Unsupervised deep neural network construction method based on ensemble learning

Abstract

The invention discloses an unsupervised deep neural network construction method based on integrated learning, and particularly relates to the technical field of neural networks, and the method comprises the steps: obtaining a continuous waveform record of an underground detector, obtaining a discrete waveform sample through preprocessing, distributing an index, constructing a pseudo-label pool, and initializing a parallel neural network; executing forward correction training based on hidden state consistency, dynamically correcting a pseudo tag and updating a network weight; sending a new round of samples into the updated network, integrating decision results and positioning an uncertain source network; a diagnostic feedback packet is generated based on the source network, and pseudo-tag pool parameters are adjusted to optimize subsequent training. According to the method, a parallel neural network architecture fusing residual connection, U-shaped coding and decoding and an attention mechanism is constructed, and a forward correction training and dynamic pseudo-label optimization strategy based on hidden state consistency is combined, so that the core technical problems of many interference signals and weak effective micro-seismic signals in the underground monitoring environment of the fractured well are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of neural network technology, and in particular to a method for constructing an unsupervised deep neural network based on ensemble learning. Background Technology

[0002] In reservoir stimulation monitoring during hydraulic fracturing, microseismic signal identification and first arrival pickup are particularly crucial. This involves using downhole geophones to capture the weak vibration signals generated when the formation is fracturing by high-pressure fluid, distinguishing these effective signals from interference such as pump noise and casing vibration, and then using algorithms to accurately find the first arrival time of P-waves and S-waves. Because there are many environmental interferences in well monitoring and the spacing between detectors is small, multiple channels must be used for joint processing to ensure the accuracy of the pickup. This accuracy directly affects the location of subsequent microseismic sources, as well as the judgment of formation fracture morphology and extension range. Ultimately, it can provide solid reference for the real-time adjustment of on-site fracturing technology and the evaluation of reservoir stimulation effect.

[0003] The development of neural network technology has provided important support for the technological upgrade of microseismic signal identification and first arrival picking in fractured wells. This technology field covers various computational models from basic neuron models to complex deep network structures. It uses multi-layer nonlinear transformation as the core implementation method and can simulate intelligent decision-making processes by learning and mining high-level feature representations of data. It also systematically encompasses various mainstream network architectures and classic training paradigms, and its application scope covers a wide range of technical directions such as image recognition, natural language processing, and signal analysis. Combining neural network technology with microseismic signal identification and first arrival acquisition can fully leverage the former's ability to mine features and intelligently analyze complex data. This is well-suited to the complex environment and diverse interference issues in well monitoring, making signal identification and first arrival acquisition more aligned with the actual engineering needs of hydraulic fracturing. Consequently, it improves the accuracy of the entire fracturing monitoring process from a technical perspective, enabling the monitoring results to more effectively support on-site construction and reservoir stimulation effect evaluation.

[0004] However, existing neural networks rely on labeled data for training, resulting in high data acquisition costs. During training, they lack a dynamic adjustment mechanism for monitoring signals. When faced with diverse interferences and complex formation conditions within the well, decision-making biases are difficult to trace accurately, and the stability of signal identification and first arrival picking is insufficient. This affects the accuracy of subsequent source location and the targeted adjustment of fracturing technology. For example, under complex interference, there is a tendency for first arrival time judgment errors to occur, leading to inaccurate assessment of the formation fracture extension range. Summary of the Invention

[0005] The main objective of this invention is to provide a method for constructing unsupervised deep neural networks based on ensemble learning, which can effectively solve the problems involved in the background art mentioned above.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for constructing an unsupervised deep neural network based on ensemble learning includes the following steps: Continuous waveform records from downhole geophones are acquired, and discrete waveform samples are obtained after preprocessing. An index is assigned to each discrete waveform sample, and a pseudo-label pool for storing parameters of each sample is constructed based on the index. Parallel neural networks with residual connections, U-shaped encoding / decoding, and attention mechanisms are initialized. For discrete waveform samples, forward correction training based on hidden state consistency is performed. The hidden states are extracted by a parallel neural network and mapped to aligned feature vectors. The difference between networks is calculated based on the aligned feature vectors and compared with an adaptive threshold. The pseudo-labels in the pseudo-label pool are dynamically corrected as supervision signals. The output error of each network is calculated based on the supervision signals. The weight parameters of the parallel neural network are updated by the backpropagation algorithm. The discrete waveform samples of the new round of downhole detectors are obtained and fed into the updated three parallel neural networks to output decision results respectively. After integrating the decision results, the uncertainty contribution of each network is decomposed to locate the source network of decision uncertainty. Based on the source network located by this uncertainty contribution, key features are extracted through causal analysis to generate a diagnostic feedback package. The parameters of the corresponding samples in the pseudo-label pool are adjusted based on the diagnostic feedback package to dynamically optimize the subsequent training batch data.

