Magnetic resonance random noise reduction system and method for petroleum hydrocarbon pollution detection

Abstract

The present invention relates to the field of geophysical detection of petroleum hydrocarbon pollution, and is a magnetic resonance random noise reduction system and method suitable for petroleum hydrocarbon pollution detection. The system comprises a petroleum hydrocarbon pollution random noise reduction network, comprising a deep residual denoiser and a classification discriminator, wherein the deep residual denoiser is configured to process a measured random noise-containing signal according to a nonlinear mapping from a random noise-containing signal to random noise to obtain a predicted random noise, and then remove the predicted random noise from the measured random noise-containing signal to obtain a denoised signal; and the classification discriminator is configured to learn the characteristics of a pure signal in a pure signal set, discriminate the denoised signal input to the classification discriminator as a denoised signal or a pure signal, and feed back the classification result of the classification discriminator to the deep residual denoiser. The deep residual denoiser optimizes network parameters based on the feedback, thereby achieving intelligent denoising.

Description

Technical Field

[0001] The present invention relates to the field of geophysical detection of petroleum hydrocarbon pollution, and in particular to a magnetic resonance random noise reduction system and method suitable for petroleum hydrocarbon pollution detection. Background Art

[0002] To strengthen the risk management of petroleum hydrocarbon contamination in groundwater, it is crucial to improve petroleum hydrocarbon contamination detection technology and enhance the level of petroleum hydrocarbon contamination detection. Magnetic resonance detection technology for groundwater petroleum hydrocarbon contamination stimulates Larmor precession of hydrogen protons in groundwater and petroleum hydrocarbons by applying an alternating current, and then collects the magnetic resonance signals released during the relaxation process. By analyzing the signal characteristics based on the differences in the relaxation properties of hydrogen protons in water and petroleum hydrocarbons, qualitative identification and quantitative assessment of groundwater petroleum hydrocarbon contamination can be achieved, and this technology has been widely used in the field of geological pollution detection. However, in actual detection, weak magnetic resonance signals are often submerged in high-intensity random noise, resulting in an extremely low signal-to-noise ratio. Unlike other magnetic resonance noise, random noise characteristics cannot be characterized by mathematical formulas. It is widely distributed, random, and unpredictable in the time and frequency domains, making it difficult to effectively remove. Therefore, it is of great significance to develop an efficient and accurate random noise reduction method to provide high-quality petroleum hydrocarbon contamination detection data for subsequent inversion interpretation.

[0003] Existing random noise reduction methods, such as those leveraging the decomposition properties of the EMD algorithm, decompose the full-wave MRI signal into its eigenmode components. The TFPF algorithm then encodes the dominant modal component of the signal into the instantaneous frequency of a unit-amplitude analytical signal. This method exploits the characteristic of the analytical signal's time-frequency distribution, which is concentrated along the instantaneous frequency, to suppress random noise. However, this method requires manual adjustment of filter parameters and relies on prior human experience.

[0004] The architecture based on convolutional neural networks is generally a convolutional neural network, which contains a large number of convolutional layers, which can easily cause gradient disappearance and long training time. Summary of the Invention

[0005] An embodiment of the present invention provides a magnetic resonance random noise reduction system suitable for petroleum hydrocarbon pollution detection, which solves the gradient vanishing problem in convolutional neural networks.

[0006] Another aspect of the present invention provides a magnetic resonance random noise reduction method suitable for petroleum hydrocarbon pollution detection.

[0007] According to a first aspect of an embodiment of the present invention, a magnetic resonance random noise reduction system suitable for petroleum hydrocarbon contamination detection includes a petroleum hydrocarbon contamination random noise reduction network, wherein the petroleum hydrocarbon contamination random noise reduction network includes a deep residual denoiser and a classification discriminator, wherein:

[0008] A deep residual denoiser is used to process the measured random noise signal according to a nonlinear mapping from the random noise signal to the random noise to obtain a predicted random noise, and to remove the predicted random noise from the measured random noise signal to obtain a denoised signal;

[0009] A classification discriminator is used to learn the characteristics of the clean signal in the clean signal set, identify the denoised signal input to the classification discriminator as a denoised signal or a clean signal, and use the classification result of the classification discriminator to feed back to the deep residual denoiser, which optimizes the network parameters according to the feedback.

