a comprehensive classification data set of asteroid materials; Model building module, building deep learning model AsterNet and random forest model RF; A training module, which trains the deep learning model AsterNet and the random forest model RF based on the asteroid material comprehensive classification dataset to obtain their respective optimal models; The integration module integrates the two optimal models to obtain the AsterRF model, and classifies asteroids based on the AsterRF model.

10. An electronic device, characterized in that: The electronic device comprises: A memory storing executable instructions; A processor, wherein the processor runs the executable instructions in the memory to implement the method according to any one of claims 1 to 8.

Images