Method for constructing mineral deposit prediction index system based on knowledge graph

By screening and selecting neurons in the network clustering model and adjusting distance, the problem of low classification accuracy of deposit text data is solved, the prediction effect of the deposit prediction index system is improved, and the structure and mining effect of the knowledge graph are enhanced.

CN118427349BActive Publication Date: 2024-11-19CHINA UNIV OF GEOSCIENCES (BEIJING)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410597779.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-14
Publication Date
2024-11-19
Estimated Expiration
2044-05-14

AI Technical Summary

Technical Problem

In the prior art, the accuracy of the classification results of ore deposit text data is low, resulting in poor prediction effect of the deposit prediction index system. This is mainly due to the random location and quantity distribution of important characteristic words in the text vector, resulting in the classification results being unfocused or misclassified.

Method used

By obtaining the deposit text vector and filtering out the selection neurons in the preset network clustering model, combining the universality and special characteristics of vector elements, adjusting the distance between vectors, obtaining the correction distance to screen the winning neurons, and building knowledge-based The deposit prediction index system of the map.

Benefits of technology

It improves the classification effect of deposit text vectors, enhances the structure and mining effect of the knowledge graph, and improves the accuracy of deposit prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118427349B_ABST
    Figure CN118427349B_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of mineral deposit data mining, and in particular to a method for constructing a mineral deposit prediction index system based on a knowledge graph. The present invention obtains all mineral deposit text vectors in mineral deposit data; inputs the mineral deposit text vectors into a preset network clustering model for model training; obtains the selected neurons and all special vector dimensions of each mineral deposit text vector, and then obtains the degree of belonging of the mineral deposit text vector relative to each selected neuron; adjusts the distance between the mineral deposit text vector and the weight vector according to the degree of belonging, obtains the winning neuron of each input mineral deposit text vector, obtains a trained preset network clustering model, and then constructs a mineral deposit prediction index system. The present invention obtains the special dimensions and non-special dimensions of the mineral deposit text vector by analysis, and then evaluates the degree of belonging of each mineral deposit text vector relative to the corresponding selected neuron, thereby improving the classification effect of the mineral deposit text vector, thereby improving the mining effect and prediction effect of the mineral deposit data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of mineral deposit data mining, and in particular to a method for constructing a mineral deposit prediction index system based on a knowledge graph. Background Art

[0002] The mineral deposit prediction index system is an index system that comprehensively applies multi-source data such as geology, geophysics, and remote sensing technology to analyze the genesis and type of mineral deposits, evaluate the potential of mineral deposits, and predict the existence of mineral deposits. By establishing associations between mineral deposit data, the knowledge graph can integrate various mineral deposit data into a structured mineral deposit knowledge network, which is convenient for analyzing the formation mechanism and characteristics of mineral deposits, so as to build a mineral deposit prediction index system, mine relevant mineral deposit data, and improve prediction results.

[0003] Since mineral deposit survey data are usually in the form of text reports, natural language processing technology is usually used to perform entity recognition and relationship extraction on text data. The concept hierarchy is constructed through expert-designed knowledge system, and then entity extraction is performed. This method has strong manual dependence and is difficult to determine the granularity and boundaries. This may lead to inefficient structural construction of the knowledge graph, which may subsequently affect the prediction efficiency of the mineral deposit prediction index system. Therefore, before constructing a knowledge graph based on the collected mineral deposit data, it is necessary to perform certain classification processing on the mineral deposit survey data, so as to organize similar entities or relationships together to form a clearer and more reasonable knowledge graph structure, while improving construction efficiency.

[0004] In the prior art, a neural network clustering algorithm is usually used to classify the mineral deposit text report data, thereby constructing a knowledge graph. When using a neural network clustering algorithm to classify mineral deposit text data, the distance between the text vector corresponding to the mineral deposit text and the neuron is usually directly evaluated to obtain the winning neuron corresponding to the text vector for classification. However, in this process, the position distribution and quantity distribution of important feature words in the text vector are random. Only based on the distance evaluation between the text vector and the neuron, the final text classification result may be unfocused, or the corresponding text may be mistakenly classified into the same category as the general text due to the low frequency of occurrence of important feature words, thereby affecting the accuracy of the classification result, and further affecting the prediction effect of the mineral deposit prediction index system. Summary of the invention

[0005] In order to solve the technical problem that the accuracy of classification results in the prior art is low and thus leads to poor ore deposit prediction effect, the purpose of the present invention is to provide a method for constructing an ore deposit prediction index system based on a knowledge graph. The technical solution adopted is as follows:

[0006] Get all the mineral deposit text vectors in the mineral deposit data;

[0007] Input the deposit text vector into a preset network clustering model for model training; select neurons for each deposit text vector from all neurons based on the similarity between each input deposit text vector and all activation vectors or weight vectors of each neuron in the preset network clustering model; obtain all special vector dimensions of each selected neuron based on the distribution of vector elements in all activation vectors or weight vectors of each selected neuron;

[0008] According to the difference between the vector elements in each input ore deposit text vector and the vector elements in the special vector dimension and the non-special vector dimension in all activation vectors of each selected neuron, the degree of belonging of the ore deposit text vector to each selected neuron is obtained; the distance between the ore deposit text vector and the weight vector corresponding to the selected neuron is adjusted according to the degree of belonging to obtain a corrected distance; according to the corrected distance, the winning neuron of each input ore deposit text vector is screened out from all corresponding selected neurons to obtain a trained preset network clustering model; all clusters are obtained according to the trained preset network clustering model;

[0009] A knowledge graph-based mineral deposit prediction indicator system is constructed based on all clusters.

[0010] Furthermore, the method for obtaining the mineral deposit text vector includes:

[0011] Each text sentence in the mineral deposit data is processed by Jieba word segmentation, and all the word segments in each text sentence are assigned values ​​as vector elements through the TF-IDF method to construct the mineral deposit text vector corresponding to each text sentence.

