A method for magnetic resonance random noise suppression based on adaptive auto-encoder

Abstract

The application discloses a magnetic resonance groundwater detection random noise suppression method, which comprises the following steps: constructing a noise suppression model, and suppressing the random noise of the magnetic resonance groundwater detection through the noise suppression model; adding random noise to a plurality of groups of ideal magnetic resonance signals to obtain a data set; building a model, wherein the model comprises an encoder and a decoder, and the network parameters are initialized; training the model by using a training set S, extracting features of data by using the encoder to obtain hidden variables z, and reconstructing effective signals from the hidden variables z by using the decoder; establishing a probability distribution model of the training sample from the encoder to learn the distribution rule of the signals, determining a loss function according to the deviation between the input signals and the reconstructed signals, introducing a parameter estimation error to constrain the loss function, updating the network model parameters, and obtaining an adaptive self-encoder denoising model until the loss function trend is stable; and testing the denoising effect of the model by using a test set T. The method solves the problem of limited denoising effect and improves the denoising efficiency.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of noise suppression methods for magnetic resonance sounding (MRS) signals, and particularly relates to a random noise suppression method for magnetic resonance groundwater detection. BACKGROUND

[0002] Ground magnetic resonance is a geophysical exploration method for directly detecting underground water, and can quantitatively explain hydrological information such as water content depth, thickness, unit volume water content, underground water-bearing layer content, and porosity. It is mainly applied to groundwater resource detection, water source geological disaster prediction, and groundwater pollution monitoring.

[0003] However, the magnetic resonance signal is very weak, with an order of magnitude of nanovolts. In actual detection, the magnetic resonance signal is interfered by random noise, power frequency noise, and peak noise in the environment, which affects the subsequent interpretation of hydrological information. At present, the magnetic resonance noise reduction method generally adopts a specific filtering method for different noise types. For random noise, Dalgaard published a paper entitled "Adaptive noise cancelling of multichannel magnetic resonance sounding signals" in Geophysical Journal International [2012, 191(1), 88-100], which adopts a superposition method to suppress random noise. Lin Tingting et al. published a paper entitled "Time-frequency peak filtering for random noise attenuation of magnetic resonance sounding signal" in Geophysical Journal International [2018, 213(2), 727-738], which proposes to use a time-frequency peak filtering method to suppress random noise in the magnetic resonance envelope signal.

[0004] The superposition method of the above-mentioned application is a commonly used magnetic resonance signal random noise reduction method, but the noise reduction effect is limited, and it takes a long time and has low efficiency when processing random noise with high noise level. The time-frequency peak filtering method needs manual adjustment of filtering parameters, which is not conducive to use by non-professional technicians. SUMMARY

[0005] The technical problem to be solved by the application is to provide a random noise suppression method for magnetic resonance groundwater detection, which solves the problems of limited noise reduction effect and long time and low efficiency when processing random noise with high noise level.

[0006] The application is implemented in this way,

[0007] A magnetic resonance groundwater detection random noise suppression method, comprising:

[0008] A noise suppression model is constructed, and random noise of the magnetic resonance groundwater detection is suppressed through the noise suppression model; wherein the noise suppression model is constructed, comprising:

[0009] a. Constructing a data set: adding random noise to a plurality of ideal magnetic resonance signals to obtain a data set X=[x1, x2,..., x n ], and dividing the data set into a training set S=[x1, x2,..., x m ] and a test set T=[x m+1 ,x m+2 ,...,x n ] in a ratio of 7:3;

[0010] b. Building a model: introducing a convolution layer and a deconvolution layer to build an adaptive autoencoder model, the autoencoder model comprising an encoder and a decoder, and initializing network parameters;

[0011] c. Training the model: training the adaptive autoencoder model using the training set S, extracting features of the data through the encoder to obtain hidden variables z, and reconstructing effective signals from the hidden variables z through the decoder;

[0012] d. Determining the model: making the encoder establish a probability distribution model of the training samples to learn the distribution rule of the signals, determining a loss function according to the deviation between the input signals and the reconstructed signals, introducing a parameter estimation error to constrain the loss function, updating the network model parameters, until the loss function trend is stable, and obtaining an adaptive autoencoder denoising model;

[0013] e. Testing the model: testing the denoising effect of the adaptive autoencoder denoising model using the test set T.

