Augmentation system and method for field groundwater magnetic resonance detection small sample data

Abstract

The invention relates to the field of field groundwater magnetic resonance detection methods, in particular to an augmentation system and method for field groundwater magnetic resonance detection small sample data. Comprising a generation network configured to receive random noise and generate augmented data based on the random noise; the discrimination network comprises a time domain discrimination network and a frequency domain discrimination network; the generative network and the time domain discrimination network form a first discrimination channel, the generative network and the frequency domain discrimination network form a second discrimination channel, and in the second discrimination channel, augmented data generated by the generative network and the noisy data are jointly input into the frequency domain discrimination network; the frequency domain discrimination network is used for discriminating whether the source of input data is a generation network or noisy data in a frequency domain. According to the method, a magnetic resonance detection small sample noisy data set is established, and multi-scale features in an amplitude spectrum and a phase spectrum are captured through the frequency domain discrimination network; the collaborative driving generation network generates a high-quality sample which is close to small sample noisy data in waveform form and spectral characteristics.

Description

Technical Field

[0001] This disclosure relates to the field of magnetic resonance sounding (MRS) data processing technology, specifically to an augmentation system and method for small sample data from field groundwater magnetic resonance sounding. It is suitable for augmenting MRS data in small sample scenarios, providing sufficient high-quality data for subsequent noise reduction, inversion and other processing. Background Technology

[0002] With water scarcity, efficient and accurate groundwater resource detection is crucial. Magnetic resonance sounding (MRS) is a geophysical method that directly utilizes the nuclear magnetic resonance effect of hydrogen nuclei in water molecules within the Earth's magnetic field to receive spin echo signals and obtain aquifer information. It offers advantages in direct water detection and quantitative assessment. However, MRS faces two major challenges in field exploration: first, severe electromagnetic noise interference results in extremely low signal-to-noise ratios; second, the long exploration cycle and high cost limit the availability of high-quality, effective data samples. This hinders the application of deep learning methods, which rely on large amounts of data, in MRS noise reduction and inversion. Therefore, introducing magnetic resonance sounding data augmentation technology has become a key approach to addressing the data shortage. This technology effectively augments existing sparse, low-signal-to-noise-ratio MRS signals with feature-preserving data to increase the diversity and quantity of data samples, thereby improving the model's generalization ability.

[0003] Existing data augmentation methods suffer from limitations in functionality and applicability. Traditional generative adversarial networks (GANs) are often used for augmenting two-dimensional datasets such as images. However, when applied to one-dimensional MRS signals, the one-dimensional signal usually needs to be converted into a two-dimensional image before being input into the GAN. This conversion can only map to a limited number of values, and accuracy is inevitably lost during post-training reconstruction. While Long Short-Term Memory (LSTM) networks, which are specifically designed for processing time series, can capture temporal features, they inevitably discard more and more information when the MRS signal sequence is too long, making it easier to lose key information of the sequence. Although the Transformer architecture (a neural network architecture based on self-attention) is good at capturing long-distance dependencies due to its attention mechanism, it has limitations such as a large number of parameters, the need for more computational resources, and the need for more data for effective optimization during training. Summary of the Invention

[0004] The purpose of this disclosure is to address the shortcomings of the prior art by providing an augmentation system and method for small sample data of magnetic resonance imaging of groundwater in the field.

[0005] The first aspect of this disclosure provides an augmentation system for small sample data from magnetic resonance imaging of groundwater in the field, comprising: A network is generated, configured to receive random noise and generate augmented data based on the random noise; Discriminant networks include time-domain discriminant networks and frequency-domain discriminant networks; The generator network and the time-domain discrimination network constitute a first discrimination channel, and the generator network and the frequency-domain discrimination network constitute a second discrimination channel. In the first discrimination channel, the augmented data generated by the generator network and the acquired noisy data are jointly input to the time-domain discrimination network, which is used to determine in the time domain whether the source of the input data is the generator network or the noisy data; In the second discrimination channel, the augmented data generated by the generating network and the noisy data are jointly input to the frequency domain discrimination network, which is used to determine in the frequency domain whether the source of the input data is the generating network or the noisy data; The training of the augmentation system is complete when the discrimination network is unable to determine whether the data originates from the generating network or from noisy data.

