Effective signal extraction method and system based on CVAE

Abstract

This application discloses an effective signal extraction method and system based on CVAE, relating to the field of digital signal processing. The specific steps are as follows: constructing a conditional variational autoencoder (CVAE), which includes an encoder, a reparameterization module, and a decoder; constructing a training dataset based on an observation dataset and a corresponding set of conditional variables; training the CVAE using the training dataset and optimizing its parameters using a loss function based on physical consistency constraints; the loss function based on physical consistency constraints includes a reconstruction error term, a KL divergence loss term, and a physical constraint loss term; obtaining the data to be processed and the corresponding conditional variables; and inputting the data to be processed and the conditional variables into the trained CVAE to obtain the effective signal. This application introduces physical consistency constraints through a CVAE, effectively reducing high-frequency noise and pseudo-frequency components, making the reconstructed signal closer to the real signal.

Description

Technical Field

[0001] This application relates to the field of digital signal processing, and in particular to an effective signal extraction method and system based on CVAE. Background Technology

[0002] Currently, in terms of effective signal extraction and noise suppression, existing methods mainly focus on traditional time-frequency signal processing algorithms, including methods based on traditional filtering (such as low-pass filters and band-pass filters), methods based on time-frequency transforms (such as wavelet transforms and short-time Fourier transforms), methods based on adaptive decomposition (such as empirical mode decomposition, ensemble empirical mode decomposition, and variational mode decomposition), and methods based on statistical learning (such as principal component analysis and independent component analysis). Among these, methods based on traditional filtering are straightforward but have limited ability to process non-stationary signals and complex noise environments, often resulting in the weakening of useful information. Methods based on time-frequency transforms have limited time-frequency resolution and are also limited by... Basis function selection is highly dependent on human experience. Adaptive decomposition methods suffer from problems such as mode mixing, endpoint effects, and the need for pre-setting parameters. Statistical learning-based methods rely on prior assumptions for separation effectiveness and are insufficiently adaptable to complex nonlinear signals. In summary, these methods often fail to effectively balance signal fidelity and noise suppression when dealing with complex noisy environments and non-stationary signals, and lack the ability to extract features for specific tasks, resulting in insufficient interpretability and stability of the extraction results. Therefore, how to develop a method that can adaptively extract effective signals, suppress noise, and enhance interpretability and stability is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0003] In view of this, the present invention provides an effective signal extraction method and system based on CVAE, which overcomes the above-mentioned defects.

[0004] To achieve the above objectives, this application provides the following solution: Firstly, this application provides an effective signal extraction method based on CVAE, the specific steps of which are as follows: A conditional variational autoencoder is constructed, which includes an encoder, a reparameterization module, and a decoder; A training dataset is constructed based on the observation dataset and the corresponding set of condition variables; The conditional variational autoencoder is trained using the training dataset, and its parameters are optimized using a loss function based on physical consistency constraints. The loss function based on physical consistency constraints includes a reconstruction error term, a KL divergence loss term, and a physical constraint loss term. Obtain the data to be processed and the corresponding condition variables; The data to be processed and the conditional variables are input into the trained conditional variational autoencoder to obtain an effective signal.

[0005] Optionally, the training steps of the conditional variational autoencoder are as follows: The encoder takes the observed data sample and the condition variable sample as inputs and outputs the mean vector and standard deviation of the latent distribution. The mean vector and the standard deviation are input into the reparameterization module, which outputs a latent vector. The latent vector is input into the decoder, which outputs the reconstructed signal. The loss function is calculated based on the observed data sample, the conditional variable sample, and the reconstructed signal. The parameters of the conditional variational autoencoder are iteratively optimized based on the loss function to obtain the trained conditional variational autoencoder.

[0006] Optionally, the encoder includes a gated recurrent unit branch and a multi-scale convolutional feature fusion branch.

[0007] Optionally, the gated loop unit branch includes an update gate module, a reset gate module, and a hidden state calculation module; the update gate module is used to control the degree of inheritance of the current hidden state from the previous hidden state; The reset gate module is used to adjust the degree of mixing between the current input and the hidden state of the previous time. The state calculation module is used to calculate the candidate hidden state based on the current input and the output of the reset gate module, and to generate the current hidden state based on the output of the update gate module, the hidden state of the previous time and the candidate hidden state.

