Hyperspectral multi-feature fusion identification method and system for molybdenite main mineral

Abstract

The invention relates to a molybdenite main mineral hyperspectral multi-feature fusion identification method and system, and belongs to the technical field of mineral identification and spectral analysis. By collecting multi-angle spectral data and extracting hyperspectral wave band characteristics, a ground object spectrum library is constructed, the defect that traditional single-angle spectral data is insufficient in characterization of anisotropic reflection characteristics of mineral crystals is effectively overcome, and completeness and distinction degree of spectral characteristics of molybdenite of different grades are remarkably improved; according to the constructed rapid recognition model, the Euclidean distance, the cosine similarity and the spectral data divergence are fused with the weighted score to serve as a curve similarity evaluation criterion, the weight can be adjusted in a self-adaptive mode, the problem that a traditional method depends on a single weight or a fixed index is solved, rapid recognition of minerals is achieved, and the recognition accuracy is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of mineral identification and spectral analysis, and particularly relates to a molybdenite main mineral hyperspectral multi-feature fusion identification method and system. BACKGROUND

[0002] Mineral identification is the core basis of geological exploration, resource assessment and beneficiation process optimization, and its accuracy directly affects the efficiency of ore classification, the design of flotation parameters and the utilization rate of resources. Molybdenite (MoS2) is the main carrier mineral of molybdenum metal, and its rapid and accurate identification is of great value to the optimization of beneficiation process and the improvement of resource recovery rate. However, due to the complex composition and similar spectral characteristics of molybdenite, traditional chemical analysis methods have defects such as long time consumption, high cost and sample destruction, which cannot meet the needs of rapid and non-destructive identification on site. The present application relates to a visible-short wave infrared (VNIR-SWIR) spectral-based rapid identification method for molybdenite main minerals, which belongs to the technical field of mineral spectral detection.

[0003] A multi-feature fusion mineral identification method and its ore sorting machine are proposed in Chinese patent CN 118570555 A, which constructs a beneficiation identification model by collecting color images, near-infrared images, XRT images, dual-energy XRT images, ultraviolet fluorescence images and black-and-white grayscale images of ores. This method requires collecting a large amount of pre-sequenced data from different sources, constructing a large multi-source database of ore samples, and lacks specificity in research on certain minerals, which is costly and inefficient for real-time identification. Furthermore, this method does not involve SWIR band hyperspectral information, ignoring the characteristic differences of different ores in the hyperspectral band.

[0004] A mineral identification method, device, system and medium based on hyperspectral technology are proposed in Chinese patent CN 117372870 A, which introduces hyperspectral technology in ore identification, conducts deep learning according to the strong linear identification waveband and characteristic peak of each kind of mineral sample in the hyperspectral reference database, and establishes a hyperspectral network identification model. However, it still has the following problems: First, the feature selection mode is single, i.e., relying on the static feature extraction strategy of "strong linear identification waveband + characteristic peak", which is easily disturbed in the mixed mineral spectral superposition scene. For example, when molybdenite and pyrite are associated, their reflection characteristics in the VNIR band are similar, and the distinguishing degree is limited. Second, the environmental adaptability is insufficient, the preprocessing process only smooths and normalizes the noise, and does not mention the spectral offset compensation mechanism for mine dust scattering and humidity, which raises doubts about the model generalization ability in actual working conditions. Finally, the model has a real-time bottleneck, i.e., using a general deep learning model (such as CNN) for end-to-end training, without model compression optimization for embedded devices, which cannot meet the millisecond-level response requirement of the mine sorting line.

[0005] The existing document Michael J. Smith et al. (Remote Sensing, 2021, 15(3): 456-472) adopts a laboratory-level hyperspectral imaging device in combination with an SVM algorithm to achieve a classification accuracy of 96.5%, but is limited by the device size and operation complexity, and is difficult to cope with the similarity problem between complex components of molybdenite and real-time processing requirements; the spectral angle matching optimization method proposed by Wang et al. (Journal of Applied Remote Sensing, 2021, 8(2): 025401) has insufficient recognition of mixed mineral spectra.

[0006] Therefore, in view of the above problems, the present application provides a multi-feature fusion recognition method for molybdenite main minerals. SUMMARY

[0007] The present application aims to overcome the problems in the prior art such as long time consumption, high cost, single mineral analysis, insufficient specificity research, and dependence on single weight or fixed index, and provides a hyperspectral multi-feature fusion recognition method and system for molybdenite main minerals, which realizes rapid and non-destructive accurate recognition of molybdenite main minerals and is suitable for real-time detection in mine sites.