[0006] Furthermore, the generative network includes: The first feature extraction module is configured to receive random noise and use a two-layer long short-term memory network to extract temporal features from the random noise, and output the first temporal feature. The second feature extraction module, connected to the first feature extraction module, is configured to perform global dependency modeling on the first temporal features using a two-layer Transformer encoder and output enhanced features. Each layer of the Transformer encoder includes a multi-head attention mechanism and a feedforward fully connected layer, which are connected through residuals. The upsampling module is configured to sample the enhanced features to the target dimension to generate augmented data.

[0007] Furthermore, the temporal discriminant network includes at least one downsampling layer, wherein the downsampling layer includes multiple fully connected layers that map the dimensional features of the input data to a (1,1)-dimensional first discriminant score.

[0008] Furthermore, the frequency domain discrimination network includes: The time-frequency transformation module is configured to perform a fast Fourier transform on the input data, output amplitude components and phase components, and stack the amplitude components and phase components into a dual-channel feature map; The convolutional feature extraction module includes multiple sequentially connected convolutional layers, which are configured to sequentially transform the number of channels and the frequency domain dimension of the input feature map; The fully connected discriminant module includes a first fully connected layer and a second fully connected layer connected in sequence, which maps the feature map output by the convolutional feature extraction module to a (1,1)-dimensional second discriminant score.

[0009] Furthermore, the discrimination network also includes a fusion module, which is used to perform a fusion operation on the first discrimination score and the second discrimination score, and output the final discrimination result.

[0010] Furthermore, the total loss function for training the augmented system includes: The first loss function is used to constrain the mapping of the generator network from random noise to noisy data. The first loss function is expressed as: negative time-domain discriminant expectation term minus product term, wherein the product term is the frequency-domain loss weight coefficient multiplied by the frequency-domain discriminant expectation term. The second loss function is used to constrain the temporal discriminant network's discrimination of random noise and the temporal component sources of noisy data. The second loss function includes a first loss term and a second loss term. The first loss term is the discrimination result of the temporal discriminant network on the noisy data. The logarithm of the discrimination result is taken and then the negative value is taken. The second loss term is the discrimination result of the temporal discriminant network on the augmented data. The complementary value is obtained by subtracting the discrimination result from 1. The logarithm of the complementary value is taken and then the negative value is taken. The third loss function is used to constrain the frequency domain discriminant network's discrimination of random noise and the source of the frequency domain components of noisy data. The third loss function is expressed as the sum of the third loss term and the fourth loss term. The third loss term is the discrimination result of the frequency domain discriminant network on the noisy data. The logarithm of the discrimination result is taken and then the negative value is taken. The fourth loss term is the discrimination result of the frequency domain discriminant network on the augmented data. The complementary value is obtained by subtracting the discrimination result from 1. The logarithm of the complementary value is taken and then the negative value is taken.

[0011] Furthermore, the augmented system trains a total loss function to establish constraints for the generative adversarial mode, requiring the generative network to minimize the first loss function and the discriminator network to maximize the second and third loss functions.

[0012] A second aspect of this disclosure provides an augmentation method for small sample data from field groundwater magnetic resonance detection, comprising: constructing a noisy dataset including a clean signal set and a noise set from a small sample of field groundwater magnetic resonance detection; A data augmentation network is constructed, comprising a generation network and a discrimination network. The discrimination network consists of a time-domain discrimination network and a frequency-domain discrimination network. The generation network is connected to the time-domain discrimination network to form a first discrimination channel. Random noise flows through the generation network and, after being output, flows into the time-domain discrimination network along with the noisy data to be identified as a source. The generation network is connected to the frequency-domain discrimination network to form a second discrimination channel. Random noise flows through the generation network and, after being output, flows into the frequency-domain discrimination network along with the noisy data to be identified as a source. The training of the data augmentation network is complete when the discrimination network cannot determine whether the input data comes from the generator network or noisy data. Generative network augmentation using data augmentation networks is employed to augment small sample data from magnetic resonance imaging of groundwater in the field.

