In-situ leaching sandstone type uranium ore quantitative gamma value prediction method and device and electronic equipment

Abstract

The invention discloses an in-situ leaching sandstone type uranium ore quantitative gamma value prediction method. The method comprises the following steps: a training stage: acquiring logging curve data and a quantitative gamma data sample set of a research block; and performing depth correction processing on the logging curve data and the quantitative gamma data, constructing an in-situ leaching sandstone type uranium mine quantitative gamma training data set, and training an improved LSTM model based on the training data set. And a prediction stage: collecting logging curve data of a to-be-predicted target block, inputting the logging curve data into the trained improved LSTM model for quantitative gamma value prediction, and outputting a quantitative gamma value prediction result of the to-be-predicted target block. According to the in-situ leaching sandstone type uranium ore quantitative gamma value prediction method based on the improved LSTM model, accurate prediction of the in-situ leaching sandstone type uranium ore quantitative gamma value can be achieved, and then accurate positioning and efficient development of uranium ore distribution are achieved.

Description

Technical Field

[0001] The invention relates to the field of geophysics, and in particular to a method and a device for predicting the quantitative gamma value of in-situ leaching sandstone-type uranium ore. Background Art

[0002] In recent years, with the increasing demand for energy, the oil resources in major old oil fields in China are decreasing. With the gradual deepening of oil development, the proportion of low-efficiency, non-efficiency and negative-efficiency wells in oil field wells has been increasing year by year, becoming an important factor affecting the social and economic benefits of oil fields. To this end, the oil field has a long-term perspective, increases revenue and reduces expenditure, and actively promotes the development and utilization of new energy sources such as geothermal and uranium mines. As a national strategic resource, uranium mines were previously explored and developed by enterprises in the nuclear industry system. The discovery of the Tongliao uranium deposit is a major breakthrough in my country's search for in-situ leachable sandstone-type uranium deposits in oil fields. It is also an example of interdisciplinary research and has opened up a new prospecting model: using oil exploration radioactive logging (natural gamma curve) data to select radioactive survey areas and then drill and verify to carry out uranium exploration, thereby submitting a uranium mine development base and ultimately developing sandstone-type uranium mines.

[0003] In the field of oil, gas and uranium exploration, natural gamma logging is an important geophysical exploration method. Natural gamma logging reflects the content and distribution of radioactive elements in the formation by measuring the intensity of gamma rays produced by the decay of radioactive elements in the formation. In existing literature, such as "Research on uranium resource evaluation methods based on natural gamma logging data-taking the Pengyang uranium mining area in the Ordos Basin as an example (2022)", it introduces the significant abnormal response of natural gamma in the sandstone-type uranium reservoir in the Pengyang uranium mining area. By comparing and analyzing the quantitative gamma in 9 verification holes and the natural gamma logging data in the oil field boreholes, the natural gamma logging regression model of the quantitative gamma value of the ore layer in the Pengyang uranium mining area is studied, and the regional uranium ore layer identification and grade evaluation method based on natural gamma logging information is established. "Application of Coalfield Natural Gamma Logging Curves in Sandstone Uranium Deposit Exploration (2024)" introduces the use of coalfield drilling data in the Heishui area of ​​the Pingzhuang Basin as the basis, and uses the correlation between the content of radioactive elements in rocks and the intensity of natural gamma radiation to provide important guidance for the delineation of uranium anomaly areas and the deployment of exploration work in the exploration of sandstone uranium deposits.

[0004] Therefore, there are often the following problems in predicting the quantitative gamma value of in-situ leaching sandstone uranium deposits: (1) The natural gamma curve of oil well logging is measured by a large-size, high-sensitivity probe, and its dead time is generally long and variable, and the general API calibration system. The natural gamma logging instruments used in different eras are different, and the recording data units are not uniform. The natural gamma curve obtained by oil well logging can be used as a clue to find uranium anomaly information, but it is impossible to quantitatively calculate uranium deposits. These characteristics lead to false anomalies in the natural gamma curves of many oil wells, which makes it impossible to directly use the radioactivity of the natural gamma data of the oil field. It is necessary to convert it into quantitative gamma first to effectively judge the radioactivity intensity of the uranium deposit; (2) Traditional gamma prediction methods mainly rely on statistical analysis. These methods can reflect the quantitative gamma of uranium deposits to a certain extent, but there are many limitations. For example, statistical analysis methods are often based on historical data and empirical formulas, which are difficult to accurately describe the quantitative gamma of uranium deposits under complex geological conditions.

[0005] With the rapid development of artificial intelligence technology, deep learning, as a powerful data analysis tool, has achieved remarkable results in many fields. By building a multi-layer neural network model, deep learning can automatically learn complex features in data and make predictions and classifications based on these features. In the field of geological exploration, deep learning has begun to be used in seismic and logging data interpretation, lithology identification, oil and gas reservoir prediction, etc., and has shown high prediction accuracy and generalization ability.

[0006] Therefore, it is necessary to develop a deep learning-based quantitative gamma value prediction method for in-situ leaching sandstone-type uranium deposits to achieve accurate prediction of quantitative gamma values ​​in order to precisely locate and efficiently develop uranium ore distribution. Summary of the invention

[0007] The object of the present invention is to provide a method, device and electronic equipment for predicting the quantitative gamma value of in-situ leaching sandstone-type uranium ore based on an improved LSTM model, so as to solve the problems raised in the above-mentioned background technology.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] In a first aspect, a method for predicting the quantitative gamma value of an in-situ leaching sandstone-type uranium ore is provided, comprising:

[0010] Training phase:

[0011] Step S1: Collect the well logging curve data and quantitative gamma data sample set of the study block through experiments and well logging field operations;

[0012] Step S2: performing a first preprocessing on the logging curve data and the quantitative gamma data, wherein the first preprocessing includes: depth correction, wherein the depth correction is to perform depth correction on the logging curve data based on the quantitative gamma depth;

[0013] Step S3: constructing a quantitative gamma training data set for in-situ leaching sandstone-type uranium deposits based on the well logging data and quantitative gamma data after the first preprocessing;

[0014] Step S4: using the quantitative gamma training data set of in-situ leaching sandstone-type uranium deposits to train an improved LSTM model;

[0015] Prediction stage:

[0016] Step S5: collecting logging curve data of the target block to be predicted through logging field operations;

[0017] Step S6: Based on the improved LSTM model trained in step S4, the well logging curve data of the target block to be predicted is input into the improved LSTM model trained in step S4 to perform quantitative gamma value prediction;

[0018] Step S7: outputting the prediction result of the quantitative gamma value of the target block to be predicted.

[0019] Furthermore, in step S4, the improved LSTM model is trained using the quantitative gamma training data set of in-situ leaching sandstone-type uranium ore. The improved LSTM model is improved on the LSTM model, and the model includes:

[0020] The first convolutional layer, the first ReLU layer, the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, the fourth cycle LSTM layer, the fourth encoding layer, the third encoding layer, the second encoding layer, the first encoding layer, the first Dense layer, and the first Sigmoid layer. The output of the first convolutional layer is connected to the first ReLU layer, the output of the first ReLU layer is connected to the first cycle LSTM layer, the output of the first cycle LSTM layer is connected to the second cycle LSTM layer, and the output of the second cycle LSTM layer is connected to the third cycle LSTM layer; the output of the third cycle LSTM layer is connected to the fourth cycle LSTM layer, the output of the fourth cycle LSTM layer is connected to the fourth encoding layer, the output of the fourth encoding layer is connected to the third encoding layer, the output of the third encoding layer is connected to the second encoding layer, the output of the second encoding layer is connected to the first encoding layer, the output of the first encoding layer is connected to the first Dense layer, and the output of the first Dense layer is connected to the first Sigmoid layer. The data flow is input from the first convolutional layer and finally output by the first Sigmoid layer.

