CN-121998017-A - Unsupervised deep neural network construction method based on ensemble learning

CN121998017ACN 121998017 ACN121998017 ACN 121998017ACN-121998017-A

Abstract

The invention discloses an unsupervised deep neural network building method based on ensemble learning, which particularly relates to the technical field of neural networks, and comprises the steps of obtaining continuous waveform records of an underground detector, obtaining discrete waveform samples through preprocessing, distributing indexes, constructing a pseudo tag pool, initializing a parallel neural network, performing forward correction training based on hidden state consistency, dynamically correcting the pseudo tag and updating network weights, sending a new round of samples into the updated network, integrating decision results and positioning an uncertainty source network, generating a diagnosis feedback packet based on the source network, and adjusting parameters of the pseudo tag pool to optimize subsequent training. The invention effectively solves the core technical problems of more interference signals and weak effective microseism signals in the underground monitoring environment of the fracturing well by constructing a parallel neural network architecture integrating residual connection, U-shaped encoding and decoding and attention mechanisms and combining forward correction training and dynamic pseudo-tag optimization strategies based on hidden state consistency.

Inventors

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

Assignees

云南民族大学

Dates

Publication Date: 20260508
Application Date: 20260304

Claims (9)

1. An unsupervised deep neural network building method based on ensemble learning is characterized by comprising the following steps: Acquiring continuous waveform records of a downhole detector, preprocessing to obtain discrete waveform samples, distributing indexes for each discrete waveform sample, constructing a pseudo-tag pool for storing parameters of each sample based on the indexes, and initializing a parallel neural network respectively provided with residual connection, U-shaped encoding and decoding and an attention mechanism; Performing forward correction training based on hidden state consistency for discrete waveform samples, extracting hidden states through a parallel neural network and mapping the hidden states into alignment feature vectors, calculating the difference degree between networks based on the alignment feature vectors and comparing the difference degree with an adaptive threshold, dynamically correcting pseudo labels in a pseudo label pool to serve as supervision signals, calculating output errors of each network based on the supervision signals, and updating weight parameters of the parallel neural network through a back propagation algorithm; obtaining a discrete waveform sample of a new round of downhole detectors, sending the discrete waveform sample into three updated parallel neural networks to output decision results respectively, and decomposing uncertainty contribution of each network after integrating the decision results to position a source network of decision uncertainty; Based on the source network with uncertainty contribution degree positioning, key features are extracted through traceability analysis to generate a diagnosis feedback packet, corresponding sample parameters in the pseudo tag pool are adjusted based on the diagnosis feedback packet, and subsequent training batch data are dynamically optimized.
2. The method for constructing the unsupervised deep neural network based on ensemble learning according to claim 1, wherein the preprocessing step comprises the steps of dividing a continuous waveform into discrete waveform samples according to a fixed time window, setting the time window length according to the periodic characteristics of the waveform, removing direct current components in the samples after division to eliminate baseline drift interference, uniformly mapping all sample amplitudes to a [0,1] interval through amplitude normalization operation, and distributing a global unique index for each preprocessed sample; The pseudo tag pool adopts a key value storage structure, takes a sample index as a unique key, and the initial storage content comprises a randomly generated pseudo tag vector, a confidence score initialized to 0.5 and a called count with an initial value of 0, and reserves a metadata domain for each index for subsequent state recording; and outputting a data set of the standardized discrete waveform sample after the data preprocessing is completed, and providing data support for subsequent training.
3. The method for constructing an unsupervised deep neural network based on ensemble learning according to claim 1, wherein the neural network is a parallel structure of a residual network, a U-type codec network and a U-type codec network with a attention mechanism; the residual network relieves the gradient vanishing problem in deep training through a shortcut path, the U-shaped encoding and decoding network extracts global features through downsampling of an encoder and restores detail features through upsampling of a decoder, and the U-shaped encoding and decoding network with an attention mechanism strengthens the weight of key waveform features through a channel attention module so as to further improve the pertinence of feature extraction; The initial value of each network parameter is set by adopting a He initialization mode, the bias item is initialized to 0, the initial value of the learning rate is uniformly set to 1e-4, the structural independence of the three networks is kept in the training process, and the collaborative updating of the parameters is realized only through hidden state interaction.