[0010] Furthermore, the depth residual denoiser comprises:

[0011] The signal extraction module is used to output the predicted random noise through convolution operation based on the nonlinear mapping from the random noise signal to the random noise, and to remove the predicted random noise from the measured random noise signal to obtain a preliminary denoised signal;

[0012] The detail recovery module is used to perform a deconvolution operation on the preliminary denoised signal to obtain a denoised signal.

[0013] Furthermore, the signal extraction module includes a first convolutional layer, a first ReLU activation function and three residual blocks connected in sequence, each residual block includes a second convolutional layer and a third convolutional layer connected in series and then jump-connected to a fourth convolutional layer, a second ReLU activation function is set between the second convolutional layer and the third convolutional layer, and the output is through a third ReLU activation function.

[0014] Furthermore, the detail recovery module includes four serially connected deconvolution layers, and a ReLU activation function is set after each deconvolution layer.

[0015] Furthermore, the deep residual denoiser and the classification discriminator engage in adversarial training. During adversarial training, the deep residual denoiser takes minimizing the petroleum hydrocarbon loss function as its training objective. As the petroleum hydrocarbon loss function gradually decreases, the network parameters of the deep residual denoiser are updated. The probability distribution of the denoised signal generated by the deep residual denoiser gradually becomes similar to that of the pure signal, making it impossible for the classification discriminator to distinguish between the pure signal and the denoised signal.

[0016] The classification discriminator is trained with the goal of maximizing the adversarial loss function. By maximizing the adversarial loss function, it improves the ability to distinguish between pure signals and denoised signals, and is used to guide the optimization of the network parameters of the deep residual denoiser.

[0017] Furthermore, the petroleum hydrocarbon pollution random noise reduction network is used to reduce noise in actual measurement of random noise signals after training. The training process includes:

[0018] Construct a signal set containing random noise;

[0019] Initialize the network parameters, input the random noise signal set into the petroleum hydrocarbon pollution random noise reduction network, establish a nonlinear mapping from the random noise signal to the random noise by the deep residual denoiser, output the predicted random noise and remove the predicted random noise from the random noise signal to obtain the denoised signal, and construct the denoised signal set; input the denoised signal set and the clean signal set into the classification discriminator, the classification discriminator learns the characteristics of the clean signal, and identifies the signal input to the classification discriminator as a denoised signal or a clean signal;

[0020] An adversarial learning strategy is adopted between the deep residual denoiser and the classification discriminator, and a petroleum hydrocarbon loss function is designed; the deep residual denoiser minimizes the petroleum hydrocarbon loss function to update the network parameters and optimize the denoising results, while the classification discriminator maximizes the adversarial loss to improve the discrimination ability and constrain the random noise reduction results of the deep residual denoiser.

[0021] Furthermore, the petroleum hydrocarbon loss function is:

[0022] ,

[0023] Where, is the petroleum hydrocarbon loss function, is the mean square error loss, is the adversarial loss between the classification discriminator and the deep residual denoiser, is the balance factor, is the number of denoised signals and pure signals, is the F norm, It is a pure signal The probability distribution of is the noise cancellation signal The probability distribution of is a function In the probability distribution The mathematical expectation under They are functions In the probability distribution The mathematical expectation under represents the nonlinear mapping corresponding to the deep residual denoiser, Represents the nonlinear mapping corresponding to the classification discriminator.

[0024] According to a second aspect of the present invention, a method for reducing magnetic resonance random noise using a magnetic resonance random noise reduction system suitable for petroleum hydrocarbon pollution detection includes:

[0025] Processing the measured signal containing random noise according to a nonlinear mapping from the signal containing random noise to random noise to obtain predicted random noise, and removing the predicted random noise from the measured signal containing random noise to obtain a denoised signal;

[0026] The features of the clean signal in the clean signal set are learned, the denoised signal is identified as the denoised signal or the clean signal, and the classification result of the classification discriminator is used to feed back to the deep residual denoiser, which optimizes the network parameters according to the feedback.

