Ore prospecting prediction method based on AE-LightGBM model

Abstract

The invention discloses a prospecting prediction method based on an AE-Light GBM model, and relates to the technical field of prospecting prediction, the method comprises the following steps: S1, obtaining research area data comprising ore deposits, non-ore deposits and unknown ore deposits, and constructing an original data set; s2, training an auto-encoder based on the label-free data to obtain an encoded label data set; s3, determining an optimal hyper-parameter combination of the Light GBM model by using a Bayesian optimization algorithm on the encoded labeled data set, and training an optimal Light GBM prediction model according to the optimal hyper-parameter combination; and S4, performing prospecting prediction and evaluation on the research area by using the optimal LightGBM prediction model. According to the prospecting prediction method based on the AE-Light GBM model, rich information of label-free data can be fully utilized, meanwhile, guidance of label data is combined, the precision, generalization ability and stability of the model are remarkably improved, and the prospecting prediction method based on the AE-Light GBM model has great significance in the field of prospecting prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of prospecting prediction, and in particular to a prospecting prediction method based on an AE-LightGBM model. Background Art

[0002] Mineral resources are an important foundation for national economic development and are of great significance to modern industry and social progress. With the continuous development and consumption of mineral resources, the difficulty of prospecting and prediction is also increasing. With the development of information technology, many machine learning technologies are gradually being applied in the field of prospecting and prediction to improve the efficiency and accuracy of prospecting.

[0003] In the task of prospecting and prediction, there are often a large amount of unlabeled data and a small amount of labeled data. Using only a small amount of labeled data for supervised learning makes it difficult to fully capture the potential patterns and regularities of the data, which may lead to poor generalization of the model. In contrast, combining the advantages of unsupervised learning and supervised learning can make more effective use of all data. Unsupervised learning can extract useful feature representations from a large amount of unlabeled data and automatically discover structures and patterns in the data, while supervised learning can use a small amount of labeled data for accurate prediction. In this way, the rich information of unlabeled data can be fully utilized, and combined with the guidance of labeled data, the accuracy, generalization ability and stability of the model can be significantly improved. Summary of the invention

[0004] The purpose of the present invention is to provide a prospecting prediction method based on the AE-LightGBM model to solve the problems raised in the above-mentioned background technology, make full use of the rich information of unlabeled data, and combine with the guidance of labeled data to significantly improve the accuracy, generalization ability and stability of the model.

[0005] To achieve the above object, the present invention provides a prospecting prediction method based on the AE-LightGBM model, comprising the following steps:

[0006] Step S1, obtaining the research area data including mining points, non-mining points and unknown mining points, and performing preprocessing operations to construct an original data set;

[0007] Step S2: train an autoencoder based on unlabeled data, compress the original data set into a low-dimensional latent space through the encoder, and then restore the data through the decoder, automatically learn the most representative feature patterns in the data, and use the trained encoder to encode all mining points to obtain an encoded labeled data set;

[0008] Step S3: On the encoded labeled data set, use the Bayesian optimization algorithm to determine the optimal hyperparameter combination of the LightGBM model, and use it to train the best LightGBM prediction model;

[0009] Step S4: Use the best LightGBM prediction model to predict and evaluate the mineral exploration in the study area.

[0010] Preferably, in step S1, the data attributes of the study area data are obtained by drilling data extraction and interpolation calculation, specifically including:

[0011] Spatial attributes: coordinates; geological attributes: strata, lithology and faults; geochemical attributes: grade, elements; geophysical attributes: polarizability, resistivity and magnetic susceptibility;

[0012] Each data corresponds to a spatial coordinate of the study area, where the number of mineral points in the study area is at least 100 groups and the number of unknown mineral points is at least 10,000.

[0013] Preferably, the preprocessing operation in step S1 includes integrating attributes, labels and spatial coordinate information of the data; scaling the attribute values; and removing duplicate samples.

[0014] Preferably, the labeled data set in step S2 is represented as [X L :Y L :I L ], the unlabeled dataset is represented as [X U :I U ], where X is the attribute matrix, Y is the label matrix, and I is the spatial coordinate matrix.

[0015] Preferably, in step S2, the autoencoder is composed of an encoder and a decoder, the decoder is composed of a multi-layer fully connected neural network, the encoder is composed of an embedding layer and a multi-layer fully connected neural network, and the embedding layer can map the categorical variable into a numerical vector;

[0016] The structures of the fully connected networks of the encoder and decoder are symmetrical. The number of neurons in each layer of the encoder gradually decreases, compressing the high-dimensional input data into a low-dimensional potential representation. The decoder increases the number of neurons layer by layer through a symmetrical structure, gradually restoring the low-dimensional representation to the original data dimension.

