A method for suppressing power frequency noise of magnetic resonance signals based on recurrent neural network

Abstract

The present invention relates to the field of power frequency noise suppression of magnetic resonance sounding signals, and is a method for suppressing power frequency noise of magnetic resonance signals based on a recurrent neural network. A network model is constructed using a recurrent layer and a bilinear layer. First, the noisy signal is divided into several segments according to the signal length, and the positive and negative bidirectional inputs are input to the input end of the network, and the simulated signal is used as the ideal output. Then, the input front-end noisy signal is preprocessed to initialize the state parameters of the recurrent layer. Then, the endpoints of each segment are bidirectionally extended, and the recurrent layer calculation is performed using redundant nodes. The global parameters are then updated using the error back propagation method until the loss function is stable to obtain a noise reduction model. By constructing a complex mapping relationship between the noisy signal and the power frequency noise, the purpose of noise reduction is achieved. In view of the characteristics of the magnetic resonance signal, a recurrent neural network is used to retain the correlation in signal time series, thereby better mapping the complex relationship between the noisy signal and the power frequency noise, and obtaining a higher signal-to-noise ratio.

Description

Technical Field

[0001] The present invention relates to the field of methods for suppressing power frequency noise of magnetic resonance sounding (MRS) signals, and in particular to a method for suppressing power frequency noise of magnetic resonance signals based on a recurrent neural network. Background Art

[0002] Nuclear magnetic resonance (NMR) groundwater detection, a geophysical method capable of qualitatively and quantitatively detecting groundwater, has experienced rapid development in recent years, from theoretical research to instrument development. However, due to the weak magnetic resonance signal, it is susceptible to strong electromagnetic interference in the environment during actual detection. This can lead to the inclusion of power-frequency noise in the signal, resulting in an extremely low signal-to-noise ratio. Therefore, filtering of the measured data is necessary to prevent it from affecting subsequent data inversion and interpretation.

[0003] Patent CN104614778A discloses a "nuclear magnetic resonance groundwater detection signal noise elimination method based on ICA". This method performs Fourier transform on the data to determine the power frequency harmonics contained near the nuclear magnetic resonance center frequency of each set of data. Then, sine and cosine functions with the same frequency as the power frequency harmonics and the same length as the nuclear magnetic resonance corresponding data are constructed, and the observation signal is formed together with the nuclear magnetic resonance response data. The data is reconstructed to eliminate the interference of the power frequency harmonics.

[0004] In the paper "Study on Magnetic Resonance Signal Denoising and Extraction Method Based on Harmonic Modeling and Autocorrelation" published in the Chinese Journal of Geophysics [2018, 61(2), 767-780], Tian Baofeng et al. proposed a method combining harmonic modeling and autocorrelation to perform denoising and extract signal characteristic parameters. First, a power frequency harmonic model was constructed, and an adaptive scanning search scheme was adopted to greatly improve the search accuracy and speed of the harmonic frequency, effectively eliminating the power frequency harmonics.

[0005] In the paper "Removal of Powerline Noise in Geophysical Datasets With a Scientific Machine-Learning Based Approach" published in IEEE Transactions on Geoscience and Remote Sensing [2022, 60, 5923410], Larsen et al. proposed using convolutional neural networks to calculate the reference frequency of power frequency harmonics and then eliminate power frequency noise by constructing a power frequency harmonic model.

[0006] The above-mentioned inventive methods all solve for the fundamental frequency of the power frequency harmonics and then use the power frequency modeling method to remove the power frequency noise. In the process of solving the fundamental frequency, a lot of time is often wasted. In addition, when the harmonic frequency offset of the noise is large, it is often impossible to obtain a good power frequency noise model. However, the noise reduction method based on recurrent neural network does not need to solve the fundamental frequency, and directly constructs a complex mapping relationship between the noisy signal and the power frequency noise, thereby achieving the purpose of noise reduction. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a method for suppressing power frequency noise of magnetic resonance signals based on a recurrent neural network. The method is based on a recurrent neural network and directly maps the complex relationship between the noisy signal and the power frequency noise to obtain a higher signal-to-noise ratio. Compared with traditional methods, the method is more efficient and does not require the fundamental frequency to be solved for power frequency modeling.