4. The method for constructing an unsupervised deep neural network based on ensemble learning according to claim 1, wherein the hidden state processing in forward correction training includes intercepting an output tensor as a hidden state representation at an 8 th layer of a network with residual connection, a 3 rd decoding layer of a U-type codec network, and an attention output layer of an attention mechanism network according to feature extraction capability of each network, wherein a dimension of the hidden state tensor is [ batch size, number of feature channels, time step ]; Performing global average pooling operation on hidden state tensors of each network, and calculating a mean value along a time step dimension to compress the dimension to obtain hidden state vectors with the dimension of [ batch size, characteristic channel number ]; And uniformly mapping hidden state vectors with different dimensions to 256-dimensional feature space through a linear transformation layer which is generated randomly by parameters and is fixed in training, and generating an aligned feature vector.
5. The method for constructing an unsupervised deep neural network based on ensemble learning according to claim 4, wherein the feature difference degree is calculated by respectively solving L2 norm distances of the first and second, first and third, second and third alignment feature vectors in a 256-dimensional feature space, and each distance value is calculated by square sum of differences of elements of the vectors; And calculating an arithmetic average value of the three distance values to serve as characteristic difference degree among networks, wherein the self-adaptive threshold value is initially 0.5, the difference degree average value and the standard differential state of the front 10 rounds of training are adjusted when the training round number is more than 11 in the dynamic optimization process, and a threshold updating formula is that the self-adaptive difference degree threshold value=the front average value+0.8xfront standard deviation, the label correction is triggered when the difference degree is larger than the threshold value, and the original pseudo label is maintained unchanged when the difference degree is smaller than or equal to the threshold value.
6. The method for setting up the unsupervised deep neural network based on ensemble learning as set forth in claim 2, wherein reliability is improved through ensemble decision for the updated output result of the neural network, specifically, the first arrival time estimated values output by the three networks are sorted in ascending order, the maximum value and the minimum value which possibly have abnormal deviation are removed, the median is selected as an ensemble decision value to improve stability of the result, absolute difference values of each network estimated value and the ensemble decision value are calculated to serve as decision deviation values, sample history records with confidence scores greater than 0.7 are screened from a pseudo tag pool metadata domain, historical decision deviation values or historical decision preset values of corresponding networks in the records are extracted, statistical average values and standard deviations are calculated based on the historical decision deviation values or the historical decision preset values, and twice standard deviation is defined as an expected deviation upper limit for the reliability of current network decisions.
7. The method for building the unsupervised deep neural network based on ensemble learning as claimed in claim 6, wherein the uncertainty contribution decomposition is specifically that if the network decision deviation value is greater than the expected deviation upper limit, the ratio of the deviation value to the upper limit is calculated as an original anomaly score, the greater the ratio is to indicate that the network decision uncertainty is higher, if the deviation value is less than or equal to the upper limit, the original anomaly score is set to 0, the three original anomaly scores are normalized by adopting a Softmax function, so that the sum of components after normalization is 1, the normalized uncertainty contribution is obtained, and the contribution component with the largest value and the corresponding source network thereof are positioned, wherein the network is the main source of the current decision uncertainty and is the locking object for the subsequent traceability analysis.
8. The method for constructing the unsupervised deep neural network based on ensemble learning, which is characterized in that the diagnostic feedback packet generation comprises the steps of extracting a feature map output by convolution of the last layer in an inference process of a source network corresponding to the maximum contribution degree, enabling the dimension of the feature map to be consistent with the time step of an input waveform sample, calculating the absolute value of each channel activation value of the feature map, taking the maximum value along the channel direction to generate a single-channel significant map, highlighting the region with the strongest feature response, multiplying the significant map with the original input waveform sample point by point to obtain an attributive thermodynamic diagram, intuitively reflecting the key region causing decision uncertainty in the waveform, locating three coordinate points with the highest value in the thermodynamic diagram, recording the time position and amplitude value corresponding to the original waveform to form an attributive feature coordinate set, and finally combining the normalized uncertainty contribution degree, the source network identification, the attributive feature coordinate set and the sample index into a structured diagnostic feedback packet in a JSON format, so that subsequent analysis and processing are facilitated.
9. The method for constructing an unsupervised deep neural network based on ensemble learning according to claim 8, wherein the confidence level is adjusted by subtracting a maximum normalized uncertainty contribution component value in a diagnosis feedback packet from 1, the attenuation factor range is [0,1], the original confidence level score is multiplied by the attenuation factor, and the confidence level score of a corresponding sample in a pseudo tag pool is updated to realize linkage adjustment of the confidence level and decision reliability.