[0008] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions: In a first aspect, the present application provides a hyperspectral multi-feature fusion recognition method for molybdenite main minerals, comprising the following steps: Collecting multi-angle spectral data of molybdenite with different grades, and pre-processing the multi-angle spectral data; Extracting hyperspectral band features of the pre-processed spectral data, and constructing a ground object spectral library based on the extracted hyperspectral band features; Based on a dynamic weighting strategy, screening potential targets in the ground object spectral library through Euclidean distance, introducing cosine similarity to measure the robustness of spectral curve shape matching, and combining spectral information divergence to analyze spectral information entropy difference to construct a fast recognition model; After pre-processing the spectral data of the molybdenite to be tested, calculating the similarity between the spectral data and the features in the ground object spectral library through the constructed fast recognition model to obtain a similarity result, and identifying the category of the molybdenite to be tested based on the similarity result.

[0009] In the step of collecting multi-angle spectral data of molybdenite with different grades and pre-processing the multi-angle spectral data, the method of pre-processing the multi-angle spectral data is as follows: Eliminate noise interference in the spectral data after removing outliers through an adaptive sliding window fitting strategy to obtain spectral absorption characteristics; The abnormal values in the spectral absorption features are removed by a multi-scale waveband coupling IQR algorithm to obtain the spectral absorption features with the abnormal values removed. The spectral absorption features with the abnormal values removed are enhanced by spectral de- enveloping.

[0010] In the step of eliminating the noise interference in the spectral data with the abnormal values removed by the adaptive sliding window fitting strategy to obtain the spectral absorption features, the specific method is as follows: The length of the adaptive sliding window and the polynomial order are determined; The data points in the window are least square fitted to generate smooth values and obtain the spectral absorption features.

[0011] In the step of removing the abnormal values in the spectral data by the multi-scale waveband coupling IQR algorithm to obtain the spectral absorption features with the abnormal values removed, the specific method is as follows: The hyperspectral waveband J in the spectral curve is divided into a plurality of overlapping subintervals, and each overlapping subinterval is denoted as waveband j; For each spectral curve, the reflectivity distribution quartiles and IQR of each waveband j are calculated; The standard IQR is introduced for dynamic threshold calculation to define the upper limit and lower limit of the abnormal values; Whether the IQR value of each waveband j is between the upper and lower limits of the abnormal values is judged one by one, if yes, it is determined as normal, if not, it is determined as abnormal value.

[0012] The specific formula of the standard IQR introduced for dynamic threshold calculation to define the upper limit and lower limit of the abnormal values is as follows:

[0013]

[0014]

[0015] Wherein, The standard IQR of waveband j is denoted as IQR j, The number of waveband j is denoted as N, The IQR of waveband j is denoted as IQR j; The abnormal upper limit of waveband j is denoted as U j, The third quartile of waveband j is denoted as Q3 j; The abnormal lower limit of waveband j is denoted as L j, The first quartile of waveband j is denoted as Q1 j; The average IQR of waveband J is denoted as IQR avg; The noise adaptive adjustment factor is denoted as a, The noise standard deviation of subinterval k is denoted as σ k.

[0016] The hyperspectral band features of the extracted pre-processed spectral data are extracted, and in the step of constructing the geo-physical spectral library based on the extracted hyperspectral band features, the following specific method is used: A multi-scale sliding window extreme value search algorithm is used to extract the absorption features of the hyperspectral data to generate candidate bands; The effective region of the absorption band of the candidate band is determined, the wavelength peak position, absorption band width and absorption band depth matrix are calculated, and the candidate band features are obtained; The geo-physical spectral library is constructed based on the calculated candidate band features.

[0017] The method for determining the effective region of the absorption band of the candidate band, calculating the wavelength peak position, absorption band width and absorption band depth matrix, and obtaining the candidate band features is as follows: The first derivative is introduced The second derivative is introduced The envelope line is introduced The boundary is corrected, and the effective region of the absorption band is defined as In the effective region, the extreme points of the second derivative are used to determine the absorption band width, and the formula is as follows:

[0018]

[0019]

[0020] wherein, represents the wavelength, represents the absorption peak center, represents the wavelength of the absorption peak center; R( ) represents the reflectivity at the wavelength , represents the first wavelength, u represents the first wavelength, u+ represents the first wavelength, u- represents the wavelength step, that is, the discrete interval in the derivative calculation, represents the empirical threshold value for controlling the sensitivity of the absorption band boundary; FWHM represents the absorption band width, which is used for mineral feature analysis; represents the wavelength on the right side of the characteristic band, represents the wavelength on the left side of the characteristic band, represents the absorption band determination condition; The absorption band depth is dynamically mapped, and the absorption band depth matrix is calculated, and the specific formula is as follows: ​