[0014] Further, the specific steps of building the model in step b include:

[0015] An encoder in the adaptive autoencoder is built using a convolution layer, network parameters φ are initialized, and the encoder network is denoted as Q(φ);

[0016] A decoder in the adaptive autoencoder is built using a deconvolution layer, network parameters θ are initialized, and the decoder network is denoted as P(θ).

[0017] Further, the determination of the model in step d specifically includes:

[0018] 1) By measuring the distribution of the hidden variables z relative to the input signals x and the reconstructed signals x rThe information loss distribution distance is calculated using variational inference, x and x r The distribution law is:

[0019]

[0020] Where p θ (z|x r ) and q φ (z|x) are the distributions estimated by the encoder and the decoder probability models, respectively.

[0021] 2) From the Bayesian formula and step 1), we get:

[0022]

[0023] Let J(θ,φx;,D KL (q φ (z|x)p θ (z|x r ))≥0, J(θ,φ;x)≤logp(x r );

[0024] To maximize the likelihood probability p(x r ), we maximize J(θ,φ;x), which is the loss function of the network

[0025]

[0026] Assuming that q φ (z|x) and p(z) are both normal distributions, they are subject to N(μ1,σ1), N(0,1), respectively, then

[0027]

[0028] The reconstruction effect of the decoder to generate the reconstructed signal;

[0029] 3) Maximizing J(θ,φ;x) is converted to:

[0030]

[0031] minD KL (q φ (z|x)p θ (z))

[0032] 4) Take the complex envelope of the model input x and the output signal x r , denoted as u(x) and u(x r ), respectively, and the gap between them is u(x)-u(x r) as a constraint model loss function, then the loss function of the model is improved as

[0033]

[0034] Optimize the loss function, update the encoder-decoder parameters φ and θ, until J(θ, φ; x) is stable, and determine the adaptive autoencoder denoising model.

[0035] Further, the random noise of the magnetic resonance groundwater detection is suppressed by the noise suppression model, comprising: the input noisy signal is subjected to variational inference by the encoder, the data x∈S is input into the encoder, and the output of the Lth layer is

[0036] a L =W1 L *a L-1 +b1 L

[0037] Wherein, a L is the output of the Lth convolutional layer, W1 L , b1 L is the convolution kernel and bias of the Lth layer, and * represents convolution operation, so that the network output after the encoder is the latent variable z

[0038] z=Q(φ; x)

[0039] Wherein, φ={W1, b1};

[0040] The decoder reconstructs the signal according to the latent variable z, which has the same distribution rule as the input signal, and the latent variable z is input into the deconvolution layer, and the output of the Lth layer is

[0041] c L =W2 L *c L-1 +b2 L

[0042] Wherein, c L is the output of the Lth deconvolutional layer, W2 L , b2 L is the convolution kernel and bias of the Lth layer, and * represents convolution operation. Then, the signal reconstructed by the network according to the latent variable z after the decoder is

[0043] x r =P(θ; z)

[0044] Wherein, θ={W2, b2}.

[0045] A magnetic resonance groundwater detection random noise suppression model construction system, comprising:

[0046] Dataset construction module: Random noise is added to multiple sets of ideal magnetic resonance signals to obtain a dataset X = [x1, x2, ..., x n The dataset is divided into training sets S = [x1, x2, ..., x3] in a 7:3 ratio. m ] and test set T = [x m+1 ,x m+2 ,...,x n ];

[0047] Model building module: Convolutional and deconvolutional layers are introduced to build an adaptive autoencoder model. The autoencoder model includes an encoder and a decoder, and the network parameters are initialized.

[0048] Model training module: The adaptive autoencoder model is trained using the training set S. The encoder extracts features from the data to obtain latent variables z, and the decoder reconstructs the effective signal from the latent variables z.