[0017] The autoencoder adopts a batch regularization strategy, uses the Adam optimizer to adjust the learning rate, and uses PReLU() as the activation function. The formula is as follows:

[0018]

[0019] Among them, α is the parameter of automatic learning, and x is the input value;

[0020] The training data of the autoencoder is X in the unlabeled data U , the loss function is the mean square error MSE, the formula is as follows:

[0021]

[0022] Among them, n is the number of samples of data, x i is the i-th sample of the input, is the i-th sample reconstructed by the autoencoder.

[0023] Preferably, in step S2, the trained encoder is used to encode the data attributes of all data samples, and the input X L and X U , the output is the encoded attribute X' L and X' U , construct the encoded labeled dataset [X' L :Y L ] and randomly divided into training and test sets.

[0024] Preferably, in step S3, the optimal hyperparameter combination of the LightGBM model is determined by a Bayesian optimization algorithm, and the objective function is set to the average F1 score of the k-fold cross validation on the training set, and the formula is as follows:

[0025]

[0026] Among them, k is the number of cross-validation folds, F1 (i) is the F1 score of the i-th fold.

[0027] Preferably, the LightGBM hyperparameters optimized by Bayesian optimization in step S3 include: number of iterations, learning rate, upper limit of leaves of decision tree, upper limit of depth of decision tree, category of gradient boosting, minimum number of samples on leaves, minimum split gain, regularization coefficient and feature sampling ratio;

[0028] The optimization number is no less than 100 times to obtain the best hyperparameter combination.

[0029] Preferably, in step S3, the hyperparameters of the LightGBM model are set according to the optimal hyperparameter combination, and training is performed on the training set. The loss function is set to cross-entropy, and the formula is as follows:

[0030]

[0031] Where n is the number of samples, y i is the true label of the i-th sample, p i is the probability that the model predicts a positive sample;

[0032] The best LightGBM model is obtained and a preliminary test is performed on the test set.

[0033] Preferably, the specific steps of step S4 are as follows:

[0034] Step S41: The attributes X' of all the samples in the study area after encoding L and X' U , as input;

[0035] Step S42, using the best LightGBM model to perform prospecting prediction and obtain the mineralization probability of the entire study area;

[0036] Step S43, combining the predicted probability with the corresponding spatial coordinates to obtain the spatial distribution of the mineralization probability of the study area;

[0037] Step S44: Evaluate the results.

[0038] Therefore, the present invention adopts the above-mentioned mineral prospecting prediction method based on the AE-LightGBM model, which can make full use of the rich information of unlabeled data, and at the same time combine the guidance of labeled data to significantly improve the accuracy, generalization ability and stability of the model, which is of great significance in the field of mineral prospecting prediction.

[0039] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a structural schematic diagram of an embodiment of a prospecting prediction method based on the AE-LightGBM model of the present invention;

[0041] Figure 2 A schematic diagram of a neural network architecture of an automatic encoder for a mineral prospecting prediction method based on an AE-LightGBM model of the present invention;

[0042] Figure 3 It is a ROC curve diagram of a comparison model of a prospecting prediction method based on the AE-LightGBM model of the present invention;

[0043] Figure 4 The figure is a ROC curve diagram of a model trained by a mineral prospecting prediction method based on the AE-LightGBM model of the present invention. DETAILED DESCRIPTION

[0044] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0045] Unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0046] Example

[0047] See also Figure 1-4 The present invention provides a prospecting prediction method based on the AE-LightGBM model. In this embodiment, a manganese ore test park is used as the research area, and prospecting prediction is performed based on the collected geophysical and chemical data and other multi-source data.

[0048] like Figure 1 As shown, it is a flow chart of the prospecting prediction method based on the AE-LightGBM model of the present invention, which specifically includes the following steps:

[0049] Step S1, collects research area data including 647 mineral points, 723 non-mineral points and nearly 3.5 million unknown mineral points. The data attributes include spatial attributes (coordinates), geological attributes (strata, faults, folds), geochemical attributes (grades of 13 elements such as Mn, Cr, Ni, etc.) and labels, among which strata and labels are categorical variables, and the remaining attributes are numerical variables.

[0050] Various attributes are spliced ​​and integrated, and the numerical attributes except coordinates are discretized and scaled to obtain the original data set, which is divided into unlabeled data sets and labeled data sets according to labels. The unlabeled data set is deduplicated, and the number of samples after processing is 137,018. The labeled data is deduplicated and balanced, and the number of samples after processing is 1,036. Some of the data after preprocessing are shown in Table 1.