[0027] Furthermore, the measured signal containing random noise is processed according to the nonlinear mapping from the signal containing random noise to random noise to obtain predicted random noise, and the predicted random noise is removed from the measured signal containing random noise to obtain a denoised signal, including:

[0028] According to the nonlinear mapping from the random noise signal to the random noise, the predicted random noise is output through the convolution operation, and the predicted random noise is removed from the measured random noise signal to obtain a preliminary denoised signal;

[0029] Perform deconvolution operation on the preliminary denoised signal to obtain the denoised signal.

[0030] The present invention has the following advantages and beneficial effects:

[0031] The embodiments of the present invention address the problems of traditional methods requiring manual adjustment of filter parameters, reliance on prior assumptions, and cumbersome processing procedures, and achieve intelligent denoising with simple and convenient operation. To address the problems of excessive convolutional layers in ordinary convolutional denoising networks leading to gradient vanishing and long training time, the convolutional layers are reduced in the petroleum hydrocarbon contaminated magnetic resonance random noise reduction network structure, solving the gradient problem and shortening the training time. To address the problem of limited denoising effect of ordinary convolutional denoising networks, the petroleum hydrocarbon loss function is used to optimize network parameters, and an adversarial learning strategy is adopted, introducing a classification discriminator to constrain the network denoising effect, thereby enhancing the network generalization ability and improving the signal-to-noise ratio and signal fidelity of the denoising results. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A schematic structural diagram of a magnetic resonance random noise reduction system for petroleum hydrocarbon pollution detection provided by an embodiment of the present invention;

[0033] Figure 2 This is a structural block diagram of a residual block provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0034] 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.

[0035] The first, second, etc. mentioned in the embodiments of the present invention are only for distinguishing different positions. For example, the first convolutional layer and the second convolutional layer both belong to convolution structures with the same function, and the first ReLU activation function and the second ReLU activation function both belong to ReLU activation functions with the same function.

[0036] In this embodiment, a magnetic resonance random noise reduction system suitable for petroleum hydrocarbon pollution detection is provided. Figure 1 As shown, it includes: a petroleum hydrocarbon pollution random noise reduction network, which includes a deep residual denoiser and a classification discriminator, wherein:

[0037] A deep residual denoiser is used to process the measured random noise signal according to a nonlinear mapping from the random noise signal to the random noise to obtain a predicted random noise, and to remove the predicted random noise from the measured random noise signal to obtain a denoised signal;

[0038] A classification discriminator is used to learn the characteristics of the clean signal in the clean signal set, identify the denoised signal input to the classification discriminator as a denoised signal or a clean signal, and use the classification result of the classification discriminator to feed back to the deep residual denoiser, which optimizes the network parameters according to the feedback.

[0039] It's understandable that the petroleum hydrocarbon pollution random noise reduction network is a trained network with optimized parameters. The nonlinear mapping from random noise signals to random noise in the deep residual denoiser is a trained nonlinear mapping. When processing measured random noise signals, the predicted random noise is directly obtained based on this trained nonlinear mapping.

[0040] The clean signal set input to the classification discriminator is a signal without any noise. The classification discriminator can distinguish whether the denoised signal obtained by the deep residual denoiser is a clean signal with noise removed or a signal containing noise. This can guide the further optimization of the network parameters of the deep residual denoiser.

[0041] In one embodiment, the deep residual denoiser comprises:

[0042] The signal extraction module is used to output the predicted random noise through convolution operation based on the nonlinear mapping from the random noise signal to the random noise, and to remove the predicted random noise from the measured random noise signal to obtain a preliminary denoised signal;

[0043] The signal extraction module includes the first convolutional layer 1, the first ReLU activation function 2, and three residual blocks 3 connected in sequence. The connection here refers to the relationship between layers according to the direction of data flow, see Figure 2As shown, each residual block 3 includes a second convolutional layer 5 and a third convolutional layer 7 connected in series, which are then jump-connected to a fourth convolutional layer 9. A second ReLU activation function 6 is set between the second convolutional layer 5 and the third convolutional layer 7, and the output is output through a third ReLU activation function 8. The data of each residual block 3 is input to the second convolutional layer 5 and the fourth convolutional layer 9. The output of the second convolutional layer 5 passes through the second ReLU activation function 6, the third convolutional layer 7, and the third ReLU activation function 8 in sequence. The output of the fourth convolutional layer 9 passes through the third ReLU activation function 8.