[0008] The present invention is achieved in this way:

[0009] A method for suppressing power frequency noise of magnetic resonance signals based on a recurrent neural network comprises the following steps:

[0010] a. Simulate the magnetic resonance signal X∈RM×N collected by the nuclear magnetic resonance groundwater detector, where N is the signal length and M is the number of signals, and mix it with power frequency noise S(n) to obtain the noisy magnetic resonance signal Y∈RM×N to form a training data set;

[0011] b. Segment the data set, then extend the endpoints of each segment, and perform forward and reverse bidirectional input on the processed data set;

[0012] c. Design the RNN network structure, preprocess the first N segments of input, and initialize the state parameters of the recurrent layer;

[0013] d. Input the training data set, use the back-propagation algorithm, and adopt the gradient descent method to optimize the RNN network parameters to minimize the loss function and complete the model training;

[0014] e. Performing segment processing on the actually acquired magnetic resonance signals in step b, then performing endpoint extension on each segment, and performing positive and negative bidirectional input on the processed data;

[0015] f. De-redundancy processing is performed on the endpoints of each segment of the magnetic resonance signal processed in step e, and the data is spliced to obtain a noise-removed signal.

[0016] Furthermore, the step b specifically includes:

[0017] b1. For the noisy magnetic resonance signal Y=Y[y1,y2,...,y N] is segmented and the signal is evenly divided into K segments with each segment length k. The training data set is Y=[S1,S2,...,S K ], where S n =[y n ,y n+1 ,...,y n+k ];

[0018] b2. Extend r points at both ends of the segmented data, where the segmented data S n =[y n ,y n+1 ,...,y n+k ] Extension to If the segmented data is the data at both ends of the signal, then copy r endpoint data S1=[y1,y2,...,y k ] Extension to

[0019] b3. Perform forward and reverse bidirectional input on the extended data set. The reverse input mode is in

[0020] Furthermore, the step c specifically includes:

[0021] c1. Set the network structure, including the number of bidirectional loop layers L r , the number of hidden units N r , the first linear layer and the number of hidden layers L l1 , the first linear layer and the number of hidden layers L l1 The number of neurons N l1 , and the second linear layer has the number of hidden layers L l2 , the second linear layer has the number of hidden layers L l2 The number of neurons N l2 ;

[0022] c2. The recurrent layer uses the Bi-RNN framework, and adds the first linear layer to the output of the gated recurrent unit (GRU) in both the forward and reverse directions, and then connects the output of the linear layer in both directions to the second linear layer;

[0023] c3. Initialize the loop layer state S r (n), the first N segments of the data set are input into the recurrent layer and encoded into the encoder part of the seq2seq model as follows:

[0024]

[0025] Among them, the function f describes the transformation done by the Encoder.

[0026] Furthermore, the step d specifically includes:

[0027] d1, the processed data set The segmented data is forward-inputted into the built recurrent neural network in sequence to obtain the recurrent layer output of each segment of data

[0028] O n =φ(W y S r (n)+b r );

[0029] S r (n) = W s S r (n-1)+W x y n ;

[0030] Among them, W x 、W y 、W s is the weight parameter of the recurrent layer, b r is the bias parameter, φ is the activation function of the cyclic layer; similarly, the dataset The segmented data is input into the built recurrent neural network in reverse order to obtain the recurrent layer output of each segment of data

[0031] d2, each segment output through the first linear layer is

[0032] L n =σ(W l1 O n +b l1 );

[0033] The output of each segment after the second linear layer is:

[0034]

[0035] Among them, W l1 is the weight parameter of the first linear layer, W l2 、W l3 is the weight parameter of the second linear layer, b l1 is the bias parameter, σ is the activation function;

[0036] d3. Determine the loss function E based on the mean square error loss function d , and then use the stochastic gradient descent method to solve the derivatives of the model parameters with respect to the loss function, and continuously update the parameters in the direction of decreasing loss function to reduce the error:

[0037]

[0038] Where w and b are the model weights and bias, η is the learning rate, and B is the batch size. The above formula is iterated repeatedly. When the change in the weights of all layers is less than the stopping threshold ε, the gradient descent method stops and the model training is completed.

[0039] Furthermore, the step f specifically includes:

[0040] f1. After processing each segment of data C n =[A n-r ,...,A n+k+r ] The endpoint performs redundancy processing,

[0041]

[0042] in, is the same point in the previous data segment;

[0043] f2. Splice the data after redundancy processing to obtain the noise-removed signal

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This method, targeting the characteristics of magnetic resonance signals, processes them segmentedly. Using a recurrent neural network, the method continuously optimizes network parameters to construct a complex mapping relationship between noisy signals and power-frequency noise, deriving effective signals from the noisy signals and achieving power-frequency noise suppression in magnetic resonance imaging. Compared with traditional methods, this method eliminates the need for power-frequency modeling based on fundamental frequency solution, improving data processing efficiency. It also addresses the problem of suppressing noise caused by large harmonic frequency offsets, a problem encountered in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A flow chart of the method of the present invention provided as an embodiment;

[0047] Figure 2 A block diagram of the noise reduction structure of the method of the present invention provided in an embodiment;

[0048] Figure 3 This is a comparison diagram of the time domain and frequency domain effects before and after noise reduction using the method of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions 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 intended to limit the present invention.