[0008] Optionally, the multi-scale convolutional feature fusion branch includes multiple convolutional branches of different scales set in parallel. Each convolutional branch extracts local features from the input data. After pooling and activation function transformation, the multiple local features are concatenated to generate multi-scale fusion features.

[0009] Optionally, the expression for the loss function based on physical consistency constraints is: ; In the formula, It is an L2 norm; For the input observation data; The reconstructed signal output by the model; All are weighting coefficients; The Kullback-Leibler divergence between the encoder's approximate distribution and the prior distribution; To provide a given input observation data With condition variables In the case of latent variables The probability distribution; For a given condition variable In the case of latent variables The prior probability distribution; This is the physical constraint loss term.

[0010] Optionally, the expression for the physical constraint loss term is: ; In the formula, It is an L2 norm; Fourier transform; For the input observation data; The reconstructed signal output by the model; These are the weighting coefficients; This refers to the signal frequency range.

[0011] Secondly, this application provides an efficient signal extraction system based on CVAE, comprising: The model building module is used to build a conditional variational autoencoder, which includes an encoder, a reparameterization module, and a decoder. The dataset construction module is used to build a training dataset based on the observation dataset and the corresponding set of condition variables. The model training module is used to train the conditional variational autoencoder using the training dataset and to optimize the parameters of the conditional variational autoencoder using a loss function based on physical consistency constraints; the loss function based on physical consistency constraints includes a reconstruction error term, a KL divergence loss term, and a physical constraint loss term. The data acquisition module is used to acquire the data to be processed and the corresponding condition variables; The data output module is used to input the data to be processed and the conditional variables into the trained conditional variational autoencoder to obtain an effective signal.

[0012] According to the specific embodiments provided in this application, this application has the following technical effects: (1) Stronger noise suppression capability: By introducing physical consistency constraints through conditional variational autoencoder (CVAE), high-frequency noise and pseudo-frequency components are effectively reduced, making the reconstructed signal closer to the real signal.

[0013] (2) Targeted and accurate feature extraction: By introducing conditional variables (labels) and physical consistency constraints, the model not only depends on the input data, but also performs guided feature extraction according to the task objectives. At the same time, global temporal dependencies and local detail features are extracted through GRU units and multi-scale convolution, which improves the stability and reliability of signal separation.

[0014] (3) Improve interpretability and stability: The loss function is improved by physical prior constraints, avoiding the "black box" problem of traditional deep learning models and ensuring that the generated results conform to the physical laws of signal propagation.

[0015] (4) Fast processing speed: The combined improved algorithm based on CVAE can quickly infer and reconstruct signals after training, reducing processing time and making it suitable for application scenarios that require fast signal analysis or processing. Furthermore, it adopts adaptive reparameterization techniques to optimize computational efficiency. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the overall method flow in one embodiment of this application; Figure 2 This is a schematic diagram of the CVAE data processing flow in one embodiment of this application; Figure 3 This is a schematic diagram of the Encoder network structure according to an embodiment of this application; Figure 4 This is a schematic diagram comparing the input signal and the reconstructed effective signal according to an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] This application discloses an effective signal extraction method based on CVAE, such as... Figure 1 As shown, the specific steps are as follows: Step 1: Construct a conditional variational autoencoder, which includes an encoder, a reparameterization module, and a decoder; Step 2: Construct a training dataset based on the observation dataset and the corresponding set of condition variables; Step 3: Train the conditional variational autoencoder using the training dataset, and optimize the parameters of the conditional variational autoencoder using a loss function based on physical consistency constraints; the loss function based on physical consistency constraints includes a reconstruction error term, a KL divergence loss term, and a physical constraint loss term. Step 4: Obtain the data to be processed and the corresponding condition variables; Step 5: Input the data to be processed and the conditional variables into the trained conditional variational autoencoder to obtain the effective signal.

[0021] In one embodiment, the training steps of the conditional variational autoencoder are as follows: Input the observed data sample and the condition variable sample into the encoder, and output the mean vector and standard deviation of the latent distribution; Input the mean vector and standard deviation into the reparameterization module to output the latent vector; The latent vector is input into the decoder, and the output is the reconstructed signal; The loss function is calculated based on the observed data samples, conditional variable samples, and reconstructed signals. The parameters of the conditional variational autoencoder are then iteratively optimized based on the loss function to obtain a trained conditional variational autoencoder.