[0021] The first cycle LSTM layer structure includes:

[0022] The first LSTM-skip (LS) layer, the first Dropout layer, and the first normalization (LN) layer; the output of the first LSTM-skip (LS) layer is connected to the first Dropout layer, and the output of the first Dropout layer is connected to the first normalization (LN) layer. The first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer have the same structure; the first encoding layer is a Transformer encoder (TE), which specifically includes:

[0023] Self-attention layer (SA), second normalization layer, feedforward layer (FFW), second Dropout layer, third normalization layer; and the output of the self-attention layer (SA) is connected to the second normalization layer, the output of the second normalization layer is connected to the feedforward layer, the output of the feedforward layer is connected to the second Dropout layer, and the output of the second Dropout layer is connected to the third normalization layer; and two residual connections are established before the second normalization layer and the third normalization layer respectively, the first residual connection is located before the second normalization layer, and the input data is the data entering the Transformer encoder; the second residual connection is located before the third normalization layer, and the input data is the output data of the second Dropout layer; and the structures of the fourth encoding layer, the third encoding layer, the second encoding layer, and the first encoding layer are the same;

[0024] The first jump connection unit is used to establish a jump connection between the first period LSTM layer and the first encoding layer. Specifically, the output of the first period LSTM layer is superimposed with the output of the first encoding layer and input into the first Dense layer;

[0025] The second jump connection unit is used to establish a jump connection between the second period LSTM layer and the second encoding layer, specifically, the output of the second period LSTM layer is superimposed with the output of the second encoding layer and inputted into the first encoding layer;

[0026] The third jump connection unit is used to establish a jump connection between the third period LSTM layer and the third encoding layer, specifically, the output of the third period LSTM layer is superimposed with the output of the third encoding layer and inputted into the second encoding layer;

[0027] The fourth jump connection unit is used to establish a jump connection between the fourth period LSTM layer and the fourth encoding layer, specifically, the output of the fourth period LSTM layer is superimposed with the output of the fourth encoding layer and input into the third encoding layer;

[0028] The jump connection modes of the first, second, third and fourth jump connection units are all the same.

[0029] Furthermore, the first cycle LSTM layer is:

[0030] The period parameter p and the multiple m are introduced into the time step of the LSTM layer. The period parameter p represents the number of hidden layer units skipped, and the period parameter p is multiplied by the multiple m to enhance the extraction of longer period deep sequence features in the LSTM layer.

[0031] Furthermore, the skip connection method is a fixed weighted hybrid skip connection, which is:

[0032] Step S11: Setting weight coefficients: setting weight coefficients w1, w2, w3, and w4 for the output results of the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer, respectively;

[0033] Step S22: Mixing and superposition: PLSTM fuse =w1×PLSTM1+w2×PLSTM2+w3×PLSTM3+w4×PLSTM4, where PLSTM fuse is the result after mixed superposition, PLSTM1, PLSTM2, PLSTM3, and PLSTM4 are the output results of the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer in the kth iteration, respectively; w1, w2, w3, and w4 are the weight coefficients assigned to the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer, respectively;

[0034] Step S33: Skip connection processing: PLSTM the result after mixing and superimposing in step S22 fuse Respectively: superimposed with the output of the first coding layer and input into the first Dense layer; superimposed with the output of the second coding layer and input into the first coding layer; superimposed with the output of the third coding layer and input into the second coding layer; superimposed with the output of the fourth coding layer and input into the third coding layer.

[0035] Furthermore, the skip connection method is a dynamic weighted hybrid skip connection, which is as follows:

[0036] Step S111: convolution: performing convolution operations on the outputs of the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer based on 1D conv with a convolution kernel size of K;

[0037] Step S222: Sum: sum the output of the first cycle LSTM layer after 1D conv, the output of the second cycle LSTM layer, the output of the third cycle LSTM layer, and the output of the fourth cycle LSTM layer;

[0038] P k =conv K(PLSTM1 k )+conv K (PLSTM2 k )+conv K (PLSTM3 k )+conv K (PLSTM4 k ), where P k For global information, PLSTM1 k 、PLSTM2 k 、PLSTM3 k 、PLSTM4 k are the output results of the first cycle LSTM layer, the output results of the second cycle LSTM layer, the output results of the third cycle LSTM layer, and the output results of the fourth cycle LSTM layer in the kth iteration, respectively. K It is a 1D convolution with a kernel size of K;

[0039] Step S333: Set weight coefficient: Calculate the weight coefficient w1 at the current iteration time k k 、w2 k 、w3 k 、w4 k :

[0040] w1 k =conv K (PLSTM1 k ) / [conv K (PLSTM1 k )+conv K (PLSTM2 k )+conv K (PLSTM3 k )+conv K (PLSTM4 k )]

[0041] w2 k =conv K (PLSTM2 k ) / [conv K (PLSTM1 k )+conv K (PLSTM2 k )+conv K (PLSTM3 k )+conv K (PLSTM4 k )]

[0042] w3 k =conv K (PLSTM3k ) / [conv K (PLSTM1 k )+conv K (PLSTM2 k )+conv K (PLSTM3 k )+conv K (PLSTM4 k )]

[0043] w4 k =conv K (PLSTM4 k ) / [conv K (PLSTM1 k )+conv K (PLSTM2 k )+conv K (PLSTM3 k )+conv K (PLSTM4 k )]

[0044] Step S444: Mixed superposition: The output result of the first cycle LSTM layer, the output result of the second cycle LSTM layer, the output result of the third cycle LSTM layer, and the output result of the fourth cycle LSTM layer are mixed and superimposed at the current iteration time k. The result is as follows: PLSTM fuse k =w1 k ×PLSTM1 k +w2 k ×PLSTM2 k +w3 k ×PLSTM3 k +w4 k ×PLSTM4 k ;

[0045] Step S555: skip connection processing: the result of mixing and superimposing in step S444 is PLSTM fuse k Respectively: superimposed with the output of the first coding layer and input into the first Dense layer; superimposed with the output of the second coding layer and input into the first coding layer; superimposed with the output of the third coding layer and input into the second coding layer; superimposed with the output of the fourth coding layer and input into the third coding layer.

[0046] Furthermore, the loss function used in the improved LSTM model is:

[0047]

[0048] Among them, f(xi ) and σ(x i ) 2 Represents a specific input x i The output prediction mean and variance, y i Represents the training data label value, and N represents the number of sample points.

[0049] Furthermore, the loss function used in the improved LSTM model is:

[0050]

[0051] Among them, α and β are control parameters. α controls the common weight of positive samples and negative samples on the total loss, and β controls the weight of easy-to-classify samples and difficult-to-classify samples. Y represents the label value. represents the predicted value, Y i ∈Y, N represents the sample points.