[0012] Furthermore, the method for obtaining the selected neurons includes:

[0013] According to the similarity between each input ore deposit text vector and all activation vectors or weight vectors of each neuron in the preset network clustering model, the preliminary matching degree between each ore deposit text vector and each neuron is obtained;

[0014] The input mineral deposit text vector and the preliminary matching degree of each neuron are sequentially sorted to construct a preliminary matching degree sequence; a differential sequence of the preliminary matching degree series is obtained, and the preliminary matching degree sequence is segmented based on the maximum differential value in the differential sequence; the mean of the preliminary matching degree in each segment is obtained, and all neurons corresponding to the preliminary matching degree in the segment with the largest mean are used as all selected neurons corresponding to the mineral deposit text vector.

[0015] Furthermore, the method for obtaining the preliminary matching degree includes:

[0016] The mean of the vector elements in the input mineral deposit text vector is taken as the first element mean; the mean of the vector elements in all activation vectors of each neuron is taken as the second element mean; the difference between the first element mean and the second element mean is normalized by negative correlation mapping to obtain the preliminary matching degree between each mineral deposit text vector and each neuron; wherein, when a neuron has no corresponding activation vector, the preliminary matching degree between the input mineral deposit text vector and the corresponding neuron is zero.

[0017] Furthermore, the method for obtaining the special vector dimension includes:

[0018] Taking a vector element on any vector dimension as a target element, the kurtosis of all the target elements corresponding to all the activation vectors of each selected neuron is obtained. When the kurtosis is greater than a preset threshold, the target element is taken as a special vector element, and the vector dimension corresponding to the special vector element is a special vector dimension.

[0019] Furthermore, the method for obtaining the degree of belonging includes:

[0020] Obtain the universal attribution parameter of the deposit text vector corresponding to the selected neuron according to the element difference between the vector elements of the non-special vector dimension in each input deposit text vector except the special vector dimension and the vector elements of the non-special vector dimension in all activation vectors of each selected neuron;

[0021] According to the vector elements of the special vector dimension in each input ore deposit text vector and the element distribution between the vector elements of the non-special vector dimension in all activation vectors of each selected neuron, obtain the special attribution parameter of the ore deposit text vector corresponding to the selected neuron;

[0022] The general attribution parameter and the special attribution parameter are multiplied, combined and normalized to obtain the attribution degree of the deposit text vector corresponding to the selected neuron.

[0023] Furthermore, the calculation formula of the general attribution parameter includes:

[0024] Among them, i is the serial number of the input mineral deposit text vector; k is the serial number of the selected neuron; k i is the serial number of the kth selected neuron corresponding to the ith input mineral deposit text vector; The kth corresponding mineral deposit text vector of the i-th input i The general attribution parameter of the selected neurons; sigmoid{} is the S-type function; x is the dimension number of the non-special vector dimension; X is the kth corresponding to the i-th input deposit text vectori The total number of dimensions of all non-special vectors in all activation vectors of the selected neurons; b i,x is the vector element on the xth non-special vector dimension in the ith input ore deposit text vector; n is the kth vector corresponding to the ith input ore deposit text vector. i The number of the activation vector of the selected neuron; N is the kth corresponding to the ith input deposit text vector. i The total number of activation vectors of the selected neurons; The kth ore deposit text vector corresponding to the i-th input i The vector element in the xth non-special vector dimension in the nth activation vector of the selected neuron.

[0025] Furthermore, the method for obtaining the special attribution parameter includes:

[0026] Obtaining reference vector elements of each of the special vector dimensions according to element distribution of vector elements of all activation vectors of each of the selected neurons corresponding to the deposit text vector;

[0027] The special attribution parameter is obtained according to the calculation formula of the special attribution parameter; the calculation formula of the special attribution parameter is:

[0028] Where i is the serial number of the input mineral deposit text vector; k is the serial number of the selected neuron; k i is the serial number of the kth selected neuron corresponding to the i-th input mineral deposit text vector; The kth corresponding to the ith input mineral deposit text vector i special attribution parameters of the selected neurons; exp() is an exponential function with the natural constant e as the base; y is the dimension number of the special vector dimension; Y is the kth vector corresponding to the ith input deposit text vector. i The total number of dimensions of all special vectors in all activation vectors of the selected neurons; b i,y is the vector element of the yth special vector dimension in the ith input mineral deposit text vector; c i,y The reference vector element of the yth special vector dimension of the i-th input deposit text vector.

[0029] Furthermore, the method for obtaining the reference vector element includes:

[0030] Obtain a frequency histogram of the vector elements of all activation vectors of each of the selected neurons corresponding to the mineral deposit text vector; in the frequency histogram, use the vector elements corresponding to the non-special vector dimensions corresponding to the peak values ​​closest to each of the special vector dimensions as reference vector elements for each of the special vector dimensions.

[0031] Furthermore, the method for obtaining the corrected distance includes:

[0032] The deposit text vector is negatively correlated with the attribution degree of each selected neuron and normalized, and the normalized value is multiplied by the Euclidean distance between the deposit text vector and the weight vector of each selected neuron to obtain the corrected distance.