[0050] See also Figure 1 Combine Figure 2As shown, a method for suppressing power frequency noise of magnetic resonance signals based on a recurrent neural network includes the following steps:

[0051] a. Magnetic resonance signal X∈R collected by simulation nuclear magnetic resonance groundwater detector M×N , where N is the signal length, M is the number of signals, and the power frequency noise S(n) is mixed in to obtain the noisy magnetic resonance signal Y∈R M×N , forming a training data set;

[0052] b. Segment the dataset and then perform endpoint extension on each segment. The processed dataset is fed with both forward and reverse inputs. The next step, step c, uses an RNN network structure. This structure accumulates errors as the number of loops increases. Using only one-way input results in significant errors at the end of the data. Using two-way input in step b reduces this error.

[0053] c. Design the RNN network structure, preprocess the first N segments of input, and initialize the state parameters of the recurrent layer;

[0054] d. Input the training data set, use the back-propagation algorithm, and adopt the gradient descent method to optimize the RNN network parameters to minimize the loss function and complete the model training;

[0055] e. Use the trained model to perform noise reduction on the actual collected magnetic resonance signals. The magnetic resonance signals must first undergo the same preprocessing process as step b.

[0056] f. After processing each segment of data, perform redundancy processing on the endpoints and splice them together to obtain the noise-removed signal.

[0057] The specific steps of constructing the data set in step a are:

[0058] According to the basic principle of magnetic resonance detection, the time domain expression of the magnetic resonance signal is:

[0059]

[0060] Where e0 is the initial amplitude, f L is the Larmor frequency, is the relaxation time, is the initial phase, t=(0:N-1) / f s , f s is the sampling rate, N is the data length. The power frequency noise S(n) is mixed into the simulation signal to obtain the noisy magnetic resonance signal Y∈R M×N , M is the number of signals, forming a training data set.

[0061] The specific steps of step b are:

[0062] b1. For the noisy magnetic resonance signal Y=Y[y1,y2,...,y N ] is segmented and the signal is evenly divided into K segments with a length of k for each segment. The training data set is Y = [S1, S2, ..., S K ], where S n =[y n ,y n+1 ,...,y n+k ];

[0063] b2. Extend r points at both ends of the segmented data, where the segmented data S n =[y n ,y n+1 ,...,y n+k ] Extension to If the segmented data is the data at both ends of the signal, then copy r endpoint data S1=[y1,y2,...,y k ] Extension to

[0064] b3. Perform forward and reverse bidirectional input on the extended data set. The reverse input mode is in

[0065] See also Figure 2 The specific steps of building the model in step c are:

[0066] c1. Set the network structure, including the number of bidirectional loop layers L r , the number of hidden units N r , the first linear layer and the number of hidden layers L l1 , and the number of neurons N l1 , and the second linear layer has the number of hidden layers L l2 , and the number of neurons N l2 ;

[0067] c2. The recurrent layer adopts the Bi-RNN framework, and adds the first linear layer to the output of the gated recurrent unit GRU in both the forward and reverse directions, and then connects the output of the linear layer in both directions to the second linear layer.

[0068] The specific steps of step d are:

[0069] d1, the processed data set The segmented data is forward-inputted into the built recurrent neural network in sequence to obtain the recurrent layer output of each segment of data

[0070] O n =φ(W y S r(n)+b r );

[0071] S r (n) = W s S r (n-1)+W x y n ;

[0072] Among them, W x 、W y 、W s is the weight parameter of the recurrent layer, b r is the bias parameter, φ is the activation function of the recurrent layer. Similarly, the dataset The segmented data is input into the built recurrent neural network in reverse order to obtain the recurrent layer output of each segment of data

[0073] d2, each segment output through the first linear layer is

[0074] L n =σ(W l1 O n +b l1 );

[0075] The output of each segment after the second linear layer is:

[0076]

[0077] Among them, W l1 is the weight parameter of the first linear layer, W l2 、W l3 is the weight parameter of the second linear layer, b 11 is the bias parameter and σ is the activation function.

[0078] d3. Determine the loss function E based on the mean square error loss function d , and then use the stochastic gradient descent method to solve the derivatives of the model parameters with respect to the loss function, and continuously update the parameters in the direction of decreasing loss function to reduce the error.

[0079]

[0080] Where w and b are the model weights and bias, η is the learning rate, and B is the batch size. The above formula is iterated repeatedly. When the change in the weights of all layers is less than the stopping threshold ε, the gradient descent method stops and the model training is completed.