[0007] Preferably, the preprocessing step specifically involves: dividing the continuous waveform into discrete waveform samples according to a fixed time window, with the time window length set according to the periodic characteristics of the waveform; removing the DC component from the samples after segmentation to eliminate baseline drift interference; and then mapping the amplitude of all samples to the [0,1] interval through amplitude normalization operation, and assigning a globally unique index to each preprocessed sample. The pseudo-label pool adopts a key-value storage structure, with the sample index as the unique key. The initial storage content includes a randomly generated pseudo-label vector, a confidence score initialized to 0.5, and a call count initialized to 0. At the same time, a metadata field is reserved for each index for subsequent status recording. After data preprocessing, a dataset of standardized discrete waveform samples is output to provide data support for subsequent training.

[0008] Preferably, the neural network is a parallel structure of a residual network, a U-shaped encoder-decoder network, and a U-shaped encoder-decoder network with an attention mechanism; Among them, the residual network alleviates the gradient vanishing problem in deep training through the shortcut path, the U-shaped encoder-decoder network extracts global features through encoder downsampling and restores detailed features through decoder upsampling, and the U-shaped encoder-decoder network with attention mechanism strengthens the weight of key waveform features through channel attention module, further improving the targeting of feature extraction; Each network parameter is initialized using the He initialization method, the bias term is initialized to 0, the initial learning rate is uniformly set to 1e-4, and the structural independence of the three networks is maintained during training, with parameter collaborative updates achieved only through hidden state interaction.

[0009] Preferably, the hidden state processing in the forward correction training includes: according to the feature extraction capabilities of each network, extracting the output tensor as the hidden state representation from the 8th layer of the residual connection network, the 3rd decoding layer of the U-shaped encoder-decoder network, and the attention output layer of the attention mechanism network, with the hidden state tensor having dimensions of [batch size, number of feature channels, time step]; Perform global average pooling on the hidden state tensor of each network, calculate the mean along the time step dimension to compress the dimension, and obtain a hidden state vector with dimension [batch size, number of feature channels]. By using a linear transformation layer with randomly generated parameters and fixed parameters during training, hidden state vectors of different dimensions are uniformly mapped to a 256-dimensional feature space to generate aligned feature vectors.

[0010] Preferably, the feature difference is calculated as follows: in the 256-dimensional feature space, the L2 norm distances between the first and second, the first and third, and the second and third aligned feature vectors are calculated respectively, and each distance value is obtained by taking the square root of the sum of the squares of the differences between the elements of the vectors; The arithmetic mean of the three distance values ​​is used as the feature difference between the networks. The adaptive threshold is initially set to 0.5. During the dynamic optimization process, when the number of training rounds is greater than 11, it is dynamically adjusted based on the mean and standard deviation of the difference in the previous 10 training rounds. The threshold update formula is: Adaptive difference threshold = previous mean + 0.8 × previous standard deviation. When the difference is greater than the threshold, label correction is triggered. When it is less than or equal to the threshold, the original pseudo label is kept unchanged.

[0011] Preferably, for the updated neural network output, reliability is improved through ensemble decision-making. Specifically, the initial arrival time estimates of the three network outputs are sorted in ascending order, and the maximum and minimum values ​​that may have abnormal biases are removed. The median is selected as the ensemble decision value to improve the stability of the results. The absolute difference between each network estimate and the ensemble decision value is calculated as the decision bias value. At the same time, historical records of samples with confidence scores greater than 0.7 are selected from the pseudo-label pool metadata domain, and the historical decision bias values ​​or historical decision preset values ​​of the corresponding networks in these records are extracted. The statistical mean and standard deviation are calculated based on the historical decision bias values ​​or historical decision preset values. The mean plus twice the standard deviation is defined as the upper limit of the expected deviation, which is used to measure the reliability of the current network decision.

[0012] Preferably, the uncertainty contribution decomposition specifically involves: 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 the original anomaly score, and the larger the ratio, the higher the network decision uncertainty; if the deviation value is less than or equal to the upper limit, the original anomaly score is set to 0; the Softmax function is used to normalize the three original anomaly scores so that the sum of each component after normalization is 1, thus obtaining the normalized uncertainty contribution; the contribution component with the largest value and its corresponding source network are located, and this network is the main source of the current decision uncertainty, which is the target for subsequent causal analysis.

[0013] Preferably, the diagnostic feedback package generation includes: extracting the feature map of the last convolutional layer output during the inference process of the source network corresponding to the maximum contribution, with the feature map dimension consistent with the time step of the input waveform sample; calculating the absolute value of the activation value of each channel of the feature map, taking the maximum value along the channel direction to generate a single-channel saliency map, highlighting the region with the strongest feature response; multiplying the saliency map point by point with the original input waveform sample to obtain an attribution heatmap, which intuitively reflects the key regions in the waveform that lead to decision uncertainty; locating the three coordinate points with the highest values ​​in the heatmap, recording their time positions and the amplitude values ​​of the corresponding original waveforms to form an attribution feature coordinate set; and finally combining the normalized uncertainty contribution, source network identifier, attribution feature coordinate set, and sample index into a structured diagnostic feedback package in JSON format for easy subsequent parsing and processing.

