An unpaired learning method for magnetic resonance signal features of underground organic pollution

Abstract

The present invention relates to the field of groundwater organic pollution magnetic resonance signal detection methods, and is a non-paired learning method for magnetic resonance signal features of underground organic pollution. The method constructs a non-paired learning network for underground organic pollution magnetic resonance signal features by forming an actual noisy data set and an effective signal data set, wherein the non-paired learning network includes a regression noise reduction network and a feature discrimination network, and the non-paired learning network is trained to establish a mapping relationship between the domain where the actual collected noisy data set is located and the domain where the effective signal data set is located, and transfer learning of the data domain is performed, thereby achieving the goal of extracting the features of underground organic pollution magnetic resonance signals, achieving the feature extraction of actual underground organic pollution magnetic resonance signals, reducing the dependence of deep learning methods on data, and enhancing the generalization ability of the network.

Description

Technical Field

[0001] The present invention relates to the field of magnetic resonance (Magnetic Resonance Sounding) signal detection methods for groundwater organic pollution, and in particular to a non-paired learning method for magnetic resonance signal features of underground organic pollution. Background Art

[0002] With the rapid development of the economy, a large number of environmental pollution incidents have occurred frequently. Among them, the organic pollution of groundwater caused by the leakage and improper discharge of chemical products, petroleum hydrocarbons and other organic substances has seriously threatened the ecological environment balance and people's health and safety. Groundwater organic pollution magnetic resonance detection technology is a geophysical exploration method that directly uses the nuclear magnetic resonance effect produced by the difference in the relaxation properties of the atomic nuclei of different substances, that is, it uses the difference in the relaxation characteristics of hydrogen nuclei in water and organic pollutants to identify groundwater organic pollutants. It has the advantages of non-destructive, qualitative and quantitative detection and has received widespread attention in recent years. With the development of ground magnetic resonance organic pollution in-situ detection equipment, composite pulse sequence measurement of groundwater and organic pollutants magnetic resonance spin echo signals has been realized. However, due to the strong noise interference of the actual site and the constraints of the actual complex medium conditions, the effective identification of organic pollutants lacks a certain degree of reliability, and there are few reports on the research on the feature extraction method of groundwater organic pollution magnetic resonance echo signals.

[0003] Existing magnetic resonance detection of oil pollution has problems such as slow speed, high cost, and cumbersome procedures, and cannot achieve large-scale delineation of underground organic pollution. The magnetic resonance detection device and method for shallow groundwater contaminated by hydrocarbons use permanent magnets to improve signal strength, but do not consider the strong noise interference in the actual detection environment. The convolutional neural network magnetic resonance groundwater noise suppression method is mainly aimed at groundwater signals. The magnetic resonance signal generated by organic pollution is weaker than the water signal. The number of convolution layers needs to be increased to effectively extract the signal, but the increase in network depth will cause the gradient to disappear. In addition, this method requires a large amount of labeled data. The input and output of the network need to be paired during the training phase to establish a nonlinear mapping relationship of the network. However, the actual acquired signals are often unlabeled, and sometimes the ideal denoising effect cannot be achieved. Summary of the invention

[0004] The purpose of the present invention is to address the deficiencies in the above-mentioned prior art and provide a non-paired learning method for magnetic resonance signal features of underground organic pollution. The method does not rely on paired label data and does not require a paired data set. By building an unsupervised training system of a discriminant loss convolutional neural network based on a two-way training mode, the mapping from the actual noisy data domain to the effective signal data domain is learned, and the feature extraction of the actual underground organic pollution magnetic resonance signal is realized, which reduces the dependence of the deep learning method on data and enhances the generalization ability of the network.

[0005] The present invention is achieved in this way.

[0006] A non-paired learning method for magnetic resonance signal features of underground organic pollution, the method comprising:

[0007] S1 respectively builds the noisy data set actually collected by the magnetic resonance organic pollution detection system , and valid signal data set , ,in, Represents each noisy data, Represents each valid signal data, i is the number of data, , the effective signal data consists of simulated pure spin echo signals:

[0008] ,

[0009] in, is the echo amplitude observed at time t, For the The signal size of the relaxation component at time zero, For the The transverse relaxation time of the relaxation component, Take the exponent of 2;