[0013] Furthermore, the total loss function for training the data augmentation network includes: The first loss function is used to constrain the mapping of the generator network from random noise to noisy data. The first loss function is expressed as: negative time-domain discriminant expectation term minus product term, wherein the product term is the frequency-domain loss weight coefficient multiplied by the frequency-domain discriminant expectation term. The second loss function is used to constrain the temporal discriminant network's discrimination of random noise and the temporal component sources of noisy data. The second loss function includes a first loss term and a second loss term. The first loss term is the discrimination result of the temporal discriminant network on the noisy data. The logarithm of the discrimination result is taken and then the negative value is taken. The second loss term is the discrimination result of the temporal discriminant network on the augmented data. The complementary value is obtained by subtracting the discrimination result from 1. The logarithm of the complementary value is taken and then the negative value is taken. The third loss function is used to constrain the frequency domain discriminant network's discrimination of random noise and the source of the frequency domain components of noisy data. The third loss function is expressed as the sum of the third loss term and the fourth loss term. The third loss term is the discrimination result of the frequency domain discriminant network on the noisy data. The logarithm of the discrimination result is taken and then the negative value is taken. The fourth loss term is the discrimination result of the frequency domain discriminant network on the augmented data. The complementary value is obtained by subtracting the discrimination result from 1. The logarithm of the complementary value is taken and then the negative value is taken.

[0014] Furthermore, the augmented system trains a total loss function to establish constraints for the generative adversarial mode, requiring the generative network to minimize the first loss function and the discriminator network to maximize the second and third loss functions.

[0015] Compared with the prior art, this disclosure has at least the following beneficial effects: This disclosure addresses the limitations of traditional GANs, which lose key sequence information. Using Transformer alone has the drawbacks of large parameter count, high computational resource requirements, and the need for more data for effective optimization during training. It learns the macroscopic waveform patterns of the signal through a time-domain discriminant network and captures multi-scale features in the amplitude and phase spectra through a frequency-domain discriminant network. Together, they drive the generator network to produce high-quality samples that closely approximate the real MRS signal in terms of waveform morphology and spectral characteristics. Attached Figure Description

[0016] Figure 1 This is a network architecture diagram of an augmentation system for small sample data of magnetic resonance detection of groundwater in the field, provided in an embodiment of this disclosure. Figure 2 A network architecture diagram of the generated network provided in this embodiment of the disclosure; Figure 3A network architecture diagram of the discrimination network provided in the embodiments of this disclosure; Figure 4 A flowchart of an augmentation method for small sample data of magnetic resonance detection of groundwater in the field, provided in an embodiment of this disclosure. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this disclosure clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this disclosure.

[0018] See Figure 1 Combination Figure 2 and Figure 3 As shown, an augmentation system for small sample data from magnetic resonance imaging of groundwater in the field includes: A network is generated, configured to receive random noise and generate augmented data based on the random noise; Discriminant networks include time-domain discriminant networks and frequency-domain discriminant networks; The generator network and the time-domain discrimination network constitute a first discrimination channel, and the generator network and the frequency-domain discrimination network constitute a second discrimination channel. The time-domain discrimination network is configured to discriminate between the augmented data and the constructed noisy data at the time-domain waveform level, learning the macroscopic variation law of the signal in time sequence, including the signal envelope shape, attenuation trend and oscillation characteristics; the frequency-domain discrimination network is configured to discriminate between the augmented data and the constructed noisy data at the frequency domain level, capturing local detail information of the signal at different frequency scales by analyzing the energy distribution characteristics of the amplitude spectrum and the continuity characteristics of the phase spectrum; the generation network is trained adversarially under the cooperative constraints of the first discrimination channel and the second discrimination channel, so that the generated augmented data simultaneously approximates the distribution characteristics of the noisy data in both the time-domain waveform structure and the frequency-domain spectral characteristics.

