Deep complex pyroclastic rock identification method based on neural network

Abstract

The invention provides a deep complex volcaniclastic rock identification method based on a neural network. The method comprises the following steps: extracting feature data; performing label coding and data normalization processing on the logging data sample; establishing a deep complex volcaniclastic rock recognition model based on the neural network; and optimizing the deep complex volcaniclastic rock recognition model of the neural network. According to the deep complex volcaniclastic rock identification method based on the neural network, a Dropout mechanism of a multi-dimensional activation function and a hierarchical prediction method are introduced into the BP neural network model, and an over-fitting phenomenon is avoided by reducing the complexity of the neural network model and refining a step-by-step prediction standard; a new prediction model is established for deep complex volcaniclastic rock reservoir lithology identification, and the lithology identification method and system which are more accurate, objective, high in feasibility and credibility and lower in economic cost are obtained.

Description

Technical Field

[0001] The invention relates to the technical field of oil and gas exploration, and in particular to a method for identifying deep complex volcaniclastic rocks based on a neural network. Background Art

[0002] Pyroclastic oil and gas reservoirs have become a new field in the exploration and development of oil and gas fields in the world. The continuous and accurate identification of pyroclastic rocks is the basis for the exploration of pyroclastic oil and gas. Considering the cost issues in the exploration and development process, the core data obtained are relatively small, the sources of cores and cuttings are limited, and the construction of lithology sequences of the entire well section cannot be achieved using only geological samples. In recent years, predecessors have successively carried out lithology identification research based on conventional logging data. However, the logging response of deep pyroclastic rocks is affected by the comprehensive influence of rock composition, structural structure and diagenesis. In areas with complex lithology, the recognition rate of conventional crossplot method is significantly low; in the deep drilling process, the complexity of geological structure and the ambiguity of logging parameter distribution often lead to large errors in lithology identification methods based on traditional logging. BP neural network can construct lithology identification mode through its own unique sample learning ability, which overcomes the defects of fuzzy mathematics method, gray clustering method and multivariate statistical method. BP neural network has the characteristics of self-organization, self-learning, self-adaptation, fault tolerance and anti-interference, and the lithology identification results are objective and reliable.

[0003] However, due to the overfitting phenomenon, the BP neural network has a small training error during the training process, but the error becomes very large in the prediction stage. Therefore, the universality of using BP neural network for lithology prediction is poor. Summary of the invention

[0004] In view of the above technical problems, the purpose of the present invention is to provide a method for identifying deep complex volcaniclastic rocks based on neural networks, introduce the Dropout mechanism and hierarchical prediction method of multidimensional activation functions into the BP neural network model, and avoid overfitting by reducing the complexity of the neural network model and refining the step-by-step prediction standards. A new prediction model is established for the lithology identification of deep complex volcaniclastic reservoirs, and a more accurate, objective, feasible and reliable lithology identification method and system with lower economic cost are obtained.

[0005] To achieve the above object, the present invention provides a method for identifying deep complex pyroclastic rocks based on a neural network, comprising the following steps:

[0006] S1, extract feature data;

[0007] S2, label encoding and data normalization processing of logging data samples;

[0008] S3. Establish a deep complex pyroclastic rock identification model based on neural network;

[0009] S4. Optimized neural network model for deep complex pyroclastic rock recognition.

[0010] Optionally, in step S1, a RESNET suitable for optimizing geometric parameters of radio frequency devices is designed and outputted, and the RESNET is composed of residual blocks whose main components are composed of convolutional layers, and the first bits of each residual block are connected.

[0011] Optionally, step S2 includes the following steps:

[0012] S21. The lithology of pyroclastic rocks is divided into three categories according to the SiO2 content: basic, neutral and acidic.

[0013] S22, standardizing the logging data.

[0014] Optionally, step S2 includes the following steps:

[0015] S31, establish a three-layer model structure of input layer, hidden layer and output layer;

[0016] S32, the input layer neurons are normalized logging data sets, i.e., natural gamma, compensated neutron, density, acoustic time difference, and deep lateral resistivity corresponding to the same depth, and the output layer is the pyroclastic rock lithology corresponding to the same depth;

[0017] S33, the input well logging data set is processed by multiple fully connected layers, activation functions and Dropout layers, and finally the output lithology label is obtained;

[0018] S34, by back propagating the comparison between the lithology label and the output lithology, the weight of the neuron is continuously optimized and then the back propagation error is calculated layer by layer to minimize the error signal;

[0019] S35. Build a deep complex pyroclastic rock recognition model based on neural network in pytorch environment.