Description

Unsupervised deep neural network construction method based on ensemble learning Technical Field The invention relates to the technical field of neural networks, in particular to an unsupervised deep neural network building method based on ensemble learning. Background In reservoir reconstruction monitoring of hydraulic fracturing, microseism signal identification and first arrival pickup are a particularly key ring, weak vibration signals generated when a stratum is fractured by high-pressure fluid are captured by using a downhole detector, the effective signals are distinguished from interference such as pump body noise and sleeve vibration, and then the time point of first arrival of P waves and S waves is accurately found through an algorithm; Because the environmental interference of monitoring in the well is large, the distance between detectors is small, and the accuracy of pickup is guaranteed by a plurality of joint treatments, the accuracy directly influences the positioning of the subsequent microseism focus, and the judgment of the formation fracture morphology and the extension range, and finally, a real-time reference basis can be provided for the real-time adjustment of the on-site fracturing process and the evaluation of the reservoir transformation effect. The development of the neural network technology provides an important support for the technology upgrading of the identification and first arrival pickup of the micro seismic signals of the fractured wells. The technical field covers various calculation models from a basic neuron model to a complex depth network structure, takes multi-layer nonlinear transformation as a core implementation means, can simulate an intelligent decision process by learning and mining high-level characteristic representation of data, systematically covers various main stream network architectures and classical training paradigms, and also widely covers a plurality of technical directions such as image recognition, natural language processing, signal analysis and the like in an application range; The neural network technology is combined with microseism signal identification and first-arrival pickup, the characteristic mining and intelligent analysis capability of the neural network technology on complex data can be fully exerted, the complex environment and various interference problems monitored in the well are well adapted, the signal screening and the first-arrival time pickup are enabled to be more fit with the engineering actual requirements of hydraulic fracturing, the accuracy of the whole fracturing monitoring is further improved from the technical level, and the monitoring result can be more effectively used for providing support for site construction and reservoir transformation effect evaluation. However, in the building process of the conventional neural network, training is carried out depending on labeling data, so that the data acquisition cost is high, a dynamic adjustment mechanism for monitoring signals is lacking in the training process, when various interference and complex stratum conditions in a well are faced, decision deviation is difficult to trace accurately, the stability of signal discrimination and first-arrival pickup is insufficient, the accuracy of subsequent seismic source positioning and pertinence of fracturing process adjustment are further affected, for example, first-arrival time judgment deviation is easy to occur under complex interference, and the stratum fracture extension range is inaccurate to evaluate. Disclosure of Invention The invention mainly aims to provide an unsupervised deep neural network building method based on ensemble learning, which can effectively solve the problems involved in the background technology. In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: an unsupervised deep neural network building method based on ensemble learning comprises the following steps: Acquiring continuous waveform records of a downhole detector, preprocessing to obtain discrete waveform samples, distributing indexes for each discrete waveform sample, constructing a pseudo-tag pool for storing parameters of each sample based on the indexes, and initializing a parallel neural network respectively provided with residual connection, U-shaped encoding and decoding and an attention mechanism; Performing forward correction training based on hidden state consistency for discrete waveform samples, extracting hidden states through a parallel neural network and mapping the hidden states into alignment feature vectors, calculating the difference degree between networks based on the alignment feature vectors and comparing the difference degree with an adaptive threshold, dynamically correcting pseudo labels in a pseudo label pool to serve as supervision signals, calculating output errors of each network based on the supervision signals, and updating weight parameters of the parallel neura