[0044] The detail recovery module is used to perform a deconvolution operation on the preliminary denoised signal to obtain a denoised signal.

[0045] The detail recovery module includes four serially connected deconvolution layers 4, and a ReLU activation function is set after each deconvolution layer 4.

[0046] The classification discriminator consists of two concatenated convolutional layers and one fully connected layer; each convolutional layer is connected with a LeakyReLU activation function, and the fully connected layer is connected with a sigmoid activation function.

[0047] In one embodiment, the deep residual denoiser and the classification discriminator engage in adversarial training. During adversarial training, the deep residual denoiser takes minimizing the petroleum hydrocarbon loss function as its training objective. As the petroleum hydrocarbon loss function gradually decreases, the network parameters of the deep residual denoiser are updated. The probability distribution of the denoised signal generated by the deep residual denoiser gradually becomes similar to that of the pure signal, making it impossible for the classification discriminator to distinguish between the pure signal and the denoised signal.

[0048] The classification discriminator is trained with the goal of maximizing the adversarial loss function. By maximizing the adversarial loss function, it improves the ability to distinguish between pure signals and denoised signals, and is used to guide the optimization of the network parameters of the deep residual denoiser.

[0049] The petroleum hydrocarbon pollution random noise reduction network in the embodiment of the present application includes a deep residual denoiser and a classification discriminator, and the training strategy is adversarial learning; the deep residual denoiser minimizes the petroleum hydrocarbon loss to update the network parameters, so that the optimal nonlinear mapping from the random noise signal to the random noise can be established, thereby reducing the random noise and generating a denoised signal close to the pure signal, so that the classification discriminator cannot distinguish between the pure signal and the denoised signal; the classification discriminator maximizes the adversarial loss to improve the discrimination ability, guides the optimization of the network parameters of the deep residual denoiser, further improves the random noise suppression effect of the deep residual denoiser, and makes the denoised signal have a high signal-to-noise ratio and high signal fidelity.

[0050] Before use, the petroleum hydrocarbon pollution random noise reduction network needs to be trained to form a trained petroleum hydrocarbon pollution random noise reduction network. The trained petroleum hydrocarbon pollution random noise reduction network has the optimal network parameters and the optimal nonlinear mapping from random noise signals to random noise. The training process includes:

[0051] Construct a signal set containing random noise;

[0052] Initialize the network parameters, input the random noise signal set into the petroleum hydrocarbon pollution random noise reduction network, establish a nonlinear mapping from the random noise signal to the random noise by the deep residual denoiser, output the predicted random noise and remove the predicted random noise from the random noise signal to obtain the denoised signal, and construct the denoised signal set; input the denoised signal set and the clean signal set into the classification discriminator, the classification discriminator learns the characteristics of the clean signal, and identifies the signal input to the classification discriminator as a denoised signal or a clean signal;

[0053] An adversarial learning strategy is adopted between the deep residual denoiser and the classification discriminator, and a petroleum hydrocarbon loss function is designed; the deep residual denoiser minimizes the petroleum hydrocarbon loss function to update the network parameters and optimize the denoising results, while the classification discriminator maximizes the adversarial loss to improve the discrimination ability and constrain the random noise reduction results of the deep residual denoiser.