[0049] Model determination module: Enables the autoencoder to establish a probability distribution model of training samples to learn the distribution law of the signal, determines the loss function based on the deviation between the input signal and the reconstructed signal, introduces parameter estimation error to constrain the loss function, updates the network model parameters until the loss function trend stabilizes, and obtains the adaptive autoencoder denoising model;

[0050] Model testing module: Uses test set T to test the denoising effect of the adaptive autoencoder denoising model.

[0051] Furthermore, the specific steps for building the model building module include:

[0052] The encoder in the adaptive autoencoder is built using convolutional layers. The network parameters φ are initialized, and the encoder network is denoted as Q(φ).

[0053] The decoder in the adaptive autoencoder is built using deconvolutional layers. The network parameters θ are initialized, and the decoder network is denoted as P(θ).

[0054] Furthermore, the specific training process of the model training module includes:

[0055] The input noisy signal is subjected to variational inference by an encoder. The data x∈S is input into the encoder, passed through a convolutional layer, and the output of the Lth layer is:

[0056] a L =W1 L *a L-1 +b1 L

[0057] Among them, a L W1 is the output of the Lth convolutional layer. L b1 LLet be the convolution kernel and bias of the Lth layer, respectively, and * represent the convolution operation. After passing through the encoder, the network output, the latent variable z, is:

[0058] z = Q(φ; x)

[0059] Where φ = {W1, b1};

[0060] The decoder reconstructs the signal based on the latent variable z, which has the same distribution pattern as the input signal. The latent variable z is used as input, and after passing through the deconvolution layer, the output of the Lth layer is:

[0061] c L =W2 L *c L-1 +b2 L

[0062] Among them, c L W2 is the output of the Lth deconvolution layer. L b2 L Let z be the convolution kernel and bias of the Lth layer, respectively, and * represent the convolution operation. After the decoder, the signal reconstructed by the network based on the latent variable z is:

[0063] x r =P(θ;z)

[0064] Where θ = {W2, b2}.

[0065] Furthermore, the specific determination process of the model determination module includes:

[0066] 1) By measuring the distribution of the latent variable z relative to the input signal x and relative to the reconstructed signal x r Information loss due to the distribution, distribution distance, and variational inference are used to calculate x and x. r Deviation in the distribution pattern:

[0067]

[0068] Where, p θ (z|x r ) and q φ (z|x) represent the distributions of z estimated by the encoder and decoder probability models, respectively;

[0069] 2) By Bayes' theorem And from step 1), we get:

[0070]

[0071] Will Denoted as J(θ,φx;,D KL (q φ (z|x)p θ(z|x)) >= 0, J(theta, phi; x) <= log p(x r );

[0072] To maximize the likelihood probability p(x r ), maximize J(theta, phi; x), that is, the loss function of the network

[0073]

[0074] Assume that q φ (z|x) and p(z) are normal distribution, respectively obey N(mu1, sigma1), N(0, 1), then

[0075]

[0076] The reconstruction effect of the reconstructed signal generated by the decoder;

[0077] 3) Maximize J(theta, phi; x) is converted to:

[0078]

[0079] minD KL (q φ (z|x)p θ (z))

[0080] 4) Take the complex envelope of the model input x, the output signal x r , expressed as u(x) and u(x r ), and the gap u(x)-u(x r ) between the two is used as a constraint on the model loss function, then the model loss function is improved to

[0081]

[0082] Optimize the loss function, update the encoder-decoder parameters phi and theta, until J(theta, phi; x) is stable, and determine the adaptive autoencoder denoising model.

[0083] Compared with the prior art, the present application has the beneficial effects that: the method of the present application is an unsupervised learning method, which reconstructs an intelligent denoising algorithm of the effective signal by training an adaptive autoencoder, and the method does not need a large number of labels compared with the supervised method, and can solve the problem of lack of labeled data in the ground magnetic resonance water finding direction. In addition, the method introduces a convolutional neural network to build two probability distribution functions, extracts features of the noisy signal through the convolutional layer in the structure, reconstructs the effective signal of the same probability distribution through the deconvolutional layer, and introduces probability and variational inference as the loss function of the model, and optimizes the network model. BRIEF DESCRIPTION OF DRAWINGS

[0084] Figure 1 Flow chart of a magnetic resonance groundwater detection random noise suppression method based on an adaptive autoencoder

