Method for nmr interpretation of organic pollution based on multi-parametric semi-supervised classification

Abstract

The application relates to the technical field of geophysical exploration, in particular to an organic pollution nuclear magnetic resonance interpretation method based on a multi-parameter semi-supervised classification, which comprises the following steps: adopting an unsupervised SVM classification network to mark characteristic spectra without labels in the field; performing a semi-supervised process on the characteristic spectra marked by the unsupervised SVM classification network to obtain a classification result, and supplementing the classification result to a labeled sample set to continue training the unsupervised SVM classification network to obtain a semi-supervised SVM classification network; and adopting the semi-supervised SVM classification network to adaptively distinguish the types of pollution of characteristic spectra of field nuclear magnetic resonance responses to be classified. The problems that the types of pollutants cannot be effectively distinguished by traditional nuclear magnetic resonance technology and a large amount of labeled data is required by traditional machine learning are solved, the adaptability to complex pollution scenes is improved, and the dependence on laboratory samples is reduced.

Description

Technical Field

[0001] This application relates to the field of geophysical exploration technology, specifically to a nuclear magnetic resonance interpretation method for organic pollution based on multi-parameter semi-supervised classification. Background Technology

[0002] With the deepening of industrialization and the acceleration of urbanization, industries such as chemical, metallurgical and petroleum processing have developed rapidly. As raw materials, intermediate products or products in the production and processing process, organic matter inevitably leaks in residues during storage, processing and transportation. It seeps into groundwater, causing organic pollution of groundwater, which seriously endangers the environment and human health.

[0003] Traditional methods for detecting organic pollution include chromatography, electrochemical methods, and spectroscopic methods. While chromatography can detect extremely low concentrations of pollutants, it requires on-site sampling, making real-time, in-situ detection difficult. Electrochemical methods, although enabling real-time detection with portable instruments, struggle to distinguish specific organic pollutants, and their electrodes are prone to aging, requiring frequent calibration and maintenance. Spectroscopic methods require no sample pretreatment and typically do not damage the sample.

[0004] Existing detection methods are all point-to-point, making it difficult to achieve rapid, non-destructive screening of organic pollution over a large area. Nuclear magnetic resonance (NMR) technology, by laying hundreds-meter-scale coils on the ground, artificially excites hydrogen nuclei, causing their spin direction to deflect. After excitation stops, the resonance signals of the hydrogen nuclei are collected. Based on the different relaxation times, hydrogen-containing substances such as water and organic matter can be distinguished. This allows for non-destructive, in-situ detection of hydrogen-containing substances within a range of tens to hundreds of meters underground, directly obtaining the spatial distribution of water and organic pollutants. However, traditional NMR technology can only obtain the average relaxation time parameter, making it difficult to effectively distinguish the types of pollutants. Summary of the Invention

[0005] This patent proposes a nuclear magnetic resonance interpretation method for organic pollution based on multi-parameter semi-supervised classification, which solves the problems of insufficient effective range of single detection in traditional pollution monitoring methods, inability of traditional nuclear magnetic resonance technology to accurately identify pollutant types, and heavy dependence of traditional machine learning on the amount of labeled data.

[0006] According to an embodiment of this application, a nuclear magnetic resonance interpretation method for organic pollution based on multi-parameter semi-supervised classification is provided, the method comprising:

[0007] An unsupervised SVM classification network is used to classify unlabeled data in the field. The feature spectra are labeled, and the unsupervised SVM classification network is trained using a labeled sample set. For longitudinal relaxation time, This refers to the lateral relaxation time;

[0008] Labeling after passing through an unsupervised SVM classification network The feature spectrum is solved in a semi-supervised process to obtain the classification result. The classification result is then added to the labeled sample set to continue training the unsupervised SVM classification network, thus obtaining the semi-supervised SVM classification network.

[0009] A semi-supervised SVM classification network was used to classify the field NMR responses. The characteristic spectrum is used to adaptively distinguish the types of pollution.