[0010] S2 builds an unpaired learning network for underground organic pollution magnetic resonance signal characteristics, the unpaired learning network includes a regression noise reduction network and a feature discrimination network, and trains the unpaired learning network to establish an actual collected noisy data set Located domain to valid signal data set Location The mapping relationship between domains is used to perform transfer learning in the data domain. The objective function of the unpaired learning network training is expressed as:

[0011] ,

[0012] in, is the objective function of the first regression denoising network, is the objective function of the second regression denoising network, is the norm operator, is the first regression denoising network, A mirror network representing a valid signal data set To the actual collected noisy data set The mapping network between Domain mapping relationship; is the second regression denoising network, used for fitting Domain to Domain mapping, is the first feature discrimination network, which is used to judge whether the generated valid signal data distribution is consistent with the real valid signal. is the second feature discriminant network, used to judge whether the generated noisy data is consistent with the real noisy data;

[0013] Set the learning rate to train the unpaired learning network, and execute multiple iterations to make the unpaired learning network converge. After the unpaired learning network training is completed, the regression denoising network is used as a denoising network to process the noisy measured data:

[0014] ,

[0015] in, It represents the effective underground organic pollution magnetic resonance signal extracted after processing. It represents the process of processing noisy measured data by regression denoising network.

[0016] Furthermore, the first regression denoising network Connect the second regression denoising network Composed of the first channel, effective signal data set Flows between the two regression denoising networks in the first channel and passes through the first feature discriminant network Make judgments;

[0017] The second regression denoising network Connect the first regression denoising network Composing the second channel, noisy data set Flows between the two regression denoising networks in the first channel and passes through the second feature discriminant network Make judgments;

[0018] When the feature discrimination network cannot determine whether the data is a simulated signal or a processed signal, the unpaired learning network training is completed.

[0019] Furthermore, the first regression denoising network and the second regression denoising network They all include multiple convolutional layers, conversion layers and upsampling layers. Multiple convolutional layers extract feature vectors from the input, and the output feature vectors are input to the conversion layer. The conversion layer is built with multiple residual modules and is used to convert the feature vectors between two domains. The output of the conversion layer is sent to the upsampling layer, and the upsampling layer restores the feature vectors to signals.

[0020] Furthermore, the first feature discrimination network And the second feature discriminant network They all include multi-layer convolution operation layers and batch normalization layers, which are used to continuously extract features from the input data and determine whether the input data is actual data or generated data. The batch normalization layer is used to make the input distribution of each layer tend to be consistent.

[0021] Furthermore, the loss function used by the unpaired learning network is:

[0022] For the second regression denoising network Build from Domain to The first loss function of domain mapping:

[0023] ,

[0024] in, Representation function In distribution The mathematical expectation under Representation function In distribution The mathematical expectation under

[0025] For the first regression denoising network Build from Domain to The second loss function for domain mapping:

[0026] ,

[0027] Establish a bidirectional training mode to constrain the training process and require the simultaneous learning of the second regression denoising network And the first regression denoising network Two mappings, and ;

[0028] Establish the L1 norm under the constraint of a bidirectional training mode between the actual noisy data and the flowing data in the network:

[0029] ,

[0030] in, represents the L1 norm, and The meaning is the same, and The meaning is consistent;

[0031] The final loss function of the unpaired learning network is established based on the first loss function, the second loss function and the L1 norm under the constraints of the bidirectional training mode:

[0032] ,

[0033] Among them, λ is the adjustment coefficient under the constraint of the bidirectional training mode.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] The present invention aims at the problem that the training process of supervised learning methods such as traditional convolutional neural networks relies on paired data sets, and it is difficult to mark effective signals from the actual data obtained. By building an unsupervised training system of a discriminant loss convolutional neural network based on a two-way training mode, the mapping from the actual noisy data domain to the effective signal data domain is learned, and the feature extraction of the actual underground organic pollution magnetic resonance signal is realized, which improves the weak dependence of the data-driven deep learning method on data and enhances the generalization ability of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 An architecture diagram of a non-paired learning network provided by an embodiment of the present invention;

[0037] Figure 2 A structural diagram of a regression denoising network provided by an embodiment of the present invention;

[0038] Figure 3 A structural diagram of a feature discrimination network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0040] See also Figure 1 Combination Figure 2 and Figure 3 As shown, a non-paired learning method for magnetic resonance signal features of underground organic pollution includes the following steps:

[0041] S1 respectively builds the noisy data set actually collected by the magnetic resonance organic pollution detection system , and valid signal data set , ,in, Represents each noisy data, Represents each valid signal data, i is the number of data, , the effective signal data consists of simulated pure spin echo signals:

[0042] ,

[0043] Among them, x(t) is the echo amplitude observed at time t, is the signal size of the i-th relaxation component at time zero, is the transverse relaxation time of the i-th relaxation component, Taking an exponent of 2, thereby generating enough variation of clean spin echo data to close the gap with the actual potential clean valid signal;

[0044] S2 builds an unpaired learning network for the magnetic resonance signal characteristics of underground organic pollution. The unpaired learning network includes a regression denoising network And feature discriminant network , train the unpaired learning network to establish the actual collected noisy data set Located domain to valid signal data set Location The mapping relationship between domains is used to perform transfer learning in the data domain; the objective function of the unpaired learning network training is expressed as:

[0045] ,

[0046] in, is the objective function of the first regression denoising network, is the objective function of the second regression denoising network, is the norm operator, is the first regression denoising network, A mirror network representing a valid signal data set To the actual noisy data set collected The mapping network between Domain mapping relationship; is the second regression denoising network, used for fitting Domain to Domain mapping, is the first feature discrimination network, which is used to judge whether the generated valid signal data distribution is consistent with the real valid signal. is the second feature discriminant network, used to judge whether the generated noisy data is consistent with the real noisy data;

[0047] Set the learning rate to train the unpaired learning network, and execute multiple iterations to make the unpaired learning network converge. After the unpaired learning network training is completed, the regression denoising network is used as a denoising network to process the noisy measured data:

[0048] ,

[0049] in, It represents the effective underground organic pollution magnetic resonance signal extracted after processing. It represents the process of processing noisy measured data by regression denoising network.

[0050] The unpaired learning network was trained using the Adam algorithm, the learning rate was set to 1e-3, and 100 iterations were performed to ensure that the unpaired learning network was fully converged. After the unpaired learning network training was completed, the regression denoising network was used as a denoising network to process the noisy measured data.

[0051] See also Figure 1 , the first regression denoising network in the unpaired learning network built Connect the second regression denoising network Composed of the first channel, effective signal data set Flows between the two regression denoising networks in the first channel and passes through the first feature discriminant network Make true / false judgments;

[0052] The second regression denoising network Connect the first regression denoising network Composing the second channel, noisy data set Flows between the two regression denoising networks in the first channel and passes through the second feature discriminant network Make true / false judgments;

[0053] When the feature discrimination network cannot determine whether the data is a simulated signal or a processed signal, the unpaired learning network training is completed.

[0054] See also Figure 2 Both regression denoising networks include multiple convolutional layers, conversion layers and upsampling layers. Multiple convolutional layers are connected in sequence and are composed of convolution operations / non-linear activation functions Conv+ReLU. The conversion layer is composed of n residual modules ResnetBlock1, Resnet Block2...Resnet Blockn. The residual module is composed of two groups of convolution Conv plus non-linear activation functions ReLU, and residual connections are added. The upsampling layer is composed of inverse convolution operations / non-linear activation functions Deconv+ReLU and feature addition / non-linear activation functions addition+ReLU. Multiple convolutional layers extract feature vectors from the input, and the output feature vectors are input to the conversion layer. The conversion layer is built with multiple residual modules to convert feature vectors between two domains. The output of the conversion layer is to the upsampling layer, and the upsampling layer restores the feature vectors to signals.

[0055] See also Figure 3The feature discrimination network includes the first layer of convolution operation / batch normalization Conv+ELU, the middle layers of convolution operation / batch normalization layer / activation function Conv+BN+ELU, and a layer of convolution operation Conv at the output end. It is used to continuously extract features from the input data and determine whether the input data is actual data or generated data. Batch normalization is used to make the input distribution of each layer basically consistent.

[0056] The specific steps for constructing the objective function of unpaired learning network training include:

[0057] a, is the second regression denoising network Build from Domain to The first loss function of domain mapping:

[0058] ,

[0059] in, Representation function In distribution The mathematical expectation under Representation function In distribution The mathematical expectation under

[0060] b, is the first regression denoising network Build from Domain to The second loss function for domain mapping:

[0061] ,

[0062] c. Establish a two-way training mode to constrain the training process and require simultaneous learning and Two mappings, and ; Establish the L1 norm under the constraint of the bidirectional training mode between the actual noisy data and the flowing data in the network:

[0063] ,

[0064] in, represents the L1 norm, and The meaning is the same, and The meaning is consistent;

[0065] d. The final loss function of the unpaired learning network is established based on the first loss function, the second loss function and the L1 norm under the constraints of the bidirectional training mode:

[0066] ,

[0067] Among them, λ is the adjustment coefficient under the constraint of the bidirectional training mode. It is set to 15, so that during the training process, the unpaired learning network retains the effective signal features in the actual noisy input data and obtains the denoised output.