[0019] In the first discrimination channel, the augmented data generated by the generator network and the acquired noisy data are jointly input to the time-domain discrimination network, which is used to determine in the time domain whether the source of the input data is the generator network or the noisy data; In the second discrimination channel, the augmented data generated by the generator network and the acquired noisy data are jointly input to the frequency domain discrimination network, which is used to determine in the frequency domain whether the source of the input data is the generator network or the noisy data; The time-domain discrimination network and the frequency-domain discrimination network output discrimination results respectively, and feed the discrimination results back to the generation network. The generation network adjusts the generation strategy according to the discrimination results of the time-domain discrimination network and the frequency-domain discrimination network to optimize the time-domain waveform features and frequency-domain spectrum features of the augmented data, making it difficult to distinguish from noisy data under dual-domain discrimination.

[0020] In the second discrimination channel, the augmented data generated by the generating network and the noisy data are jointly input to the frequency domain discrimination network, which is used to determine in the frequency domain whether the source of the input data is the generating network or the noisy data; The augmentation system is trained until the discrimination network cannot determine whether the data originates from the generator network or from noisy data. The training dataset is a constructed noisy dataset consisting of a small sample of clean signals and a noisy dataset from field groundwater magnetic resonance imaging.

[0021] In one embodiment, clean signal sets for field groundwater magnetic resonance detection are constructed respectively. , and noise set The clean signal is superimposed with noise to obtain a mixed noisy dataset. .in This represents pure data. Represents a noisy data point. This represents a noisy magnetic resonance imaging (MRI) data point. For the number of data points, A sample size of 96 is chosen to suit the scenario of small sample sizes, verifying that this method can generate a sufficient number of high-quality samples that approximate real MRS signals. The clean signal set is constructed based on free inductive attenuation: , in, For pure signal amplitude, The initial amplitude, The average relaxation time, For the Lamo frequency, For the initial phase, For time; the noise set is constructed by collecting real environmental noise (including relatively pure environmental noise in the wild, complex electromagnetic environmental noise in the city, etc.).

[0022] The training process uses the Adam optimizer, sets the learning rate, and alternately updates the parameters of the generator and discriminator networks to fit random noise to noisy data.

[0023] In one embodiment, see Figure 2 As shown, the generator network includes: The first feature extraction module is configured to receive random noise and use a two-layer long short-term memory network to extract temporal features from the random noise, and output the first temporal feature. The second feature extraction module, connected to the first feature extraction module, is configured to perform global dependency modeling on the first temporal features using a two-layer Transformer encoder and output enhanced features. Each layer of the Transformer encoder includes a multi-head attention mechanism and a feedforward fully connected layer, which are connected through residuals. The upsampling module is configured to sample the enhanced features to the target dimension to generate augmented data.

[0024] In one example, the Long Short-Term Memory (LSTM) network controls the flow of information through three gating mechanisms: the forget gate determines which information to discard from the previous state; the input gate determines which new information to store in the state unit; and the output gate determines which information to output from the state unit. Through the synergistic effect of these three gating mechanisms, the two-layer dual-layer LSM network can effectively capture temporal dependencies in random noise.

[0025] In one example, the input random noise has a dimension of 512. After being processed by a two-layer long short-term memory network, the first temporal feature output has a dimension of 512.

[0026] Each Transformer encoder layer includes an 8-head attention mechanism and a feedforward fully connected layer. In the attention mechanism, self-attention weights are calculated using the query matrix Q, key matrix V, and value matrix K, capturing feature relationships between positions in the input sequence in parallel from eight different representation subspaces. The feedforward fully connected layer performs a non-linear transformation on the attention output, with a hidden dimension of 512. The two-layer Transformer encoder receives the first temporal feature (input dimension 512), and after global dependency modeling, the output augmented feature dimension remains unchanged at 512.

[0027] The upsampling module is connected to the second feature extraction module and consists of multiple components connected in series. It is used to sample the enhanced features to the target dimension.

[0028] The upsampling module specifically includes a Leaky_ReLU activation layer, a batch normalization layer, a first Dropout layer, a first fully connected layer, a second Dropout layer, and a second fully connected layer connected in sequence.