[0010] Further, obtain the The process of characteristic spectrum is as follows:

[0011] The system acquires various responses generated by a measurement sequence, which includes a DC pulse, an adiabatic full-wave pulse, and an adiabatic half-wave pulse. Each response includes a nuclear magnetic resonance FID signal, an inversion recovery FID signal, and a spin echo signal.

[0012] An inversion objective function is constructed for each response generated by the measurement sequence. Solving the inversion objective function yields the hydrogen nucleus content of organic pollutants, the average relaxation time, the longitudinal relaxation time, and the transverse relaxation time. This process generates the inversion results. Characteristic spectrum.

[0013] Furthermore, the responses generated by the measurement sequence are acquired, including:

[0014] Nuclear magnetic resonance (FID) signals were acquired by sequentially transmitting DC pulses, adiabatic full-wave pulses, and adiabatic half-wave pulses; and the data were collected at intervals after excitation. Repeatedly transmit adiabatic half-wave pulses and acquire inverted recovery FID signals; then at intervals of time... , emits adiabatic full-wave pulses, and at intervals of Repeatedly emit adiabatic full-wave pulses and collect spin echo signals.

[0015] Furthermore, the inversion objective function is constructed for each response generated by the measurement sequence, including:

[0016] Construct the objective function for inverting nuclear magnetic resonance FID signals;

[0017] The objective function for inverting nuclear magnetic resonance (FID) signals is expressed as an iterative format;

[0018] Solve the objective function for the inversion of the nuclear magnetic resonance (FID) signal to obtain the spatial distribution information of the hydrogen nucleus content and average relaxation time of organic pollutants;

[0019] Using the hydrogen nucleus content and average relaxation time of organic pollutants as inputs, an inversion objective function for the reverse recovery FID signal is established, and the spatial distribution of longitudinal relaxation time is obtained by solving the problem.

[0020] Using the longitudinal relaxation time as input, an inversion objective function for the spin echo signal is established and converted into an iterative scheme for solution, yielding the spatial distribution of the transverse relaxation time.

[0021] Combined organic pollution hydrogen nucleus content, longitudinal relaxation time, and transverse relaxation time generation Characteristic spectrum.

[0022] Furthermore, the objective function for inverting the nuclear magnetic resonance FID signal is:

[0023] ,

[0024] in, The hydrogen nucleus content of organic pollutants, The average relaxation time, This is an FID signal from nuclear magnetic resonance imaging. For adiabatic half-wave pulse excitation pulse moment, Indicates different locations underground. For regularization parameters, The smoothness matrix, For the FID response sensitivity kernel function of organic pollution detection, The objective function for inverting nuclear magnetic resonance FID signals;

[0025] The iterative form of the objective function for inverting nuclear magnetic resonance (FID) signals is:

[0026] ,

[0027] in, Let Jacobian matrix be the Jacobian matrix of the sensitivity kernel function corresponding to the nuclear magnetic resonance FID signal to be solved. Represents matrix transpose. This represents the current iteration number. It is the identity matrix. To update the step size, The simulated FID signal response is updated iteratively.

[0028] Furthermore, an inversion objective function for recovering the inverted FID signal is established, and the longitudinal relaxation time is obtained by solving for it. The inversion objective function for recovering the inverted FID signal is as follows:

[0029] ,

[0030] in, For longitudinal relaxation time, It is an inverted recovery of the FID signal. The interval time, To determine the sensitivity kernel function for the reverse recovery FID response to organic pollution detection, the longitudinal relaxation time is iteratively updated, and the sensitivity kernel function for the reverse recovery FID response to organic pollution detection is recalculated. The spatial distribution of the longitudinal relaxation time is successively approximated and finally solved. The objective function for inverting and recovering the FID signal is denoted as .

[0031] Furthermore, the objective function for inverting the spin echo signal is established:

[0032] ,

[0033] in, For the lateral relaxation time, It is a spin echo signal. The excitation pulse distance for the adiabatic full-wave, A kernel function for detecting spin echo response sensitivity of organic pollutants. The objective function for inverting the spin echo signal is denoted as .