[0014] Preferably, the confidence adjustment is achieved by subtracting the maximum normalized uncertainty contribution component in the diagnostic feedback package from 1 to obtain the confidence decay factor, which is in the range of [0,1]. The original confidence score is multiplied by the decay factor to update the confidence score of the corresponding sample in the pseudo-label pool, thereby realizing the linkage adjustment between confidence and decision reliability.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention effectively solves the core technical challenges of numerous interference signals and weak effective microseismic signals in the downhole monitoring environment of fractured wells by constructing a parallel neural network architecture that integrates residual connections, U-shaped encoding and decoding, and attention mechanisms, combined with forward correction training based on hidden state consistency and dynamic pseudo-label optimization strategies. Through multi-network parallel feature extraction, the ability to distinguish complex signals is enhanced, and the accurate correction of pseudo-labels is achieved by leveraging hidden state alignment and dynamic difference judgment. This not only allows for efficient model training without relying on a large amount of labeled data, but also significantly improves the accuracy of P-wave and S-wave first arrival picking by integrating decision-making to eliminate abnormal biases and enhance the stability of first arrival time picking. This provides reliable data support for the accurate determination of subsequent microseismic source location, formation fracture morphology, and extension range, ensuring the quality of core data for fracture monitoring from the source.

[0016] 2. This invention, by introducing a decision uncertainty contribution decomposition and cause-finding analysis mechanism, can accurately locate the source network and key waveform regions leading to decision bias. Through structured diagnostic feedback packages, it achieves dynamic optimization of the pseudo-label pool and targeted adjustment of subsequent training data, enabling the model to continuously adapt and optimize, significantly improving its robustness under different fracturing conditions and complex formations. This scheme deeply integrates ensemble learning with unsupervised deep neural networks, fully leveraging the advantages of neural networks in high-level feature mining of complex signals and the decision reliability of ensemble learning. It not only adapts to the practical engineering needs of small-spacing in-well geophones and multi-channel joint processing, but also provides more valuable technical support for real-time adjustment of on-site fracturing processes and scientific evaluation of reservoir stimulation effects through precise signal recognition and first-arrival picking, helping to improve the targeting and effectiveness of reservoir stimulation and demonstrating significant engineering application value. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall steps of the present invention; Figure 2 This is a flowchart of the data preprocessing process of the present invention; Figure 3 This is a flowchart of the alignment feature vector acquisition process of the present invention; Figure 4 This is a flowchart of the network decision reliability determination process of the present invention; Figure 5 This is a flowchart of the uncertainty source network localization process of the present invention; Figure 6 This is a flowchart of the diagnostic feedback package acquisition process of the present invention. Detailed Implementation

[0018] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0019] like Figure 1 As shown, this technical solution is mainly applied to the identification and first arrival picking of microseismic signals in fractured wells. It achieves efficient processing and accurate first arrival time extraction of microseismic signals by constructing an unsupervised deep neural network based on ensemble learning. Specifically, it involves: acquiring continuous waveform records from downhole geophones; preprocessing these records to obtain discrete waveform samples and assigning indices; constructing a pseudo-label pool; initializing a parallel neural network; performing forward correction training based on hidden state consistency; dynamically correcting pseudo-labels and updating network weights; feeding a new round of samples into the updated network; integrating decision results and locating the uncertainty source network; generating a diagnostic feedback packet based on the source network; and adjusting the pseudo-label pool parameters to optimize subsequent training. The technical solution of this invention will be further elaborated below with actual data.

[0020] Example 1: This example aims to complete the preprocessing of microseismic signals from fractured wells, construct a traceable pseudo-label pool, and initialize a parallel neural network with a specific structure, providing standardized data and a basic network architecture for subsequent model training. In the specific implementation process, a fractured well in the XX block of a certain oilfield was selected as the application object. The fractured well has a depth of 3200m. An HS-100 three-component geophone was deployed downhole, and the sampling rate of the geophone was set to 2000Hz. Microseismic signals were continuously collected for 24 hours, and a total of 172,800,000 raw sampling points (2000Hz×24h×3600s) were obtained. The signal amplitude range was -0.8V~0.9V, including effective microseismic signals, formation interference signals and equipment noise.