[0029] Finally, the output layer outputs an augmented signal with 2048 dimensions. This augmented data is the time-domain signal sample of the final output of the generator network, which is used as input to the discriminator network for source determination.

[0030] See Figure 3As shown, the discriminant network includes a parallel time-domain discriminant network and a frequency-domain discriminant network. The time-domain discriminant network includes at least one downsampling layer, which comprises multiple fully connected layers that map the dimensional features of the input data to a (1,1)-dimensional first discriminant score. Each fully connected layer undergoes a nonlinear transformation via activation functions, progressively compressing the feature space to extract higher-level discriminative information. The input to the downsampling layer is a 2048-dimensional augmented signal, which is reduced to 1024 dimensions by the first fully connected layer, with nonlinearity introduced using the Leaky ReLU activation function. The second fully connected layer further compresses the features to 512 dimensions, also using Leaky ReLU activation. The third fully connected layer outputs 256-dimensional features, which are stabilized by the Tanh activation function. Finally, the fourth fully connected layer maps to a single-dimensional output, which is normalized using the Sigmoid function to obtain the first discriminant score. The first discriminant score represents the true probability distribution of the augmented data in the time domain.

[0031] The frequency domain discrimination network is connected to the second feature extraction module and is used to extract and discriminate frequency domain features from the augmented signal.

[0032] In one embodiment, the frequency domain discrimination network includes: The time-frequency transformation module is configured to perform a fast Fourier transform on the input data, output amplitude components and phase components, and stack the amplitude components and phase components into a dual-channel feature map; The convolutional feature extraction module includes multiple sequentially connected convolutional layers, which are configured to sequentially transform the number of channels and the frequency domain dimension of the input feature map; The fully connected discriminant module, comprising a third fully connected layer and a fourth fully connected layer connected in sequence, maps the feature map output by the convolutional feature extraction module to a (1,1)-dimensional second discriminant score.

[0033] Specifically, the time-frequency transformation module receives the 2048-dimensional augmented data output by the generator network. It first converts the time-domain signal to a frequency-domain representation using a Fast Fourier Transform (FFT), then calculates the corresponding amplitude and phase spectra. The amplitude component reflects the energy distribution of each frequency component of the signal, while the phase component characterizes the initial phase relationship of different frequency components; together, they constitute the complete frequency domain features of the signal. The amplitude and phase components are stacked along the channel dimension to form a dual-channel feature map of size (2, 1024), where 1024 is the number of frequency domain sampling points, and 2 corresponds to the amplitude and phase channels.

[0034] The convolutional feature extraction module comprises multiple sequentially connected convolutional layers, employing one-dimensional convolutional operations to extract features along the frequency domain dimension. The first convolutional layer has 64 kernels with a kernel size of 7, a stride of 2, and padding of 3, expanding the input 2-channel feature map to 64 channels. Simultaneously, the frequency domain dimension is downsampled from 1024 to 512. Each convolutional layer is followed by a batch normalization layer and a Leaky ReLU activation function to accelerate convergence and enhance non-linear expressive power. The second convolutional layer uses 128 kernels with a kernel size of 5, a stride of 2, and padding of 2, increasing the number of channels from 64 to 128 and further compressing the frequency domain dimension to 256. The third convolutional layer has 256 kernels with a kernel size of 3, a stride of 2, and padding of 1, outputting 256 channels and reducing the frequency domain dimension to 128. The fourth convolutional layer serves as a transition layer, using 512 convolutional kernels with a kernel size of 3, a stride of 2, and padding of 1. It transforms the feature map into a high-dimensional abstract representation with 512 channels and a frequency domain dimension of 64. Finally, the spatial dimension is compressed to (512, 1) through a global average pooling operation.

