Nuclear magnetic resonance signal noise suppression method and device based on convolutional neural network

Abstract

The invention discloses a nuclear magnetic resonance signal noise suppression method and device based on a convolutional neural network. The method comprises the following steps: configuring a CPMG pulse sequence in nuclear magnetic resonance equipment; carrying out nuclear magnetic resonance testing on the shale sample based on the CPMG pulse sequence, setting a plurality of groups of echo intervals by utilizing a multi-time micro-variation echo interval technology, and generating a time-aligned multi-dimensional data set; segmenting a multi-dimensional data set by using a sliding window technology, performing segmented data processing through a pre-trained CNN network, removing noise components and performing signal reconstruction; a reinforcement learning mechanism is introduced, and according to the quality of de-noised echo signals, pulse sequence configuration parameters and a sliding window segmentation strategy are reversely optimized, so that the continuous and stable noise suppression effect is ensured. According to the method, the signal acquisition and processing process is optimized, the interference of noise on the fast relaxation component signal is reduced by using the convolutional neural network, the recognition accuracy of the signal is improved, and technical support is provided for accurate representation of the reservoir.

Description

Technical Field

[0001] This invention relates to the field of reservoir nuclear magnetic resonance detection technology, and in particular to a method and apparatus for suppressing nuclear magnetic resonance signal noise based on a convolutional neural network. Background Technology

[0002] Nuclear magnetic resonance (NMR) is a non-destructive testing technique that characterizes the internal structure and fluid properties of samples by detecting the relaxation properties of hydrogen nuclei. It is widely used in oil and gas exploration to assess reservoir porosity and permeability. However, in practical applications, the quality of NMR signals is often significantly affected by noise, especially when measuring fast-relaxing components (such as clay-bound water and solid organic matter). These components have extremely short transverse relaxation times (T2), typically less than 1 ms, and their signals decay almost entirely within the first few echo points of the CPMG pulse sequence. Furthermore, factors such as ambient noise, instrument thermal noise, and radio frequency interference can significantly reduce the signal-to-noise ratio (SNR) of the front-end echo signal. In low-field NMR equipment, the SNR may fall below 20, severely limiting the effective extraction of signals and the accuracy of subsequent analysis.

[0003] To address this challenge, existing technologies have proposed various noise suppression strategies. For example, increasing the number of signal accumulations can improve the signal-to-noise ratio (SNR), but this significantly prolongs the experimental time and reduces measurement efficiency. Traditional signal processing methods such as wavelet transform and Fourier filtering can separate noise to some extent, but their suppression effect is limited for non-stationary noise or noise that overlaps with the signal spectrum, and they are prone to losing the weak characteristics of fast relaxation components. While Multi-Contrast Tuning Echo Measurement (MCTEM) can provide more data points, it cannot fundamentally solve the problem of noise masking the front-end signal, especially in the low SNR environment of complex reservoirs such as shale.

[0004] Therefore, this invention proposes a method and apparatus for suppressing noise in nuclear magnetic resonance signals based on convolutional neural networks. By optimizing the signal acquisition and processing process, the method utilizes the convolutional learning capability of CNNs to reduce noise interference in fast relaxation component signals, providing technical support for the accurate characterization of complex reservoirs such as shale oil and gas. Summary of the Invention

[0005] In view of this, the present invention provides a method and apparatus for suppressing nuclear magnetic resonance signal noise based on convolutional neural networks, in order to solve the technical problem that existing methods have limited suppression effects on non-stationary noise or noise overlapping with the signal spectrum when measuring fast relaxation components, resulting in low accuracy of signal extraction and subsequent analysis, and failing to accurately characterize shale oil and gas.