[0052] In a second aspect, a device for predicting the quantitative gamma value of an in-situ leaching sandstone-type uranium ore is provided, comprising:

[0053] Training Unit:

[0054] The first acquisition module: through experiments and logging field operations, the logging curve data and quantitative gamma data sample sets of the research block are collected;

[0055] A preprocessing module: performs a first preprocessing on the logging curve data and the quantitative gamma data, wherein the first preprocessing includes: depth correction, wherein the depth correction is to perform depth correction on the logging curve data based on the quantitative gamma depth; a construction module: constructs a quantitative gamma training data set for in-situ leaching sandstone-type uranium ore based on the logging curve data and the quantitative gamma data after the first preprocessing;

[0056] Training and learning module: using the quantitative gamma training data set of in-situ leaching sandstone-type uranium deposits to train and improve the LSTM model;

[0057] Prediction Unit:

[0058] The second acquisition module: collects the logging curve data of the target block to be predicted through logging field operations;

[0059] Quantitative prediction module: Based on the improved LSTM model trained in step S4, the well logging curve data of the target block to be predicted is input into the improved LSTM model trained in step S4 to perform quantitative gamma value prediction;

[0060] Output module: outputs the quantitative gamma value prediction results of the target block to be predicted.

[0061] In a third aspect, an electronic device is also provided, comprising: a memory for storing a computer program; and a processor for implementing the steps of a method for predicting a quantitative gamma value of an in-situ leaching sandstone-type uranium ore when executing the computer program.

[0062] In a fourth aspect, a computer-readable storage medium is further provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method for predicting the quantitative gamma value of in-situ leaching sandstone-type uranium ore as described in any one of claims 1 to 7 are implemented.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] 1. The present invention provides an improved LSTM model. The improved LSTM model can exert efficient prediction capabilities even when there is relatively little monitoring data, especially when data is scarce, which is of great significance for the exploration and development of in-situ leaching sandstone-type uranium deposits.

[0065] 2. The present invention provides an improved skip connection mechanism, which can more effectively transmit information in the deep network and enhance the feature learning ability of the model. This new skip connection method can improve the accuracy and robustness of the model in predicting the quantitative gamma value of in-situ leaching sandstone-type uranium deposits.

[0066] 3. The loss function specially designed for the quantitative gamma value prediction of in-situ leaching sandstone-type uranium ore can more accurately measure the difference between the predicted value and the actual value, thereby guiding the model to learn more accurate prediction capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0068] Figure 1 A flowchart of a method for predicting the quantitative gamma value of in-situ leaching sandstone-type uranium ore according to an embodiment of the present invention;

[0069] Figure 2 An improved LSTM deep learning model for quantitative gamma value prediction provided according to an embodiment of the present invention;

[0070] Figure 3 A composition diagram of a first period LSTM block provided according to an embodiment of the present invention;

[0071] Figure 4 A composition diagram of a first coding layer Transformer encoder (TE) provided according to an embodiment of the present invention;

[0072] Figure 5 A schematic diagram of an improved LSTM deep learning hybrid weighted skip connection used for quantitative gamma value prediction provided according to an embodiment of the present invention;

[0073] Figure 6 This is a graph showing the variation of F1-score with Epoch for three skip connection modes provided according to an embodiment of the present invention;

[0074] Figure 7 The graphs of IoU changes with Epoch for three skip connection modes provided according to an embodiment of the present invention;

[0075] Figure 8 A graph showing changes in Precision with Epoch for three skip connection modes provided according to an embodiment of the present invention;

[0076] Fig. 9 This is a result diagram of the first embodiment of quantitative gamma value prediction of in-situ leaching sandstone-type uranium ore provided by an embodiment of the present invention;

[0077] Fig.10 This is a result diagram of the second embodiment of quantitative gamma value prediction of in-situ leaching sandstone-type uranium ore provided by an embodiment of the present invention;

[0078] Fig.11 This is a result diagram of the third embodiment of quantitative gamma value prediction of in-situ leaching sandstone-type uranium ore provided by an embodiment of the present invention;

[0079] Fig.12 This is a diagram of a quantitative gamma value prediction device for in-situ leaching sandstone-type uranium ore provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0080] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0081] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0082] Figure 1 The figure shows a flowchart of a five-in-one oil and uranium prospecting method according to an embodiment of the present invention. For ease of description, only the part related to the embodiment of the present invention is shown, which is described in detail as follows:

[0083] A method for predicting the quantitative gamma value of an in-situ leaching sandstone-type uranium ore, comprising:

[0084] Training phase:

[0085] Step S1: Collect the well logging curve data and quantitative gamma data sample set of the study block through experiments and well logging field operations;

[0086] Furthermore, the logging curve data used in the present invention may include: resistivity logging curve: deep lateral resistivity (RD), shallow lateral resistivity (RS); porosity logging curve: compensated neutron (CNL), density (DEN), acoustic wave (AC); natural gamma logging curve: natural gamma (GR); other logging curves: spontaneous potential (SP), well diameter (CALC), uranium-free gamma (KTh), energy spectrum natural gamma (GRSL), uranium (U), thorium (Th), potassium (K), bottom gradient resistivity (R250), well inclination angle (DEV), well inclination azimuth (DPAZ), logging velocity (LSPD), coal seam mark (CI), etc.

[0087] Step S2: performing a first preprocessing on the logging curve data and the quantitative gamma data, wherein the first preprocessing includes: depth correction, wherein the depth correction is to perform depth correction on the logging curve data based on the quantitative gamma depth;

[0088] Step S3: constructing a quantitative gamma training data set for in-situ leaching sandstone-type uranium deposits based on the well logging data and quantitative gamma data after the first preprocessing;

[0089] Step S4: using the quantitative gamma training data set of in-situ leaching sandstone-type uranium deposits to train an improved LSTM model;

[0090] Prediction stage:

[0091] Step S5: collecting logging curve data of the target block to be predicted through logging field operations;

[0092] Step S6: Based on the improved LSTM model trained in step S4, the well logging curve data of the target block to be predicted is input into the improved LSTM model trained in step S4 to perform quantitative gamma value prediction;

[0093] Step S7: outputting the prediction result of the quantitative gamma value of the target block to be predicted.

[0094] Figure 2-Figure 5 The improved LSTM deep learning model used for quantitative gamma value prediction provided by an embodiment of the present invention, the composition diagram of the first cycle LSTM block, the composition diagram of the first encoding layer Transformer encoder (TE), and the schematic diagram of the improved LSTM deep learning hybrid weighted jump connection are respectively described; the improved LSTM deep learning model used for quantitative gamma value prediction of the present invention and the composition and operation principle of each module of the model are described in detail below:

[0095] The present invention mainly extracts important depth sequence features from the input logging curve data, and establishes a nonlinear mapping relationship between the logging curve data and the quantitative gamma value to achieve quantitative gamma value prediction. When establishing the nonlinear mapping relationship, the focus is on the memory of long-term features as well as the memory of short-term and global features.

[0096] The proposed quantitative gamma value prediction model is developed based on the LSTM model, which consists of two structures, namely, the period LSTM (PLSTM, PL) block and the encoding layer (Transformer encoder, TE) block. The first layer is a one-dimensional convolutional layer (1D convolutional layer), and the present invention uses zero padding in the one-dimensional convolutional layer to maintain the size of the feature map.

[0097] The second layer is a ReLU layer with a rectified linear unit (ReLU), which is used to extract the short-term features of the input logging data. Its output is expressed as follows:

[0098] v c =ReLU(W c *X+b c )

[0099] Among them, W c 、b c are weights and biases, * represents convolution operation, X represents the input depth sequence data, and ReLU(x)=max(0,x) is a nonlinear activation function.

[0100] The third layer is the first period LSTM (PL) layer, namely the Period LSTM (PL) layer, which has an LSTM-skip (LS) layer, a Dropout layer, and a layer normalization (LN) layer. The short-term features extracted by the ReLU layer with a linear rectified unit (ReLU) in the second layer are input into the first period LSTM (PL) layer in the third layer to further extract long-term features.