[0035] The fully connected discriminant module receives a 512-dimensional feature vector output from the convolutional feature extraction module. The third fully connected layer maps the 512-dimensional features to a 128-dimensional hidden space using the Leaky ReLU activation function. The fourth fully connected layer further compresses the 128-dimensional features into a single-dimensional output, which is then normalized using the Sigmoid function to obtain a second discriminant score. This second discriminant score represents the true probability distribution of the augmented data in the frequency domain. The first and second discriminant scores output by the time-domain discriminant network and the frequency-domain discriminant network are combined using a weighted fusion strategy to form a comprehensive discriminant result. This guides the generator network to optimize the quality of the augmented signal, making it approximate the characteristic distribution of the real groundwater magnetic resonance signal in both its time-domain waveform structure and frequency-domain energy distribution.

[0036] The discrimination scores of the time-domain and frequency-domain discrimination networks are combined using a weighted fusion strategy to generate a comprehensive discrimination result. For example, the time-domain weight is set to 0.6 and the frequency-domain weight to 0.4 to balance the temporal continuity of the signal and the physical rationality of the spectrum. The fused discrimination score is fed back to the generator network to guide the update of the generator parameters, forming an adversarial training closed loop.

[0037] In one embodiment, the total loss function for training the augmentation system includes: The first loss function, used to constrain the mapping of the generator network from random noise to noisy data, is expressed as: the negative time-domain discriminant expectation term minus the product term, where the product term is the frequency-domain loss weight coefficient multiplied by the frequency-domain discriminant expectation term; the time-domain discriminant expectation term is: the mathematical expectation of random noise under the random noise distribution in the time domain, where the mathematical expectation is calculated as the logarithm of the discrimination result of the time-domain discriminant network on the augmented data output by the generator network; the frequency-domain discriminant expectation term is: the mathematical expectation of random noise under the random noise distribution in the frequency domain, where the mathematical expectation is calculated as the logarithm of the discrimination result of the frequency-domain discriminant network on the frequency-domain representation of the augmented data output by the generator network after performing a fast Fourier transform.

[0038] Represented as: , in, For the first loss function, Representation function In random noise distribution The expected value of the following mathematical expression It is random noise. This indicates that the output of the generator network will be subjected to a Fast Fourier Transform. This is the frequency domain loss weighting coefficient, set to 0.5; For the output of the time-domain discrimination network, For the output of the frequency domain discrimination network, This is to generate the network's output, i.e., augmented data.

[0039] The second loss function constrains the temporal discriminant network's ability to distinguish the sources of random noise and noisy data in the temporal domain. The second loss function includes a first loss term and a second loss term. The first loss term is the discriminant result of the temporal discriminant network on the noisy data; the logarithm of this result is taken, and then the negative value is applied. The second loss term is the discriminant result of the temporal discriminant network on the augmented data; the complementary value is obtained by subtracting 1 from this result, and the logarithm of the complementary value is taken, and then the negative value is applied. It is expressed as: , in, For the second loss function, For function In noisy data distribution The expected value of the following mathematical expression This indicates noisy data.

[0040] The third loss function constrains the frequency domain discriminant network's ability to distinguish the sources of random noise and noisy data frequency domain components. This third loss function is expressed as the sum of the third and fourth loss terms. The third loss term is the discrimination result of the frequency domain discriminant network on the noisy data; the logarithm of this result is taken, and then the negative value is applied. The fourth loss term is the discrimination result of the frequency domain discriminant network on the augmented data; the complementary value is obtained by subtracting 1 from this result, and the logarithm of this complementary value is taken, and then the negative value is applied. It is expressed as follows: , in, This is the third loss function.

[0041] The constrained training process for generative adversarial mode is established, requiring the generative network to minimize a first loss function and the discriminator network to maximize a second and a third loss function. Based on the first, second, and third loss functions, the final total loss function of the data augmentation network is then established. : .

[0042] On the other hand, see Figure 4 As shown, this disclosure provides an augmentation method for small sample data from field groundwater magnetic resonance detection, including: constructing a noisy dataset comprising a clean signal set and a noise set from a small sample of field groundwater magnetic resonance detection; A data augmentation network is constructed, comprising a generation network and a discrimination network. The discrimination network consists of a time-domain discrimination network and a frequency-domain discrimination network. The generation network is connected to the time-domain discrimination network to form a first discrimination channel. Random noise flows through the generation network and, after being output, flows into the time-domain discrimination network along with the noisy data to be identified as a source. The generation network is connected to the frequency-domain discrimination network to form a second discrimination channel. Random noise flows through the generation network and, after being output, flows into the frequency-domain discrimination network along with the noisy data to be identified as a source. The training of the data augmentation network is complete when the discrimination network cannot determine whether the input data comes from the generator network or noisy data. Generative network augmentation using data augmentation networks is employed to augment small sample data from magnetic resonance imaging of groundwater in the field.