[0022] Furthermore, the Conditional Variational Autoencoder (CVAE) is a deep learning method based on a probabilistic generative model. It not only maps the input data to a low-dimensional continuous latent space for compact representation, but also introduces conditional constraints during the latent space construction process, so that the latent vector can maintain consistency with the given conditional variables while expressing the core features of the original signal.

[0023] Furthermore, because the latent space performs probabilistic compression and regularization on the input data during modeling, high-frequency invalid components such as noise are often weakened during the encoding stage, thus achieving a natural noise reduction effect. In the decoding stage, CVAE reconstructs the signal based on the smooth and structured latent representation, combined with conditional information, reconstructing only the main patterns relevant to the task. This ensures that the final output retains effective signal features while suppressing random noise interference. In this latent space, not only are redundant features removed and compressed, but it also becomes a controllable and structured probabilistic space where the model can perform conditional sampling and generate an output that meets specific task requirements and exhibits significant noise suppression through the decoder.

[0024] In one embodiment, such as Figure 2 As shown, the core structure of CVAE consists of four parts: encoder, reparameterization trick, decoder, and loss function. Their specific implementation is as follows: (1) Encoder: In CVAE, the encoder receives input data (i.e., observation data). and the corresponding condition variables Through a deep neural network, the encoder maps the input data to a latent space and outputs two parameters of the latent distribution. and Its expression is: ; In the formula, Let be the mean vector of the latent variable distribution; The standard deviation of the latent variable distribution; For the input observation data; This is conditional information; it is introduced through... The encoder can preserve condition-related effective features during feature compression, thereby enhancing the model's task adaptability.

[0025] (2) Reparameterization: Since directly sampling the high-dimensional latent distribution would prevent the model from backpropagating, CVAE employs a reparameterization technique to decouple randomness from the distribution parameters. However, the main features of the input data contained in the latent space of a traditional Conditional Variational Autoencoder (CVAE) cannot be dynamically adjusted based on the loss function during training. Therefore, this embodiment adds an adjustment coefficient to dynamically adjust the standard deviation of the variable distribution to expand the range of the latent space. When used for effective signal extraction, if the signal reconstruction error is large, indicating that the model has not accurately captured the main features of the signal, the range of the latent space is expanded to encompass more features of the input signal; if the signal reconstruction error is small, indicating that the model has accurately captured most of the main features of the signal, the range of the latent space is shrunk to remove redundant signal features and optimize computational efficiency. The specific expression is as follows: ; In the formula, The latent vector carries a low-dimensional representation of the input data in the latent space; Let be the mean vector of the latent variable distribution; This is an adjustment factor, with a value ranging from 0 to 2; The standard deviation of the latent variable distribution; Each component of the vector is independent and identically distributed Gaussian noise.

[0026] To ensure a smooth adjustment of the potential space, the adjustment coefficient is used. The design is based on the hyperbolic tangent function, and the specific formula is as follows: ; In the formula, The adjustment coefficient has a value range of 0 to 2. It is the hyperbolic tangent function; This is the loss function for CVAE.

[0027] (3) Decoder: The decoder receives the latent vector and condition variables And reconstructed data of the input data is recovered through nonlinear mapping. Its expression is: ; In the formula, The reconstructed signal output by the model; The latent vector carries a low-dimensional representation of the input data in the latent space; Indicates conditional information.

[0028] This conditional constraint mechanism can effectively avoid excessive randomness in the generated results and improve the relevance and consistency of signal generation.

[0029] (4) Loss Function: In this embodiment, to ensure that the extracted effective signal conforms to actual physical laws during the processing of complex signals (such as seismic signals during excavation) and to guarantee the physical consistency of the model, an improved loss function is used, incorporating a physical consistency constraint. In the formula of this embodiment, this is expressed as... (Band-Constrained ConsistencyRegularization) represents this part, and the mathematical expression of the loss function is: ; In the formula, It is an L2 norm; For the input observation data; The reconstructed signal output by the model; The Kullback-Leibler divergence between the encoder approximation and the prior distribution is used to constrain the latent space distribution; Given input data With condition variables In the case of latent variables The probability distribution; For a given condition variable In the case of latent variables The prior probability distribution; This refers to the physical consistency constraint part; These are all weighting coefficients, representing the weights of the KL divergence and physical consistency constraints, respectively.