[0021] See Figure 2 First, microseismic signals are preprocessed. Continuous waveforms are divided into discrete waveform samples by a fixed time window. The time window length is set according to the periodic characteristics of the waveform. The periodic characteristics of microseismic signals are 0.1s to 0.5s. Considering the signal recognition requirements, the fixed time window length is set to 0.3s. Each time window contains 600 sampling points (2000Hz × 0.3s). The continuous waveform is divided by a sliding time window (sliding step size of 0.1s), resulting in a total of 144,000 discrete waveform samples. The specific calculation process is: (24h × 3600s - 0.3s) / 0.1s + 1 ≈ 144,000. After segmentation, the DC component in the sample is removed to eliminate baseline drift interference. Each discrete waveform sample has a DC component caused by detector baseline drift. By calculating the mean amplitude of each sample and subtracting the mean, the baseline drift interference is eliminated. For example, the mean amplitude of the sample with index 000001 is 0.023V. Subtracting 0.023V from the amplitude values ​​of all sampling points of this sample, the amplitude range of the sample is adjusted to -0.812V~0.877V after processing. Then, amplitude normalization is performed to map all sample amplitudes to the [0,1] interval. The min-max normalization algorithm is used, and the normalization formula is: ,in This is the original amplitude value. The minimum amplitude value of the sample. For the sample with index 000001, the maximum amplitude value is given. , The original amplitude at a certain sampling point was 0.5V, and after normalization it was... ; Finally, a globally unique index is assigned to each preprocessed sample to achieve traceability management. The index increases sequentially from 000001 to 144000. The index enables full lifecycle traceability management of the sample, facilitating subsequent pseudo-label updates and historical record queries. After preprocessing, a standardized sample dataset containing 144000 discrete waveform samples is output. The sample format is [index, 600 normalized sampling points], providing data support for subsequent training.

[0022] Subsequently, a pseudo-label pool is constructed. The pseudo-label pool adopts a key-value storage structure, with the sample index as the unique key. The initial storage content includes a randomly generated pseudo-label vector, a confidence score initialized to 0.5, and a call count initialized to 0. At the same time, a metadata field is reserved for each index for subsequent status recording. The pseudo-label pool is implemented through a Redis key-value database. For example, the initial pseudo-label vector corresponding to the index "000001" is [0.32, 0.45, 0.23] (corresponding to 3 types of target recognition results: effective microseismic signal, formation interference signal, and equipment noise), with a confidence score of 0.5, a call count of 0, and an initial value of an empty string in the metadata field.

[0023] Finally, parallel neural network initialization is performed. The initialized neural network is a parallel structure of residual network, U-shaped encoder-decoder network and U-shaped encoder-decoder network with attention mechanism. The residual network adopts a simplified version of ResNet-18 architecture, which contains 8 convolutional layers. The gradient vanishing problem in deep training is alleviated by using a shortcut path (directly passing the output of the first convolutional layer to the third layer, and the output of the fifth layer to the seventh layer). Gradient vanishing refers to the gradient value of backpropagation approaching 0 when the neural network is trained in a deep layer, which makes it impossible to update the parameters effectively. The shortcut path can make the gradient propagate directly in the backpropagation, avoiding gradient decay. The encoder of the U-shaped encoder-decoder network contains 4 downsampling layers (compressing the feature map size and increasing the number of channels through convolution and pooling operations), and the decoder contains 4 upsampling layers (restoring the feature map size through transposed convolution). Global features are extracted through encoder downsampling and detailed features are restored through decoder upsampling. The U-shaped encoder-decoder network with attention mechanism adds a channel attention module to the second layer of the U-Net decoder. By calculating the importance weight of each feature channel (e.g., the weight of the microseismic signal feature channel is 0.8, and the weight of the noise channel is 0.2), the weight of key waveform features is strengthened, and the targeting of feature extraction is further improved. Each network parameter is initialized using the He initialization method, meaning the parameters have a mean of 0 and a variance of 0. normal distribution ( The number of input channels for the convolutional layer is 1 (single-channel waveform signal). For example, if the number of input channels for the first convolutional layer of the residual network is 1, then the parameter variance is 2 / 1=2. The initial parameter values ​​are in the range of [-0.5, 0.5]. The bias term is initialized to 0. The initial learning rate is uniformly set to 1e-4. During training, the structural independence of the three networks is maintained, and parameter co-update is only achieved through subsequent hidden state interactions.

[0024] Example 2: Based on the standardized discrete waveform sample dataset and the initialized parallel neural network output in Example 1, this example further performs forward correction training based on hidden state consistency. By dynamically correcting the pseudo-labels in the pseudo-label pool to provide supervision signals, the weight parameters of the parallel neural network are updated.