[0006] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for suppressing noise in nuclear magnetic resonance signals based on a convolutional neural network, comprising: Configure CPMG pulse sequences in nuclear magnetic resonance equipment; Nuclear magnetic resonance tests were performed on shale samples based on the configured CPMG pulse sequence. Multiple sets of echo intervals were set using the technique of multiple micro-variable echo intervals to generate time-aligned multidimensional datasets. The multidimensional dataset is segmented using the sliding window technique, and the segmented data is processed by a pre-trained CNN network to remove noise components and perform signal reconstruction. A reinforcement learning mechanism is introduced to optimize the pulse sequence configuration parameters and sliding window segmentation strategy in reverse based on the quality of the denoised echo signal, so as to ensure the continuous stability of the noise suppression effect.

[0007] Furthermore, configuring the CPMG pulse sequence includes adjusting the echo interval, pulse width, and repetition time parameters.

[0008] Furthermore, the method of setting multiple sets of echo intervals using multiple micro-variable echo interval techniques to generate a time-aligned multidimensional dataset includes: Set multiple echo intervals to cover the decay process of fast relaxation components; The CPMG pulse sequence is run independently for each echo interval to obtain the echo data points corresponding to each echo interval. Based on the trigger timestamp, data with different echo intervals are aligned to form a three-dimensional data structure of sample × echo point × echo interval channel, which is used to describe the transverse relaxation time characteristics and noise variation of the sample.

[0009] Furthermore, the segmentation of the multidimensional dataset using the sliding window technique includes: The window size and step size are dynamically adjusted based on the frequency characteristics and noise distribution of the signal.

[0010] Furthermore, the pre-trained CNN network adopts a three-layer stacked convolutional layer architecture, with the convolutional kernel width expanding layer by layer; The first convolutional layer is used to slide along the signal time axis to extract short-term time-domain fluctuation features and achieve local noise suppression; The second convolutional layer is used to fuse the noise output from the first layer across echo interval channels, analyze the correlation of power frequency interference under different echo intervals, and generate a composite noise mode representation. The third convolutional layer is used to capture the coupled noise pattern across echo points, separating noise from the signal in the overlapping spectrum to eliminate global noise.

[0011] Furthermore, a reinforcement learning mechanism is introduced to inversely optimize the configuration parameters and sliding window strategy based on the quality of the denoised window signal, including: Using signal-to-noise ratio, phase stability, noise spectrum, and environmental parameters as the state space, the echo interval, sampling frequency, pulse width, and repetition time are dynamically adjusted to achieve adaptive optimization of signal acquisition parameters in response to changes in sample characteristics or environment.

[0012] Furthermore, the method also includes: adjusting configuration parameters based on signal-to-noise ratio feedback when the sample or environment changes, so as to achieve synchronous adjustment of the acquisition process and noise suppression effect.

[0013] Secondly, the present invention also provides a nuclear magnetic resonance signal noise suppression device based on a convolutional neural network, comprising: The initial setup module configures the CPMG pulse sequence in the MRI equipment; The echo signal acquisition module is used to perform nuclear magnetic resonance testing on shale samples based on the configured CPMG pulse sequence. It uses multiple micro-variable echo interval technology to set multiple sets of echo intervals and generate time-aligned multidimensional datasets. The signal reconstruction module is used to segment the multidimensional dataset using the sliding window technique, process the segmented data through a pre-trained CNN network, predict noise components, and perform signal reconstruction. The feedback optimization module is used to introduce a reinforcement learning mechanism to optimize the pulse sequence configuration parameters and sliding window segmentation strategy in reverse based on the quality of the denoised window signal, so as to ensure the continuous stability of the noise suppression effect.

[0014] Furthermore, the signal reconstruction module and the feedback optimization module are integrated into the embedded hardware system of the nuclear magnetic resonance device to achieve real-time signal denoising and parameter optimization.

[0015] Thirdly, the present invention also provides an electronic device, including a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, it implements the nuclear magnetic resonance signal noise suppression method based on convolutional neural networks described in the above technical solution.

[0016] Compared with existing technologies, the nuclear magnetic resonance signal noise suppression method and apparatus based on convolutional neural networks proposed in this invention have the following advantages: 1) By combining multi-echo interval data with CNN, a deep suppression strategy for complex noise patterns is constructed to improve the processing accuracy under low signal-to-noise ratio conditions. 2) Introduce reinforcement learning to dynamically adjust the acquisition parameters, forming an adaptive noise suppression framework to enhance the applicability of the method to diverse samples and environments.

