CN-121980273-A - Meta-learning training algorithm for first arrival pickup of microseism phase in fracturing well

CN121980273ACN 121980273 ACN121980273 ACN 121980273ACN-121980273-A

Abstract

The invention discloses a meta-learning training algorithm for first arrival pickup of microseism phases in a fracturing well, in particular to the technical field of oil gas well fracturing and microseism monitoring, which comprises the steps of constructing a training data set for first arrival pickup of the microseism phases of the fracturing well, the accurate label sample accounts for 5% -10% and the rest is the inaccurate label sample, the accurate label sample is input into a deep neural network, after training by adopting a double-cycle element learning framework, preprocessing actual measurement data is input, and a vibration phase first arrival pickup result is output through model reasoning. The method combines STA/LTA automation to generate the inaccurate label, improves the training precision and stability of the model through double-cycle element learning framework and weight matrix optimization, can dynamically adjust the label weight, enhances the robustness to the inaccurate label error, improves the training efficiency through standardized preprocessing, reasonable super-parameter setting and GPU acceleration, improves the scene generalization capability of the model through a fine tuning strategy, adapts to low signal-to-noise ratio and geological region crossing requirements, and ensures the accuracy of the first arrival pickup of the microseism earthquake phase.

Inventors

ZHANG YILUN
CHEN FANGHUANG
CHEN XIFAN
LAN QIANQIAN
Leng Jiaxuan
YU ZHICHAO
HE CHUAN
CHEN SHIJIE

Assignees

云南民族大学

Dates

Publication Date: 20260505
Application Date: 20260304

Claims (8)

1. A meta-learning training algorithm for first arrival pickup of microseism phases in a fracturing well comprises the following steps: S1, constructing a training data set picked up by a micro-seismic phase of a fracturing well in the first arrival, wherein the training data set consists of an inaccurate label data set and an accurate label data set, and the sample size of the accurate label data set accounts for 5% -10% of the total sample size of the training data; S2, inputting the training data set into a deep neural network model, training the deep neural network model by adopting a double-cycle element learning optimization framework, wherein the double-cycle element learning optimization framework comprises an outer cycle and an inner cycle, the outer cycle uses the inaccurate label data set as a training sample to update the trainable parameters of the deep neural network model, and the inner cycle uses the accurate label data set as a guide sample to optimize a weight matrix matched with a label dimension, and low weight is given to the inaccurate label data set sample and high weight is given to the accurate label data set sample through the weight matrix; s3, inputting the preprocessed fracturing well microseism actual measurement data into a depth neural network model after training, and reasoning and outputting a microseism phase first arrival pickup result through the model.
2. The meta-learning training algorithm for the first arrival picking of microseism in a fractured well according to claim 1, wherein the imprecise tag data set is generated by performing first arrival picking annotation on original microseism data of the fractured well through an automatic algorithm, the precise tag data set is generated by performing first arrival picking correction annotation on the original microseism data of the fractured well manually, and the automatic algorithm is a short-time average/long-time average STA/LTA algorithm.
3. The meta-learning training algorithm for first arrival pickup of microseismic phases in a fracturing well according to claim 1, wherein the single training iteration flow of the dual-cycle meta-learning optimization framework is characterized in that an outer cycle is firstly executed to complete primary updating of trainable parameters of a deep neural network model, an inner cycle is then executed to complete optimal solution of the weight matrix, a second outer cycle is finally executed, the optimized weight matrix is incorporated into loss calculation and secondary updating of trainable parameters of the model is completed, a random gradient descent SGD algorithm with momentum is adopted in the outer cycle to conduct parameter optimization, and a cross entropy loss function is selected as a loss function.
4. The meta-learning training algorithm for first arrival pickup of microseismic phases in a fracturing well according to claim 1, wherein the optimization updating method of the weight matrix in the inner loop is characterized in that an average gradient is calculated based on small batches of samples of an accurate tag data set, the weight matrix is updated by the average gradient, non-negative correction and normalization processing are sequentially carried out on the updated weight matrix, influence of learning rate on a model training process is eliminated, and training stability is guaranteed.
5. The meta-learning training algorithm for first arrival picking of microseismic phases in a fractured well according to claim 1, wherein the pre-step S1 is characterized by further comprising a step of preprocessing original microseismic data of the fractured well, specifically comprising the steps of removing non-valid tracks in the original microseismic data, cutting off an earthquake section without valid events to an adaptive input size of a deep neural network model, and normalizing amplitude values of all data to a [0,1] interval by adopting a mapping technology, wherein the adaptive input size is 64 tracks multiplied by 512 sampling points.
6. The meta-learning training algorithm for the first arrival pickup of microseismic phases in a fracturing well according to claim 1, wherein the deep neural network model is any one of a convolutional neural network CNN, a U-Net network and a U-Net++ network, when the U-Net network is adopted, the U-Net network comprises three modules of a compression path and an expansion path, each module of the compression path consists of three convolution layers and a maximum pooling layer, each module of the expansion path realizes feature map size amplification and channel number reduction through deconvolution operation, jump connection is arranged between the expansion path and the compression path, and the activation functions of all the convolution layers are LeakyReLU.
7. The meta-learning training algorithm for the first arrival pickup of microseismic phases in a fracturing well according to claim 1 is characterized in that the super-parameters of the deep neural network model in the step S2 are set to be 20-40 in small batches of sample numbers, the training iteration times are not less than 1000, the model learning rate is selected to be 1 e-4-7 e-3, the small batches of sample numbers are 30, the model learning rate is preferably 1 e-3-5 e-3, random seeds are fixed in the training process, and GPU (graphic processing unit) is adopted for parallel calculation to accelerate model training.
8. The meta-learning training algorithm for the first arrival picking of microseismic phases in a fractured well according to claim 1, wherein when the model is trained in the step S2, when the first arrival picking error of the microseismic phases of the imprecise label data set is within 20% of a basic error, the loss weight of the imprecise label sample is dynamically adjusted through the weight matrix, so that the model keeps robustness to the label error; The training is completed, the model generalization capability improvement step further comprises the steps of taking a deep neural network pre-training model obtained through element learning training as a basis, selecting fracturing well microseism data with low signal to noise ratio or cross geological areas as target domain data, performing secondary training on the pre-training model by adopting a fine tuning strategy of all layer parameters outside a frozen output layer or all layer parameters after full thawing, and outputting a micro earthquake phase first arrival pickup model adapting to cross scenes, wherein the basic error is an inherent error of a labeling result of an automatic labeling algorithm relative to a manual correction labeling result.