[0020] Optionally, step S4 includes the following steps:

[0021] S41. The training set is input to the input end of the deep complex pyroclastic rock recognition model of the neural network, and lithology prediction learning is performed through forward propagation. The lithology prediction learning labels are first divided into large-category prediction learning, i.e., basic, neutral, and acidic prediction learning, and then the small-category lithology prediction learning of the large category is performed. The lithology and predicted lithology are compared through back propagation to automatically optimize the neuron weights. Each learning is defined as an iteration.

[0022] S42, the test set is tested, and the data group in the test set is input into the input end of the deep complex pyroclastic rock recognition model of the optimized neural network to predict the corresponding lithology label and perform accuracy evaluation.

[0023] From the above, the deep complex pyroclastic rock identification method based on neural network of the present invention has the following beneficial effects:

[0024] 1. The present invention introduces the Dropout mechanism of multidimensional activation function and the hierarchical prediction method into the BP neural network model, and avoids the overfitting phenomenon by reducing the complexity of the neural network model and refining the prediction criteria at each level. A new prediction model is established for the lithology identification of deep complex volcaniclastic reservoirs, and a more accurate, objective, feasible and reliable lithology identification method and system with low economic cost are obtained.

[0025] 2. The deep complex pyroclastic rock identification model of the neural network trained by the present invention has a high accuracy rate in lithology identification in new work areas and basically meets production needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flow chart of the deep complex pyroclastic rock identification method based on neural network of the present invention. DETAILED DESCRIPTION

[0027] Below, see Figure 1 The neural network-based deep complex pyroclastic rock identification method of the present invention is described in detail.

[0028] The neural network-based deep complex pyroclastic rock identification method of the present invention comprises the following steps:

[0029] S1. Get feature data

[0030] The logging data sensitive to lithology, including natural gamma (GR), compensated neutron (CNL), density (DEN), acoustic time difference (AC), and deep lateral resistivity (RLLD) collected by logging instruments are discretized into data with a sampling interval of 0.2m after depth correction as input features. The lithology results of drilling cores are used as labels to establish a corresponding relationship with the above logging data. The data samples of thin layers, lithology mutation sections, and wellbore collapse sections are eliminated, the sample data are cleaned, the stable lithology section data are retained, and the neural network data samples are established.

[0031] S2. Label encoding and data normalization of logging data samples

[0032] S21. The lithology of pyroclastic rocks is divided into three categories according to the SiO2 content: basic, neutral and acidic, with corresponding codes 0, 1, and 2; volcanic breccia, tuff, welded breccia, welded tuff, brecciated lava, tuff lava, sedimentary volcanic breccia, and sedimentary tuff are divided into twenty-four subcategories according to the SiO2 content, corresponding to 00, 01, 02, 03, 04, 05, 06, 07; 10, 11, 12, 13, 14, 15, 16, 17; 20, 21, 22, 23, 24, 25, 26, and 27.

[0033] S22. Normalize the logging data to [0, 1]. The formula is:

[0034]

[0035] Among them, Y is the normalized result of the logging curve; X is the value of a logging curve at a certain depth; X min is the minimum value of the same logging curve; X max It is the maximum value of the same logging curve.

[0036] S3. Establish a deep complex pyroclastic rock identification model based on neural network

[0037] S31. Establish a three-layer model structure of input layer, hidden layer, and output layer. The number of neurons in the input layer is 5, the hidden layer contains two layers, the number of neurons in the first layer is 100, the number of neurons in the second layer is 50, and the number of neurons in the output layer is 1.

[0038] S32, the input layer neurons are normalized logging data sets, i.e., natural gamma (GR), compensated neutron (CNL), density (DEN), acoustic time difference (AC), and deep lateral resistivity (RLLD) corresponding to the same depth, and the output layer is the pyroclastic rock lithology corresponding to the same depth.

[0039] S33, forward propagation process from input layer to output layer

[0040] In the neural network model of the present invention, the input well logging data set X is processed by multiple fully connected layers, activation functions and Dropout layers to finally obtain the output lithology label.