[0017] 3) By directly integrating the pre-trained CNN module with the MRI signal acquisition hardware system (such as an embedded processor), real-time noise suppression of the front-end echo signal is achieved, with a processing delay of less than a single echo interval (<0.8ms), ensuring that the signal is denoised before storage; This invention optimizes the signal acquisition and processing flow, thereby reducing noise interference in fast relaxation component signals (T2 < 1ms) and improving the characterization efficiency of complex reservoirs. Attached Figure Description

[0018] Figure 1 A flowchart illustrating the nuclear magnetic resonance signal noise suppression method based on a convolutional neural network provided by this invention; Figure 2 A flowchart illustrating the practical application of the noise suppression method provided by this invention; Figure 3 A schematic diagram of the structure of the nuclear magnetic resonance signal noise suppression device based on a convolutional neural network provided by the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided by the present invention. Detailed Implementation

[0019] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0020] Please see Figure 1 This embodiment provides a method for suppressing noise in nuclear magnetic resonance signals based on convolutional neural networks, including: Step S101: Configure the CPMG pulse sequence in the nuclear magnetic resonance equipment; Step S102: Perform nuclear magnetic resonance testing on shale samples based on the configured CPMG pulse sequence, and use multiple micro-variable echo interval technology to set multiple sets of echo intervals to generate time-aligned multidimensional datasets; Step S103: Use the sliding window technique to segment the multidimensional dataset, process the segmented data through a pre-trained CNN network, remove noise components, and perform signal reconstruction. Step S104: Introduce a reinforcement learning mechanism to optimize the pulse sequence configuration parameters and sliding window segmentation strategy in reverse based on the quality of the denoised echo signal, so as to ensure the continuous stability of the noise suppression effect.

[0021] The method in this embodiment utilizes a configured CPMG pulse sequence to perform nuclear magnetic resonance testing on shale samples. The multidimensional dataset generated by different echo intervals more comprehensively reflects the characteristics and state of the samples, making signal acquisition more detailed and diverse, thus enhancing data accuracy. Signal processing via a sliding window segmentation method more accurately captures signal characteristics within different time periods, thereby improving noise suppression. A CNN network accurately predicts noise components and reconstructs the signal, ensuring the efficiency and accuracy of the denoising process and further improving the quality of the test results. Through a reinforcement learning mechanism, feedback based on the quality of the denoised signal is provided, dynamically adjusting the pulse sequence configuration parameters to continuously improve signal quality and ensure the stability of the noise suppression effect.

[0022] In a preferred embodiment, in step S101, configuring the CPMG pulse sequence includes adjusting the echo interval, pulse width, and repetition time parameters.

[0023] As a specific implementation, a CPMG pulse sequence is configured in the nuclear magnetic resonance (NMR) device, and the echo interval (TE), pulse width, and repetition time (TR) parameters are adjusted to enhance the acquisition capability of fast relaxation component signals. Simultaneously, a pre-trained convolutional neural network (CNN) module is embedded in the acquisition system for real-time noise suppression of the front-end echo signal. By finely adjusting the parameters of the CPMG pulse sequence (e.g., shortening the TE to below 0.6 ms and optimizing the pulse width to within 10 μs), the sampling density of the fast relaxation component signal in the initial echo stage is increased, thereby providing more sufficient raw data for noise suppression. Meanwhile, the pre-trained CNN module, embedded in the acquisition system, utilizes its convolutional operations to quickly identify noise components in the front-end echo signal and perform real-time suppression before the signal enters storage or subsequent processing, thereby reducing the masking of fast relaxation features by noise.