[0054] Among them, constructing a signal set containing random noise: simulating and generating a pure signal to construct a pure signal set; mixing the simulated random noise with the measured noise to form a random noise to construct a random noise set; superimposing the pure signal set and the random noise set to form a signal containing random noise to construct a signal set containing random noise as a training data set, and preprocessing the signal containing random noise;

[0055] Training the petroleum hydrocarbon pollution random noise reduction network: The training dataset is input into the petroleum hydrocarbon pollution random noise reduction network, and the deep residual denoiser establishes a nonlinear mapping from random noise signals to random noise, outputs the predicted random noise and removes it from the random noise signal to obtain the denoised signal, and constructs the denoised signal set; the denoised signal set and the clean signal set are input into the classification discriminator together, and the classification discriminator learns the characteristics of the clean signal and identifies the signal input to the classification discriminator as a denoised signal or a clean signal; an adversarial learning strategy is adopted between the deep residual denoiser and the classification discriminator, and a petroleum hydrocarbon loss function is designed; the deep residual denoiser minimizes the petroleum hydrocarbon loss function to update the network parameters and optimize the denoising results, and the classification discriminator maximizes the adversarial loss to improve the identification ability, further constraining the random noise reduction results of the deep residual denoiser;

[0056] Testing the petroleum hydrocarbon pollution random noise reduction network: Collect measured random noise signals, construct a measured random noise signal set, and preprocess the spike noise and power frequency harmonic noise reduction; input the preprocessed measured random noise signal set as the test data set into the trained petroleum hydrocarbon pollution random noise reduction network to perform random noise reduction and obtain the denoising result; invert the denoising result, and analyze the actual distribution of petroleum hydrocarbon pollution based on the inverted signal relaxation time spectrum.

[0057] In one embodiment, a signal set containing random noise is input into a deep residual denoiser of a petroleum hydrocarbon contamination random noise reduction network;

[0058] The signal extraction module of the deep residual denoiser establishes a nonlinear mapping from random noise signals to random noise, and predicts the random noise through convolution operation output:

[0059] ,

[0060] Where, is the number of convolutional layers in the signal extraction module, It is The predicted random noise output by the convolutional layer, It is Convolutional layer input layer predicts random noise, and They are The convolution kernel and bias of the convolution layer, * is the convolution operation, Is the activation function ReLU operation; when the Layer is the signal extraction module When the last convolution layer is completed, the next step is to The network parameters of the signal extraction module;

[0061] Remove the predicted random noise from the signal containing random noise to obtain the preliminary denoised signal:

[0062] ,

[0063] Where, is a signal containing random noise, is the initial noise elimination signal, After the multi-layer convolution operation of the signal extraction module, the signal information is damaged; then, after the deconvolution operation of the detail recovery module, the effective information of the signal is largely restored;

[0064] The deconvolution operation is as follows:

[0065] ,

[0066] ,

[0067] Where, is the number of deconvolution layers in the detail recovery module, and The first and second The denoised signal output by the deconvolution layer is and The first and second The input of the deconvolution layer, and The first and second The transpose of the convolution kernel of the deconvolution layer, and The first and second The bias of the deconvolution layer; when the Layer is the detail recovery module The last deconvolution layer removes the noise signal Recorded as ,

[0068] The nonlinear mapping constructed by the deep residual denoiser is , are the network parameters of the deep residual denoiser;

[0069] The denoised signal set and the clean signal set are input into the classification discriminator, and the clean signal is used as the label; the classification discriminator extracts the input denoised signal through two convolutional layers and one fully connected layer Features and classify the input denoised signal; the convolutional layer of the classification discriminator is operated as follows:

[0070] ,

[0071] and are the output and input of the rth convolutional layer, and It is The convolution kernel and bias of the convolution layer, It is the LeakyReLU activation function operation;

[0072] The fully connected layer of the classification discriminator operates as follows:

[0073] ,

[0074] and are the output and input of the fully connected layer, is the fully connected layer operation, It is the Sigmoid activation function operation; when the classification discriminator inputs the denoised signal Identified as a noise-cancelled signal, the identification result is the input noise-cancelled signal When there is random noise in the input signal, D is 0; when the classification discriminator Classified as a pure signal, the identification result is the input noise-removed signal When there is no random noise in is 1;

[0075] The nonlinear mapping constructed by the classification discriminator is , represents the network parameters of the classification discriminator;

[0076] Designing a petroleum hydrocarbon loss function , The mean square error loss and classification discriminator , Deep Residual Denoiser The adversarial loss between Organic combination is defined as follows:

[0077] ,

[0078] Where, is the number of denoised signals and pure signals, for norm, It is a pure signal The probability distribution of is the noise cancellation signal The probability distribution of is a function In the probability distribution The mathematical expectation under They are functions In the probability distribution The mathematical expectation under represents the nonlinear mapping corresponding to the deep residual denoiser, Represents the nonlinear mapping corresponding to the classification discriminator.