[0030] It represents the mean square error between the original signal and the reconstructed signal in the time domain, and is used to measure the error between the original signal and the reconstructed signal; The Kullback-Leibler divergence between the encoder's approximate distribution and the prior distribution is used to constrain the latent space distribution. This is used to ensure that the generated result conforms to the physical laws of signal generation, and it is based on the original signal. (i.e., the input observed signal) and the generated signal The frequency domain characteristics are compared, and the expression is: ; In the formula, It is an L2 norm; Fourier transform; For the input observation data; The reconstructed signal output by the model; The signal frequency range; is the weighting coefficient, representing the weight of the pseudo-frequency component signal.

[0031] in, The mid-frequency range is This represents the mean square error between the spectra of the original and reconstructed signals in the frequency domain. By limiting the spectral differences between the generated and original signals within this range, the frequency range of the generated signal deviates from reality. Furthermore, This represents the energy generated by the signal outside the aforementioned effective frequency range, and it is used to add an energy attenuation regularity to avoid invalid frequency components; using The role of balanced frequency band consistency and out-of-band suppression.

[0032] In one embodiment, the encoder includes a gated recurrent unit branch and a multi-scale convolutional feature fusion branch.

[0033] Furthermore, in this embodiment, the encoder network adopts a hybrid structure design that combines temporal modeling with multi-scale convolution, simultaneously capturing both global temporal features and local detail features of the input signal, making the main features captured in the latent space more accurate and comprehensive. The encoder includes gated recurrent unit (GRU) branches and convolutional feature branches at different scales, such as... Figure 3 As shown, the final step involves activating and amplifying salient features while suppressing weakly correlated components.

[0034] In one embodiment, the gated loop unit branch includes an update gate module, a reset gate module, and a hidden state calculation module; the update gate module is used to control the degree of inheritance of the current hidden state from the previous hidden state; The reset gate module is used to adjust the degree of blending between the current input and the hidden state of the previous time step; The state calculation module is used to calculate the candidate hidden state based on the current input and the output of the reset gate module, and to generate the current hidden state based on the output of the update gate module, the hidden state of the previous time step, and the candidate hidden state.

[0035] Furthermore, the gated recurrent unit (GRU) is an improved structure of recurrent neural network (RNN). It mainly adds update gates and reset gates to the RNN, enabling the model to adaptively choose to retain or forget information.

[0036] The update gate controls how much of the previous time step's hidden state is inherited from the current time step, and is output through the Sigmoid function. The values ​​between 0 and 1 represent the complete discarding of the previous state information and the complete inheritance of the previous state information, respectively. The mathematical expression for this is: ; In the formula, To update the door; For the Sigmoid function; To update the gate weight matrix; This is the hidden state from the previous moment; This is the input for the current moment; To update the gate offset.

[0037] The reset gate determines the mixed state of the current input and the previous state, and outputs it through the Sigmoid function. The values ​​are between 0 and 1, where 0 indicates that the previous state information is completely ignored and 1 indicates that the previous state information is fully utilized. The mathematical expression for this is: ; In the formula, To reset the door; For the Sigmoid function; To reset the gate weight matrix; This is the hidden state from the previous moment; This is the input for the current moment; To reset the door offset.

[0038] The candidate hidden state is calculated based on the current input and the result of resetting the gate, representing possible new hidden states. Its mathematical expression is as follows: ; In the formula, The candidate hidden state is calculated based on the current input and the result of resetting the gate; It is the hyperbolic tangent function; Candidate state weight matrix; To reset the door; This is the hidden state from the previous moment; This is the input for the current moment; Candidate state bias; This is element-wise multiplication.

[0039] The expression for updating the hidden state is: ; In the formula, The current hidden state; To update the door; This is element-wise multiplication; This is the hidden state from the previous moment; To update the door; The candidate hidden state is calculated based on the current input and the result of resetting the gate.

[0040] Through the above mechanism, the gated recurrent unit (GRU) can effectively obtain global dependencies across time steps while ensuring structural simplicity, and adaptively select information to retain or forget.