[0024] In a preferred embodiment, in step S102, the process of setting multiple sets of echo intervals using multiple micro-variable echo interval technology to generate a time-aligned multidimensional dataset includes: Set multiple echo intervals to cover the decay process of fast relaxation components; The CPMG pulse sequence is run independently for each echo interval to obtain the echo data points corresponding to each echo interval. Based on the microsecond-level timestamps (error ≤ 1μs) generated by the built-in hardware timer of the nuclear magnetic resonance spectrometer, data with different echo intervals are aligned to form a three-dimensional data structure of sample × echo point × echo interval channel, and the distribution of noise as echo interval changes is obtained, which is used to reflect the relaxation characteristics inside the sample.

[0025] As a specific implementation, the Multiple Micro-Variable Echo Interval (MCTEM) technique is employed. Multiple echo intervals (e.g., 0.6ms, 0.8ms, 1.0ms, acquired three times, with 1000 echo points acquired each time) are set to acquire the NMR signal of the sample, thereby forming a multidimensional dataset containing diverse noise characteristics. This multidimensional data not only reflects the relaxation characteristics within the sample but also captures the distribution pattern of noise as the echo interval changes, providing rich input information for CNNs for subsequent noise suppression processing.

[0026] In a preferred embodiment, in step S103, the segmentation of the multidimensional dataset using the sliding window technique includes: dynamically adjusting the window size and step size by combining the frequency characteristics of the signal and the noise distribution.

[0027] As a specific implementation, the sliding window technique divides the front-end echo signal into multiple time segments (e.g., each window covers 5 echo points, with a step size of 2 echo points), extracting local features segment by segment. The window size is dynamically adjusted according to the signal frequency (e.g., narrowing the window for high-frequency bands and expanding the window for low-frequency bands) and noise distribution (e.g., shortening the step size in high-noise regions) to ensure the capture of details of noise variations. The CNN module processes each segment of data, specifically suppressing the effects of random and system noise, thus improving the overall signal clarity. By processing segmented data through CNN, noise interference in different time periods is specifically suppressed.

[0028] It should be noted that for high-frequency noise capture: short TE (0.6ms) dense sampling (approximately 1.67kHz) is mainly used to capture power frequency interference and residual radio frequency pulses. For low-frequency noise separation: long TE (1.0ms) covers the entire relaxation decay range, mainly used to enhance the spectral separation of white noise and residual signals. By accurately capturing and separating noise, the image quality and signal-to-noise ratio of MRI can be effectively improved, interference can be reduced, and the accuracy of subsequent data processing can be enhanced.

[0029] In a preferred embodiment, the pre-trained CNN network adopts a three-layer stacked convolutional layer architecture, with the convolutional kernel width expanding layer by layer; The first convolutional layer is used to slide along the signal time axis to extract short-term time-domain fluctuation features and achieve local noise suppression; The second convolutional layer is used to fuse the noise response across echo interval channels of the first layer output, analyze the correlation of power frequency interference under different echo intervals, and generate a composite noise mode representation. The third convolutional layer is used to capture the coupled noise pattern across echo points, separating noise from the signal in the overlapping spectrum to eliminate global noise.

[0030] In a preferred embodiment, step S104 introduces a reinforcement learning mechanism to inversely optimize configuration parameters and sliding window strategy based on the quality of the denoised window signal, including: Using signal-to-noise ratio, phase stability, noise spectrum, and environmental parameters as the state space, the echo interval, sampling frequency, pulse width, and repetition time are dynamically adjusted to achieve adaptive optimization of signal acquisition parameters in response to changes in sample characteristics or environment.

[0031] As a specific implementation, a reinforcement learning (RL) mechanism is introduced, using echo interval (TE), sampling frequency, and pulse sequence repetition count as adjustable parameters. Based on real-time signal-to-noise ratio feedback of the acquired signal, these parameters are dynamically adjusted to optimize signal acquisition quality and adapt to different sample characteristics and environmental conditions. Furthermore, a pre-trained CNN module is integrated with the reinforcement learning mechanism into the NMR instrument hardware system, running in real-time during signal acquisition. The acquisition parameters (such as TE and TR) and processing flow are dynamically adjusted through a real-time feedback mechanism to adapt to sample changes or environmental variations, ensuring the continuous stability of noise suppression performance.