[0025] See Figure 3 First, hidden state extraction and alignment feature vector generation are performed. Based on the feature extraction capabilities of each network, the output tensors of the 8th layer (the last convolutional layer) of the residual network, the 3rd decoding layer of the U-shaped encoder-decoder network, and the attention output layer of the U-shaped encoder-decoder network with attention mechanism are truncated as hidden state representations. The dimensions of the hidden state tensors are all [32, 64, 600] (batch size 32, number of feature channels 64, time step 600, consistent with the number of sampling points of the input samples). Global average pooling is performed on the hidden state tensor of each network, and the mean is calculated along the time step dimension to compress the dimension, resulting in a hidden state vector of dimension [32, 64]. For example, the hidden state tensor of a sample in the residual network has values ​​of [0.12, 0.15, ..., 0.09] (600 values) along the time step dimension, with a mean of... Finally, the hidden state vector of this sample is [0.11, 0.13, ..., 0.10] (64 dimensions). Then, through a linear transformation layer with randomly generated parameters and fixed during training (weight matrix size 64×256, bias vector size 256×1), the hidden state vectors of different dimensions are uniformly mapped to the 256-dimensional feature space to generate aligned feature vectors. For example, if the first element of the hidden state vector of a sample is 0.11, multiplying it by the parameter 0.02 in the first row and first column of the weight matrix, and adding the product terms of other elements and the bias term 0.005, the first element of the aligned feature vector is 0.032.

[0026] Next, the inter-network dissimilarity is calculated and compared with an adaptive threshold. In the 256-dimensional feature space, the L2 norm distances of the aligned feature vectors of the residual network and the U-shaped encoder / decoder network, the residual network and the U-shaped encoder / decoder network with attention mechanism, and the U-shaped encoder / decoder network and the U-shaped encoder / decoder network with attention mechanism are calculated respectively. Each distance value is obtained by taking the square root of the sum of the squares of the differences between the elements of the vector. The formula for the L2 norm distance is: ,in , Let A, B, and C be the i-th elements of the two vectors, respectively. The alignment feature vectors of the three networks are calculated as follows: ; ; ; The arithmetic mean of the three distance values ​​is used as the feature dissimilarity between networks, i.e. The adaptive threshold is initially preset to 0.5. During dynamic optimization, when the number of training rounds > 11, it is dynamically adjusted based on the mean and standard deviation of the difference in the previous 10 training rounds. The threshold update formula is: adaptive difference threshold = previous mean + 0.8 × previous standard deviation. In this embodiment, the number of training rounds is 12, the mean difference in the previous 10 rounds is 0.12, and the standard deviation is 0.03. Therefore, the adaptive difference threshold = 0.12 + 0.8 × 0.03 = 0.144. When the difference is greater than the threshold, label correction is triggered. When it is less than or equal to the threshold, the original pseudo label remains unchanged. The current difference of 0.15 > the adaptive threshold of 0.144, triggering pseudo label correction.

[0027] Finally, pseudo-label correction and network weight updates are performed. For the samples in this training batch (index 000001~000032), based on the consistency judgment of the hidden states of the three networks, the pseudo-label vector of index 000001 is corrected from [0.32, 0.45, 0.23] to [0.41, 0.38, 0.21] (strengthening the label weight of the effective microseismic signal), and the call count of this sample is updated to 1. Using the corrected pseudo-label as the supervision signal, the output error of each network is calculated (using the cross-entropy loss function). For example, the output vector of the residual network is [0.39, 0.40, 0.21], the pseudo-label vector is [0.41, 0.38, 0.21], and the loss value is... The weight parameters of the parallel neural network are updated using the backpropagation algorithm. The weight update formula is as follows: (η is the learning rate, taking a value of 1e-4), for example, the parameters of a certain convolution kernel in a residual network. ,gradient ,but Repeat the above training process up to 50 rounds to complete the forward correction training of the parallel neural network.

[0028] Example 3: Based on the updated parallel neural network in Example 2, this example inputs a new round of fracturing well microseismic signal samples, improves the reliability of initial arrival picking through integrated decision-making, decomposes the uncertainty contribution of each network, and locates the source network of decision uncertainty.

[0029] In the specific implementation of this embodiment, the microseismic signal of the fractured well for the following hour is obtained. After the preprocessing process of Embodiment 1, 3600 discrete waveform samples (index 144001~147600) are obtained. These samples are then fed into three updated parallel neural networks in batches of size 32. Each network outputs the estimated first arrival time of the corresponding sample (the first arrival time refers to the starting time of the microseismic signal arriving at the detector, which is a key parameter for signal identification). Taking the sample with index 144001 as an example, the estimated first arrival times of the three networks are as follows: residual network (network 1) 0.08s, U-shaped encoding and decoding network (network 2) 0.07s, and U-shaped encoding and decoding network with attention mechanism (network 3) 0.12s.