[0101] In particular, the present invention introduces a period parameter p in the LSTM-skip (LS) layer (referring to introducing the period parameter p for skipping), and multiplies the period parameter p by the multiple m to enhance the extraction of the longer period depth sequence features in the LS layer. The period parameter p represents the number of hidden layer units skipped. The period parameter p and the multiple m are added to the time step of the LS layer. The specific method is as follows:

[0102] In the first step of the LS layer, the output of the previous layer is filtered through the forget gate, such as the cell state C from the previous step. t-mp After one-dimensional convolution and extraction, v t is input into the forget gate and then combined with the hidden state u of the previous time step t-mp t-mp Finally, the present invention uses the sigmoid activation function to output the comprehensive result as 0 or 1, and then passes it to C t-mp , where 0 means deletion and 1 means retention. t It is expressed as:

[0103] f t =σ(W fv v t +W fu u t-mp +b f )

[0104] Among them, W fu and W fv represents the weight matrix to be trained, b f represents the bias vector added in the forget gate, σ represents sigmoid(x)=(1 / 1+e -x ) activation function.

[0105] In the second step of the LS layer, the input gate determines the cell state C t This is divided into two steps: (1) using the sigmoid function to determine the retention and update of information; (2) creating a new unit state. The present invention uses the ReLU function to update information and transfer the information to the unit state C at the current time t t . Input gate i t The expression is: t =σ(W iv v t +W iu u t-mp +b i )

[0106]

[0107] in, represents the weight matrix to be trained, Represents the bias added in the forget gate.

[0108] After these two steps, the current cell state C t The previous cell state C will be merged t-mp and the new cell state computed by the input gate Current cell status C t The calculation formula is:

[0109]

[0110] in, Represents the element-by-element product.

[0111] In the third step of the LS layer, the output gate is calculated: the unit uses the sigmoid function to determine the output value, and then uses the ReLU function to combine the calculation result with the unit state C t Multiply. Unit Status C t The output gate o t And the final output z t They are: o t =σ(W ov v t +W ou u t-mp +b o )

[0112]

[0113] Among them, W ou , W ov represents the weight matrix to be trained, b o Represents the bias added in the forget gate. The advantage of the LSTM with period p and multiple m proposed in the present invention is that it is very easy to extract important features of long spans of well logging curve data. The period parameter p can be easily adjusted in the LS layer, that is, adjusted according to the logging time interval of the input logging data.

[0114] After the LS layer, the present invention uses the Layer normalization (LN) algorithm in the prior art to accelerate the convergence speed and improve the training efficiency. The LN algorithm can prevent the internal covariate shift caused by the accumulation of errors during the loop iteration process and reduce overfitting. The output of the LN layer is expressed as:

[0115]

[0116] Among them, γ and β represent the scale parameter and offset parameter of the LN layer respectively, and ∈ represents a small value. Then, a dropout layer (the filtering rate is set to 0.2) is used to further alleviate the overfitting problem.

[0117] After the present invention uses four periodic LSTM (PL) layers to extract long-term deep sequence features, specifically, the fourth, fifth, and sixth layers in the present invention are the second, third, and fourth periodic LSTM (PL) layers, respectively, and their structures and principles are consistent with the first LSTM (PL) layer structure and principle. Of course, the number of periodic LSTM (PL) layers can also be determined according to the actual needs of the logging data. Four Transformer encoders (TE) are further used to further extract the features of the logging data, and the number of Transformer encoders (TE) is the same as the number of periodic LSTM (PL) layers.

[0118] The seventh layer is the fourth TE layer. For the TE layer, the most important structure is the attention mechanism (Self-attention (SA)), which focuses on important features by reallocating different weights to the input sequence, thereby improving the relevance of position embedding. Different from the multi-head SA used in the prior art, the present invention uses a single-head attention mechanism (SA). The detailed structure includes:

[0119] Self-attention layer (SA), second normalization layer, feedforward layer (FFW), second Dropout layer, third normalization layer; and the output of the self-attention layer (SA) is connected to the second normalization layer, the output of the second normalization layer is connected to the feedforward layer, the output of the feedforward layer (FFW) is connected to the second Dropout layer (FFW), and the output of the second Dropout layer is connected to the third normalization layer; and two residual connections are established before the second normalization layer and the third normalization layer, respectively. The first residual connection is located before the second normalization layer, and the input data is the data entering the Transformer encoder; the second residual connection is located before the third normalization layer, and the input data is the output data of the second Dropout layer; that is, the TE block has two residual connections before the two LN layers to extract long-term features. The important features of the previous layer will generate three feature spaces when input to SA, and the weights will be redistributed. The three feature space matrices are Q, K and value V, which are expressed as follows:

[0120] Q=LW Q ,K=LW K , V=LW V

[0121] Among them, W V W K W Q Represents the weight matrix, and L represents the input of the SA mechanism. SA assigns weights through dot product attention, that is, after the dot product of Q and each K is divided by c, expressed as:

[0122]

[0123] Among them, c represents the dimension of the key vector. is a probabilistic activation function used to calculate the similarity score and normalize the output to (0,1). After that, the output of the SA layer is input to LN through a residual connection with the input L, which can be expressed as:

[0124] LN1(L)=L+SA(L)

[0125] For the TE layer, each TE layer contains a FFW network with two hidden FFW layers. The output of FFW can be expressed as:

[0126] FFW(LN1)=ReLU(LN1W1+b1)W2+b2

[0127] Where W1, W2, b1, b2 are the weights and bias parameters that FFW needs to train. Before inputting the second LN layer, the present invention adds a dropout layer to avoid overfitting. In addition, the second residual connection in the TE block is connected between the output of LN1 and the output of the dropout layer, and the final output LN2 is expressed as:

[0128] LN2(L)=LN1+FFW(LN1)

[0129] Similarly, the present invention uses four encoding layers (TE layers) to extract highly robust features. Specifically, the eighth, ninth, and tenth layers in the present invention are the third, second, and first Transformer encodings, respectively. Their structure and principle are consistent with the structure and principle of the fourth encoding layer (TE layer). Of course, the number of Transformer encoders (TE layers) is the same as the number of periodic LSTM (PL) layers. In particular, the present invention adds four skip connections between PL and TE blocks to enhance the learning representation ability from low-level to high-level features, specifically:

[0130] The first jump connection unit is used to establish a jump connection between the first period LSTM layer and the first encoding layer. Specifically, the output of the first period LSTM layer is superimposed with the output of the first encoding layer and input into the first Dense layer;

[0131] The second jump connection unit is used to establish a jump connection between the second period LSTM layer and the second encoding layer, specifically, the output of the second period LSTM layer is superimposed with the output of the second encoding layer and inputted into the first encoding layer;

[0132] The third jump connection unit is used to establish a jump connection between the third period LSTM layer and the third encoding layer, specifically, the output of the third period LSTM layer is superimposed with the output of the third encoding layer and inputted into the second encoding layer;

[0133] The fourth jump connection unit is used to establish a jump connection between the fourth period LSTM layer and the fourth encoding layer, specifically, the output of the fourth period LSTM layer is superimposed with the output of the fourth encoding layer and input into the third encoding layer;

[0134] Finally, in the present invention, the output of the first encoding layer is connected to the first Dense layer, and the output of the first Dense layer is connected to the first Sigmoid layer to output the predicted quantitative gamma value.