[0032] The reinforcement learning mechanism monitors the signal-to-noise ratio (SNR) of the acquired signal in real time (e.g., calculating SNR every 10 seconds) and uses TE, sampling frequency (e.g., adjusting from 500kHz to 1MHz), and repetition time (TR) (e.g., adjusting from 1s to 2s) as control variables; it dynamically adjusts these parameters based on changes in the quality of the feedback signal. For example, when the SNR is below 20, the TE is shortened and the sampling frequency is increased to improve the acquisition quality of the fast relaxation component, thereby creating more favorable conditions for noise suppression.

[0033] As a specific implementation, a CNN module and reinforcement learning mechanism are deployed into the embedded system of an NMR instrument to ensure synchronous operation during signal acquisition. Each time a sample is changed or the environment changes (such as temperature fluctuations or increased electromagnetic interference), the system adjusts the acquisition parameters and processing flow through real-time feedback (SNR or noise power spectrum). For example, when a high-noise environment is detected, the system automatically shortens the TE and increases the window density to maintain the stability of the noise suppression effect.

[0034] To better illustrate the application of this method, a specific implementation example is provided below to demonstrate the method of the present invention in detail.

[0035] 1) Preparatory work for the experiment First, obtain the sample to be tested. Extract a representative sample (such as shale) from the target oil and gas reservoir, selecting a cylindrical core with a diameter of 2.5 cm and a height of 5 cm to ensure that the sample reflects the geological characteristics and mineral distribution of the reservoir. Clean and wipe the sample surface clean. Then, place the sample in a vacuum drying oven (temperature 120℃, vacuum degree 0.01MPa) and dry for 48 hours to ensure that the sample is in a water-free and oil-free state. At the same time, select a low-field nuclear magnetic resonance instrument (operating frequency 2MHz, magnetic field strength 0.05T) and set up the experimental environment in a constant temperature laboratory (temperature 25±1℃).

[0036] The instrument's radio frequency pulse parameters were calibrated, with the 90° pulse width set to 10μs and the 180° pulse width set to 20μs. The initial signal-to-noise ratio (SNR) was tested using a standard sample (such as distilled water) and found to be above 50 to ensure stable instrument operation.

[0037] 2) Data Acquisition and Preprocessing An improved CPMG pulse sequence was used to acquire the signal. The initial echo interval (TE) was set to 0.8 ms, the repetition time (TR) to 1.5 s, and 1000 echoes were acquired. The pulse amplitudes were 5V for 90° pulses and 10V for 180° pulses to enhance the signal strength of the fast relaxation component. A pre-acquisition experiment was run, and the signal amplitudes of the first 20 echo points were recorded to confirm that the attenuation curve met expectations (e.g., initial amplitude > 50 μV, attenuation to < 10 μV by the 10th echo point).

[0038] After acquisition, signal processing software (such as MATLAB) is used to remove outliers (data points that exceed the mean ± 3 times the standard deviation) and normalize the signal amplitude to the [0, 1] interval.

[0039] To simulate noise interference in a real environment, the spectral characteristics of environmental noise (such as radio frequency interference and thermal noise) of a low-field nuclear magnetic resonance device were measured to generate matched Gaussian white noise (mean 0, variance 0.01-0.05). The Kolmogorov-Smirnov test (KS test) was used to verify the statistical consistency between the simulated noise and the real noise, ensuring the practical applicability of the training data.

[0040] 3) Embedding the CNN module and initial configuration A pre-trained convolutional neural network (CNN) module was embedded into the signal acquisition system of the MRI instrument via a high-speed data interface (1MHz sampling rate) and configured for real-time processing. The CNN module's input was the first 20 echo points of the front-end echo signal, and its output was the denoised signal sequence. The module's operating delay was tested to ensure that the processing time was less than the interval between individual echoes (<0.8ms), and the synchronization of the input and output signals was monitored using an oscilloscope (delay error <0.1ms).