[0034] Furthermore, the labeled sample set includes typical laboratory contaminated samples. Characteristic spectra and corresponding category labels and known contaminated samples from field sites Feature spectrum and corresponding category labels.

[0035] Furthermore, the marked - The characteristic spectrum is solved through a semi-supervised process, including:

[0036] The predicted pollution category label is obtained by solving an optimization problem through a semi-supervised process, where the optimization problem is:

[0037] ,

[0038] in, Let be the normal vector of the hyperplane. It is the hyperplane bias term. For the first One slack variable, , and All are labeled samples Feature spectrum and unlabeled samples Compromise parameters between characteristic spectra This represents the number of uncontaminated samples from field sites. To label the number of samples, To predict pollution category labels.

[0039] The beneficial effects of this application are as follows: This application combines multi-parameter inversion of organic pollution by nuclear magnetic resonance with machine learning classification, which can achieve adaptive differentiation of organic pollution components over a wide range without damage. Furthermore, it expands the training set with unlabeled samples of unknown pollution in the field, thereby improving the generalization ability of the model. This solves the problems that traditional nuclear magnetic resonance technology is difficult to effectively distinguish pollutant types and that traditional machine learning requires a large amount of labeled data, thus improving the adaptability to complex pollution scenarios and reducing the dependence on laboratory samples. Attached Figure Description

[0040] Figure 1 This is an overall flowchart of the method provided in the embodiments of this application;

[0041] Figure 2 A flowchart of a method for training a semi-supervised SVM classification network provided in an embodiment of this application; Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0043] See Figure 1 Combination Figure 2 A method for interpreting organic pollution using nuclear magnetic resonance based on multi-parameter semi-supervised classification includes the following steps:

[0044] S1 uses an unsupervised SVM classification network to classify unlabeled data in the field. The feature spectra are labeled, and the unsupervised SVM classification network is trained using a labeled sample set. For longitudinal relaxation time, This refers to the lateral relaxation time;

[0045] S2 will be labeled by the unsupervised SVM classification network. The feature spectrum is solved in a semi-supervised process to obtain the classification result. The classification result is then added to the labeled sample set to continue training the unsupervised SVM classification network, thus obtaining the semi-supervised SVM classification network.

[0046] S3 uses a semi-supervised SVM classification network to classify the field nuclear magnetic resonance responses. The characteristic spectrum is used to adaptively distinguish the types of pollution.

[0047] SVM stands for Support Vector Machine. An unsupervised SVM classification network is trained using a labeled sample set to obtain a model. The training process includes initializing the SVM parameters and employing the Support Vector Machine algorithm. Figure 2After processing the labeled sample set (represented by SVM in Chinese), the SVM parameters are updated. The training ends either by determining whether the error condition is met and outputting the trained model, or by reprocessing the labeled sample set using the Support Vector Machine algorithm. The output trained model is then used... express.

[0048] To ensure the quality of training samples under different response intensities and background conditions Feature spectrum training consistency requires consistency with the input. Characteristic spectrum normalization processing.

[0049] The sample set used for training includes the source domain dataset and the target domain dataset.

[0050] Source domain dataset Defined as:

[0051] ,

[0052] in, This is a characteristic spectrum of typical contaminated samples from the laboratory. This represents the number of typical contaminated samples in the laboratory. For source domain multi-pollutant component category labels, the definition is:

[0053] ,

[0054] The target domain dataset is Known contaminated samples from field sites that have undergone indoor shimmed nuclear magnetic resonance analysis Feature spectrum, as a labeled sample set :

[0055] ,

[0056] in, For known contaminated samples from field sites Characteristic spectrum The number of known contaminated samples at the field site. For labeled sample sets with multiple contaminant category labels, the definition is: .