[0079] The deep residual denoiser in the petroleum hydrocarbon pollution random noise reduction network competes with the classification discriminator. In the adversarial training, the deep residual denoiser minimizes the petroleum hydrocarbon loss function. As the training goal, Gradually decrease, network parameters are updated, and the denoised signal generated by the deep residual denoiser With pure signal The probability distribution of is gradually similar, making it impossible for the classification discriminator to distinguish between the pure signal and the noise-removed signal; while the classification discriminator maximizes the adversarial loss function As the training goal, by maximizing Improve identification capabilities and distinguish pure signals and noise cancellation signal , guiding the optimization of the network parameters of the deep residual denoiser, and further improving the random noise suppression effect of the deep residual denoiser;

[0080] The adversarial training process is as follows:

[0081] ,

[0082] Where, and They represent the optimal solutions of the deep residual denoiser and classification discriminator in adversarial training respectively;

[0083] When the classification discriminator cannot distinguish between the pure signal and the denoised signal, the denoising result of the petroleum hydrocarbon pollution random noise reduction network is ideal and the network training is completed.

[0084] On the other hand, an embodiment of the present application provides a magnetic resonance random noise reduction method suitable for petroleum hydrocarbon contamination detection, comprising: processing a measured random noise-containing signal according to a nonlinear mapping from a random noise-containing signal to random noise to obtain a predicted random noise, and removing the predicted random noise from the measured random noise-containing signal to obtain a denoised signal;

[0085] The features of the clean signal in the clean signal set are learned, the denoised signal is identified as the denoised signal or the clean signal, and the classification result of the classification discriminator is used to feed back to the deep residual denoiser, which optimizes the network parameters according to the feedback.

[0086] In one embodiment, processing a measured signal containing random noise according to a nonlinear mapping from a signal containing random noise to random noise to obtain predicted random noise, and removing the predicted random noise from the measured signal containing random noise to obtain a denoised signal includes:

[0087] According to the nonlinear mapping from the random noise signal to the random noise, the predicted random noise is output through the convolution operation, and the predicted random noise is removed from the measured random noise signal to obtain a preliminary denoised signal;

[0088] Perform a deconvolution operation on the preliminary denoised signal to obtain the denoised signal.

[0089] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

[0091] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A magnetic resonance random noise reduction system suitable for petroleum hydrocarbon pollution detection, characterized in that: The system includes a petroleum hydrocarbon contamination random noise reduction network, wherein the petroleum hydrocarbon contamination random noise reduction network includes a deep residual denoiser and a classification discriminator, wherein: A deep residual denoiser is used to process the measured random noise signal according to a nonlinear mapping from the random noise signal to the random noise to obtain a predicted random noise, and to remove the predicted random noise from the measured random noise signal to obtain a denoised signal; A classification discriminator is used to learn the characteristics of the clean signal in the clean signal set, identify the denoised signal input to the classification discriminator as a denoised signal or a clean signal, and use the classification result of the classification discriminator to feed back to the deep residual denoiser, which optimizes the network parameters according to the feedback.

2. The magnetic resonance random noise reduction system for petroleum hydrocarbon pollution detection according to claim 1, characterized in that: The deep residual denoiser comprises: The signal extraction module is used to output the predicted random noise through convolution operation based on the nonlinear mapping from the random noise signal to the random noise, and to remove the predicted random noise from the measured random noise signal to obtain a preliminary denoised signal; The detail recovery module is used to perform a deconvolution operation on the preliminary denoised signal to obtain a denoised signal.

3. The magnetic resonance random noise reduction system for petroleum hydrocarbon pollution detection according to claim 2, characterized in that: The signal extraction module includes a first convolutional layer, a first ReLU activation function and three residual blocks connected in sequence. Each residual block includes a second convolutional layer and a third convolutional layer connected in series and then jump-connected to a fourth convolutional layer. A second ReLU activation function is set between the second convolutional layer and the third convolutional layer, and the output is through a third ReLU activation function.