[0041] Specifically, in the initial configuration, the CNN is set to process one batch of data (20 echo points) at a time, with a sampling frequency of 1MHz, to ensure hardware compatibility with the acquisition system and to identify noise features through the pre-acquired data verification module.

[0042] 4) Segmented signal processing via sliding window A sliding window technique is applied to the acquired front-end echo signals. The initial window size is set to 3 to 7 echo points, with a step size of 1 to 3 echo points, dynamically adjusted according to signal frequency and noise intensity. Signal frequency and noise power spectrum are monitored in real time. The window size is adjusted according to signal frequency; for example, it is reduced to 3 echo points for high-frequency bands (>5kHz) to capture rapid changes, and increased to 7 echo points for low-frequency bands (<1kHz) to cover complete features. The step size is adjusted according to noise intensity; for example, it is reduced to 1 echo point when noise power >50μW to improve resolution. The segmented data is sequentially input into a CNN module to process noise interference (e.g., random noise peak >10μV) in each signal segment, outputting denoised signal segments, which are then spliced ​​into a complete echo sequence.

[0043] 5) Training of the noise suppression model (CNN) The first 20 echo points of the front-end echo signal are selected as input to the CNN to construct a simplified CNN model containing three convolutional layers. The first convolutional layer uses a 1×3 kernel with a stride of 1, padding of 0, and a kernel count of 16, with ReLU activation. The second convolutional layer uses a 1×3 kernel with a stride of 1, padding of 0, and a kernel count of 32. The third convolutional layer uses a 1×5 kernel with a stride of 1, padding of 0, and a kernel count of 64. A linear output layer is then constructed to output the denoised signal.

[0044] The mean squared error (MSE) is used as the loss function, with the goal of minimizing the difference between the denoised signal and the true signal. The Adam optimizer is used with a learning rate of 0.001, and the training is conducted for 500 epochs with 32 samples per batch, enabling the model to learn the ability to suppress signal noise.

[0045] The training data comes from the noisy dataset in step 2. The input is a noisy signal, and the target output is the original noiseless signal (generated by simulation). The training process is completed on a high-performance computer, and each iteration takes about 0.5 seconds.

[0046] 6) Real-time noise suppression processing A pre-trained CNN model is deployed in the NMR instrument. During signal acquisition, the front-end echo data (the first 20 echo points) is input to the CNN module in real time for noise suppression. A reinforcement learning (RL) module is introduced to monitor the signal-to-noise ratio (SNR, calculated every 10 seconds, target value >30) in real time, with TE (range 0.6-1.0 ms), sampling frequency (range 500 kHz-1 MHz), and TR (range 1.2-2.0 s) as adjustable parameters. For example, when the SNR is below 25, TE is shortened to 0.6 ms and the sampling frequency is increased to 1 MHz to improve signal quality. Each time the sample is changed (e.g., from shale to sandstone) or the environment changes (e.g., the temperature rises to 30°C), the RL module adjusts the parameters based on the SNR feedback, for example, increasing TR to 2.0 s to enhance signal recovery. The processed signal is output to the display in real time, ensuring that the noise suppression effect is synchronized with the acquisition.

[0047] Please see Figure 2 , Figure 2 This diagram illustrates the specific application process of the noise suppression method provided in this embodiment.

[0048] like Figure 3 As shown, this embodiment of the invention also provides a nuclear magnetic resonance signal noise suppression device 300 based on a convolutional neural network, comprising: Initial setup module 301 configures the CPMG pulse sequence in the nuclear magnetic resonance device; The echo signal acquisition module 302 is used to perform nuclear magnetic resonance testing on shale samples based on the configured CPMG pulse sequence, and to set multiple sets of echo intervals using the multiple micro-variable echo interval technology to generate a time-aligned multidimensional dataset. The signal reconstruction module 303, which is mounted on the FPGA real-time processing unit of the nuclear magnetic resonance equipment, performs parallel calculations on segmented data through a pre-trained CNN network to achieve noise component prediction and signal reconstruction. The feedback optimization module 304 is integrated into the device's ARM architecture main control chip. It runs a reinforcement learning mechanism in real time through embedded algorithms and dynamically adjusts the pulse sequence parameters and sliding window strategy according to the signal quality to ensure the continuous stability of the noise suppression effect.