[0057] For labeled sample sets and hyperplane ,in, It is a normal vector. It is the bias term, the data sample points. geometric interval for:

[0058] ,

[0059] The distance from the support vectors to the hyperplane is defined as:

[0060] ,

[0061] make Setting it to 1 transforms the problem of finding the maximum separating hyperplane into an optimization problem:

[0062] ,

[0063] ,

[0064] Introducing the Lagrange operator, constructing the Lagrange function :

[0065] ,

[0066] in, Let Lagrange operator vectors be used. For the first A Lagrange operator, For the first A Lagrange operator is obtained. The optimal Lagrange operator is obtained by solving the dual problem, and finally the optimal separating hyperplane and classification decision function are solved, completing the initial training of the unsupervised SVM classification network.

[0067] An unsupervised SVM classification network was used to classify unlabeled samples in the field. Label-free samples composed of characteristic spectra are labeled for normalization with more unknown contamination samples in the field. Feature spectra are used as a set of unlabeled samples to construct a semi-supervised learning framework to improve the training of the classification network. To address the problem of extremely limited known contaminated sample data and insufficient sample quantity, an unsupervised SVM classification network is employed to classify unlabeled samples from the field. After labeling the unlabeled samples composed of feature spectra, if the error condition is met, they are used as labeled sample sets. It is then determined whether the classification accuracy is met. If it is, the classification result is output and placed into the labeled sample set. If it is not, it is returned to the unlabeled sample set.

[0068] Label-free sample set obtained by inversion of unknown pollution data in the wild Represented as:

[0069] ,

[0070] in, For samples of unknown contamination in the field Characteristic spectrum This represents the number of uncontaminated samples from field sites.

[0071] The unsupervised SVM classification network after initial training is used to classify SVM. l Label-free sample set obtained by inversion of unknown pollution data in the field Label the samples and define , A compromise parameter between labeled and unlabeled samples; slack variables .set up The optimization problem is then solved using a semi-supervised process:

[0072] ,

[0073] in, Let be the normal vector of the hyperplane. It is the hyperplane bias term. For the first One slack variable,

[0074] , and All are labeled samples Feature spectrum and unlabeled samples Compromise parameters between characteristic spectra This represents the number of uncontaminated samples from field sites. To indicate the number of labeled samples;

[0075] This will give you the predicted pollution category labels for the unlabeled field pollution spectrum:

[0076]

[0077] in, To predict pollution category labels.

[0078] In one embodiment, unknown contamination samples from field sites The process of obtaining the feature spectrum is as follows:

[0079] The system acquires various responses generated by the measurement sequence, which includes DC pulses, adiabatic full-wave pulses, and adiabatic half-wave pulses, and each response includes nuclear magnetic resonance FID signals, inversion recovery FID signals (R-FID), and spin echo signals (SE).

[0080] An inversion objective function is constructed for each response generated by the measurement sequence. The inversion objective function is solved to obtain the hydrogen nucleus content of organic pollutants. Mean relaxation time Longitudinal relaxation time and lateral relaxation time Inversion generation Characteristic spectrum.

[0081] Acquire the responses generated by the measurement sequence, including:

[0082] Nuclear magnetic resonance (FID) signals were acquired by sequentially transmitting DC pulses, adiabatic full-wave pulses, and adiabatic half-wave pulses; and after excitation... At a certain time interval, the adiabatic half-wave pulse is repeatedly transmitted, and the inverted recovery FID signal is acquired; then at another time interval... , emits adiabatic full-wave pulses, and at intervals of Repeatedly emit adiabatic full-wave pulses and collect spin echo signals.

[0083] To distinguish the content and composition of underground organic pollution, inversion objective functions were constructed by combining the corresponding expressions of each response generated from the measurement sequence.