[0085] Figure 2 Denoising structure block diagram of a magnetic resonance groundwater detection random noise suppression method based on an adaptive autoencoder. DETAILED DESCRIPTION

[0086] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0087] The present application will be further described in detail below with reference to the accompanying drawings and embodiments:

[0088] Referring to Figure 1 In combination Figure 2 As shown in the figure, a magnetic resonance groundwater detection random noise suppression method comprises:

[0089] A noise suppression model is constructed, and the magnetic resonance groundwater detection random noise is suppressed through the noise suppression model; wherein the noise suppression model comprises:

[0090] a. Constructing a data set: random noise is added to a plurality of ideal magnetic resonance signals to obtain a data set X=[x1, x2,..., x n ], and the data set is divided into a training set S=[x1, x2,..., x m ] and a test set T=[x m+1 ,x m+2 ,...,x n ] in a ratio of 7:3;

[0091] b. Building a model: introducing a convolution layer and a deconvolution layer to build an autoencoder model, the model comprising an encoder and a decoder, and initializing network parameters;

[0092] c. Training the model: training the model using the training set S, extracting features of the data through the encoder to obtain hidden variables z, and reconstructing the effective signal from the hidden variables z through the decoder;

[0093] d. Determining the model: making the autoencoder establish a probability distribution model of the training sample to learn the distribution rule of the signal, determining a loss function according to the deviation between the input signal and the reconstructed signal, introducing a parameter estimation error to constrain the loss function, updating the network model parameters until the loss function trend is stable, and obtaining an adaptive autoencoder denoising model;

[0094] e. Testing the model: testing the denoising effect of the model using the test set T.

[0095] The specific steps of building the model in step b are:

[0096] First, use the convolutional layer to build the encoder in the adaptive autoencoder, initialize the network parameters φ, and the encoder network is denoted as Q(φ);

[0097] Then, use the deconvolutional layer to build the decoder in the adaptive autoencoder, initialize the network parameters θ, and the decoder network is denoted as P(θ).

[0098] The specific steps of training the model in step c are:

[0099] First, perform variational inference on the input noisy signal through the encoder, input the data x∈S into the encoder, pass through the convolutional layer, and the output of the Lth layer is

[0100] a L =W1 L *a L-1 +b1 L

[0101] where a L is the output of the Lth convolutional layer, W1 L , b1 L is the convolution kernel and bias of the Lth layer, and * represents convolution operation. Then, after passing through the encoder, the network output is the latent variable z

[0102] z=Q(φ;x)

[0103] where φ={W1,b1};

[0104] Second, the decoder reconstructs the signal according to the latent variable z, which has the same distribution as the input signal. The latent variable z is input, and the output of the Lth layer after passing through the deconvolutional layer is

[0105] c L =W2 L *c L-1 +b2 L

[0106] where c L is the output of the Lth deconvolutional layer, W2 L , b2 L is the convolution kernel and bias of the Lth layer, and * represents convolution operation. Then, after passing through the decoder, the signal reconstructed by the network according to the latent variable z is

[0107] x r =P(θ;z)

[0108] where θ={W2,b2}.

[0109] The specific steps of determining the model in step d are:

[0110] 1) To estimate whether the signal reconstructed by the hidden layer z is close to the true signal, the distribution distance between the two is calculated using variational inference:

[0111]

[0112] where p θ (z|x) and q φ (z|x) are the distributions of z estimated by the encoder and decoder probability models, respectively;

[0113] 2) From the Bayes formula and 1), we have

[0114]

[0115] Let J(θ,φ;x) be because D KL (q φ (z|x)p θ (z|x)) ≥ 0, so

[0116] J(θ,φ;x) ≤ log p(x)

[0117] To maximize the likelihood probability p(x), we can maximize J(θ,φ;x), which is the loss function of the network

[0118]

[0119] 3) To facilitate calculation, assume that q φ (z|x) and p(z) are both normal distributions, subject to N(μ1,σ1), N(0,1), respectively, then

[0120]

[0121] 4) And is the reconstruction effect of the decoder to generate the reconstructed signal

[0122] 5) Maximizing J(θ,φ;x) is converted to:

[0123]

[0124] min D KL (q φ (z|x)p θ (z)

[0125] 6) Take the complex envelope of the model input x and the output signal x r , denoted as u(x) and u(x r), the difference u(x)-u(x r ) between the two as a further constraint on the model loss function, then the loss function of the model is improved as

[0126]

[0127] By optimizing the loss function by the above formula, update the encoder-decoder parameters φ and θ until J(θ, φ; x) is stable, determine the adaptive self-encoder denoising model.