[0021] wherein, denotes the wavelength of the absorption band, exp() denotes the exponential function, and a is the depth enhancement coefficient, SNR( ) denotes the local signal-to-noise ratio at the wavelength ; the wavelength peak position , the absorption band width, and the absorption band depth matrix are unified into a vector as a characteristic value for constructing the ground object spectrum library, and the specific formula is as follows:

[0022] wherein, denotes the derivative curvature feature.

[0023] In the step of constructing the fast recognition model based on the dynamic weighting strategy, the potential targets in the ground object spectrum library are screened through the Euclidean distance, the robustness of the cosine similarity in measuring the shape matching of the spectral curve is introduced, and the spectral information entropy difference is analyzed by combining the spectral information divergence, and the constructed fast recognition model is as follows:

[0024] wherein, FS denotes the similarity score of the characteristic to be detected molybdenite and the characteristic in the ground object spectrum library, is the dynamically allocated weight, and the sum is 1; denotes the Euclidean distance normalization result, denotes the cosine similarity, denotes the spectral information divergence normalization result.

[0025] The smaller the values of the Euclidean distance normalization result and the spectral information normalization result, the higher the similarity of the spectral curve.

[0026] In the second aspect, the present application provides a molybdenite main mineral hyperspectral multi-feature fusion recognition system, comprising: a data acquisition and preprocessing module for acquiring multi-angle spectral data of molybdenite with different grades and preprocessing the multi-angle spectral data; a ground object spectrum library construction module for extracting hyperspectral band features of the preprocessed spectral data and constructing a ground object spectrum library based on the extracted hyperspectral band features; a fast recognition model construction module for screening potential targets in the ground object spectrum library through the Euclidean distance based on a dynamic weighting strategy, introducing the robustness of the cosine similarity in measuring the shape matching of the spectral curve, and constructing a fast recognition model by combining the spectral information entropy difference analyzed by the spectral information divergence; The ore recognition module is used for calculating the similarity of the to-be-tested molybdenite with features in the ground object spectrum library through a constructed rapid recognition model after pre-processing the spectrum data of the to-be-tested molybdenite, obtaining a similarity result, and recognizing the category of the to-be-tested molybdenite based on the similarity result.

[0027] Compared with the prior art, the present application has the following beneficial effects: The present application provides a molybdenite main mineral hyperspectral multi-feature fusion recognition method, comprising the following steps: collecting multi-angle spectrum data of molybdenite of different grades, and pre-processing the multi-angle spectrum data; extracting hyperspectral band features of the pre-processed spectrum data, and constructing a ground object spectrum library based on the extracted hyperspectral band features; based on a dynamic weighting strategy, screening potential targets in the ground object spectrum library through Euclidean distance, introducing cosine similarity to measure the robustness of spectral curve shape matching, and combining spectral information divergence to analyze spectral information entropy difference to construct a rapid recognition model; after pre-processing the spectrum data of the to-be-tested molybdenite, calculating the similarity of the to-be-tested molybdenite with features in the ground object spectrum library through the constructed rapid recognition model, obtaining a similarity result, and recognizing the category of the to-be-tested molybdenite based on the similarity result. By collecting multi-angle spectrum data and extracting hyperspectral band features, the ground object spectrum library is constructed, effectively overcoming the defect of the traditional single-angle spectrum data in insufficiently representing the anisotropic reflection characteristics of mineral crystals, and significantly improving the completeness and distinguishability of the spectrum features of molybdenite of different grades; the constructed rapid recognition model uses Euclidean distance, cosine similarity and spectral data divergence fusion weighted score as the curve similarity evaluation standard, can adaptively adjust the weight, overcomes the problem of the traditional method relying on a single weight or fixed index, realizes the rapid recognition of minerals, and enhances the accuracy of recognition. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 The present application is a method flowchart; Figure 2 The present application is a system structure diagram. DETAILED DESCRIPTION

[0029] In order to further understand the content of the present application, the present application will be described in detail below in combination with the drawings and specific embodiments. It should be understood that the embodiments are only used to explain and not limit the present application.