[0043] This method uses a noisy dataset and a generative network to balance the local temporal dynamics and global structural dependencies of small noisy data. It introduces a time-domain discriminant network and a frequency-domain discriminant network. The time-domain discriminant network learns the macroscopic waveform patterns of the signal, and the frequency-domain discriminant network captures multi-scale features in the amplitude spectrum and phase spectrum. Together, they drive the generative network to generate high-quality samples that approximate small noisy data in terms of waveform morphology and spectral characteristics.

[0044] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. An augmentation system for small sample data from magnetic resonance imaging of groundwater in the field, characterized in that, include: A network is generated, configured to receive random noise and generate augmented data based on the random noise; Discriminant networks include time-domain discriminant networks and frequency-domain discriminant networks; The generator network and the time-domain discrimination network constitute a first discrimination channel, and the generator network and the frequency-domain discrimination network constitute a second discrimination channel. In the first discrimination channel, the augmented data generated by the generator network and the acquired noisy data are jointly input to the time-domain discrimination network, which is used to determine in the time domain whether the source of the input data is the generator network or the noisy data; In the second discrimination channel, the augmented data generated by the generating network and the noisy data are jointly input to the frequency domain discrimination network, which is used to determine in the frequency domain whether the source of the input data is the generating network or the noisy data; The training of the augmentation system is complete when the discrimination network is unable to determine whether the data originates from the generating network or from noisy data.

2. The augmentation system for small sample data from magnetic resonance imaging of groundwater in the field, as described in claim 1, is characterized in that... The generative network includes: The first feature extraction module is configured to receive random noise and use a two-layer long short-term memory network to extract temporal features from the random noise, and output the first temporal feature. The second feature extraction module is connected to the first feature extraction module and is configured to perform global dependency modeling on the first temporal features using a two-layer Transformer encoder to output enhanced features. Each Transformer encoder layer includes a multi-head attention mechanism and a feedforward fully connected layer, which are connected through residuals. The upsampling module is configured to sample the enhanced features to the target dimension to generate augmented data.

3. The augmentation system for small sample data from magnetic resonance imaging of groundwater in the field, as described in claim 1, is characterized in that... The temporal discriminant network includes at least one downsampling layer, wherein the downsampling layer includes multiple fully connected layers that map the dimensional features of the input data to a (1,1)-dimensional first discriminant score.

4. The augmentation system for small sample data from magnetic resonance imaging of groundwater in the field, as described in claim 3, is characterized in that... The frequency domain discrimination network includes: The time-frequency transformation module is configured to perform a fast Fourier transform on the input data, output amplitude components and phase components, and stack the amplitude components and phase components into a dual-channel feature map; The convolutional feature extraction module includes multiple sequentially connected convolutional layers, which are configured to sequentially transform the number of channels and the frequency domain dimension of the input feature map; The fully connected discriminant module includes a first fully connected layer and a second fully connected layer connected in sequence, which maps the feature map output by the convolutional feature extraction module to a (1,1)-dimensional second discriminant score.

5. The augmentation system for small sample data from magnetic resonance imaging of groundwater in the field, as described in claim 4, is characterized in that... The discriminant network also includes a fusion module, which is used to perform a fusion operation on the first discriminant score and the second discriminant score, and output the final discriminant result.