[0135] It can be seen that the present invention first uses a ReLU layer to preliminarily extract the short-term depth sequence features of the input logging data, and then inputs it into multiple superimposed PL blocks with period parameters p and multiples m to extract important long-span features. Afterwards, multiple superimposed TE blocks with SA and two residual connection mechanisms are used to extract highly robust features from the input sequence features. Furthermore, the present invention also adds multiple residual connections between the PL and TE blocks to enhance the learning ability from low-level to high-level feature representation. In specific applications, the present invention adjusts the parameters and analyzes their sensitivity before testing the prediction performance on a real data set. For the one-dimensional convolutional layer, the kernel size depends on the type of input logging data. For example, when applied to a real data set, the kernel size can be set to 3, because the input well logging curve data is acoustic well logging curve (AC), density well logging curve (DEN) and GR well logging curve (GR). Of course, the kernel size can also be selected according to the actual well logging curve. The one-dimensional convolution layer has 48 kernels, the kernel size is 3×3, and the step size is set to 1. In each PL block in this model, the LS layer contains 36 units, the period p and multiple m of the LS layer are set to 8 and 1 respectively, and there is a dropout layer with a dropout rate of 0.2 in front of the LN layer. For another block, the number of TE blocks is consistent with the LP block. Each TE block contains a single-head SA block, a FFW network, a dropout layer and two LN layers. The number of units in the FFW network is 64, and the dropout rate is 0.2. There are multiple corresponding residual connections between the LP block and the TE block for sharing low-level and high-level features.

[0136] Furthermore, the present invention finds that directly connected skip connections may lead to information redundancy and may amplify noise interference, thereby reducing the expressiveness of features. The present invention further uses a weighted hybrid skip connection unit for skip connections, so that the model selectively focuses on important features of different PL modules and ignores irrelevant features.

[0137] (I) Fixed Weighted Hybrid Skip Connection

[0138] In order to further enhance the learning representation ability from low-level to high-level features, weight coefficients are assigned to the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer respectively, and then mixed and superimposed. The result after mixed and superimposed is:

[0139] PLSTM fuse =w1×PLSTM1+w2×PLSTM2+w3×PLSTM3+w4×PLSTM4, where PLSTM fuseis the result after mixed superposition, PLSTM1, PLSTM2, PLSTM3, and PLSTM4 are the output results of the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer in the i-th iteration respectively, and w1, w2, w3, and w4 are the weight coefficients assigned to the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer respectively.

[0140] Then, the superimposed result PLSTM fuse respectively:

[0141] Superimpose the output of the first encoding layer and input it into the first Dense layer;

[0142] The output of the second coding layer is superimposed and input into the first coding layer;

[0143] The output is superimposed with the output of the third coding layer and input into the second coding layer;

[0144] The output is superimposed with the output of the fourth coding layer and input into the third coding layer;

[0145] Table 1 below shows the weight coefficient values ​​adopted, where w1, w2, w3, and w4 are the weight coefficients assigned to the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer, respectively.

[0146] Table 1: Fixed weighted hybrid jump connection weight coefficient values

[0147]

[0148] The above method can realize fixed weighted hybrid jump connection.

[0149] (II) Dynamic Weighted Hybrid Skip Connection

[0150] The present invention finds that respectively allocating fixed weight coefficients to the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer and then mixing and superimposing them cannot accurately reflect the long-term depth sequence characteristics of the logging curve data extracted by the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer in each iteration process during the iteration cycle, thereby reducing the expressiveness of the long-term depth sequence characteristics of the logging curve data. Therefore, the present invention further proposes to dynamically allocate weight coefficients with each iteration, and the method for dynamically allocating weight coefficients is as follows: Step A: Convolution: Convolution operations are performed on the outputs of the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer based on 1D conv with a convolution kernel size of N; the convolution kernel size can be selected according to the actual logging data, and the embodiment of the present invention adopts N=3;

[0151] Step B: Sum: Sum the output of the first cycle LSTM layer after 1D conv, the output of the second cycle LSTM layer, the output of the third cycle LSTM layer, and the output of the fourth cycle LSTM layer;

[0152] P k =conv K (PLSTM1 k )+conv K (PLSTM2 k )+conv K (PLSTM3 k )+conv K (PLSTM4 k ), where P k For global information, PLSTM1 k 、PLSTM2 k 、PLSTM3 k 、PLSTM4 k are the output results of the first cycle LSTM layer, the output results of the second cycle LSTM layer, the output results of the third cycle LSTM layer, and the output results of the fourth cycle LSTM layer in the kth iteration, respectively. K It is a 1D convolution with a kernel size of K;

[0153] Step C: Calculate the weight coefficient w1 at the current iteration time k k 、w2 k 、w3 k 、w4 k :

[0154] w1 k =conv K (PLSTM1 k) / [conv K (PLSTM1 k )+conv K (PLSTM2 k )+conv K (PLSTM3 k )+conv K (PLSTM4 k )]

[0155] w2 k =conv K (PLSTM2 k ) / [conv K (PLSTM1 k )+conv K (PLSTM2 k )+conv K (PLSTM3 k )+conv K (PLSTM4 k )]

[0156] w3 k =conv K (PLSTM3 k ) / [conv K (PLSTM1 k )+conv K (PLSTM2 k )+conv K (PLSTM3 k )+conv K (PLSTM4 k )]

[0157] w4 k =conv K (PLSTM4 k ) / [conv K (PLSTM1 k )+conv K (PLSTM2 k )+conv K (PLSTM3 k )+conv K (PLSTM4 k )]

[0158] Step D: At the current iteration time k, the output results of the first cycle LSTM layer, the output results of the second cycle LSTM layer, the output results of the third cycle LSTM layer, and the output results of the fourth cycle LSTM layer are mixed and superimposed. The results are as follows: PLSTM fusek =w1 k ×PLSTM1 k +w2 k ×PLSTM2 k +w3 k ×PLSTM3 k +w4 k ×PLSTM4 k ;

[0159] Step E: The mixed superposition results are respectively: superimposed with the output of the first coding layer and input into the first Dense layer; superimposed with the output of the second coding layer and input into the first coding layer; superimposed with the output of the third coding layer and input into the second coding layer; and superimposed with the output of the fourth coding layer and input into the third coding layer.

[0160] The above method can realize fixed weighted hybrid jump connection.

[0161] The present invention is described below with respect to the loss function:

[0162] In the fields of machine learning, deep learning, and optimization, the loss function plays a vital role. It is a mathematical expression that measures the difference or error between the model's prediction results and the actual observed values, and is the basis for guiding parameter adjustment during model training. The choice of loss function directly affects the performance of the model, the efficiency of training, and the accuracy of the final result. The basic idea of ​​the loss function is to quantify the inaccuracy of the model's prediction by calculating a scalar value (i.e., the loss value). In supervised learning tasks, there is usually a set of known training data, including input features and target outputs. The loss function calculates the loss value based on the difference between the model's predicted output and the true target output. Ideally, the model parameters are continuously adjusted through iterative optimization algorithms (such as gradient descent) so that the loss value gradually decreases until it reaches an acceptably low level, at which point the model is considered to have fitted the training data well. Common types of loss functions include mean squared error,

[0163] Researchers can design custom loss functions based on the needs of specific tasks, such as Mean Squared Error (MSE), Mean Absolute Error (MAE), Cross-Entropy Loss, and others. For example, researchers have proposed many improved and more adaptable loss functions, such as Huber loss (combining the advantages of MSE and MAE), Focal Loss (used to solve the problem of class imbalance), etc., to meet the challenges in different scenarios and to better reflect the nature of the problem or optimize the goal.