[0033] The present invention has the following beneficial effects:

[0034] The present invention first obtains all deposit text vectors in deposit data; inputs the deposit text vectors into a preset network clustering model for model training; selects neurons for each deposit text vector from all neurons according to the similarity between each input deposit text vector and all activation vectors or weight vectors of each neuron in the preset network clustering model, so as to facilitate the subsequent selection of winning neurons from all selected neurons in combination with the universality and particularity characteristics of the vector elements; obtains all special vector dimensions of each selected neuron according to the distribution of vector elements in all activation vectors or weight vectors of each selected neuron, and analyzes the vector approximation characteristics between the deposit text vector and the activation vector or weight vector of the selected neuron on the special vector dimension, so as to accurately evaluate the degree of attribution, which preliminarily reflects the probability that each selected neuron is a winner of the deposit text vector. The possibility of winning neurons, the greater the degree of belonging, the greater the possibility of winning neurons, so as to adjust the distance between vectors in the subsequent acquisition of winning neurons; according to the difference between the vector elements in each input ore deposit text vector and the special vector dimension and non-special vector dimension in all activation vectors of each selected neuron, the degree of belonging of the ore deposit text vector relative to each selected neuron is obtained; according to the degree of belonging, the distance between the ore deposit text vector and the weight vector of the corresponding selected neuron is adjusted to obtain the corrected distance, thereby avoiding the situation that misclassification is prone to occur only based on the distance evaluation between vectors; according to the corrected distance, the winning neuron of each input ore deposit text vector is accurately screened out from all corresponding selected neurons to obtain a trained preset network clustering model; according to the trained preset network clustering model, all clusters are obtained, and then a ore deposit prediction index system is constructed. The present invention obtains the special dimension and non-special dimension of the ore deposit text vector by analyzing, and then evaluates the degree of belonging of each ore deposit text vector relative to the corresponding selected neuron in different dimensions, thereby improving the classification effect of the ore deposit text vector, thereby improving the structurality of the knowledge map, improving the mining effect of related ore deposit data, and improving the ore deposit prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are 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 these drawings without paying creative work.

[0036] Figure 1 A flowchart of a method for constructing a mineral deposit prediction index system based on a knowledge graph provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0037] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the method for constructing a mineral deposit prediction index system based on a knowledge graph proposed by the present invention, its specific implementation method, structure, characteristics and effects, in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0038] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0039] The following is a detailed description of a method for constructing a mineral deposit prediction index system based on a knowledge graph provided by the present invention in conjunction with the accompanying drawings.

[0040] See also Figure 1 , which shows a flow chart of a method for constructing a mineral deposit prediction index system based on a knowledge graph provided by an embodiment of the present invention, specifically comprising:

[0041] Step S1, obtaining all mineral deposit text vectors in mineral deposit data.

[0042] In one embodiment of the present invention, a mineral deposit text report including the mineral deposit to be predicted, i.e., mineral deposit data, is obtained through geological survey, geophysical prospecting, remote sensing, chemical analysis and other technologies, and then each text sentence in the mineral deposit data is processed by Jieba word segmentation, and all the words in each text sentence are assigned values ​​as vector elements through the TF-IDF method to construct a mineral deposit text vector corresponding to each text sentence.

[0043] It should be noted that the TF-IDF method analyzes the importance of each word by analyzing the word frequency and inverse document frequency of the word in the mineral deposit data. The lower the TF-IDF value, the more common and general the corresponding word is. Conversely, the rarer and more special the word is, the more likely it is to be an important feature word. The implementer may also map the TF-IDF value to a range of 1-100, and then round it up or down to facilitate subsequent analysis of the mineral deposit text vector. Jieba word segmentation processing and TF-IDF method are both prior arts well known to those skilled in the art and will not be described in detail here. In other embodiments of the present invention, the implementer may also adopt other word segmentation methods, and assign values ​​to the words by constructing a preset word segmentation dictionary through manual annotation, but it is necessary to ensure that the more common and general the word is, the lower the corresponding dictionary value is, so as to construct the mineral deposit text vector.

[0044] Step S2, inputting the deposit text vector into the preset network clustering model for model training; selecting neurons for each deposit text vector from all neurons based on the similarity between each input deposit text vector and all activation vectors or weight vectors of each neuron in the preset network clustering model; obtaining all special vector dimensions of each selected neuron based on the distribution of vector elements in all activation vectors or weight vectors of each selected neuron.

[0045] In the prior art, the deposit text vector is usually input into the neural network clustering model for training, so as to obtain the winning neuron corresponding to each input deposit text vector, and further self-organize and update the weight vector between the winning neuron and its neighboring neurons, so that similar deposit text vectors are close to each other in the network, thereby obtaining the final clustering result. Therefore, the embodiment of the present invention inputs the deposit text vector into the preset network clustering model for model training, and then adjusts the distance between the input deposit text vector and the neuron in combination with the universality and specificity analysis of the text words during each training process, so as to improve the accuracy of obtaining the winning neuron of the input deposit text vector.

[0046] In one embodiment of the present invention, the preset network clustering model is a square topology structure, that is, a grid-like topology structure, each node in the topology structure is a neuron, and there are a total of 35×35 neurons in the preset network clustering model. The weights of the neurons are then randomly initialized, that is, a weight vector is randomly assigned to each neuron in the preset network clustering model, and then the SOM network clustering algorithm is specifically used for training; randomly initialized weights and SOM network clustering algorithms are already prior arts well known to those skilled in the art and will not be described in detail here; implementers can also construct preset network clustering models of other topological structures according to specific implementation scenarios.

[0047] It should be noted that the vector dimension of the weight vector, that is, the number of vector elements in the weight vector, needs to be consistent with the vector dimension of the input ore deposit text vector in order to evaluate the distance between vectors; in one embodiment of the present invention, taking into account the inconsistent length of each text sentence, that is, the vector dimensions between different ore deposit text vectors may be different, the vector dimension of the weight vector is first made consistent with the vector dimension of the ore deposit text vector input for the first time, and in the subsequent input process, the vector dimensions between the input ore deposit text vector and the weight vector are continuously compared to adjust them to be consistent.

[0048] In one embodiment of the present invention, the weight vector is DTW matched with the mineral deposit text vector input each time. For example, the input text vector is (a, b, c, e, f, g) and the weight vector is (A, C, F). There may be a one-to-many or one-to-one matching relationship, where the matching relationship may be a, b matches A, c matches C, and e, f, g matches F. The vector dimension of the input text vector corresponding to the largest vector dimension of the two is taken as the target, and the nearest neighbor interpolation algorithm is used to interpolate the weight vector, so that the interpolated weight vector is (A, A, C, F, F, F), which is consistent with the vector dimension of the input text vector.