[0030] Embodiment 1 A molybdenite main mineral hyperspectral multi-feature fusion recognition method, comprising the following steps: S1: collecting multi-angle spectrum data of molybdenite of different grades, and pre-processing the multi-angle spectrum data; S2: extracting hyperspectral band features of the pre-processed spectrum data, and constructing a ground object spectrum library based on the extracted hyperspectral band features; S3: Based on the dynamic weighting strategy, the potential targets in the feature spectrum library are screened by the Euclidean distance, the cosine similarity is introduced to measure the robustness of spectral curve shape matching, and the spectral information entropy difference is analyzed by combining the spectral information divergence to construct a fast recognition model; S4: After preprocessing the spectral data of the molybdenite to be tested, the similarity between the spectral data of the molybdenite to be tested and the features in the feature spectrum library is calculated by using the constructed fast recognition model to obtain a similarity result, and the category of the molybdenite to be tested is recognized based on the similarity result.

[0031] Specifically, in S1, multi-angle spectral data of molybdenite with different grades is collected, and the multi-angle spectral data is preprocessed. Molybdenite samples for constructing a fast recognition standard library: The molybdenite powder samples obtained are from different process stages of one roughing, three cleanings and three scavengings. In order to unify the experimental environment and ensure that the sampling process eliminates environmental factors, a total of 5 different grade molybdenite powder samples are collected, which are 1.993%, 13.76%, 22.95%, 38.25% and 48.25% respectively. Since molybdenite and other metal minerals are scattered, fine-grained and spotted disseminated in the ore matrix composed of quartz, potassium feldspar and other silicate minerals, samples of quartz, mica, potassium feldspar and other silicate minerals commonly associated with molybdenum deposits are also collected, and samples of pyrite in sulfides are also collected.

[0032] Using the ATH9100-17 feature spectrometer of Optosky, multi-point average sampling is performed on the obtained molybdenite samples, and twenty multi-angle spectral data of molybdenite concentrate powder samples with different grades and associated ore powder samples are collected under the same lighting conditions.

[0033] In order to eliminate the influence of external noise factors such as light source environment on the spectral data, the collected multi-angle spectral data is preprocessed before rapid identification of different ores: 1) First, the adaptive sliding window fitting strategy is used to suppress noise interference and retain spectral absorption characteristics, eliminate high-frequency noise, and obtain spectral absorption characteristics. The specific method is as follows: First, the adaptive sliding parameter window is optimized, and the network search is used to determine the optimal window length and polynomial order: The optimal window length calculation method is as follows:

[0034] The polynomial order d=3; where K is the sliding window half-width, i.e. the center point is expanded to the left and right by K waveband numbers, which is initialized to 5 wavelengths here.

[0035] Secondly, the local polynomial fitting is used, that is, the data points in the window are least square fitted to generate smooth values to smooth the noise and preserve the spectral absorption features:

[0036] where is the fitting coefficient matrix, which is solved by:

[0037] where X represents the polynomial basis matrix of local polynomial fitting, which is used to convert the band reflectance data within the sliding window into a matrix in polynomial form; represents the matrix of the data point set within the window; represents the value of the m-th polynomial basis of the M-th data point in the polynomial basis matrix X, which is a specific element in the matrix.

[0038] 2) Secondly, the multiscale band coupled IQR algorithm is used to remove outliers in the spectral data, eliminate spectral data outliers caused by external light source interference or equipment noise during data acquisition, and constrain the characteristic spectrum within a stable range to ensure the accuracy of subsequent data analysis and modeling.

[0039] The multiscale band coupled IQR algorithm captures the local continuity features of the hyperspectral curve through overlapping band cluster division, avoids the "fragmentation" misjudgment caused by independent band statistics, and dynamically introduces the k factor to dynamically scale the IQR multiple according to the noise intensity of the sub-interval. In low signal-to-noise ratio bands (such as SWIR 1400nm water absorption area), the threshold is relaxed, and in high signal-to-noise ratio bands (such as VNIR 500-800nm), the threshold is tightened. By setting a reliable threshold to filter out abnormal curves, an accurate molybdenite main mineral spectral standard library under different grades is established to ensure the stability and reliability of the data source for subsequent similarity analysis. The specific improvements are as follows: Firstly, based on the spectral continuity assumption, the hyperspectral VNIR-SWIR (Visible and Near-Infrared, VNIR-Short-Wave Infrared, SWIR) band J is divided into several overlapping sub-intervals (such as 400-600nm, 550-750nm, etc.), each overlapping sub-interval is denoted as band j, and the local noise mutation is suppressed to interfere with the global determination; then for each spectral curve, the reflectance distribution quartiles and IQR (Interquartile Range) of each band j are calculated, and the calculation formula is as follows:

[0040] where, represents the third quartile, represents the first quartile.