[0041] In one embodiment, the multi-scale convolutional feature fusion branch includes multiple convolutional branches of different scales set in parallel. Each convolutional branch extracts local features from the input data. After pooling and activation function transformation, the multiple local features are concatenated to generate multi-scale fused features.

[0042] Furthermore, convolutional feature branches of different scales aim to extract local detail features and multi-level structural information from the input signal, thereby compensating for the shortcomings of recurrent neural networks in capturing rapidly changing local features. By setting 1×1, 1×3, 1×5, and 1×7, convolutional kernels of different scales can extract feature information under different receptive fields, overcoming the defect of losing features by a single convolutional scale. The expression is as follows: ; In the formula, It is a non-linear activation function (such as ReLU); This is a pooling operation; The weights of the convolution kernel are of scale k; This represents the convolution operation; Input data; The bias is of scale k.

[0043] In one embodiment, the parameter settings and training steps for applying the conditional variational autoencoder are as follows: Step 1: Set the initial parameters of the Conditional Variational Autoencoder (CVAE), including the number of convolutional branches, the number of convolutional branch layers, the convolutional kernel size, the pooling function selection, the pooling function parameters, the number of recurrent neural network layers, and the activation function selection in the encoder; and the number of fully connected layers, the input and output parameters of the fully connected function, and the activation function selection in the decoder, i.e., define the network structure.

[0044] Step 2: Preprocess the input data and label the signal type. Perform one-hot encoding and combine with the original signal Sequences are combined and used as joint inputs to the encoder.

[0045] Step 3: Set the hyperparameters involved in the training process, namely, batch size, number of training epochs, learning rate, and optimizer type.

[0046] Step 4: Input the preprocessed data into the encoder, extract global and local features through gated recurrent units (GRU) and multi-scale convolutional structures, and map them into the latent space, outputting the latent space distribution parameters: mean vector. and standard deviation Combining adaptive reparameterization techniques and introducing adjustment coefficients (The adjustment factor is set to 1 by default when there is no loss function) and noise vector Generate the potential vector z.

[0047] Step 5: Transfer the latent vector With signal type label The input is to the decoder, which outputs a reconstructed signal through a fully connected layer and nonlinear activation operations. Signal.

[0048] Step 6: Calculate the loss function and update the model parameters. Specifically, the mean squared error (MSE) method is used to calculate the time-domain error between the original signal and the reconstructed signal; the Kullback-Leibler divergence is used to partially constrain the signal to ensure that the reconstructed signal has a continuous structure; the loss function based on the above-mentioned physical constraint improvement is used to ensure that the reconstructed signal conforms to physical laws. The loss function is calculated and iterated by combining the three parts.

[0049] Step 7: After training is complete, save the model parameters to form a conditional variational autoencoder model for prediction. Input the signal to be processed and the corresponding conditional variables into the trained CVAE model. The model output is the effective signal extracted by the algorithm. The comparison diagram between the effective signal and the signal to be processed is shown in the figure below. Figure 4 As shown.

[0050] This embodiment also discloses an efficient signal extraction system based on CVAE, including: The model building module is used to build a conditional variational autoencoder, which includes an encoder, a reparameterization module, and a decoder. The dataset construction module is used to build a training dataset based on the observation dataset and the corresponding set of condition variables. The model training module is used to train the conditional variational autoencoder using the training dataset and to optimize the parameters of the conditional variational autoencoder using a loss function based on physical consistency constraints. The loss function based on physical consistency constraints includes a reconstruction error term, a KL divergence loss term, and a physical constraint loss term. The data acquisition module is used to acquire the data to be processed and the corresponding condition variables; The data output module is used to input the data to be processed and the conditional variables into the trained conditional variational autoencoder to obtain effective signals.

[0051] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0052] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. In summary, the content of this specification should not be construed as a limitation of this application.