[0128] The magnetic resonance groundwater detection random noise is suppressed by the noise suppression model, including: the input noisy signal is subjected to variational inference by the encoder, the data x∈S is input into the encoder, and the output of the Lth layer is

[0129] a L =W1 L *a L-1 +b1 L

[0130] Wherein, a L is the output of the Lth convolutional layer, W1 L , b1 L is the convolution kernel and bias of the Lth layer, and * represents convolution operation, so that the network output after the encoder is the latent variable z

[0131] z=Q(φ; x)

[0132] Wherein, φ={W1, b1};

[0133] The decoder reconstructs the signal according to the latent variable z, which has the same distribution rule as the input signal, and the latent variable z is input into the deconvolution layer, and the output of the Lth layer is

[0134] c L =W2 L *c L-1 +b2 L

[0135] Wherein, c L is the output of the Lth deconvolutional layer, W2 L , b2 L is the convolution kernel and bias of the Lth layer, and * represents convolution operation. Then, the signal reconstructed by the network according to the latent variable z after the decoder is

[0136] x r =P(θ; z)

[0137] Wherein, θ={W2, b2}.

[0138] A magnetic resonance groundwater detection random noise suppression model construction system, comprising:

[0139] Dataset construction module: Random noise is added to multiple sets of ideal magnetic resonance signals to obtain a dataset X = [x1, x2, ..., x n The dataset is divided into training sets S = [x1, x2, ..., x3] in a 7:3 ratio. m ] and test set T = [x m+1 ,x m+2 ,...,x n ];

[0140] Model building module: Convolutional and deconvolutional layers are introduced to build an adaptive autoencoder model. The autoencoder model includes an encoder and a decoder, and the network parameters are initialized.

[0141] Model training module: The adaptive autoencoder model is trained using the training set S. The encoder extracts features from the data to obtain latent variables z, and the decoder reconstructs the effective signal from the latent variables z.

[0142] Model determination module: Enables the autoencoder to establish a probability distribution model of training samples to learn the distribution law of the signal, determines the loss function based on the deviation between the input signal and the reconstructed signal, introduces parameter estimation error to constrain the loss function, updates the network model parameters until the loss function trend stabilizes, and obtains the adaptive autoencoder denoising model;

[0143] Model testing module: Uses test set T to test the denoising effect of the adaptive autoencoder denoising model.

[0144] The specific steps for building the model building module include:

[0145] The encoder in the adaptive autoencoder is built using convolutional layers. The network parameters φ are initialized, and the encoder network is denoted as Q(φ).

[0146] The decoder in the adaptive autoencoder is built using deconvolutional layers. The network parameters θ are initialized, and the decoder network is denoted as P(θ).

[0147] The specific training process of the model training module includes:

[0148] The input noisy signal is subjected to variational inference by an encoder. The data x∈S is input into the encoder, passed through a convolutional layer, and the output of the Lth layer is:

[0149] a L =W1 L *a L-1 +b1 L

[0150] Among them, a L W1 is the output of the Lth convolutional layer. L b1 LLet be the convolution kernel and bias of the Lth layer, respectively, and * represent the convolution operation. After passing through the encoder, the network output, the latent variable z, is:

[0151] z = Q(φ; x)

[0152] Where φ = {W1, b1};

[0153] The decoder reconstructs the signal based on the latent variable z, which has the same distribution pattern as the input signal. The latent variable z is used as input, and after passing through the deconvolution layer, the output of the Lth layer is:

[0154] c L =W2 L *c L-1 +b2 L

[0155] Among them, c L W2 is the output of the Lth deconvolution layer. L b2 L Let z be the convolution kernel and bias of the Lth layer, respectively, and * represent the convolution operation. After the decoder, the signal reconstructed by the network based on the latent variable z is:

[0156] x r =P(θ;z)

[0157] Where θ = {W2, b2}.