[0049] It should be noted that the text vector elements A, B, C...a, b, c... in the above examples are all numerical values, and symbols are used as examples here to facilitate distinction and understanding. In other embodiments of the present invention, the implementer may also insert the mean of two vector elements between every two vector elements in the weight vector (A, C, F) to make it consistent with the vector dimension of the deposit text vector; the interpolation method may also be adjusted by comparing the difference in DTW distance before and after interpolation, and interpolation may be performed while minimizing the impact on the DTW distance between the two vectors; the DTW algorithm and the nearest neighbor interpolation algorithm are both prior arts well known to those skilled in the art and will not be described in detail here.

[0050] In another embodiment of the present invention, after obtaining the ore deposit text vector, the implementer obtains the maximum vector dimension, and interpolates and supplements all the ore deposit text vectors through DTW matching and nearest neighbor interpolation algorithm, so that the vector dimensions of all input text vectors are consistent with the maximum vector dimension, avoiding subsequent continuous adjustment of matching weight vectors and ore deposit text vectors.

[0051] It should be noted that directly matching and aligning all deposit text vectors with the vector dimension can avoid the continuous adjustment of the weight vector to improve the training efficiency of the preset network clustering model, but it may cause semantic distortion of some deposit text vectors, thereby resulting in a decrease in training performance; and continuously adjusting the vector dimensions of the matched and aligned input deposit text vectors and weight vectors during each input process can alleviate semantic distortion to a certain extent, but its training efficiency decreases; implementers can choose the corresponding processing method according to specific needs.

[0052] It should be noted that, in one embodiment of the present invention, taking into account the situation that some neurons in the preset network clustering model have no corresponding activation vectors during the input of the ore deposit text vector, a preset number of ore deposit text vectors are first randomly selected and input into the preset network clustering model, and the winning neuron of the ore deposit text vector is directly determined according to the Euclidean distance between the ore deposit text vector and the weight vector, and the remaining ore deposit text vectors are then screened and selected by neurons to re-evaluate the degree of attribution to obtain the winning neuron; the preset number is 5% of the total number of ore deposit text vectors, and the implementer can set the preset number by himself.

[0053] Taking into account that during the training process of the neural network model, the competition mechanism between neurons will cause the neuron with the highest similarity to the input deposit text vector to be activated as the winning neuron, and the input deposit text vector is the activation vector of the winning neuron; and considering that the neurons correspond to corresponding activation vectors, that is, the deposit text vector input during the historical training process obtains the corresponding winning neuron, then in the subsequent training process, the similarity between the input deposit text vector and the activation vector corresponding to each neuron can be compared, so that similar deposit text vectors are easier to be classified into one category; therefore, the embodiment of the present invention selects the selected neuron for each deposit text vector from all neurons based on the differences between all activation vectors or weight vectors of each neuron in the preset network clustering model according to each input deposit text vector, so as to facilitate the subsequent selection of the winning neuron from all selected neurons in combination with the universal characteristics of the vector elements.

[0054] It should be noted that when each input ore deposit text vector has only one selection neuron, there is no need to perform subsequent steps, and the selection neuron is directly used as the winning neuron of the input ore deposit text vector.

[0055] Preferably, in one embodiment of the present invention, considering that the selection neuron should have a high degree of similarity matching with the corresponding input ore deposit text vector, it is convenient to select the best matching neuron as the winning neuron; and considering that by sorting the similarity matching degree in descending order, it is easy to screen out relatively matching partial neurons from all neurons as selection neurons; based on this, the method for obtaining the selection neuron includes:

[0056] According to the similarity between each inputted mineral deposit text vector and all activation vectors or weight vectors of each neuron in the preset network clustering model, the preliminary matching degree between each mineral deposit text vector and each neuron is obtained;

[0057] The input mineral deposit text vector and the preliminary matching degree of each neuron are sorted in sequence to construct a preliminary matching degree sequence; the difference sequence of the preliminary matching degree series is obtained, and the preliminary matching degree sequence is segmented according to the maximum difference value in the difference sequence; the mean of the preliminary matching degree in each segment is obtained, and all neurons corresponding to the preliminary matching degree in the segment with the largest mean are used as all selected neurons corresponding to the mineral deposit text vector.

[0058] In one embodiment of the present invention, considering that the mean difference of the vector elements between the input ore deposit text vector and all activation vectors of each neuron can reflect the similarity between the vectors to a certain extent, the smaller the mean difference, the higher the similarity; based on this, the method for obtaining the preliminary matching degree includes:

[0059] The mean of the vector elements in the input mineral deposit text vector is taken as the first element mean; the mean of the vector elements in all activation vectors of each neuron is taken as the second element mean; the difference between the first element mean and the second element mean is normalized by negative correlation mapping to obtain the preliminary matching degree between each mineral deposit text vector and each neuron; wherein, when a neuron has no corresponding activation vector, the preliminary matching degree between the input mineral deposit text vector and the corresponding neuron is zero.

[0060] The calculation formula for the initial matching degree is:

[0061]

[0062] Among them, s i,k is the initial matching degree between the ith input mineral deposit text vector and the kth neuron in the preset network clustering model; exp() is an exponential function with the natural constant e as the base; μ i is the mean of the first elements of all vector elements in the ith input mineral deposit text vector; j is the serial number of the activation vector of the kth neuron; c k is the total number of all activation vectors of the kth neuron; μ k,j is the element-wise mean of all vector elements in the j-th activation vector of the k-th neuron; is the mean of the second elements of all vector elements in all activation vectors of the kth neuron; || is the sign for taking the absolute value.