[0030] See Figure 4 To improve reliability, the first arrival time estimates from the three networks are integrated and decided. First, the three estimates are sorted in ascending order: [0.07s, 0.08s, 0.12s]. The maximum value (0.12s) and the minimum value (0.07s) that may have abnormal bias are removed. The median of 0.08s is selected as the integrated decision value to improve the stability of the first arrival picking results. Next, the absolute difference between each network estimate and the ensemble decision value is calculated as the decision bias value. The decision bias value of network 1 is |0.08-0.08|=0.00s, that of network 2 is |0.07-0.08|=0.01s, and that of network 3 is |0.12-0.08|=0.04s. Simultaneously, historical records of samples with confidence scores greater than 0.7 were screened from the pseudo-label pool metadata domain, resulting in 1200 samples. The historical decision deviation values ​​of each network were extracted from these records. The statistical mean of the historical decision deviation values ​​of network 3 was 0.02s, and the standard deviation was 0.01s. Based on the statistical mean and standard deviation calculated from the historical decision deviation values, the mean plus twice the standard deviation was defined as the upper limit of expected deviation, which is used to measure the reliability of the current network's decision. The upper limit of expected deviation for network 3 = 0.02 + 2 × 0.01 = 0.04s.

[0031] See Figure 5 Then, uncertainty contribution decomposition is performed to locate the source network. 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 the original anomaly score. The larger the ratio, the higher the network decision uncertainty. If the deviation value is less than or equal to the upper limit, the original anomaly score is set to 0. The decision deviation value of network 1 is 0.00s≤0.04s, so the original anomaly score is 0. The decision deviation value of network 2 is 0.01s≤0.04s, so the original anomaly score is 0. The decision deviation value of network 3 is 0.04s=0.04s, so the original anomaly score is 0 (if the deviation value of network 3 is 0.05s, then the original anomaly score = 0.05 / 0.04 = 1.25). The three original outlier scores are normalized using the Softmax function, so that the sum of the normalized components is 1, thus obtaining the normalized uncertainty contribution. The Softmax formula is as follows: ( (where the original outlier score is [0,0,0]). , , In this case, it is considered that there is no significant source of uncertainty; When the original anomaly score is [0, 0, 1.25], , The sum = 1 + 1 + 3.4903 = 5.4903, which is normalized to [0.182, 0.182, 0.636]. At this point, the component with the largest contribution value and its corresponding source network are located. Network 3 is the main source of uncertainty in the current decision and is the target for subsequent causal analysis.

[0032] Example 4: Based on the uncertainty source network located in Example 3 (taking network 3 as an example), this example extracts key features through causal analysis, generates a diagnostic feedback package, and adjusts the pseudo-label pool parameters based on the feedback package to dynamically optimize subsequent training batch data.

[0033] See Figure 6 To generate a diagnostic feedback package, a causal analysis was performed on the source network (network 3) corresponding to the largest contribution. First, the feature map output of the last convolutional layer in the inference process was extracted. The feature map dimension is [32, 64, 600], which is consistent with the time step of the input waveform sample (600 sampling points). The absolute value of the activation value of each channel of the feature map was calculated, and the maximum value was taken along the 64 channel directions to generate a single-channel saliency map (dimension [600,]), highlighting the region with the strongest feature response. For example, the activation value of the 100th sampling point of the saliency map is 0.92 (which is the maximum value of all channels at this position), corresponding to the strong response region of the microseismic signal. The attribution heatmap is obtained by multiplying the saliency map point by point with the normalized sampling points of the original input waveform sample (index 144001). This intuitively reflects the key areas in the waveform that lead to decision uncertainty. The normalized amplitude value of the 100th sampling point of the original waveform sample is 0.777, and the value of the heatmap at this position is 0.92×0.777≈0.715. The three coordinate points with the highest values ​​in the heatmap are located, with sampling indexes of 98, 100, and 102. Their corresponding time positions are calculated using the formula: "Time position = Sampling index / Sampling rate". Sampling point 98 has a time position of 98 / 2000 = 0.049s, corresponding to a normalized amplitude value of 0.68 in the original waveform. Sampling point 100 has a time position of 100 / 2000 = 0.050s, corresponding to an amplitude value of 0.777. Sampling point 102 has a time position of 102 / 2000 = 0.051s, corresponding to an amplitude value of 0.69. These coordinates form the attribution feature coordinates. The set is: [{"Time Position": 0.049, "Amplitude Value": 0.68}, {"Time Position": 0.050, "Amplitude Value": 0.777}, {"Time Position": 0.051, "Amplitude Value": 0.69}]; finally, the normalized uncertainty contribution [0.182, 0.182, 0.636], the source network identifier "U-shaped encoder-decoder network with attention mechanism", the attribution feature coordinate set and the sample index "144001" are combined into a structured diagnostic feedback package in JSON format, as follows: {"Sample Index":"144001","Source Network Identifier":"U-shaped Encoder-Decoder Network with Attention Mechanism","Normalized Uncertainty Contribution":[0.182,0.182,0.636],"Attribution Feature Coordinate Set":[{"Time Position":0.049,"Amplitude Value":0.68},{"Time Position":0.050,"Amplitude Value":0.777},{"Time Position":0.051,"Amplitude Value":0.69}]}.