[0041] Introducing standard IQR for dynamic threshold calculation, defining upper and lower limits for anomalies:

[0042]

[0043]

[0044] in, The standard IQR for band j Indicates the number of bands j. Indicates the IQR of band j; This indicates the upper limit of the anomaly for band j. This represents the third quartile of band j; This indicates the lower limit of outliers for band j. This represents the first quartile of band j; This represents the average IQR of band J.

[0045] in This is a noise adaptive adjustment factor. Let k be the noise standard deviation of subinterval k. This represents the global noise level. The formula is as follows:

[0046] In a preferred embodiment of the present invention, it is assumed that there are N spectral curves, and the reflectance value of each spectral curve in different wavelength bands is... Here, i represents the index of the spectral curve, and j represents the index of the band. For each band j, it is determined whether it is abnormal. Each band on each spectral curve is determined to be abnormal. Finally, if the proportion of abnormal bands on a certain curve exceeds 15%, it is determined to be an abnormal sample and removed to ensure that the data source for subsequent similarity analysis is stable and reliable.

[0047] 3) Finally, the spectral absorption features of outliers are enhanced by spectral de-envelope enhancement. By extracting the envelope features of the spectral curve, the depth, width and overall shape of the absorption band are highlighted, and the interference of background noise on the recognition results is reduced.

[0048] Specifically, in S2, based on the preprocessed spectral data, the main features of the VNIR-SWIR bands are extracted in depth. Based on the extracted hyperspectral band features, a ground cover spectral library is constructed, which serves as a standard library for rapid identification of molybdenite main minerals under different environments. The specific method is as follows: 1) In the feature analysis, a multi-scale sliding window extremum search algorithm is designed first. A dynamic window w is used to scan the data point by point, and the window width w changes with the local signal-to-noise ratio SNR, so that the key absorption features of hyperspectral data can be extracted quickly, and the preliminary feature band information and candidate results are provided for the subsequent steps.

[0049] 2) In order to accurately depict the absorption band boundary of the candidate band, especially the rate mutation feature at the edge of the absorption band. The first derivative and the second derivative are introduced to jointly correct the boundary, where corresponds to the absorption peak center; the envelope line is extracted, and the effective area of the absorption band width is defined as , in which the second derivative extremum point is used to determine the absorption band width (FWHM).

[0050]

[0051]

[0052]

[0053] wherein, represents the wavelength, represents the wavelength at the absorption peak center, i.e. the wavelength at ; R( ) represents the reflectivity at the wavelength ; represents the first wavelength, u represents the first wavelength, u+ represents the first wavelength, u- represents the wavelength step, i.e. the discrete interval in the derivative calculation, represents the empirical threshold value for controlling the sensitivity of the absorption band boundary; FWHM represents the absorption band width, which is used for mineral feature analysis; represents the wavelength on the right side of the feature band, represents the wavelength on the left side of the feature band, represents the absorption band determination condition.

[0054] Then, the absorption depth dynamic mapping is performed to quantify the absorption band shape, and the normalized absorption depth matrix is defined.

[0055]

[0056] wherein, represents the wavelength​​ the absorption band depth matrix, exp() represents the exponential function, a is the depth enhancement coefficient, SNR( ) represents the local signal-to-noise ratio at the wavelength ; wherein is the depth enhancement coefficient, the depth coefficient of the high signal-to-noise ratio band is increased, and the depth coefficient of the low signal-to-noise ratio band is attenuated.

[0057] Finally, the above candidate band multi-angle feature is encoded into a unified vector as a standard reference spectrum input vector of the fast identification model, and the feature composition is as follows:

[0058] wherein, is the peak position, FWHM is the absorption band width, is the normalized absorption band depth matrix, is the derivative curvature feature.

[0059] Through the peak detection, derivative-envelope joint analysis method, the spectral curve features of the main mineral of molybdenite, such as peak value, wavelength position, absorption band, etc. are systematically extracted. These methods complement each other, wherein the peak detection and derivative method accurately locate the position and boundary of the absorption band, and the spectral envelope analysis verifies the depth and morphology of the absorption band, and finally realizes the combination of local features and global information to obtain the ground object spectral information library of the main mineral of molybdenite.