[0063] In the calculation formula for the preliminary matching degree, the absolute value of the difference between the mean of the first element and the mean of the second element is negatively correlated and mapped to an exponential function to adjust the logical relationship and normalize it, so that the smaller the absolute value of the difference, the greater the preliminary matching degree. The implementer may also use other basic mathematical operations or related mapping methods, which are existing technologies well known to technicians in this field and will not be elaborated here.

[0064] Since the values ​​of the corresponding vector elements of the common word segmentations contained in the mineral deposit data are roughly the same, and the characteristic word segmentations containing important information account for a small proportion in the mineral deposit text vector, the neural network clustering algorithm is prone to misjudgment of the winning neuron when comparing the distance between the input mineral deposit text vector and the neuron; for example, the input mineral deposit text vector is [2,5,7,2,6], and a certain activation vector of the neuron is [2,4,7,2,6]. For these two vectors, the direct calculation of the vector distance may have a small difference, so it is necessary to evaluate whether the second element in the vector is an important characteristic word, and then judge the degree of belonging of the input mineral deposit text vector to the neuron based on the relevant information of the important characteristic word; based on the above logic, according to the distribution of vector elements in all activation vectors of each selected neuron, all special vector dimensions of each selected neuron are obtained, so that on the special vector dimensions, the vector approximate characteristics between the mineral deposit text vector and the activation vector or weight vector of the selected neuron are analyzed, and then the winning neuron of the mineral deposit text vector is obtained.

[0065] Preferably, in one embodiment of the present invention, considering that the text sentence formats in the deposit text report are roughly similar, that is, the vector elements corresponding to the important feature words or common text words should have roughly similar position distributions in each deposit text vector; and considering that the kurtosis can help understand the distribution form of the vector elements on the same vector dimension, the larger the kurtosis, the sharper the distribution curve of the vector element, the more likely the vector element is to be concentrated near the mean, the higher its universality, and the less likely it is to be a special vector element; based on this, the method for obtaining the special vector dimension includes:

[0066] Taking the vector element on any vector dimension as the target element, the kurtosis of all target elements corresponding to all activation vectors of each selected neuron is obtained. When the kurtosis is greater than the preset threshold, the target element is taken as a special vector element, and the vector dimension corresponding to the special vector element is the special vector dimension.

[0067] In the embodiment of the present invention, the preset threshold is 0.5, and the implementer may select other values, which are not limited here; the calculation of kurtosis is an existing technology well known to those skilled in the art and will not be described in detail here.

[0068] It should be noted that there are fewer activation vectors corresponding to each neuron, which may lead to inaccurate kurtosis evaluation results. In one embodiment of the present invention, when the number of activation vectors corresponding to each neuron is less than 30, the method for obtaining special vector elements is: taking the vector element on any vector dimension as the target element, obtaining the element mean and element median of all target elements corresponding to all activation vectors of each selected neuron, when the absolute value of the difference between the element mean and the element median is greater than the preset difference threshold, the target element is used as a special vector element to obtain all special vector elements in all activation vectors of each selected neuron. The preset difference threshold is 0.2, and the implementer can choose other values, which are not limited here.

[0069] Step S3, according to the difference between the vector elements in each input ore deposit text vector and the vector elements in the special vector dimension and the non-special vector dimension in all activation vectors of each selection neuron, obtain the degree of belonging of the ore deposit text vector relative to each selection neuron; adjust the distance between the ore deposit text vector and the weight vector of the corresponding selection neuron according to the degree of belonging to obtain the corrected distance; select the winning neuron of each input ore deposit text vector from all corresponding selection neurons according to the corrected distance to obtain the trained preset network clustering model; obtain all clusters according to the trained preset network clustering model.

[0070] Considering that if the common word segmentation in the input deposit text vector, that is, the vector element on the non-special vector dimension, is greatly different from the vector element on the non-special vector dimension in the activation vector of each selection neuron, it means that the selection neuron is less likely to be the winning neuron of the input deposit text vector; on the contrary, when the element difference of the vector element on the non-special vector dimension is small, it means that the selection neuron may be the winning neuron of the input deposit text vector, but it needs to be further evaluated in combination with the element characteristics of the vector element on the special vector dimension; based on this, the embodiment of the present invention obtains the degree of belonging of the deposit text vector relative to each selection neuron according to the difference between the vector elements in each input deposit text vector and the vector elements on the special vector dimension and the non-special vector dimension in all activation vectors of each selection neuron; the degree of belonging preliminarily reflects the possibility that each selection neuron is the winning neuron of the deposit text vector. The greater the degree of belonging, the greater the possibility of being the winning neuron, so as to subsequently adjust the distance between vectors to obtain the winning neuron.

[0071] Preferably, in one embodiment of the present invention, the method for obtaining the degree of belonging includes:

[0072] Obtain the first reference coefficient of the deposit text vector corresponding to the selected neuron according to the element difference between the vector elements of the non-special vector dimension excluding the special vector dimension in each input deposit text vector and the vector elements of the non-special vector dimension in all activation vectors of each selected neuron;

[0073] According to the vector elements of the special vector dimension in each input ore deposit text vector and the element distribution between the vector elements of the non-special vector dimension in all activation vectors of each selected neuron, the special attribution parameters of the ore deposit text vector corresponding to the selected neuron are obtained;

[0074] The general attribution parameter is multiplied and combined with the special attribution parameter, and then the product is normalized to obtain the attribution degree of the deposit text vector corresponding to the selected neuron.