6. The augmentation system for small sample data from magnetic resonance imaging of groundwater in the field, as described in claim 1, is characterized in that... The total loss function for training the augmented system includes: The first loss function is used to constrain the mapping of the generator network from random noise to noisy data. The first loss function is expressed as: negative time-domain discriminant expectation term minus product term, wherein the product term is the frequency domain loss weight coefficient multiplied by the frequency domain discriminant expectation term. The second loss function is used to constrain the temporal discriminant network's discrimination of random noise and the temporal component sources of noisy data. The second loss function includes a first loss term and a second loss term. The first loss term is the discrimination result of the temporal discriminant network on the noisy data. The logarithm of the discrimination result is taken and then the negative value is taken. The second loss term is the discrimination result of the temporal discriminant network on the augmented data. The complementary value is obtained by subtracting the discrimination result from 1. The logarithm of the complementary value is taken and then the negative value is taken. The third loss function is used to constrain the frequency domain discriminant network's discrimination of random noise and the source of the frequency domain components of noisy data. The third loss function is expressed as the sum of the third loss term and the fourth loss term. The third loss term is the discrimination result of the frequency domain discriminant network on the noisy data. The logarithm of the discrimination result is taken and then the negative value is taken. The fourth loss term is the discrimination result of the frequency domain discriminant network on the augmented data. The complementary value is obtained by subtracting the discrimination result from 1. The logarithm of the complementary value is taken and then the negative value is taken.

7. The augmentation system for small sample data from magnetic resonance imaging of groundwater in the field, as described in claim 6, is characterized in that... The augmented system trains a total loss function to establish constraints for generative adversarial mode training, requiring the generative network to minimize the first loss function and the discriminator network to maximize the second and third loss functions.

8. A method for augmenting small sample data from magnetic resonance imaging of groundwater in the field, characterized in that, include: Construct a noisy dataset that includes a small sample of clean signals and a noise set from magnetic resonance imaging of groundwater in the field; A data augmentation network is constructed, comprising a generation network and a discrimination network. The discrimination network consists of a time-domain discrimination network and a frequency-domain discrimination network. The generation network is connected to the time-domain discrimination network to form a first discrimination channel. Random noise flows through the generation network and, after being output, flows into the time-domain discrimination network along with the noisy data to be identified as a source. The generation network is connected to the frequency-domain discrimination network to form a second discrimination channel. Random noise flows through the generation network and, after being output, flows into the frequency-domain discrimination network along with the noisy data to be identified as a source. The training of the data augmentation network is complete when the discrimination network cannot determine whether the input data comes from the generator network or noisy data. Generative network augmentation using data augmentation networks is employed to augment small sample data from magnetic resonance imaging of groundwater in the field.

9. The augmentation method for small sample data from field groundwater magnetic resonance detection according to claim 8, characterized in that, The total loss function for training the data augmentation network includes: The first loss function is used to constrain the mapping of the generator network from random noise to noisy data. The first loss function is expressed as: negative time-domain discriminant expectation term minus product term, wherein the product term is the frequency domain loss weight coefficient multiplied by the frequency domain discriminant expectation term. The second loss function is used to constrain the temporal discriminant network's discrimination of random noise and the temporal component sources of noisy data. The second loss function includes a first loss term and a second loss term. The first loss term is the discrimination result of the temporal discriminant network on the noisy data. The logarithm of the discrimination result is taken and then the negative value is taken. The second loss term is the discrimination result of the temporal discriminant network on the augmented data. The complementary value is obtained by subtracting the discrimination result from 1. The logarithm of the complementary value is taken and then the negative value is taken. The third loss function is used to constrain the frequency domain discriminant network's discrimination of random noise and the source of the frequency domain components of noisy data. The third loss function is expressed as the sum of the third loss term and the fourth loss term. The third loss term is the discrimination result of the frequency domain discriminant network on the noisy data. The logarithm of the discrimination result is taken and then the negative value is taken. The fourth loss term is the discrimination result of the frequency domain discriminant network on the augmented data. The complementary value is obtained by subtracting the discrimination result from 1. The logarithm of the complementary value is taken and then the negative value is taken.

10. The augmentation method for small sample data from field groundwater magnetic resonance detection according to claim 9, characterized in that, The augmented system trains a total loss function to establish constraints for generative adversarial mode training, requiring the generative network to minimize the first loss function and the discriminator network to maximize the second and third loss functions.

Images