[0051] Table 1 Partial data sheet

[0052]

[0053]

[0054] Step S02: Use the unlabeled data after deduplication (excluding coordinates and labels) as training data for the autoencoder, where Figure 2 Schematic diagram of the neural network architecture of the autoencoder.

[0055] The autoencoder adopts a batch normalization strategy, and the batch size (batch_size) is set to 64; the number of epochs of the dataset is set to 100; the Adam optimizer is used to adjust the learning rate, and the initial learning rate is 0.001; PReLU() is used as the activation function, and the formula is as follows:

[0056]

[0057] Among them, α is a parameter that can be automatically learned, and x is the input value;

[0058] The loss function of the autoencoder is the mean square error (MSE), which is as follows:

[0059]

[0060] Where n is the number of samples in the data, x i is the i-th sample of the input, is the i-th sample reconstructed by the autoencoder.

[0061] All data are encoded using the trained encoder module, and some of the encoded data are shown in Table 2. The encoded labeled dataset is randomly divided into a training set (75%) and a test set (25%).

[0062] Table 2 Part of the data table after encoding

[0063] X Y Feature 0 Feature 1 Feature 2 Feature 3 label 887133 3120000 -4.6451054 5.75887 -7.9776893 4.170929 1 887457 3121000 -4.425543 5.350671 -7.3455424 4.205764 1 883972 3097994 -2.613967 4.137289 -6.745719 6.0801363 0 865162 3127244 -2.9887774 6.879778 -4.2890563 1.8771343 0 867112 3090974 -7.2686644 7.262398 -7.3756323 7.7210927 -1 858022 3140174 -0.54217523 2.9779255 -3.3129318 5.168526 -1

[0064] Step S03, determine the optimal hyperparameter combination of the LightGBM model through the Bayesian optimization algorithm; the objective function is set to the average F1 score of the 5-fold cross validation on the training set, and the formula is as follows:

[0065]

[0066] Among them, F1 (i) is the F1 score of the i-th fold.

[0067] The hyperparameters for optimization include the number of iterations, learning rate, the upper limit of the leaf size of the decision tree, the upper limit of the depth of the decision tree, the minimum number of samples on the leaf, the minimum split gain, the regularization coefficient and the feature sampling ratio. The range of the parameters is shown in Table 3. The number of iterations is set to 100 times. The best hyperparameter combination is shown in Table 4.

[0068] Table 3 Parameter adjustment range table (k is the step size)

[0069] Hyperparameters Search Scope n_estimators [150,400],k＝50 learning_rate [0.01,0.1],k＝0.01 num_leaves [20,150],k＝1 max_depth [5,15],k＝1 min_data_in_leaf [1,50],k＝1 min_gain_to_split [0,1] reg_alpha [0,1] reg_lambda [0,1] feature_fraction [0.6,1] bagging_fraction [0.6,1]

[0070] Table 4 Optimal parameters

[0071]

[0072]

[0073] The hyperparameters of the LightGBM model are set according to the optimal hyperparameter combination, and training is performed on the training set; the loss function is set to cross-entropy, and the formula is as follows:

[0074]

[0075] Where n is the number of samples, y i is the true label of the i-th sample, p i is the probability that the model predicts a positive sample, and the best LightGBM model is obtained.

[0076] The comparison model is introduced. Compared with the model trained by the present invention, the autoencoder is not used to learn in unlabeled data. The rest of the technical route is the same as the present invention. The comparison model and the model trained by the present invention are tested on the test set, and their ROC curves are obtained respectively, as shown in Figure 3 , 4 shown.

[0077] Step S4, the comparison model and the model trained by the present invention are used to predict all samples respectively to obtain the mineralization probability of the entire study area, and the mineralization probability is combined with the corresponding spatial coordinates to obtain the spatial distribution of the mineralization probability of the study area. The average mineralization probability of the comparison model in the test area (a known mineral area somewhere in the study area) is 0.22, and the average mineralization probability of the model trained by the present invention is 0.57.

[0078] Therefore, the present invention adopts the above-mentioned mineral prospecting prediction method based on the AE-LightGBM model, which can make full use of the rich information of unlabeled data, and at the same time combine the guidance of labeled data to significantly improve the accuracy, generalization ability and stability of the model, which is of great significance in the field of mineral prospecting prediction.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A prospecting prediction method based on AE-LightGBM model, characterized in that: The following steps are involved: Step S1, obtaining the research area data including mining points, non-mining points and unknown mining points, and performing preprocessing operations to construct an original data set; Step S2: train an autoencoder based on unlabeled data, compress the original data set into a low-dimensional latent space through the encoder, and then restore the data through the decoder, automatically learn the most representative feature patterns in the data, and use the trained encoder to encode all mining points to obtain an encoded labeled data set; Step S3: On the encoded labeled data set, use the Bayesian optimization algorithm to determine the optimal hyperparameter combination of the LightGBM model, and use it to train the best LightGBM prediction model; Step S4: Use the best LightGBM prediction model to predict and evaluate the mineral exploration in the study area.