[0041] The input X is linearly transformed through the first fully connected layer self.fc1. The activation function of the first fully connected layer is the Tanh function. The activation function performs nonlinear processing on the result after linear transformation. This step maps the input data to a new feature space, introduces nonlinear characteristics, and enables the neural network to learn more complex patterns. Then X processed by the activation function passes through the first Dropout layer self.dropout1. The Dropout layer randomly sets the output of 25% of the neurons to 0 during training to reduce the risk of overfitting. The Tanh activation function is:

[0042]

[0043] X then passes through the second fully connected layer self.fc2 for linear transformation. The Sigmoid activation function (torch.sigmoid) is applied to the result of linear transformation for nonlinear processing. The Sigmoid function compresses the output value between 0 and 1. It passes through the Dropout layer self.dropout2 for regularization again, that is:

[0044]

[0045] Finally, x passes through the third fully connected layer self.fc3 to get the final output. Apply the softmax function to convert the output into a probability distribution, and then select the category with the highest probability as the prediction result. That is:

[0046]

[0047] S34. Back propagation process from output layer to input layer

[0048] By back-propagating and comparing the lithology labels and output lithology, the weights of neurons are continuously optimized, and then the back-propagation errors are calculated layer by layer to minimize the error signal. The BP neural network model is optimized through the cross entropy loss function and the stochastic gradient descent Adam optimization algorithm. The specific method is as follows:

[0049] S341. Define the cross entropy loss function

[0050] The cross entropy loss function is used to measure the difference between the probability distribution predicted by the model and the true label. For multi-classification problems, the cross entropy loss function encourages the model to assign high probabilities to the correct category and low probabilities to the wrong category. Assume there are C categories, the true label of the sample is y, and the probability distribution predicted by the model is p. The calculation formula of the cross entropy loss function is:

[0051]

[0052] Among them, yi is the value of the class in the true label (if the sample belongs to the i-th class, then y i =1, otherwise y i =0), P i is the probability that the model predicts that it belongs to the i-th class.

[0053] S342, define the stochastic gradient descent Adam optimization algorithm

[0054] The Adam optimizer is a commonly used optimization algorithm that combines momentum and adaptive learning rate methods. In Adam, a momentum term is introduced, which takes into account the weighted average of previous gradients, making the direction of parameter updates more stable, accelerating convergence and reducing oscillations. The momentum term is calculated by taking the exponential weighted average of past gradients. In this way, when updating parameters, not only the current gradient is considered, but also the trend of previous gradients, making the update smoother. Adaptive learning rate, different parameters may require different learning rates to update during training. The Adam algorithm automatically adjusts the learning rate by estimating the first-order moment (mean) and second-order moment (variance) of each parameter. The first-order moment estimate is used to estimate the mean of the gradient, and the second-order moment estimate is used to estimate the variance of the gradient. By estimating these two moments, the Adam algorithm can calculate the adaptive learning rate for each parameter, so that the learning rate can be automatically adjusted for different parameters and different training stages, improving the convergence speed and stability of the algorithm. The initial learning rate is 0.01.

[0055] The algorithm steps are as follows:

[0056] ① Initialization parameters: The momentum term is initialized to 0, and the first-order moment and second-order moment are initialized to 1.

[0057] Calculate gradient: For the current training sample, calculate the gradient of the model's loss function with respect to the parameters.

[0058] ② Update the momentum term and the first-order moment and the second-order moment: Update their values ​​according to the current gradient and the previous momentum term, first-order moment, and second-order moment.

[0059] ③ Update the momentum term and the first-order moment and the second-order moment: Update the values ​​of the momentum term, the first-order moment, and the second-order moment according to the current gradient and the previous values. The update formula of the momentum term is: t =β1m t-1 +(1-β1)gt.

[0060] Among them, m t is the momentum term at the current moment, m t-1 is the momentum term at the previous moment, β1 is the momentum coefficient, which is 0.9, and gt is the gradient at the current moment.

[0061] The update formula of the moment is: t=β2v t-1 +(1-β2)g 2 t.

[0062] Among them, v t is the momentum term at the current moment, v t-1 is the first-order moment of the previous moment, β2 is the first-order moment coefficient, which is 0.999, g 2 t is the square of the gradient at the current moment.