[0075] The calculation formula of the general attribution parameter includes:

[0076]

[0077] Where i is the serial number of the input mineral deposit text vector; k is the serial number of the selected neuron; k i is the serial number of the kth selected neuron corresponding to the i-th input mineral deposit text vector; The kth corresponding to the ith input mineral deposit text vector i The general attribution parameter of the selected neurons; sigmoid{} is an S-type function used for normalization; x is the dimension number of the non-special vector dimension; X is the kth vector corresponding to the i-th input deposit text vector. i The total number of dimensions of all non-special vectors in all activation vectors of the selected neurons; b i,x is the vector element on the xth non-special vector dimension in the ith input ore deposit text vector; n is the kth vector corresponding to the ith input ore deposit text vector. i The number of the activation vector of the selected neuron; N is the kth corresponding to the ith input deposit text vector. i The total number of activation vectors of the selected neurons; The kth ore deposit text vector corresponding to the i-th input i The vector element in the xth non-special vector dimension in the nth activation vector of the selected neuron.

[0078] In the calculation formula of the general attribution parameter, It reflects the difference between the vector elements on the input deposit text vector and the vector elements on the activation vector of a selected neuron on the xth non-special dimension. The larger the difference is, the greater the difference is between the input deposit text vector and the general text word segmentation in the activation vector corresponding to a selected neuron, and the less likely the selected neuron is to be the winning neuron of the input deposit text vector. Therefore, the differences of the vector elements on all non-special vector dimensions are summed up and normalized by sigmoid{}, and negative correlation mapping is performed by subtracting the normalized value from 1 to adjust the logic, so that the greater the difference in the vector elements, the smaller the general attribution parameter.

[0079] The methods for obtaining special attribution parameters include:

[0080] Obtain reference vector elements of each special vector dimension according to the element distribution of vector elements of all activation vectors of each selected neuron corresponding to the deposit text vector;

[0081] The special attribution parameter is obtained according to the calculation formula of the special attribution parameter; the calculation formula of the special attribution parameter is:

[0082]

[0083] Among them, i is the serial number of the input mineral deposit text vector; k is the serial number of the selected neuron; k i is the serial number of the kth selected neuron corresponding to the ith input mineral deposit text vector; The kth corresponding mineral deposit text vector of the i-th input i The degree of belonging of the selected neurons; exp() is an exponential function with the natural constant e as the base; y is the dimension number of the special vector dimension; Y is the kth vector corresponding to the i-th input deposit text vector. i The total number of dimensions of all special vectors in all activation vectors of the selected neurons; b i,y is the vector element of the yth special vector dimension in the ith input mineral deposit text vector; c i,y The reference vector element of the yth special vector dimension of the i-th input deposit text vector.

[0084] Among them, the method for obtaining the reference vector element is: obtaining the frequency histogram of the vector elements of all activation vectors of each selected neuron corresponding to the deposit text vector; in the frequency histogram, the vector element corresponding to the non-special vector dimension corresponding to the peak closest to each special vector dimension is used as the reference vector element of each special vector dimension; the reference vector element is the vector element on the non-special vector dimension that is closest to the vector element corresponding to each special dimension and appears frequently in all activation vectors of the selected neurons. The acquisition of frequency histograms and peak values ​​is a prior art well known to those skilled in the art and will not be repeated here.

[0085] In the calculation formula of special attribution parameters, b i,y Reflects the size of the vector element value on the special vector dimension in the input deposit text vector, b i,y The larger the value is, the lower the frequency of the word corresponding to the vector element is relative to other words, and the more likely it is to be a feature word containing important information; |b i,y -c i,y | is larger, indicating that the difference between the vector element value on the special vector dimension in the input ore deposit text vector and the reference vector element is greater, and the vector element on the special vector dimension in the input ore deposit text vector is more likely to be a feature word containing important information; i,y with |b i,y -c i,y |The logic is adjusted by multiplication and merging, and then negatively correlated mapping to the exponential function. The smaller the normalized value, the smaller the corresponding special attribution parameter, which means that the greater the difference between the activation vector of the selected neuron and the input deposit text vector, the less likely the selected neuron is to be the winning neuron of the input deposit text vector.

[0086] After obtaining the degree of attribution of the input deposit text vector relative to each selection neuron, the distance between the deposit text vector and the weight vector of the corresponding selection neuron can be adjusted according to the degree of attribution to obtain the corrected distance. Therefore, the winning neuron of each input deposit text vector can be screened out from all corresponding selection neurons according to the corrected distance to obtain a trained preset network clustering model.

[0087] Preferably, in one embodiment of the present invention, considering that the greater the degree of attribution, the more likely the selected neuron is to be the winning neuron of the input ore deposit text vector, and the smaller the distance between vectors should be; based on this, the method for obtaining the corrected distance includes:

[0088] The deposit text vector is negatively correlated with the degree of attribution of each selected neuron and normalized, and the normalized value is multiplied by the Euclidean distance between the deposit text vector and the weight vector of each selected neuron to obtain the corrected distance. In one embodiment of the present invention, negative correlation mapping is performed by subtracting the degree of attribution from 1, adjusting the logic and normalizing; in other embodiments of the present invention, the implementer may adopt other negative correlation mapping methods, which are all existing technologies in the field and will not be described in detail here.

[0089] In one embodiment of the present invention, when the corrected distance between the input ore deposit text vector and its corresponding selected neuron is the smallest, the selected neuron is used as the winning neuron of the input ore deposit text vector.

[0090] All deposit text vectors are input into the preset network clustering model. After each input deposit text vector obtains the corresponding winning neuron, the model training is completed; during the training process, the topological structure of the preset network clustering model will be adjusted as the weight vector is adjusted, and finally the clustering and distribution of the data will be reflected, that is, in the finally trained preset network clustering model, similar deposit text vectors are clustered in the same neuron, thereby obtaining a clustering result; in other embodiments of the present invention, the implementer may also set the number of iterative training times, that is, the number of times the deposit text vector is input. When the set number of iterative training times is reached, the model training is considered to be completed, and then the remaining deposit text vectors are input into the model for classification to obtain a clustering result; this is a prior art well known to those skilled in the art and will not be described in detail here.