[0164] In order to adapt the proposed improved LSTM model, the present invention redesigns the loss function expression to accelerate convergence and improve prediction accuracy in the face of quantitative gamma prediction. The principle is explained as follows:

[0165] Considering that the uncertainty of the prediction actually depends on the input values, the heteroskedasticity assumption is adopted. For this purpose, the following loss function is used:

[0166]

[0167] Among them, f(x i ) and σ(x i ) 2 Represents a specific input x i The output prediction mean and variance, y i Represents the training data label value, and N represents the number of sample points. The loss function in the formula is proportional to the negative log-likelihood based on the Gaussian distribution assumption. Therefore, using the loss function in the formula to train a deep neural network is to find the model parameters that minimize the loss function.

[0168] When the predicted mean value f(x i ) and the training data yi are large, the algorithm will rely on reducing the value of the first term in the formula, which tends to make the corresponding conditional variance σ(x i ) 2 However, the presence of the second term in the formula prevents the conditional variance from being too large, which in turn helps the deep learning model learn the training input-output relationship instead of only providing the same relationship with x in a specific training dataset. i In other words, this loss function helps to find the structure of the input-output relationship, while training models with other loss functions requires additional efforts, such as cross-validation, to achieve a trade-off between bias and variance.

[0169] In addition, in order to further adapt the proposed improved LSTM model, the present invention also designs a hybrid loss function expression to accelerate convergence and improve prediction accuracy in the face of quantitative gamma prediction.

[0170] The hybrid loss function can provide smooth gradients and handle class imbalance, capturing both large-scale structures and small-scale fine structures. The expression of the hybrid loss function is as follows:

[0171] L Fusion =0.5*L1+L2+0.5*L3

[0172]

[0173] According to the above formula and simplified as follows:

[0174]

[0175] Among them, α and β are control parameters. α controls the common weight of positive samples and negative samples on the total loss, while β controls the weight of easy-to-classify samples and difficult-to-classify samples. Y represents the label value. represents the predicted value, Y i ∈Y, N represents the sample points, (·) + =max(·,0). After verification, it is found that setting α=0.25 and β=2 produces the best experimental results.

[0176] In order to accurately measure the superiority of the designed loss function, the present invention adopts several evaluation criteria commonly used in deep learning to measure it.

[0177] SSIM (Structural Similarity Index): Evaluates their similarity. The closer the SSIM value is to 1, the more similar they are. The calculation formula is: Where μ is the mean, σ is the standard deviation, and c1 and c2 are small constants to avoid zero denominator.

[0178] SSE (Sum of Squared Errors): SSE is the sum of the squares of the differences between the predicted value and the true value. The smaller the SSE value, the more accurate the model's prediction. Calculation formula where y i is the true value, is the predicted value.

[0179] SAE (Sum of Absolute Errors): is the absolute sum of the differences between the predicted value and the true value. Similar to SSE,

[0180] The smaller the SAE value, the more accurate the model's prediction. Calculation formula

[0181] The evaluation results of the loss function are shown in Table 2:

[0182] Table 2: Evaluation results of various loss functions

[0183] Loss function SSIM SSE SAE <![CDATA[L MSE ]]> 0.8917 5.6125 158.3493 <![CDATA[L MAE ]]> 0.9898 0.5530 45.0632 <![CDATA[L New ]]> 0.9933 0.4485 34.3173 <![CDATA[L BCE ]]> 0.9910 0.5868 44.5901 <![CDATA[L Huber ]]> 0.9916 0.4762 41.3430 <![CDATA[L FOCAL ]]> 0.9922 0.5161 40.3446 <![CDATA[L Fusion ]]> 0.9932 0.3379 34.6522

[0184] It can be seen from the table that L Fusion It has the highest SSIM value of 0.9932, which means it does the best job in maintaining similarity. New and L Fusion Both have lower SSE values, 0.4485 and 0.3379, respectively, which indicates that their prediction errors are smaller. FusionWith the lowest SAE value of 34.6522, L New They also have a lower SAE value of 34.3173, which further confirms their advantage in reducing prediction errors. Combining these indicators, we can see that L Fusion It performs well in all three metrics, with the highest SSIM value and the lowest SSE and SAE values, which indicates that it minimizes the prediction error while maintaining the similarity. New It also performs well, especially in terms of SSE and SAE metrics, although its SSIM value is slightly lower than that of L Fusion , but it is still close to 1, indicating that it is very effective in reducing errors. In order to accurately measure the prediction performance of the deep learning model, the embodiment of the present invention uses three accuracy evaluation indicators, namely precision, F1 score (F1_Score) and intersection over union (IoU), to demonstrate the learning ability of the model. The principles and calculation methods of precision, F1 score and intersection over union (IoU) are all prior art.

[0185] Figure 6-Figure 8 The figure shows the F1-score variation with Epoch, IoU variation with Epoch, and Precision variation with Epoch of the three skip connection methods provided according to the embodiment of the present invention; it also shows the performance comparison of three deep learning models: direct skip connection, fixed weighted skip connection, and dynamic weighted skip connection during the training process. Figure 6-Figure 8 The changes of different evaluation indicators with training rounds (Epoch) are shown respectively.

[0186] Figure 6 The changes of IoU indicators of the three models during the training process are shown. IoU is an indicator to measure the degree of overlap between the predicted area and the real area. The IoU value of the dynamic weighted skip connection grows the fastest and maintains the lead in the later stage, showing its advantage in prediction accuracy. The IoU value of the fixed weighted skip connection grows steadily, but the final value is lower than that of R2SE-Unet, indicating that its performance is slightly inferior. The IoU value of the direct skip connection grows slowly, and the final value is lower than that of the other two models. It works well in processing this task, but is not as effective as the other two models in terms of accuracy.

[0187] Figure 7The changes in the F1-score of the three models during training are shown. The F1-score is the harmonic mean of precision and recall, which is used to measure the accuracy and recall of the model. The F1-score of the dynamic weighted skip connection rises rapidly in the early stage and maintains the lead in the later stage, showing its advantage in balancing precision and recall. The F1-score of the fixed weighted skip connection grows steadily, but the final value is lower than that of R2SE-Unet, indicating that its performance is slightly inferior. The F1-score of the direct skip connection grows slowly, and the final value is lower than that of the other two models, which is consistent with its performance on IoU.

[0188] Figure 8 The changes in the precision of the three models during the training process are shown. Precision refers to the proportion of samples predicted to be positive that are actually positive. The precision of the dynamic weighted skip connection fluctuates greatly in the early stage, but then rises rapidly and maintains the lead, showing its advantage in reducing false positives. The precision of the fixed weighted skip connection grows steadily, but the final value is lower than that of R2SE-Unet, indicating that its performance is slightly inferior. The precision of the direct skip connection grows slowly, and the final value is lower than that of the other two models, which is consistent with its performance on F1-score.

[0189] Combining these three indicators, the dynamic weighted skip connection performs well in terms of loss value, IoU, F1-score and precision, showing that it has the best performance among the three models. The direct skip connection does not perform as well as the other two models in all indicators, and further adjustment or optimization is needed in practical applications. The performance of the fixed weighted skip connection is between the two.

[0190] Although stable, it is slightly inferior to the dynamic weighted jump connection in performance and can be selected based on the actual side logging curve input.