Claims

1. An effective signal extraction method based on CVAE, characterized in that, The specific steps are as follows: A conditional variational autoencoder is constructed, which includes an encoder, a reparameterization module, and a decoder; A training dataset is constructed based on the observation dataset and the corresponding set of condition variables; The conditional variational autoencoder is trained using the training dataset, and its parameters are optimized using a loss function based on physical consistency constraints. The loss function based on physical consistency constraints includes a reconstruction error term, a KL divergence loss term, and a physical constraint loss term. Acquire the data to be processed and the corresponding condition variables, wherein the data to be processed is the seismic signal during excavation; The data to be processed and the conditional variables are input into the trained conditional variational autoencoder to obtain an effective signal; The expression for the physical constraint loss term is as follows: ； In the formula, It is an L2 norm; Fourier transform; For the input observation data; The reconstructed signal output by the model; These are the weighting coefficients; The signal frequency range; The medium signal frequency range is , which represents the mean square error between the spectrum of the observed signal and the reconstructed signal in the frequency domain. By limiting the spectral difference between the reconstructed signal and the observed signal within the signal frequency range, the occurrence of the frequency range of the reconstructed signal deviating from the actual value is reduced.

2. The effective signal extraction method based on CVAE according to claim 1, characterized in that, The training steps for the conditional variational autoencoder are as follows: The encoder takes the observed data sample and the condition variable sample as inputs and outputs the mean vector and standard deviation of the latent distribution. The mean vector and the standard deviation are input into the reparameterization module, which outputs a latent vector. The latent vector is input into the decoder, which outputs the reconstructed signal. The loss function is calculated based on the observed data sample, the conditional variable sample, and the reconstructed signal. The parameters of the conditional variational autoencoder are iteratively optimized based on the loss function to obtain the trained conditional variational autoencoder.

3. The effective signal extraction method based on CVAE according to claim 1, characterized in that, The encoder includes a gated recurrent unit branch and a multi-scale convolutional feature fusion branch.

4. The effective signal extraction method based on CVAE according to claim 3, characterized in that, The gated loop unit branch includes an update gate module, a reset gate module, and a hidden state calculation module; The update gate module is used to control the degree of inheritance of the current hidden state from the previous hidden state; The reset gate module is used to adjust the degree of mixing between the current input and the hidden state of the previous time. The state calculation module is used to calculate the candidate hidden state based on the current input and the output of the reset gate module, and to generate the current hidden state based on the output of the update gate module, the hidden state of the previous time and the candidate hidden state.

5. The effective signal extraction method based on CVAE according to claim 3, characterized in that, The multi-scale convolutional feature fusion branch includes multiple convolutional branches of different scales set in parallel. Each convolutional branch extracts local features from the input data. After pooling and activation function transformation, the multiple local features are concatenated to generate multi-scale fused features.

6. The effective signal extraction method based on CVAE according to claim 1, characterized in that, The expression for the loss function based on physical consistency constraints is: ； In the formula, It is an L2 norm; For the input observation data; The reconstructed signal output by the model; All are weighting coefficients; The Kullback-Leibler divergence between the encoder's approximate distribution and the prior distribution; To provide a given input observation data With condition variables In the case of latent variables The probability distribution; For a given condition variable In the case of latent variables The prior probability distribution; This is the physical constraint loss term.

7. An effective signal extraction system based on CVAE, characterized in that, include: The model building module is used to build a conditional variational autoencoder, which includes an encoder, a reparameterization module, and a decoder. The dataset construction module is used to build a training dataset based on the observation dataset and the corresponding set of condition variables. The model training module is used to train the conditional variational autoencoder using the training dataset and to optimize the parameters of the conditional variational autoencoder using a loss function based on physical consistency constraints. The loss function based on physical consistency constraints includes a reconstruction error term, a KL divergence loss term, and a physical constraint loss term. The expression for the physical constraint loss term is: ； In the formula, It is an L2 norm; Fourier transform; For the input observation data; The reconstructed signal output by the model; These are the weighting coefficients; The signal frequency range; The medium signal frequency range is , which represents the mean square error between the spectrum of the observed signal and the reconstructed signal in the frequency domain. By limiting the spectral difference between the reconstructed signal and the observed signal within the signal frequency range, the occurrence of the frequency range of the reconstructed signal deviating from the actual value is reduced. The data acquisition module is used to acquire the data to be processed and the corresponding condition variables, wherein the data to be processed is the seismic signal during excavation; The data output module is used to input the data to be processed and the conditional variables into the trained conditional variational autoencoder to obtain an effective signal.

Images