[0060] Specifically, in S3, based on a dynamic weighting strategy, potential targets in the ground object spectral library are screened through the Euclidean distance, the cosine similarity is introduced to measure the robustness of spectral curve shape matching, and the spectral information divergence is combined to analyze the spectral information entropy difference to construct a fast identification model, which is as follows: A dynamic weighting fusion strategy is introduced to quickly screen potential targets in the ground object spectral library through the Euclidean distance ED. It is assumed that two spectral curves are and , the reflectivity values of each band are and , wherein n represents the total number of spectral bands, and the Euclidean distance between two spectral bands is calculated according to the following formula:

[0061] The cosine similarity CS is introduced to measure the similarity between the spectral curve of the potential target and the spectral curve of the ore to be measured, and the robustness of the spectral curve shape matching is enhanced. The cosine similarity calculation formula of two spectral curves and is as follows:

[0062] wherein denotes the dot product of two band reflectance spectral vectors; the denominator part denotes the dot product of two band reflectance spectral vectors. The cosine similarity ranges from [-1, 1], the closer to 1, the more similar the two curves, and vice versa.

[0063] In combination with the spectral information divergence SID, the spectral information entropy difference is deeply analyzed, a fast identification model is constructed, and a comprehensive similarity score is calculated by the fast identification model. The spectral information divergence is a method based on information theory to compare the similarity between two spectra by quantifying the difference between the two spectral data. A and B correspond to the probability vectors of the two spectra and where , From information theory, the information of A and B can be obtained:

[0064] where, denotes the information divergence of the spectral curve X , denotes the information divergence of the spectral curve Y .

[0065] Based on the above formula, the spectral curve X The relative entropy of the spectral curve Y :

[0066] The relative entropy of the spectral curve Y The relative entropy of the spectral curve X :

[0067] The spectral information divergence of the spectral curve X and the spectral curve Y :

[0068] where, denotes the vector value of one band reflectance of the spectral curve X , denotes the vector value of one band reflectance of the spectral curve Y , denotes the relative entropy of the spectral curve Y The relative entropy of the spectral curve X where one band, denotes the relative entropy of the spectral curve X The relative entropy of the spectral curve YThe relative entropy of one of the bands; the smaller the SID value, the more similar the two spectral curves are.

[0069] As a preferred embodiment of the present application, the initial weight is set as: (0.4), (0.5), (0.1) and satisfies The specific weight dynamic adjustment rules are as follows: 1. When the amplitude difference is significant ( ), wherein is the average value of ED, is the standard deviation of ED: increase to 0.5, decrease and ; 2. When the shape difference is significant ( ), wherein is the average value of CS, is the standard deviation of CS: increase to 0.6, decrease and ; 3. When the ore distribution is complex ( the dispersion coefficient CV>0.3): increase to 0.2, decrease and .

[0070] After determining the weight of each index, normalize each index and establish the calculation system of Final Score (FS): Euclidean distance ED normalization:

[0071] wherein, represents the Euclidean distance normalization result, represents the Euclidean distance to be normalized, represents the minimum Euclidean distance in the potential target, represents the maximum Euclidean distance in the potential target.

[0072] Spectral information divergence SID normalization:

[0073] wherein, represents the spectral information divergence normalization result, represents the spectral information divergence to be normalized, represents the minimum spectral information divergence in the potential target, represents the maximum spectral information divergence in the potential target.

[0074] The judgment criteria of the Euclidean distance ED and the spectral information divergence SID are that the smaller the value is, the higher the similarity of the spectral curve is, and therefore the final similarity score FS The calculation formula of the final score is as follows:

[0075] wherein FS represents the similarity score of the to-be-detected molybdenite feature and the features in the ground object spectral library, is a dynamically allocated weight, the sum of which is 1, and is dynamically adjusted according to the amplitude, shape, and complexity of the distribution of the spectral curve; represents the Euclidean distance normalization result, represents the cosine similarity, represents the spectral information divergence normalization result, and the smaller the values of the Euclidean distance normalization result and the spectral information normalization result are, the higher the similarity of the spectral curve is.

[0076] According to the FS, the candidate data are sorted, a confidence threshold of 0.8 is combined, the most matched molybdenite concentrate main mineral type is output, and the low confidence result is further analyzed.

[0077] Specifically, in S4, after the spectral data of the to-be-detected molybdenite are preprocessed, the similarity of the to-be-detected molybdenite and the features in the ground object spectral library is calculated through the constructed fast recognition model to obtain a similarity result, and the category of the to-be-detected molybdenite is recognized based on the similarity result.