4. The magnetic resonance random noise reduction system for petroleum hydrocarbon pollution detection according to claim 2, characterized in that: The detail recovery module includes four serially connected deconvolution layers, and a ReLU activation function is set after each deconvolution layer.

5. The magnetic resonance random noise reduction system for petroleum hydrocarbon pollution detection according to claim 2, characterized in that: The deep residual denoiser and the classification discriminator engage in adversarial training. During adversarial training, the deep residual denoiser takes minimizing the petroleum hydrocarbon loss function as its training objective. As the petroleum hydrocarbon loss function gradually decreases, the network parameters of the deep residual denoiser are updated. The probability distribution of the denoised signal generated by the deep residual denoiser gradually becomes similar to that of the pure signal, making it impossible for the classification discriminator to distinguish between the pure signal and the denoised signal. The classification discriminator is trained with the goal of maximizing the adversarial loss function. By maximizing the adversarial loss function, it improves the ability to distinguish between pure signals and denoised signals, and is used to guide the optimization of the network parameters of the deep residual denoiser.

6. The magnetic resonance random noise reduction system for petroleum hydrocarbon pollution detection according to claim 1, characterized in that: The petroleum hydrocarbon pollution random noise reduction network is used to reduce the noise of a measured random noise signal after training. The training process includes: Construct a signal set containing random noise; Initialize the network parameters, input the random noise signal set into the petroleum hydrocarbon pollution random noise reduction network, establish a nonlinear mapping from the random noise signal to the random noise by the deep residual denoiser, output the predicted random noise and remove the predicted random noise from the random noise signal to obtain the denoised signal, and construct the denoised signal set; input the denoised signal set and the clean signal set into the classification discriminator, the classification discriminator learns the characteristics of the clean signal, and identifies the signal input to the classification discriminator as a denoised signal or a clean signal; An adversarial learning strategy is adopted between the deep residual denoiser and the classification discriminator, and a petroleum hydrocarbon loss function is designed; the deep residual denoiser minimizes the petroleum hydrocarbon loss function to update the network parameters and optimize the denoising results, while the classification discriminator maximizes the adversarial loss to improve the discrimination ability and constrain the random noise reduction results of the deep residual denoiser.

7. The magnetic resonance random noise reduction system for petroleum hydrocarbon pollution detection according to claim 6, characterized in that: The petroleum hydrocarbon loss function is: ， Where, is the petroleum hydrocarbon loss function, is the mean square error loss, is the adversarial loss between the classification discriminator and the deep residual denoiser, is the balance factor, is the number of denoised signals and pure signals, is the F norm, It is a pure signal The probability distribution of is the noise cancellation signal The probability distribution of is a function In the probability distribution The mathematical expectation under They are functions In the probability distribution The mathematical expectation under represents the nonlinear mapping corresponding to the deep residual denoiser, Represents the nonlinear mapping corresponding to the classification discriminator.

8. A magnetic resonance random noise reduction method, using the magnetic resonance random noise reduction system for petroleum hydrocarbon pollution detection according to any one of claims 1 to 7, characterized in that: include: Processing the measured signal containing random noise according to a nonlinear mapping from the signal containing random noise to random noise to obtain predicted random noise, and removing the predicted random noise from the measured signal containing random noise to obtain a denoised signal; The features of the clean signal in the clean signal set are learned, the denoised signal is identified as the denoised signal or the clean signal, and the classification result of the classification discriminator is used to feed back to the deep residual denoiser, which optimizes the network parameters according to the feedback.

9. The method for reducing magnetic resonance random noise according to claim 8, wherein: Processing the measured signal containing random noise according to a nonlinear mapping from the signal containing random noise to random noise to obtain predicted random noise, and removing the predicted random noise from the measured signal containing random noise to obtain a denoised signal, including: According to the nonlinear mapping from the random noise signal to the random noise, the predicted random noise is output through the convolution operation, and the predicted random noise is removed from the measured random noise signal to obtain a preliminary denoised signal; Perform deconvolution operation on the preliminary denoised signal to obtain the denoised signal.

Images