[0158] The specific determination process of the model determination module includes:

[0159] 1) By measuring the distribution of the latent variable z relative to the input signal x and relative to the reconstructed signal x r Information loss due to the distribution, distribution distance, and variational inference are used to calculate x and x. r Deviation in the distribution pattern:

[0160]

[0161] Where, p θ (z|x r ) and q φ (z|x) represent the distributions of z estimated by the encoder and decoder probability models, respectively;

[0162] 2) By Bayes' theorem And from step 1), we get:

[0163]

[0164] Will Denoted as J(θ,φx;,D KL (q φ (z|x)p θ(z|x)) ≥ 0, J(0, f; x) < log p(x r );

[0165] To maximize the likelihood probability p(x r ), maximize J(0, f; x), that is, the loss function of the network

[0166]

[0167] Assume that q φ (z|x) and p(z) are both normal distributions, subject to N(0, 1), N(0, 1), respectively, then

[0168]

[0169] The reconstruction effect of the reconstructed signal generated by the decoder;

[0170] 3) Maximize J(0, f; x) is converted to:

[0171]

[0172] minD KL (q φ (z|x)p θ (z))

[0173] 4) Take the complex envelope of the model input x and the output signal x r , denoted as u(x) and u(x r ), and take the difference u(x)-u(x r ) between the two as the condition of the constraint model loss function, then the model loss function is improved to

[0174]

[0175] Optimize the loss function, update the encoder-decoder parameters f and 0, until J(0, f; x) is stable, and determine the adaptive self-encoder denoising model.

[0176] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method of random noise suppression for magnetic resonance groundwater prospecting, characterized in that, The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: a、Construction of dataset: Add random noise to multiple groups of ideal magnetic resonance signals to obtain dataset X=[x1,x2,...,x n ],Divide the dataset into training set S=[x1,x2,...,x m ] and test set T=[x m+1 ,x m+2 ,...,x n ] in the ratio of 7:3; b. Model building: introducing a convolution layer and a deconvolution layer to build an adaptive autoencoder model, the adaptive autoencoder model comprising an encoder and a decoder, and initializing network parameters; c. Model training: training the adaptive autoencoder model by using a training set S, extracting features of data by the encoder to obtain hidden variables z, and reconstructing effective signals from the hidden variables z by the decoder; d. Model determination: making the encoder establish a probability distribution model of training samples to learn distribution rules of signals, determining a loss function according to a deviation between an input signal and a reconstructed signal, introducing a parameter estimation error to constrain the loss function, updating network model parameters until the loss function tends to be stable, and obtaining an adaptive autoencoder denoising model; e. Model testing: testing a denoising effect of the adaptive autoencoder denoising model by using a test set T. The specific steps of the model building in step b. comprise: An encoder in the adaptive autoencoder is built using a convolutional layer, and network parameters are initialized The encoder network is denoted as Q ) using a deconvolutional layer to build a decoder in the adaptive autoencoder, initializing network parameters , the decoder network is denoted as P( ). The specific steps of the model determination in step d. comprise: 1) By measuring the information loss distribution distance of the distribution of the latent variable z with respect to the input signal x and with respect to the reconstructed signal x r , the distribution regularity of x and x r is calculated by using variational inference: ， wherein, and are the encoder and decoder probability model estimates of the distribution of z, respectively; 2) from the Bayes formula and step 1) gives: ， will be described below. denoted as , ， ； to maximize the likelihood probability , maximizing the loss function of the adaptive autoencoder model ， Assume and are normally distributed, with , N(0, 1), respectively, then ， to generate a reconstructed effect of the reconstructed signal for a decoder; 3) Maximize Converts to: ， ， 4) The signal x r Take the complex envelope, denoted as u(x) and u(x r ), the gap between the two As a constraint model loss function conditions, the loss function of the model is improved as ， Optimizing the loss function, updating the encoder-decoder parameters and until stable, determine adaptive self-encoder denoising model.