[0078] The molybdenite ore after once jaw breaking in a certain mining area is subjected to data collection by using the ground object spectrometer, that is, the full-waveband light source is placed at a position 30 centimeters above the to-be-detected ore. After the collected spectral curve data are preprocessed, the spectral curve data are matched with the features extracted from the ground object spectral library, the similarity score is calculated through the fast recognition model, the reaction time and the recognition result are recorded, and the identification is performed multiple times to ensure the accuracy and stability of the fast recognition model.

[0079] Embodiment 2 A molybdenite main mineral hyperspectral multi-feature fusion recognition system comprises: a data collection and preprocessing module, configured to collect multi-angle spectral data of molybdenite with different grades, and to preprocess the multi-angle spectral data; a ground object spectral library construction module, configured to extract hyperspectral band features of the preprocessed spectral data, and to construct a ground object spectral library based on the extracted hyperspectral band features; a fast recognition model construction module, configured to filter potential targets in the ground object spectral library through a Euclidean distance based on a dynamic weighting strategy, to introduce a cosine similarity to measure the robustness of shape matching of spectral curves, and to construct a fast recognition model in combination with a spectral information divergence to analyze spectral information entropy difference. The ore recognition module is used for calculating the similarity of the molybdenite to be measured with the features in the ground object spectrum library through the constructed rapid recognition model after the spectrum data of the molybdenite to be measured is preprocessed, obtaining a similarity result, and recognizing the category of the molybdenite to be measured based on the similarity result.

[0080] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, but not to limit it, although the present application has been described in detail with reference to the above examples, those skilled in the art should understand that: the specific embodiments of the present application can still be modified or replaced by the equivalent, without departing from the spirit and scope of the present application, any modification or equivalent replacement, which should be covered within the protection scope of the claims of the present application.

Claims

1. A method for identifying molybdenite main mineral high-spectrum multi-feature fusion, characterized in that, The method comprises the following steps: Collecting multi-angle spectral data of molybdenite with different grades; Pretreating the multi-angle spectral data; Extracting hyperspectral band features of the pretreated spectral data, and constructing a ground object spectral library based on the extracted hyperspectral band features; Based on a dynamic weighting strategy, potential targets in the ground object spectral library are screened through the Euclidean distance, the robustness of spectral curve shape matching is measured by introducing the cosine similarity, and a fast recognition model is constructed by combining spectral information divergence to analyze spectral information entropy difference; 2. The molybdenite main mineral hyperspectral multi-feature fusion identification method according to claim 1, characterized in that, After pretreating the spectral data of the molybdenite to be tested, the similarity between the molybdenite to be tested and the features in the ground object spectral library is calculated through the constructed fast recognition model to obtain a similarity result, and the category of the molybdenite to be tested is recognized based on the similarity result. In the step of collecting multi-angle spectral data of molybdenite with different grades and pretreating the multi-angle spectral data, the method for pretreating the multi-angle spectral data is as follows: Noise interference in the spectral data after removing outliers is eliminated through an adaptive sliding window fitting strategy to obtain spectral absorption features; Outliers in the spectral absorption features are removed through a multi-scale band coupling IQR algorithm to obtain spectral absorption features after removing outliers; 3. The molybdenite main mineral hyperspectral multi-feature fusion identification method according to claim 2, characterized in that, The spectral absorption features after removing outliers are enhanced by spectral de- enveloping. In the step of eliminating noise interference in the spectral data after removing outliers through an adaptive sliding window fitting strategy to obtain spectral absorption features, the specific method is as follows: The length of the adaptive sliding window and the polynomial order are determined; 4. The molybdenite main mineral hyperspectral multi-feature fusion identification method according to claim 2, characterized in that, The data points in the window are least square fitted to generate smooth values and obtain spectral absorption features. In the step of removing outliers in the spectral data through a multi-scale band coupling IQR algorithm to obtain spectral absorption features after removing outliers, the specific method is as follows: The hyperspectral band J in the spectral curve is divided into a plurality of overlapping subintervals, and each overlapping subinterval is denoted as band j; For each spectral curve, the reflectance distribution quartiles and IQR of each band j are calculated; A standard IQR is introduced for dynamic threshold calculation to define the upper and lower limits of outliers; Whether the IQR value of each band j is between the upper and lower limits of outliers is judged one by one, and if yes, it is determined as normal; 5. The molybdenite main mineral hyperspectral multi-feature fusion identification method according to claim 4, characterized in that, If not, it is determined as an outlier. wherein, denotes the standard IQR of waveband j, denotes the number of wavebands j, denotes the IQR of waveband j; denotes the upper outlier limit of waveband j, denotes the third quartile of waveband j; denotes the lower outlier limit of waveband j, denotes the first quartile of waveband j; denotes the average IQR of waveband j; is a noise adaptive adjustment factor, is the noise standard deviation of sub-interval k.