[0191] Fig. 9 The figure shows the result of the first embodiment of quantitative gamma value prediction of in-situ leaching sandstone type uranium ore provided by the embodiment of the present invention. For the convenience of description, only the part related to the embodiment of the present invention is shown, which is described in detail as follows: Fig. 9 The figure shows the prediction results of three logging curves, AC, DEN, and GR (improved LSTM model with direct skip connection). It can be seen from the figure that the prediction results are consistent with the trend of the GR curve, showing a better prediction result.

[0192] Fig.10 This is a result diagram of the third embodiment of quantitative gamma value prediction of in-situ leaching sandstone-type uranium ore provided in an embodiment of the present invention; for ease of description, only the part related to the embodiment of the present invention is shown, which is described in detail as follows: Fig.10The figure shows the prediction results of the four logging curves GR, CAL, LLD, and AC (improved LSTM model with fixed weighted skip connections). It can be seen from the figure that the prediction results show good prediction results.

[0193] Fig.11 This is a result diagram of the third embodiment of quantitative gamma value prediction of in-situ leaching sandstone-type uranium ore provided in an embodiment of the present invention; for ease of description, only the part related to the embodiment of the present invention is shown, which is described in detail as follows: Fig.11 The figure shows the prediction results of 8 logging curves, namely GR, DPAZ, CALC, CI, DEV, LSPD, R250, and SP (improved LSTM model with dynamic weighted skip connection). It can be seen from the figure that the prediction results show good prediction results.

[0194] Fig.12 This is a diagram of a quantitative gamma value prediction device for in-situ leaching sandstone-type uranium ore provided according to an embodiment of the present invention. For ease of description, only the parts related to the embodiment of the present invention are shown, which are described in detail as follows:

[0195] A quantitative gamma value prediction device for in-situ leaching sandstone type uranium ore, comprising:

[0196] Training Unit:

[0197] The first acquisition module: through experiments and logging field operations, the logging curve data and quantitative gamma data sample sets of the research block are collected;

[0198] Furthermore, the logging curve data used in the present invention may include: resistivity logging curve: deep lateral resistivity (RD), shallow lateral resistivity (RS); porosity logging curve: compensated neutron (CNL), density (DEN), acoustic wave (AC); natural gamma logging curve: natural gamma (GR); other logging curves: spontaneous potential (SP), well diameter (CALC), uranium-free gamma (KTh), energy spectrum natural gamma (GRSL), uranium (U), thorium (Th), potassium (K), bottom gradient resistivity (R250), well inclination angle (DEV), well inclination azimuth (DPAZ), logging velocity (LSPD), coal seam mark (CI), etc.

[0199] A preprocessing module: performs a first preprocessing on the logging curve data and the quantitative gamma data, wherein the first preprocessing includes: depth correction, wherein the depth correction is to perform depth correction on the logging curve data based on the quantitative gamma depth; a construction module: constructs a quantitative gamma training data set for in-situ leaching sandstone-type uranium ore based on the logging curve data and the quantitative gamma data after the first preprocessing;

[0200] Training and learning module: using the quantitative gamma training data set of in-situ leaching sandstone-type uranium deposits to train and improve the LSTM model;

[0201] Prediction Unit:

[0202] The second acquisition module: collects the logging curve data of the target block to be predicted through logging field operations;

[0203] Quantitative prediction module: Based on the improved LSTM model trained in step S4, the well logging curve data of the target block to be predicted is input into the improved LSTM model trained in step S4 to perform quantitative gamma value prediction;

[0204] Output module: outputs the quantitative gamma value prediction results of the target block to be predicted.

[0205] An embodiment of the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the method for predicting the quantitative gamma value of in-situ leaching sandstone-type uranium ore are implemented.

[0206] An embodiment of the present invention further provides a computer storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of a method for predicting a quantitative gamma value of an in-situ leaching sandstone-type uranium ore are implemented.

[0207] It should be understood by those skilled in the art that the embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The present invention is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, may be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the functions in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0208] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0209] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0210] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting the quantitative gamma value of in-situ leaching sandstone-type uranium ore, characterized in that: include: Training phase: Step S1: Collect the well logging curve data and quantitative gamma data sample set of the study block through experiments and well logging field operations; Step S2: performing a first preprocessing on the logging curve data and the quantitative gamma data, wherein the first preprocessing includes: depth correction, wherein the depth correction is to perform depth correction on the logging curve data based on the quantitative gamma depth; Step S3: constructing a quantitative gamma training data set for in-situ leaching sandstone-type uranium deposits based on the well logging data and quantitative gamma data after the first preprocessing; Step S4: using the quantitative gamma training data set of in-situ leaching sandstone-type uranium deposits to train an improved LSTM model; Prediction stage: Step S5: collecting logging curve data of the target block to be predicted through logging field operations; Step S6: Based on the improved LSTM model trained in step S4, the well logging curve data of the target block to be predicted is input into the improved LSTM model trained in step S4 to perform quantitative gamma value prediction; Step S7: outputting the prediction result of the quantitative gamma value of the target block to be predicted.

2. The method for predicting the quantitative gamma value of in-situ leaching sandstone-type uranium ore according to claim 1, characterized in that: In the step S4, the quantitative gamma training data set of in-situ leaching sandstone-type uranium ore is used to train the improved LSTM model, wherein the improved LSTM model is improved on the LSTM model, and the model includes: First convolutional layer, first ReLU layer, first cycle LSTM layer, second cycle LSTM layer, third cycle LSTM layer, the fourth cycle LSTM layer, the fourth encoding layer, the third encoding layer, the second encoding layer, the first encoding layer, the first Dense layer, and the first Sigmoid layer, and the output of the first convolution layer is connected to the first ReLU layer, the output of the first ReLU layer is connected to the first cycle LSTM layer, the output of the first cycle LSTM layer is connected to the second cycle LSTM layer, and the output of the second cycle LSTM layer is connected to the third cycle LSTM layer; the output of the third cycle LSTM layer is connected to the fourth cycle LSTM layer, the output of the fourth cycle LSTM layer is connected to the fourth encoding layer, the output of the fourth encoding layer is connected to the third encoding layer, the output of the third encoding layer is connected to the second encoding layer, the output of the second encoding layer is connected to the first encoding layer, the output of the first encoding layer is connected to the first Dense layer, and the output of the first Dense layer is connected to the first Sigmoid layer. The data flow is input from the first convolution layer and finally output by the first Sigmoid layer. The first cycle LSTM layer structure includes: The first LSTM-skip (LS) layer, the first Dropout layer, and the first normalization (LN) layer; the output of the first LSTM-skip (LS) layer is connected to the first Dropout layer, and the output of the first Dropout layer is connected to the first normalization (LN) layer. The first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer have the same structure; the first encoding layer is a Transformer encoder (TE), which specifically includes: Self-attention layer (SA), second normalization layer, feedforward layer (FFW), second Dropout layer, third normalization layer; and the output of the self-attention layer (SA) is connected to the second normalization layer, the output of the second normalization layer is connected to the feedforward layer, the output of the feedforward layer is connected to the second Dropout layer, and the output of the second Dropout layer is connected to the third normalization layer; and two residual connections are established before the second normalization layer and the third normalization layer respectively, the first residual connection is located before the second normalization layer, and the input data is the data entering the Transformer encoder; the second residual connection is located before the third normalization layer, and the input data is the output data of the second Dropout layer; and the structures of the fourth encoding layer, the third encoding layer, the second encoding layer, and the first encoding layer are the same; The first jump connection unit is used to establish a jump connection between the first period LSTM layer and the first encoding layer. Specifically, the output of the first period LSTM layer is superimposed with the output of the first encoding layer and input into the first Dense layer; The second jump connection unit is used to establish a jump connection between the second period LSTM layer and the second encoding layer, specifically, the output of the second period LSTM layer is superimposed with the output of the second encoding layer and inputted into the first encoding layer; The third jump connection unit is used to establish a jump connection between the third period LSTM layer and the third encoding layer, specifically, the output of the third period LSTM layer is superimposed with the output of the third encoding layer and inputted into the second encoding layer; The fourth jump connection unit is used to establish a jump connection between the fourth period LSTM layer and the fourth encoding layer, specifically, the output of the fourth period LSTM layer is superimposed with the output of the fourth encoding layer and input into the third encoding layer; The jump connection modes of the first, second, third and fourth jump connection units are all the same.