[0034] Among them, the sample index corresponds to the globally unique source identifier of the preprocessed discrete waveform sample, and the complete information of the sample in the pseudo-label pool can be queried through the index; the source network identifier clearly identifies the main source of decision uncertainty, the neural network, and provides a target object for subsequent network optimization; The normalized uncertainty contribution corresponds to the uncertainty proportion of the residual network, the U-shaped encoder-decoder network, and the U-shaped encoder-decoder network with attention mechanism, respectively, and the sum is 1. The larger the value, the stronger the influence of the network on decision uncertainty. The attribution feature coordinate set mainly records the key waveform regions that lead to decision uncertainty. Among them, "time position (s)" is the time axis coordinate of the microseismic signal (calculated by sampling point index ÷ 2000Hz sampling rate), and "normalized amplitude value" is the standardized amplitude of the waveform at the corresponding position (mapped to the [0,1] interval).

[0035] Furthermore, based on the diagnostic feedback package, the parameters of the corresponding samples in the pseudo-label pool are adjusted. The confidence level is adjusted by subtracting the maximum normalized uncertainty contribution value in the diagnostic feedback package from 1 to obtain the confidence attenuation factor. The attenuation factor ranges from [0,1], and the closer it is to 1, the higher the decision reliability. The attenuation factor calculated in this case is 1-0.636=0.364. The original confidence score is multiplied by the attenuation factor to update the confidence score of the corresponding sample in the pseudo-label pool, realizing the linkage adjustment between confidence and decision reliability. The original confidence score of this sample (index 144001) is 0.5, and the updated confidence score is 0.5×0.364=0.182. At the same time, the updated confidence score and the attribution feature coordinate set in the diagnostic feedback package are written into the metadata field of the pseudo-label pool to complete the sample parameter adjustment.

[0036] In subsequent training, samples with confidence scores greater than 0.7 in the pseudo-label pool are selected as training data first. At the same time, for the attribution feature region of the uncertainty source network (network 3), the training weight of similar microseismic signal samples is increased to achieve dynamic optimization of subsequent training batch data and continuously improve the initial arrival picking accuracy of the model.

[0037] Through the step-by-step execution of the above four embodiments, the complete application of unsupervised deep neural networks based on ensemble learning in the identification and first arrival of microseismic signals in fractured wells is achieved, realizing efficient training and accurate decision-making in unsupervised scenarios. Furthermore, through uncertainty analysis and data optimization, the robustness and reliability of the model are further improved.

[0038] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing an unsupervised deep neural network based on ensemble learning, characterized in that, Includes the following steps: Continuous waveform records from downhole geophones are acquired, and discrete waveform samples are obtained after preprocessing. An index is assigned to each discrete waveform sample, and a pseudo-label pool for storing parameters of each sample is constructed based on the index. Parallel neural networks with residual connections, U-shaped encoding / decoding, and attention mechanisms are initialized. For discrete waveform samples, forward correction training based on hidden state consistency is performed. The hidden states are extracted by a parallel neural network and mapped to aligned feature vectors. The difference between networks is calculated based on the aligned feature vectors and compared with an adaptive threshold. The pseudo-labels in the pseudo-label pool are dynamically corrected as supervision signals. The output error of each network is calculated based on the supervision signals. The weight parameters of the parallel neural network are updated by the backpropagation algorithm. The discrete waveform samples of the new round of downhole detectors are obtained and fed into the updated three parallel neural networks to output decision results respectively. After integrating the decision results, the uncertainty contribution of each network is decomposed to locate the source network of decision uncertainty. Based on the source network located by this uncertainty contribution, key features are extracted through causal analysis to generate a diagnostic feedback package. The parameters of the corresponding samples in the pseudo-label pool are adjusted based on the diagnostic feedback package to dynamically optimize the subsequent training batch data.

2. The method for constructing an unsupervised deep neural network based on ensemble learning according to claim 1, characterized in that, The preprocessing steps are as follows: the continuous waveform is divided into discrete waveform samples according to a fixed time window. The length of the time window is set according to the periodic characteristics of the waveform. After the segmentation, the DC component in the sample is removed to eliminate baseline drift interference. Then, the amplitude of all samples is uniformly mapped to the [0,1] interval through amplitude normalization operation. A globally unique index is assigned to each preprocessed sample. The pseudo-label pool adopts a key-value storage structure, with the sample index as the unique key. The initial storage content includes a randomly generated pseudo-label vector, a confidence score initialized to 0.5, and a call count initialized to 0. At the same time, a metadata field is reserved for each index for subsequent status recording. After data preprocessing, a dataset of standardized discrete waveform samples is output to provide data support for subsequent training.

3. The method for constructing an unsupervised deep neural network based on ensemble learning according to claim 1, characterized in that, The neural network is a parallel structure of a residual network, a U-shaped encoder-decoder network, and a U-shaped encoder-decoder network with an attention mechanism; Among them, the residual network alleviates the gradient vanishing problem in deep training through the shortcut path, the U-shaped encoder-decoder network extracts global features through encoder downsampling and restores detailed features through decoder upsampling, and the U-shaped encoder-decoder network with attention mechanism strengthens the weight of key waveform features through channel attention module, further improving the targeting of feature extraction; Each network parameter is initialized using the He initialization method, the bias term is initialized to 0, the initial learning rate is uniformly set to 1e-4, and the structural independence of the three networks is maintained during training, with parameter collaborative updates achieved only through hidden state interaction.