[0081] The specific steps of step f are:

[0082] f1. After processing each segment of data C n =[An-r ,...,A n+k+r ] The endpoint performs redundancy processing,

[0083]

[0084] in, The same point in the previous data segment.

[0085] f2. Splice the data after redundancy processing to obtain the noise-removed signal

[0086] Figure 3 This is a comparison diagram of the time domain and frequency domain effects before and after noise reduction by the method of the present invention. Through comparison, it can be seen that the present invention achieves the suppression of power frequency noise of magnetic resonance.

[0087] 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 scope of protection of the present invention.

Claims

1. A method for suppressing power frequency noise of magnetic resonance signals based on recurrent neural networks, characterized in that: The following steps are involved: a. Simulate the magnetic resonance signal X∈RM×N collected by the nuclear magnetic resonance groundwater detector, where N is the signal length and M is the number of signals, and mix it with power frequency noise S(n) to obtain the noisy magnetic resonance signal Y∈RM×N to form a training data set; b. Segment the data set, then extend the endpoints of each segment, and perform forward and reverse bidirectional input on the processed data set; c. Design the RNN network structure, preprocess the first N segments of input, and initialize the state parameters of the recurrent layer; d. Input the training data set, use the back-propagation algorithm, and adopt the gradient descent method to optimize the RNN network parameters to minimize the loss function and complete the model training; e. Performing segment processing on the actually acquired magnetic resonance signals in step b, then performing endpoint extension on each segment, and performing positive and negative bidirectional input on the processed data; f. De-redundancy processing is performed on the endpoints of each segment of the magnetic resonance signal processed in step e, and the data is spliced to obtain a noise-removed signal.

2. The method according to claim 1, characterized in that The step b specifically includes: b1. For the noisy magnetic resonance signal Y=Y[y1,y2,...,y N ] is segmented and the signal is evenly divided into K segments with each segment length k. The training data set is Y=[S1,S2,...,S K ], where S n =[y n ,y n+1 ,...,y n+k ]; b2. Extend r points at both ends of the segmented data, where the segmented data S n =[y n ,y n+1 ,...,y n+k ] Extension to If the segmented data is the data at both ends of the signal, then copy r endpoint data S1=[y1,y2,...,y k ] Extension to b3. Perform forward and reverse bidirectional input on the extended data set. The reverse input mode is in 3. The method according to claim 1, characterized in that The step c specifically includes: c1. Set the network structure, including the number of bidirectional loop layers L r , the number of hidden units N r , the first linear layer and the number of hidden layers L l1 , the first linear layer and the number of hidden layers L l1 The number of neurons N l1 , and the second linear layer has the number of hidden layers L l2 , the second linear layer has the number of hidden layers L l2 The number of neurons N l2 ; c2. The recurrent layer uses the Bi-RNN framework, and adds the first linear layer to the output of the gated recurrent unit (GRU) in both the forward and reverse directions, and then connects the output of the linear layer in both directions to the second linear layer; c3. Initialize the loop layer state S r (n), the first N segments of the data set are input into the recurrent layer and encoded into the encoder part of the seq2seq model as follows: Among them, the function f describes the transformation done by the Encoder.

4. The method according to claim 3, characterized in that The step d specifically includes: d1, the processed data set The segmented data is forward-inputted into the built recurrent neural network in sequence to obtain the recurrent layer output of each segment of data The n =φ(W y S r (n)+b r ); S r (n)=W s S r (n-1)+W x y n ; Among them, W x 、W y 、W s is the weight parameter of the recurrent layer, b r is the bias parameter, φ is the activation function of the cyclic layer; similarly, the dataset The segmented data is input into the built recurrent neural network in reverse order to obtain the recurrent layer output of each segment of data d2, each segment output through the first linear layer is L n =σ(W l1 The n +b l1 ); The output of each segment after the second linear layer is: Among them, W l1 is the weight parameter of the first linear layer, W l2 、W l3 is the weight parameter of the second linear layer, b l1 is the bias parameter, σ is the activation function; d3. Determine the loss function E based on the mean square error loss function d , and then use the stochastic gradient descent method to solve the derivatives of the model parameters with respect to the loss function, and continuously update the parameters in the direction of decreasing loss function to reduce the error: Where w and b are the model weights and bias, η is the learning rate, and B is the batch size. The above formula is iterated cyclically. When the change in the weights of all layers is less than the stopping iteration threshold ε, the gradient descent method is stopped and the model training is completed.

5. The method according to claim 1, characterized in that The step f specifically includes: f1. After processing each segment of data C n =[A n-r ,...,A n+k+r ] The endpoint performs redundancy processing, in, is the same point in the previous data segment; f2. Splice the data after redundancy processing to obtain the noise-removed signal

Images