6. The molybdenite main mineral hyperspectral multi-feature fusion identification method according to claim 1, characterized in that, The specific formula for introducing a standard IQR for dynamic threshold calculation to define the upper and lower limits of outliers is as follows: In the step of extracting hyperspectral band features of the pretreated spectral data and constructing a ground object spectral library based on the extracted hyperspectral band features, the specific method is as follows: A multi-scale sliding window extreme value search algorithm is used to extract absorption features of the hyperspectral data to generate candidate bands; The absorption band effective region of the candidate bands is determined, the wavelength peak position, absorption band width and absorption band depth matrix are calculated, and the candidate band features are obtained; 7. The molybdenite main mineral hyperspectral multi-feature fusion identification method according to claim 6, characterized in that, The ground object spectral library is constructed based on the calculated candidate band features. Introducing the first derivative With the second derivative Joint envelope The effective area of the absorption band is defined as Within the effective area, the width of the absorption band is determined by the extreme points of the second derivative , which is expressed as follows: wherein, denotes the wavelength, corresponds to the absorption peak center, denotes the wavelength of the absorption peak center; R( ) denotes the reflectivity at the wavelength , denotes the first u wavelength, denotes the first u+ wavelength, denotes the first u- wavelength, denotes the wavelength step, i.e. the discrete interval in the derivative calculation, denotes an empirical threshold value for controlling the sensitivity of the absorption band boundary; FWHM denotes the absorption band width for mineral feature analysis; denotes the wavelength right of the characteristic band, denotes the wavelength left of the characteristic band, denotes the absorption band determination condition; The method for determining the absorption band effective region of the candidate bands, calculating the wavelength peak position, absorption band width and absorption band depth matrix, and obtaining the candidate band features is as follows: The absorption band depth is dynamically mapped, and the absorption band depth matrix is calculated, and the specific formula is as follows: wherein represents a matrix of absorption band depth at wavelengths exp() represents the exponential function, a is a depth enhancement coefficient, SNR( ) represents the local signal-to-noise ratio at wavelengths ​ The matrix of wavelength peak position , absorption bandwidth and absorption band depth is unified as a vector, which is the characteristic value of the construction of the ground object spectrum library, and the specific formula is as follows: wherein denotes the derivative curvature feature.

8. The molybdenite main mineral hyperspectral multi-feature fusion identification method according to claim 1, characterized in that, The dynamic weighting strategy is used to filter potential targets in the ground object spectrum library through the Euclidean distance, the robustness of cosine similarity is introduced to measure the shape matching of the spectrum curve, and the spectrum information entropy difference is analyzed by combining the spectrum information divergence to construct a fast recognition model. wherein, FS represents the similarity score of the characteristic of the molybdenite to be detected and the characteristic in the ground object spectrum library, is a dynamically allocated weight, and the sum is 1; represents the Euclidean distance normalization result, represents the cosine similarity, represents the spectral information divergence normalization result.

9. The molybdenite main mineral hyperspectral multi-feature fusion recognition method according to claim 8, characterized in that, The smaller the values of the Euclidean distance normalization result and the spectrum information normalization result, the higher the similarity of the spectrum curve.

10. A molybdenite main mineral hyperspectral multi-feature fusion identification system, characterized in that, Comprise: A data acquisition and preprocessing module is configured to acquire multi-angle spectral data of molybdenite with different grades, and to preprocess the multi-angle spectral data; A ground object spectrum library construction module is configured to extract hyperspectral band features of the preprocessed spectral data, and to construct a ground object spectrum library based on the extracted hyperspectral band features; A fast recognition model construction module is configured to filter potential targets in the ground object spectrum library through the Euclidean distance based on a dynamic weighting strategy, to introduce the robustness of cosine similarity to measure the shape matching of the spectrum curve, and to construct a fast recognition model by analyzing the spectrum information entropy difference in combination with the spectrum information divergence; An ore recognition module is configured to calculate the similarity between the spectral data of the molybdenite to be tested and the features in the ground object spectrum library after preprocessing the spectral data of the molybdenite to be tested by using the constructed fast recognition model, to obtain a similarity result, and to recognize the category of the molybdenite to be tested based on the similarity result.

Images