[0063] ④ Calculate the deviation correction term: Since the estimation of momentum term and first-order moment and second-order moment may be inaccurate in the initial stage, deviation correction is required. The momentum term after deviation correction is:

[0064] in, is the bias-corrected momentum term, m t is the uncorrected momentum term, It is β1 to the power of t.

[0065] The first-order moment after deviation correction is:

[0066] in, is the bias-corrected first-order moment, v t is the uncorrected first-order moment, It is β2 to the power of t.

[0067] ⑤ Update parameters: Calculate the update amount of parameters based on the momentum term and first-order moment after deviation correction, and update the parameters of the model. The parameter update formula is:

[0068] Among them, θ t+1 is the updated parameter, θ t is the current parameter, η is the learning rate, The bias-corrected first moment, The momentum term after bias correction is 10-8.

[0069] S35. Building a deep complex pyroclastic rock recognition model based on neural network in pytorch environment

[0070] S4. Optimization and example of deep complex pyroclastic rock recognition model based on neural network

[0071] The data set normalized in step S2 is divided into a training set and a test set in a ratio of 8:2.

[0072] S41. The training set is input to the input end of the deep complex pyroclastic rock recognition model of the neural network, and lithology prediction learning is performed through forward propagation. The lithology prediction learning labels are first divided into large-category prediction learning, i.e., basic, neutral, and acidic prediction learning, and then the small-category lithology prediction learning of the large category is performed, and the lithology and predicted lithology are compared through back propagation to automatically optimize the neuron weights. Each learning is defined as an iteration. In the embodiment of the present invention, the number of iterations is 5000 to achieve the optimal effect, and the overfitting phenomenon is minimized. The optimal model after 5000 iterations is saved.

[0073] S42, the test set is tested, and the data group in the test set is input into the input end of the deep complex pyroclastic rock recognition model of the optimized neural network to predict the corresponding lithology label and perform accuracy evaluation.

[0074] The accuracy of the training set of the invention is 96.5%, and the accuracy of the test set is 94.8%. The main pyroclastic rock types can be effectively distinguished.

[0075] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying deep complex pyroclastic rocks based on neural networks, characterized in that: The steps include: S1, extract feature data; S2, label encoding and data normalization processing of logging data samples; S3. Establish a deep complex pyroclastic rock identification model based on neural network; S4. Optimized neural network model for deep complex pyroclastic rock recognition.

2. The neural network-based deep complex pyroclastic rock identification method according to claim 1, characterized in that: The step S2 comprises the following steps: S21. The lithology of pyroclastic rocks is divided into three categories according to the SiO2 content: basic, neutral and acidic. S22, standardizing the logging data.

3. The deep complex pyroclastic rock identification method based on neural network according to claim 1 is characterized in that: The step S2 comprises the following steps: S31, establish a three-layer model structure of input layer, hidden layer and output layer; S32, the input layer neurons are normalized logging data sets, i.e., natural gamma, compensated neutron, density, acoustic time difference, and deep lateral resistivity corresponding to the same depth, and the output layer is the pyroclastic rock lithology corresponding to the same depth; S33, the input well logging data set is processed by multiple fully connected layers, activation functions and Dropout layers, and finally the output lithology label is obtained; S34, by back propagating the comparison between the lithology label and the output lithology, the weight of the neuron is continuously optimized and then the back propagation error is calculated layer by layer to minimize the error signal; S35. Build a deep complex pyroclastic rock recognition model based on neural network in pytorch environment.

4. The neural network-based deep complex pyroclastic rock identification method according to claim 1, characterized in that: The step S4 comprises the following steps: S41. The training set is input to the input end of the deep complex pyroclastic rock recognition model of the neural network, and lithology prediction learning is performed through forward propagation. The lithology prediction learning labels are first divided into large-category prediction learning, i.e., basic, neutral, and acidic prediction learning, and then the small-category lithology prediction learning of the large category is performed. The lithology and predicted lithology are compared through back propagation to automatically optimize the neuron weights. Each learning is defined as an iteration. S42, the test set is tested, and the data group in the test set is input into the input end of the deep complex pyroclastic rock recognition model of the optimized neural network to predict the corresponding lithology label and perform accuracy evaluation.

Images