[0049] In a preferred embodiment, the signal reconstruction module and the feedback optimization module are integrated into the embedded hardware system of the MRI equipment (such as a real-time processing unit based on FPGA or an ARM architecture main control chip) to achieve real-time signal denoising and parameter optimization.

[0050] Deploying these modules directly within NMR equipment enables real-time signal reconstruction and optimization, avoiding the delays of transmitting data to external systems. This makes the MRI process more efficient, particularly suitable for situations requiring rapid results. Furthermore, real-time reconstruction and optimization reduce the need for post-processing, potentially lowering equipment operating costs and time consumption, and improving equipment utilization efficiency.

[0051] like Figure 4 As shown in the above-described method for suppressing nuclear magnetic resonance signal noise based on a convolutional neural network, this invention also provides an electronic device 400, which can be a mobile terminal, desktop computer, laptop, handheld computer, server, or other computing device. The electronic device includes a processor 401, a memory 402, and a display 403.

[0052] In some embodiments, memory 402 may be an internal storage unit of a computer device, such as a hard disk or RAM. In other embodiments, memory 402 may be an external storage device of a computer device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, memory 402 may include both internal and external storage units. Memory 402 is used to store application software and various types of data installed on the computer device, such as program code for installing the computer device. Memory 402 can also be used to temporarily store data that has been output or will be output. In one embodiment, memory 402 stores a program 504 for a nuclear magnetic resonance signal noise suppression method based on a convolutional neural network. This program 504 can be executed by processor 401 to implement a nuclear magnetic resonance signal noise suppression method based on a convolutional neural network according to various embodiments of the present invention.

[0053] In some embodiments, processor 401 may be a central processing unit (CPU), microprocessor, or other data processing chip, used to run program code stored in memory 402 or process data, such as executing a program for suppressing nuclear magnetic resonance signal noise based on a convolutional neural network.

[0054] In some embodiments, display 403 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 403 is used to display information on the computer device and to display a visual user interface. Components 501-503 of the computer device communicate with each other via a system bus.

[0055] This invention provides a method and apparatus for noise suppression of nuclear magnetic resonance (NMR) signals based on convolutional neural networks (CNNs). It utilizes a configured CPMG pulse sequence to perform NMR testing on shale samples, and generates multidimensional datasets with different echo intervals to more comprehensively reflect the characteristics and state of the samples, making signal acquisition more detailed and diverse, thus enhancing data accuracy. By processing the signal through a sliding window segmentation method, it can more accurately capture signal features within different time periods, thereby improving noise suppression. The CNN network accurately predicts noise components and reconstructs the signal, ensuring the efficiency and accuracy of the denoising process and further improving the quality of the test results. Through a reinforcement learning mechanism, feedback can be provided based on the quality of the denoised signal, dynamically adjusting the pulse sequence configuration parameters, which can continuously improve signal quality through iteration, ensuring the stability of the noise suppression effect. This invention optimizes the signal acquisition and processing process, utilizing the convolutional learning capability of CNNs to reduce noise interference from fast relaxation component signals, providing technical support for the accurate characterization of complex reservoirs such as shale oil and gas.

[0056] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for suppressing noise in nuclear magnetic resonance signals based on convolutional neural networks, characterized in that, include: Configure CPMG pulse sequences in nuclear magnetic resonance equipment; Nuclear magnetic resonance tests were performed on shale samples based on the configured CPMG pulse sequence. Multiple sets of echo intervals were set using the technique of multiple micro-variable echo intervals to generate time-aligned multidimensional datasets. The multidimensional dataset is segmented using the sliding window technique, and the segmented data is processed by a pre-trained CNN network to remove noise components and perform signal reconstruction. A reinforcement learning mechanism is introduced to optimize the pulse sequence configuration parameters and sliding window segmentation strategy in reverse based on the quality of the denoised echo signal, so as to ensure the continuous stability of the noise suppression effect.