[0068] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A non-paired learning method for magnetic resonance signal features of underground organic pollution, characterized in that: The method includes: S1 respectively builds the noisy data set actually collected by the magnetic resonance organic pollution detection system , and valid signal data set , ,in, Represents each noisy data, Represents each valid signal data, i is the number of data, , the effective signal data consists of simulated pure spin echo signals: ， in, is the echo amplitude observed at time t, For the The signal size of the relaxation component at time zero, For the The transverse relaxation time of the relaxation component, Take the exponent of 2; S2 builds an unpaired learning network for underground organic pollution magnetic resonance signal characteristics, the unpaired learning network includes a regression noise reduction network and a feature discrimination network, and trains the unpaired learning network to establish an actual collected noisy data set Located domain to valid signal data set Location The mapping relationship between domains is used to perform transfer learning in the data domain. The objective function of the unpaired learning network training is expressed as: ， in, is the objective function of the first regression denoising network, is the objective function of the second regression denoising network, is the norm operator, is the first regression denoising network, A mirror network representing a valid signal data set To the actual collected noisy data set The mapping network between Domain mapping relationship; is the second regression denoising network, used for fitting Domain to Domain mapping, is the first feature discrimination network, which is used to judge whether the generated valid signal data distribution is consistent with the real valid signal. is the second feature discriminant network, used to judge whether the generated noisy data is consistent with the real noisy data; Set the learning rate to train the unpaired learning network, and execute multiple iterations to make the unpaired learning network converge. After the unpaired learning network training is completed, the regression denoising network is used as a denoising network to process the noisy measured data: ， in, It represents the effective underground organic pollution magnetic resonance signal extracted after processing. It represents the process of processing noisy measured data by regression denoising network.

2. The unpaired learning method for magnetic resonance signal features of underground organic pollution according to claim 1 is characterized in that: The first regression denoising network Connect the second regression denoising network Composed of the first channel, effective signal data set Flows between the two regression denoising networks in the first channel and passes through the first feature discriminant network Make judgments; The second regression denoising network Connect the first regression denoising network Composing the second channel, noisy data set Flows between the two regression denoising networks in the first channel and passes through the second feature discriminant network Make judgments; When the feature discrimination network cannot determine whether the data is a simulated signal or a processed signal, the unpaired learning network training is completed.

3. The unpaired learning method for magnetic resonance signal features of underground organic pollution according to claim 2 is characterized in that: The first regression denoising network and the second regression denoising network They all include multiple convolutional layers, conversion layers and upsampling layers. Multiple convolutional layers extract feature vectors from the input, and the output feature vectors are input to the conversion layer. The conversion layer is built with multiple residual modules and is used to convert the feature vectors between two domains. The output of the conversion layer is sent to the upsampling layer, and the upsampling layer restores the feature vectors to signals.

4. The method according to claim 1, characterized in that The first feature discrimination network And the second feature discriminant network They all include multi-layer convolution operation layers and batch normalization layers, which are used to continuously extract features from the input data and determine whether the input data is actual data or generated data. The batch normalization layer is used to make the input distribution of each layer tend to be consistent.

5. The method according to claim 1, characterized in that The loss function used by the unpaired learning network is: For the second regression denoising network Build from Domain to The first loss function of domain mapping: ， in, Representation function In distribution The mathematical expectation under Representation function In distribution The mathematical expectation under For the first regression denoising network Build from Domain to The second loss function for domain mapping: ， Establish a bidirectional training mode to constrain the training process and require the simultaneous learning of the second regression denoising network And the first regression denoising network Two mappings, and ; Establish the L1 norm under the constraint of a bidirectional training mode between the actual noisy data and the flowing data in the network: ， in, represents the L1 norm, and The meaning is the same, and The meaning is consistent; The final loss function of the unpaired learning network is established based on the first loss function, the second loss function and the L1 norm under the constraints of the bidirectional training mode: ， Among them, λ is the adjustment coefficient under the constraint of the bidirectional training mode.

Images