3. The method for predicting the quantitative gamma value of in-situ leaching sandstone-type uranium ore according to claim 2, characterized in that: The structures of the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer are the same, wherein the first cycle LSTM layer is: The period parameter p and the multiple m are introduced into the time step of the LSTM layer. The period parameter p represents the number of hidden layer units skipped, and the period parameter p is multiplied by the multiple m to enhance the extraction of longer period deep sequence features in the LSTM layer.

4. The method for predicting the quantitative gamma value of in-situ leaching sandstone-type uranium ore according to claim 3, characterized in that: The jump connection modes of the first, second, third and fourth jump connection units are all the same, and the jump connection mode is a fixed weighted hybrid jump connection, which is: Step S11: Setting weight coefficients: setting weight coefficients w1, w2, w3, and w4 for the output results of the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer, respectively; Step S22: Mixing and superposition: PLSTM fuse =w1×PLSTM1+w2×PLSTM2+w3×PLSTM3+w4×PLSTM4, where PLSTM fuse is the result after mixed superposition, PLSTM1, PLSTM2, PLSTM3, and PLSTM4 are the output results of the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer in the kth iteration, respectively; w1, w2, w3, and w4 are the weight coefficients assigned to the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer, respectively; Step S33: Skip connection processing: PLSTM the result after mixing and superimposing in step S22 fuse Respectively: superimposed with the output of the first encoding layer and input into the first Dense layer; It is superimposed with the output of the second coding layer and input to the first coding layer; it is superimposed with the output of the third coding layer and input to the second coding layer; and it is superimposed with the output of the fourth coding layer and input to the third coding layer.

5. The method for predicting the quantitative gamma value of in-situ leaching sandstone-type uranium ore according to claim 3, characterized in that: The jump connection modes of the first, second, third and fourth jump connection units are all the same, and the jump connection mode is a dynamic weighted hybrid jump connection, which is: Step S111: convolution: performing convolution operations on the outputs of the first cycle LSTM layer, the second cycle LSTM layer, the third cycle LSTM layer, and the fourth cycle LSTM layer based on 1D conv with a convolution kernel size of K; Step S222: Sum: sum the output of the first cycle LSTM layer after 1D conv, the output of the second cycle LSTM layer, the output of the third cycle LSTM layer, and the output of the fourth cycle LSTM layer; P k =conv K (PLSTM1 k )+conv K (PLSTM2 k )+conv K (PLSTM3 k )+conv K (PLSTM4 k ), where P k For global information, PLSTM1 k 、PLSTM2 k 、PLSTM3 k 、PLSTM4 k are the output results of the first cycle LSTM layer, the output results of the second cycle LSTM layer, the output results of the third cycle LSTM layer, and the output results of the fourth cycle LSTM layer in the kth iteration, respectively. K It is a 1D convolution with a kernel size of K; Step S333: Set weight coefficient: Calculate the weight coefficient w1 at the current iteration time k k 、w2 k 、w3 k 、w4 k : w1 k =conv K (PLSTM1 k ) / [conv K (PLSTM1 k )+conv K (PLSTM2 k )+conv K (PLSTM3 k )+conv K (PLSTM4 k )] w2 k =conv K (PLSTM2 k ) / [conv K (PLSTM1 k )+conv K (PLSTM2 k )+conv K (PLSTM3 k )+conv K (PLSTM4 k )] w3 k =conv K (PLSTM3 k ) / [conv K (PLSTM1 k )+conv K (PLSTM2 k )+conv K (PLSTM3 k )+conv K (PLSTM4 k )] w4 k =conv K (PLSTM4 k ) / [conv K (PLSTM1 k )+conv K (PLSTM2 k )+conv K (PLSTM3 k )+conv K (PLSTM4 k )] Step S444: Mixed superposition: The output result of the first cycle LSTM layer, the output result of the second cycle LSTM layer, the output result of the third cycle LSTM layer, and the output result of the fourth cycle LSTM layer are mixed and superimposed at the current iteration time k. The result is as follows: PLSTM fuse k =w1 k ×PLSTM1 k +w2 k ×PLSTM2 k +w3 k ×PLSTM3 k +w4 k ×PLSTM4 k ; Step S555: skip connection processing: the result of mixing and superimposing in step S444 is PLSTM fuse k Respectively: superimposed with the output of the first coding layer and input into the first Dense layer; superimposed with the output of the second coding layer and input into the first coding layer; superimposed with the output of the third coding layer and input into the second coding layer; superimposed with the output of the fourth coding layer and input into the third coding layer.

6. The method for predicting the quantitative gamma value of in-situ leaching sandstone-type uranium ore according to claim 3, characterized in that: The loss function used in the improved LSTM model is: Among them, f(x i ) and σ(x i ) 2 Represents a specific input x i The output prediction mean and variance, y i Represents the training data label value, and N represents the number of sample points.

7. The method for predicting the quantitative gamma value of in-situ leaching sandstone-type uranium ore according to claim 3, characterized in that: The loss function used in the improved LSTM model is: Among them, α and β are control parameters. α controls the common weight of positive samples and negative samples on the total loss, and β controls the weight of easy-to-classify samples and difficult-to-classify samples. Y represents the label value. represents the predicted value, Y i ∈Y, N represents the sample points.

8. A quantitative gamma value prediction device for in-situ leaching sandstone type uranium ore, characterized in that: include: Training Unit: The first acquisition module: through experiments and logging field operations, the logging curve data and quantitative gamma data sample sets of the research block are collected; A preprocessing module: performs a first preprocessing on the logging curve data and the quantitative gamma data, wherein the first preprocessing includes: depth correction, wherein the depth correction is to perform depth correction on the logging curve data based on the quantitative gamma depth; a construction module: constructs a quantitative gamma training data set for in-situ leaching sandstone-type uranium ore based on the logging curve data and the quantitative gamma data after the first preprocessing; Training and learning module: using the quantitative gamma training data set of in-situ leaching sandstone-type uranium deposits to train and improve the LSTM model; Prediction Unit: The second acquisition module: collects the logging curve data of the target block to be predicted through logging field operations; Quantitative prediction module: Based on the improved LSTM model trained in step S4, the well logging curve data of the target block to be predicted is input into the improved LSTM model trained in step S4 to perform quantitative gamma value prediction; Output module: outputs the quantitative gamma value prediction results of the target block to be predicted.

9. An electronic device, characterized in that: include: Memory for storing computer programs; A processor is used to implement the steps of the method for predicting the quantitative gamma value of in-situ leaching sandstone-type uranium ore as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method for predicting the quantitative gamma value of in-situ leaching sandstone-type uranium ore according to any one of claims 1 to 7 are implemented.

Images