Description

Meta-learning training algorithm for first arrival pickup of microseism phase in fracturing well Technical Field The invention relates to the technical field of oil gas well fracturing and microseism monitoring, in particular to a meta-learning training algorithm for first-arrival pickup of microseism phases in a fracturing well. Background The first-arrival picking of the microseism phase of the fracturing well is a core step of oil and gas development microseism monitoring, the picking precision of the first-arrival picking is directly used for determining the accuracy of subsequent seismic source positioning, fracture morphology inversion and fracturing effect evaluation, the automatic processing of the first-arrival picking is realized by a deep learning method in the mainstream of the industry at present, and the identification and marking of the first-arrival points of microseism signals are completed by depending on the autonomous feature extraction capability of deep neural networks such as CNN, U-Net and the like. The method has the obvious defects that firstly, deep neural network training highly depends on massive high-precision manual labeling data, professionals are required to correct microseism data labels group by group, labor and time cost are extremely high, a large number of inaccurate labels cannot be guided to complete effective training through a small number of accurate labels, secondly, a targeted double-circulation training frame is lacked, a weight matrix optimization mechanism of label matching dimension is not available, differential weight distribution cannot be conducted on the inaccurate labels and the accurate labels, label errors directly affect model training precision, thirdly, error robustness of the inaccurate labels is poor, dynamic weight adjustment rules are not set for inherent errors of automatic labeling, pickup result deviation is caused by small-range label errors, fourthly, model generalization capability is weak, a fine adjustment strategy of low signal-to-noise ratio and cross-geological area data is not available, precision is greatly reduced when the labels are applied across scenes, and fifthly, the microseism data preprocessing has no standardized flow, model input size is not uniform, and training super-parameters have no reasonable value interval, so that model training efficiency is low and convergence effect is poor. Disclosure of Invention The invention mainly aims to provide a meta-learning training algorithm for first arrival pickup of microseism and earthquake phases in a fracturing well, which can effectively solve the problems in the background technology. In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: A meta-learning training algorithm for first arrival pickup of microseism phases in a fracturing well comprises the following steps: S1, constructing a training data set picked up by a micro-seismic phase of a fracturing well in the first arrival, wherein the training data set consists of an inaccurate label data set and an accurate label data set, and the sample size of the accurate label data set accounts for 5% -10% of the total sample size of the training data; S2, inputting the training data set into a deep neural network model, training the deep neural network model by adopting a double-cycle element learning optimization framework, wherein the double-cycle element learning optimization framework comprises an outer cycle and an inner cycle, the outer cycle uses the inaccurate label data set as a training sample to update the trainable parameters of the deep neural network model, and the inner cycle uses the accurate label data set as a guide sample to optimize a weight matrix matched with a label dimension, and low weight is given to the inaccurate label data set sample and high weight is given to the accurate label data set sample through the weight matrix; s3, inputting the preprocessed fracturing well microseism actual measurement data into a depth neural network model after training, and reasoning and outputting a microseism phase first arrival pickup result through the model. Preferably, the imprecise tag data set is generated by performing vibration phase first arrival pickup labeling on the original microseism data of the fractured well through an automatic algorithm, the precise tag data set is generated by manually performing vibration phase first arrival pickup correction labeling on the original microseism data of the fractured well, and the automatic algorithm is a short-time average/long-time average STA/LTA algorithm. The single training iterative process of the double-loop element learning optimization framework comprises the steps of firstly executing an outer loop to complete primary updating of trainable parameters of a deep neural network model, then executing an inner loop to complete optimal solution solving of the weight matrix, finally executing a secondary outer loop