[0091] Step S4, constructing a mineral deposit prediction index system based on the knowledge graph according to all clusters.

[0092] In one embodiment of the present invention, all similar deposit text vectors are divided into a cluster, and then in each cluster, the text sentence corresponding to the deposit text vector is restored to the text sentence state, and the text data is subjected to entity recognition and relationship extraction by using natural language processing technology, and the relationship between entities in the text is identified, wherein the entity refers to important characteristic keywords such as deposit, mining area, ore body, mining section name, place name, ore, lithology, stratum, structure, geological age and geological event, and the entity is taken as a node and the relationship is taken as an edge, and the entity and the relationship are constructed into the structure of the knowledge graph. The implementer can also add attribute information to the nodes and edges in the knowledge graph to enrich the content of the knowledge graph. Constructing the knowledge graph according to the clustering results can help optimize the organizational structure of the knowledge graph, organize similar entities or relationships together, form a clearer and more reasonable knowledge graph structure, help improve the readability and understandability of the knowledge graph, and also help to achieve knowledge fusion; in other embodiments of the present invention, the implementer can also perform knowledge fusion on similar text sentences, and fuse entities of similar nature or the same entities with different expressions into a single entity to improve the structural hierarchy construction effect of the knowledge graph. It should be noted that the construction and application of knowledge graphs are existing technologies well known to those skilled in the art and will not be elaborated here.

[0093] In one embodiment of the present invention, after constructing the knowledge graph of the mineral deposit text report, the target prediction entity and the entities related to it in the knowledge graph can be obtained based on the direct query method, and some related entities with higher confidence can be selected as prediction indicators of mineralization potential in combination with expert knowledge, and the mineralization potential of the mineral deposit can be predicted based on the deep learning algorithm or the machine learning algorithm; this is already a prior art well known to those skilled in the art and will not be elaborated here.

[0094] In another embodiment of the present invention, after constructing the knowledge graph of the mineral deposit text report, the implementer can perform knowledge reasoning and connection prediction based on the knowledge graph inference method, for example, extracting the following relationships based on natural language processing technology:

[0095] (1) “Baoban gold deposit” - mineralization type - “mixed hydrothermal alteration rock type”;

[0096] (2) The relationship between “Baoban gold deposit” – genesis – “ductile shear zone control”;

[0097] (3) “Yanshanian mineralization” – including – “late Early Cretaceous mineralization subperiod”;

[0098] (4) “Late Early Cretaceous mineralization substage” – mineralization – “fault control”;

[0099] (5) “Late Early Cretaceous mineralization substage” – mineralization type – “hydrothermal type”;

[0100] Among them, "mixed hydrothermal alteration rock type" belongs to "hydrothermal type", the two entities are highly similar, and "ductile shear zone control" and "fracture control" are semantically similar. The embodiment of the present invention can cluster "mixed hydrothermal alteration rock type" and "hydrothermal type" into the same category based on SOM network clustering, and cluster "ductile shear zone control" and "fracture control" into the same category. The relationship vector between "Baoban gold deposit" and "hydrothermal type" and "fracture control" should have a high similarity with the relationship vector between "Baoban gold deposit" and "mixed hydrothermal alteration rock type" and "ductile shear zone control". Therefore, the following relationship can be inferred without specific research data: "Baoban gold deposit" - related - "hydrothermal type", "fracture control" - mineralization period - "late Early Cretaceous mineralization sub-period" - belongs to - "Yanshanian mineralization", from which it can be inferred that the type of Baoban gold deposit is hydrothermal type, and it can be prospected in the strata of the late Early Cretaceous; similarly, the implementer can also infer other information of the deposit based on other entity relationships.

[0101] In other embodiments of the present invention, the implementer can also directly determine the target prediction index, such as the content of a certain mineral component in the ore deposit, such as the content of alum stone. The implementer can query in the knowledge graph whether the ore deposit entity is connected to the entity of "alum stone" to determine whether alum stone has an implicit relationship with the ore deposit. If so, the indicator of "alum stone content" can be selected as one of the mineralization potential prediction indicators, and the mineralization potential can be predicted using models such as neural networks. Similarly, the knowledge graph can be used to mine other entities related to mineralization, and combined with expert knowledge, prediction indicators can be selected quickly and accurately. It can also be used to detect whether the indicators selected by experts are correlated with the ore deposit to be predicted.

[0102] In summary, the present invention obtains all deposit text vectors in the deposit data; inputs the deposit text vectors into a preset network clustering model for model training; further selects the selection neurons of each deposit text vector from all neurons; further obtains all special vector dimensions of each selection neuron; further obtains the degree of attribution of the deposit text vector relative to each selection neuron; adjusts the distance between the deposit text vector and the weight vector of the corresponding selection neuron according to the degree of attribution, thereby obtaining the winning neuron of each input deposit text vector, and obtains a trained preset network clustering model; constructs a deposit prediction index system according to the trained preset network clustering model. The present invention obtains the special dimensions and non-special dimensions of the deposit text vector by analysis, and then evaluates the degree of attribution of each deposit text vector relative to the corresponding selection neuron in different dimensions, thereby improving the classification effect of the deposit text vector, and then improving the deposit prediction effect.