2. The ore prospecting prediction method based on the AE-LightGBM model according to claim 1 is characterized in that: In step S1, the data attributes of the study area data are obtained by drilling data extraction and interpolation calculation, specifically including: Spatial attributes: coordinates; geological attributes: strata, lithology and faults; geochemical attributes: grade, elements; geophysical attributes: polarizability, resistivity and magnetic susceptibility; Each data corresponds to a spatial coordinate of the study area, where the number of mineral points in the study area is at least 100 groups and the number of unknown mineral points is at least 10,000.

3. The ore prospecting prediction method based on the AE-LightGBM model according to claim 2, characterized in that: The preprocessing operation in step S1 includes integrating the attributes, labels and spatial coordinate information of the data; scaling the attribute values; and removing duplicate samples.

4. The ore prospecting prediction method based on the AE-LightGBM model according to claim 3 is characterized in that: The labeled data set in step S2 is represented as [X L :Y L :I L ], the unlabeled dataset is represented as [X U :I U ], where X is the attribute matrix, Y is the label matrix, and I is the spatial coordinate matrix.

5. The ore prospecting prediction method based on the AE-LightGBM model according to claim 4 is characterized in that: In step S2, the autoencoder is composed of an encoder and a decoder, the decoder is composed of a multi-layer fully connected neural network, the encoder is composed of an embedding layer and a multi-layer fully connected neural network, and the embedding layer can map the categorical variable to a numerical vector; The structures of the fully connected networks of the encoder and decoder are symmetrical. The number of neurons in each layer of the encoder gradually decreases, compressing the high-dimensional input data into a low-dimensional potential representation. The decoder increases the number of neurons layer by layer through a symmetrical structure, gradually restoring the low-dimensional representation to the original data dimension. The autoencoder adopts a batch regularization strategy, uses the Adam optimizer to adjust the learning rate, and uses PReLU() as the activation function. The formula is as follows: Among them, α is the parameter of automatic learning, and x is the input value; The training data of the autoencoder is X in the unlabeled data U , the loss function is the mean square error MSE, the formula is as follows: Among them, n is the number of samples of data, x i is the i-th sample of the input, is the i-th sample reconstructed by the autoencoder.

6. The ore prospecting prediction method based on the AE-LightGBM model according to claim 5 is characterized in that: In step S2, the trained encoder is used to encode the data attributes of all data samples, and the input X L and X U , the output is the encoded attribute X' L and X' U , construct the encoded labeled dataset [X' L :Y L ] and randomly divided into training and test sets.

7. The ore prospecting prediction method based on the AE-LightGBM model according to claim 6 is characterized in that: In step S3, the optimal hyperparameter combination of the LightGBM model is determined by the Bayesian optimization algorithm, and the objective function is set to the average F1 score of the k-fold cross validation on the training set, and the formula is as follows: Among them, k is the number of cross-validation folds, F1 (i) is the F1 score of the i-th fold.

8. The ore prospecting prediction method based on the AE-LightGBM model according to claim 7 is characterized in that: The LightGBM hyperparameters optimized by Bayesian optimization in step S3 include: number of iterations, learning rate, upper limit of leaves of decision tree, upper limit of depth of decision tree, category of gradient boosting, minimum number of samples on leaves, minimum split gain, regularization coefficient and feature sampling ratio; The optimization number is no less than 100 times to obtain the best hyperparameter combination.

9. The ore prospecting prediction method based on the AE-LightGBM model according to claim 8, characterized in that: In step S3, the hyperparameters of the LightGBM model are set according to the optimal hyperparameter combination, and training is performed on the training set. The loss function is set to cross-entropy, and the formula is as follows: Where n is the number of samples, y i is the true label of the i-th sample, p i is the probability that the model predicts a positive sample; The best LightGBM model is obtained and a preliminary test is performed on the test set.

10. The ore prospecting prediction method based on the AE-LightGBM model according to claim 9 is characterized in that: The specific steps of step S4 are as follows: Step S41: The attributes X' of all the samples in the study area after encoding L and X' U , as input; Step S42, using the best LightGBM model to perform prospecting prediction and obtain the mineralization probability of the entire study area; Step S43, combining the predicted probability with the corresponding spatial coordinates to obtain the spatial distribution of the mineralization probability of the study area; Step S44: Evaluate the results.

Images