4. The method for constructing an unsupervised deep neural network based on ensemble learning according to claim 1, characterized in that, The hidden state processing in the forward correction training includes: based on the feature extraction capabilities of each network, the output tensor of the 8th layer of the network with residual connection, the 3rd decoding layer of the U-shaped encoder-decoder network, and the attention output layer of the attention mechanism network is truncated as the hidden state representation. The dimensions of the hidden state tensor are [batch size, number of feature channels, time step]; Perform global average pooling on the hidden state tensor of each network, calculate the mean along the time step dimension to compress the dimension, and obtain a hidden state vector with dimension [batch size, number of feature channels]. By using a linear transformation layer with randomly generated parameters and fixed parameters during training, hidden state vectors of different dimensions are uniformly mapped to a 256-dimensional feature space to generate aligned feature vectors.

5. The method for constructing an unsupervised deep neural network based on ensemble learning according to claim 4, characterized in that, The feature difference is calculated as follows: In the 256-dimensional feature space, the L2 norm distance between the first and second, the first and third, and the second and third aligned feature vectors are calculated respectively. Each distance value is calculated by taking the square root of the sum of the squares of the differences between the elements of the vector. The arithmetic mean of the three distance values ​​is used as the feature difference between the networks. The adaptive threshold is initially set to 0.

5. During the dynamic optimization process, when the number of training rounds is greater than 11, it is dynamically adjusted based on the mean and standard deviation of the difference in the previous 10 training rounds. The threshold update formula is: Adaptive difference threshold = previous mean + 0.8 × previous standard deviation. When the difference is greater than the threshold, label correction is triggered. When it is less than or equal to the threshold, the original pseudo label is kept unchanged.

6. The method for constructing an unsupervised deep neural network based on ensemble learning according to claim 2, characterized in that, For the updated neural network output, reliability is improved through ensemble decision-making. Specifically, the initial arrival time estimates of the three networks are sorted in ascending order, the maximum and minimum values ​​that may have abnormal biases are removed, and the median is selected as the ensemble decision value to improve the stability of the results. The absolute difference between each network estimate and the ensemble decision value is calculated as the decision bias value. At the same time, historical records of samples with confidence scores greater than 0.7 are selected from the pseudo-label pool metadata domain, and the historical decision bias value or historical decision preset value of the corresponding network in these records is extracted. The statistical mean and standard deviation are calculated based on historical decision deviation values ​​or historical decision preset values. The mean plus twice the standard deviation is defined as the upper limit of expected deviation, which is used to measure the reliability of the current network decision.

7. The method for constructing an unsupervised deep neural network based on ensemble learning according to claim 6, characterized in that, The uncertainty contribution decomposition is as follows: 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 the original anomaly score. The larger the ratio, the higher the network decision uncertainty. If the deviation value is less than or equal to the upper limit, the original anomaly score is set to 0. The Softmax function is used to normalize the three original anomaly scores so that the sum of the normalized components is 1, thus obtaining the normalized uncertainty contribution. The contribution component with the largest value and its corresponding source network are located. This network is the main source of the current decision uncertainty and is the target for subsequent causal analysis.

8. The method for constructing an unsupervised deep neural network based on ensemble learning according to claim 7, characterized in that, The diagnostic feedback package generation includes: extracting the feature map of the last convolutional layer output during the inference process of the source network corresponding to the maximum contribution, with the feature map dimension consistent with the time step of the input waveform sample; calculating the absolute value of the activation value of each channel of the feature map, taking the maximum value along the channel direction to generate a single-channel saliency map, highlighting the region with the strongest feature response; multiplying the saliency map point by point with the original input waveform sample to obtain an attribution heatmap, which intuitively reflects the key regions in the waveform that lead to decision uncertainty; locating the three coordinate points with the highest values ​​in the heatmap, recording their time position and the amplitude value of the corresponding original waveform, forming an attribution feature coordinate set; and finally combining the normalized uncertainty contribution, source network identifier, attribution feature coordinate set, and sample index into a structured diagnostic feedback package in JSON format for easy subsequent parsing and processing.

9. The method for constructing an unsupervised deep neural network based on ensemble learning according to claim 8, characterized in that, The confidence level is adjusted by subtracting the maximum normalized uncertainty contribution component in the diagnostic feedback package from 1 to obtain a confidence attenuation factor, which ranges from [0,1]. The original confidence score is multiplied by the attenuation factor to update the confidence score of the corresponding sample in the pseudo-label pool, thereby realizing the linkage adjustment between confidence level and decision reliability.

Images