[0103] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0104] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

Claims

1. A method for constructing a mineral deposit prediction index system based on knowledge graph, characterized in that: The method comprises: Get all the mineral deposit text vectors in the mineral deposit data; Input the deposit text vector into a preset network clustering model for model training; select neurons for each deposit text vector from all neurons based on the similarity between each input deposit text vector and all activation vectors or weight vectors of each neuron in the preset network clustering model; obtain all special vector dimensions of each selected neuron based on the distribution of vector elements in all activation vectors or weight vectors of each selected neuron; According to the difference between the vector elements in each input ore deposit text vector and the vector elements in the special vector dimension and the non-special vector dimension in all activation vectors of each selected neuron, the degree of belonging of the ore deposit text vector to each selected neuron is obtained; the distance between the ore deposit text vector and the weight vector corresponding to the selected neuron is adjusted according to the degree of belonging to obtain a corrected distance; according to the corrected distance, the winning neuron of each input ore deposit text vector is screened out from all corresponding selected neurons to obtain a trained preset network clustering model; all clusters are obtained according to the trained preset network clustering model; Construct a mineral deposit prediction indicator system based on knowledge graph according to all clusters; The method for obtaining the selected neurons includes: According to the similarity between each input ore deposit text vector and all activation vectors or weight vectors of each neuron in the preset network clustering model, the preliminary matching degree between each ore deposit text vector and each neuron is obtained; The input ore deposit text vector and the preliminary matching degree of each neuron are sequentially sorted to construct a preliminary matching degree sequence; a difference sequence of the preliminary matching degree series is obtained, and the preliminary matching degree sequence is segmented in combination with the maximum difference value in the difference sequence; the mean of the preliminary matching degree in each segment is obtained, and all neurons corresponding to the preliminary matching degree in the segment with the largest mean value are used as all selected neurons corresponding to the ore deposit text vector; The method for obtaining the preliminary matching degree includes: The mean of the vector elements in the input ore deposit text vector is used as the first element mean; the mean of the vector elements in all activation vectors of each neuron is used as the second element mean; the difference between the first element mean and the second element mean is normalized by negative correlation mapping to obtain the preliminary matching degree between each ore deposit text vector and each neuron; wherein, when a neuron has no corresponding activation vector, the preliminary matching degree between the input ore deposit text vector and the corresponding neuron is zero; The method for obtaining the special vector dimension includes: Taking a vector element on any vector dimension as a target element, obtaining the kurtosis of all the target elements corresponding to all activation vectors of each selected neuron, when the kurtosis is greater than a preset threshold, taking the target element as a special vector element, and the vector dimension corresponding to the special vector element as a special vector dimension; The method for obtaining the degree of belonging includes: Obtain the universal attribution parameter of the deposit text vector corresponding to the selected neuron according to the element difference between the vector elements of the non-special vector dimension in each input deposit text vector except the special vector dimension and the vector elements of the non-special vector dimension in all activation vectors of each selected neuron; According to the vector elements of the special vector dimension in each input ore deposit text vector and the element distribution between the vector elements of the non-special vector dimension in all activation vectors of each selected neuron, obtain the special attribution parameter of the ore deposit text vector corresponding to the selected neuron; The general attribution parameter and the special attribution parameter are multiplied, combined and normalized to obtain the attribution degree of the deposit text vector corresponding to the selected neuron; The calculation formula of the general attribution parameter includes: Among them, i is the serial number of the input mineral deposit text vector; k is the serial number of the selected neuron; k i is the serial number of the kth selected neuron corresponding to the i-th input mineral deposit text vector; The kth corresponding to the ith input mineral deposit text vector i The general attribution parameter of the selected neurons; sigmoid{} is the S-type function; x is the dimension number of the non-special vector dimension; X is the kth corresponding to the i-th input deposit text vector i The total number of dimensions of all non-special vectors in all activation vectors of the selected neurons; b i,x is the vector element on the xth non-special vector dimension in the ith input ore deposit text vector; n is the kth vector corresponding to the ith input ore deposit text vector. i The number of the activation vector of the selected neuron; N is the kth corresponding to the ith input mineral deposit text vector. i The total number of activation vectors of the selected neurons; The kth ore deposit text vector corresponding to the i-th input i The vector element in the xth non-special vector dimension in the nth activation vector of the selected neuron; The method for obtaining the special attribution parameter includes: Obtaining reference vector elements of each of the special vector dimensions according to element distribution of vector elements of all activation vectors of each of the selected neurons corresponding to the deposit text vector; The special attribution parameter is obtained according to the calculation formula of the special attribution parameter; the calculation formula of the special attribution parameter is: Among them, i is the serial number of the input mineral deposit text vector; k is the serial number of the selected neuron; k i is the serial number of the kth selected neuron corresponding to the i-th input mineral deposit text vector; The kth corresponding to the ith input mineral deposit text vector i special attribution parameters of the selected neurons; exp() is an exponential function with the natural constant e as the base; y is the dimension number of the special vector dimension; Y is the kth vector corresponding to the ith input deposit text vector. i The total number of dimensions of all special vectors in all activation vectors of the selected neurons; b i,y is the vector element of the yth special vector dimension in the ith input mineral deposit text vector; c i,y The reference vector element of the yth special vector dimension of the i-th input deposit text vector.

2. The method for constructing a mineral deposit prediction index system based on a knowledge graph according to claim 1, characterized in that: The method for obtaining the deposit text vector includes: Each text sentence in the mineral deposit data is processed by Jieba word segmentation, and all the word segments in each text sentence are assigned values ​​as vector elements through the TF-IDF method to construct the mineral deposit text vector corresponding to each text sentence.

3. The method for constructing a mineral deposit prediction index system based on a knowledge graph according to claim 1, characterized in that: The method for obtaining the reference vector element includes: Obtain a frequency histogram of the vector elements of all activation vectors of each of the selected neurons corresponding to the mineral deposit text vector; in the frequency histogram, use the vector elements corresponding to the non-special vector dimensions corresponding to the peak values ​​closest to each of the special vector dimensions as reference vector elements for each of the special vector dimensions.

4. The method for constructing a mineral deposit prediction index system based on a knowledge graph according to claim 1, characterized in that: The method for obtaining the corrected distance includes: The deposit text vector is negatively correlated with the attribution degree of each selected neuron and normalized, and the normalized value is multiplied by the Euclidean distance between the deposit text vector and the weight vector of each selected neuron to obtain the corrected distance.