2. The magnetic resonance groundwater detection random noise suppression method of claim 1, wherein, The random noise of the magnetic resonance groundwater detection is suppressed through a noise suppression model, including: performing variational inference on the input noisy signal through an encoder, performing data The input encoder is subjected to a convolutional layer, and the output of the Lth layer is: ， where, is the output of the Lth convolutional layer, , is the convolution kernel and bias of the Lth layer, and * represents the convolution operation. Then, the network output after the encoder is the latent variable z as z = Q x), wherein = {W1, b1}; The decoder reconstructs signals according to the hidden variables z, and the reconstructed signals have the same distribution rules as the input signals, the hidden variables z are taken as inputs, and the Lth layer outputs are: ， where, is the output of the Lth deconvolutional layer, , is the convolution kernel and bias of the Lth layer, and * represents the convolution operation. Then, after the decoder, the signal reconstructed by the network according to the latent variable z is: = P( ; z), wherein, = {W2, b2}.

3. A system for constructing a magnetic resonance groundwater detection random noise suppression model, characterized in that, The method comprises the following steps: Dataset construction module: add random noise to multiple sets of ideal magnetic resonance signals to obtain dataset X=[x1,x2,...,x n ], divide the dataset into training set S=[x1,x2,...,x m ] and test set T=[x m+1 ,x m+2 ,...,x n ] according to the proportion of 7:3; The model training module trains the adaptive autoencoder model by using a training set S, extracts features of data by the encoder to obtain hidden variables z, and reconstructs effective signals from the hidden variables z by the decoder; The model determination module makes the encoder establish a probability distribution model of training samples to learn distribution rules of signals, determines a loss function according to a deviation between an input signal and a reconstructed signal, introduces a parameter estimation error to constrain the loss function, updates network model parameters until the loss function tends to be stable, and obtains an adaptive autoencoder denoising model; The model testing module tests a denoising effect of the adaptive autoencoder denoising model by using a test set T. The specific steps of the model building module comprise: The specific process of the model training module comprises: An encoder in the adaptive autoencoder is built using a convolutional layer, and network parameters are initialized The encoder network is denoted as Q ) using a deconvolutional layer to build a decoder in the adaptive autoencoder, initializing network parameters , the decoder network is denoted as P( ) The decoder reconstructs signals according to the hidden variables z, and the reconstructed signals have the same distribution rules as the input signals, the hidden variables z are taken as inputs, and the Lth layer outputs are: The data is encoded by the encoder through variational inference on the input noisy signal The input encoder is passed through a convolutional layer, and the output of the Lth layer is: ， where, is the output of the Lth convolutional layer, , are the convolutional kernel and bias of the Lth layer, respectively, and * represents the convolution operation. Then, the network output after the encoder is the latent variable z as z = Q x), wherein = {W1, b1}; The specific determination process of the model determination module comprises: ， where, is the output of the Lth deconvolutional layer, , are the convolutional kernel and bias of the Lth layer, respectively, and * denotes the convolution operation. After the decoder, the signal reconstructed by the network according to the latent variable z is = P( ; z), wherein, = {W2, b2}.

4. The system for constructing a model for random noise suppression in magnetic resonance groundwater detection according to claim 3, characterized in that, ​ 1) By measuring the distribution of the latent variable z with respect to the input signal x and the information loss distribution distance with respect to the reconstructed signal x r r the distribution of x and x​ ， wherein, and are the encoder and decoder probability model estimates of the distribution of z, respectively; 2) from the Bayes formula and step 1) gives: ， will be described below. denoted as , ， ； to maximize the likelihood probability , maximizing the loss function of the adaptive autoencoder model ， Assume and are normally distributed, with N(0,1), respectively, then ， to generate a reconstructed effect of the reconstructed signal for a decoder; 3) Maximize Converts to: ， ， 4) The signal x r Taking the complex envelope, denoted as u(x) and u(x r ), the gap between them As a constraint on the model loss function, the model's loss function is improved as ， Optimizing the loss function, updating the encoder-decoder parameters and until stable, determine adaptive self-encoder denoising model.

Images