2. The nuclear magnetic resonance signal noise suppression method based on convolutional neural networks according to claim 1, characterized in that, The configuration of the CPMG pulse sequence includes adjusting the echo interval, pulse width, and repetition time parameters.

3. The nuclear magnetic resonance signal noise suppression method based on convolutional neural networks according to claim 1, characterized in that, The method of setting multiple sets of echo intervals using multiple micro-variable echo interval techniques to generate a time-aligned multidimensional dataset includes: Set multiple echo intervals to cover the decay process of fast relaxation components; The CPMG pulse sequence is run independently for each echo interval to obtain the echo data points corresponding to each echo interval. Based on the trigger timestamp, data with different echo intervals are aligned to form a three-dimensional data structure of sample × echo point × echo interval channel, which is used to describe the transverse relaxation time characteristics and noise variation of the sample.

4. The nuclear magnetic resonance signal noise suppression method based on convolutional neural networks according to claim 1, characterized in that, The segmentation of the multidimensional dataset using the sliding window technique includes: The window size and step size are dynamically adjusted based on the frequency characteristics and noise distribution of the signal.

5. The nuclear magnetic resonance signal noise suppression method based on convolutional neural networks according to claim 1, characterized in that, The pre-trained CNN network adopts a three-layer stacked convolutional architecture, with the convolutional kernel width expanding layer by layer; The first convolutional layer is used to slide along the signal time axis to extract short-term time-domain fluctuation features and achieve local noise suppression; The second convolutional layer is used to fuse the noise output from the first layer across echo interval channels, analyze the correlation of power frequency interference under different echo intervals, and generate a composite noise mode representation. The third convolutional layer is used to capture the coupled noise pattern across echo points, separating noise from the signal in the overlapping spectrum to eliminate global noise.

6. The method for suppressing nuclear magnetic resonance signal noise based on a convolutional neural network according to claim 1, characterized in that, A reinforcement learning mechanism is introduced to optimize the configuration parameters and sliding window strategy in reverse based on the quality of the denoised window signal, including: Using signal-to-noise ratio, phase stability, noise spectrum, and environmental parameters as the state space, the echo interval, sampling frequency, pulse width, and repetition time are dynamically adjusted to achieve adaptive optimization of signal acquisition parameters in response to changes in sample characteristics or environment.

7. The nuclear magnetic resonance signal noise suppression method based on convolutional neural networks according to claim 1, characterized in that, Also includes: When the sample or environment changes, the configuration parameters are adjusted based on the signal-to-noise ratio feedback to achieve synchronous adjustment of the acquisition process and noise suppression effect.

8. A nuclear magnetic resonance signal noise suppression device based on a convolutional neural network, characterized in that, include: The initial setup module configures the CPMG pulse sequence in the MRI equipment; The echo signal acquisition module is used to perform nuclear magnetic resonance testing on shale samples based on the configured CPMG pulse sequence. It uses multiple micro-variable echo interval technology to set multiple sets of echo intervals and generate time-aligned multidimensional datasets. The signal reconstruction module is used to segment the multidimensional dataset using the sliding window technique, process the segmented data through a pre-trained CNN network, predict noise components, and perform signal reconstruction. The feedback optimization module is used to introduce a reinforcement learning mechanism to optimize the pulse sequence configuration parameters and sliding window segmentation strategy in reverse based on the quality of the denoised window signal, so as to ensure the continuous stability of the noise suppression effect.

9. The nuclear magnetic resonance signal noise suppression device based on a convolutional neural network according to claim 8, characterized in that, The signal reconstruction module and feedback optimization module are integrated into the embedded hardware system of the nuclear magnetic resonance device to achieve real-time signal denoising and parameter optimization.

10. An electronic device, characterized in that, The device includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the nuclear magnetic resonance signal noise suppression method based on a convolutional neural network as described in any one of claims 1-7.

Images