[0084] Construct the objective function for inverting nuclear magnetic resonance FID signals;

[0085] The objective function for inverting nuclear magnetic resonance (FID) signals is expressed as an iterative format;

[0086] Solve the objective function for the inversion of the nuclear magnetic resonance (FID) signal to obtain the spatial distribution information of the hydrogen nucleus content and average relaxation time of organic pollutants;

[0087] Using the hydrogen nucleus content and average relaxation time of organic pollutants as inputs, an inversion objective function for the reverse recovery FID signal is established, and the spatial distribution of longitudinal relaxation time is obtained by solving the problem.

[0088] Using the longitudinal relaxation time as input, an inversion objective function for the spin echo signal is established and converted into an iterative scheme for solution, yielding the spatial distribution of the transverse relaxation time.

[0089] Combined organic pollution hydrogen nucleus content, longitudinal relaxation time, and transverse relaxation time generation Characteristic spectrum.

[0090] Specifically, in one embodiment, the objective function for inverting the nuclear magnetic resonance FID signal is:

[0091] ,

[0092] in, The hydrogen nucleus content of organic pollutants, The average relaxation time, The acquired nuclear magnetic resonance FID signal, For adiabatic half-wave pulse excitation pulse moment, Indicates different locations underground. For regularization parameters, The smoothness matrix, For the FID response sensitivity kernel function of organic pollution detection, The objective function for inverting nuclear magnetic resonance FID signals;

[0093] The iterative form of the objective function for inverting nuclear magnetic resonance (FID) signals is:

[0094] ,

[0095] in, Let Jacobian matrix be the Jacobian matrix of the sensitivity kernel function corresponding to the nuclear magnetic resonance FID signal to be solved. Represents matrix transpose. This represents the current iteration number. It is the identity matrix. To update the step size, The simulated FID signal response is updated iteratively.

[0096] Establish the inversion objective function for recovering the FID signal and solve for the longitudinal relaxation time, including:

[0097] ,

[0098] in, For longitudinal relaxation time, It is the acquired inverted recovery FID signal. The interval time, To determine the sensitivity kernel function for the reverse recovery FID response to organic pollution detection, the longitudinal relaxation time is iteratively updated, and the sensitivity kernel function for the reverse recovery FID response to organic pollution detection is recalculated. The spatial distribution of the longitudinal relaxation time is obtained by successively approximating and finally solving it.

[0099] Establish the objective function for inverting the spin echo signal:

[0100]

[0101] in, For the lateral relaxation time, It is the acquired spin echo signal. The excitation pulse distance for the adiabatic full-wave, A kernel function for the spin echo response sensitivity of organic pollution detection is derived. This function is then transformed into an iterative solution to obtain the transverse relaxation time.

[0102] By combining FID, R-FID, and SE information on the hydrogen nucleus content, longitudinal relaxation time, and transverse relaxation time distribution of organic pollutants, a gene is generated. Characteristic spectrum, i.e., the hydrogen nucleus content of organic pollutants Different relaxation times Distribution on.

[0103] By analyzing unmarked samples in the field The characteristic spectra are labeled for normalization of more unknown contamination samples in the field. Feature spectrum, as an unlabeled sample set, is used to construct a semi-supervised learning framework to improve the training of the classification network, thus solving the problem of extremely limited known contaminated sample data and insufficient samples.

[0104] For field NMR responses that need to be classified, calculate The feature spectrum is used to achieve adaptive differentiation of pollution types by employing a pre-trained semi-supervised SVM classification network.

[0105] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A nuclear magnetic resonance interpretation method for organic pollution based on multi-parameter semi-supervised classification, characterized in that, The method includes: An unsupervised SVM classification network is used to classify unlabeled data in the field. The feature spectra are labeled, and the unsupervised SVM classification network is trained using a labeled sample set. For longitudinal relaxation time, This refers to the lateral relaxation time; Obtain the The process of characteristic spectrum analysis is as follows: The measurement sequence is acquired, including DC pulses, adiabatic full-wave pulses, and adiabatic half-wave pulses. Each response includes nuclear magnetic resonance (NMR) FID signals, inversion recovery (FID) signals, and spin echo signals. An inversion objective function is constructed from each response generated by the measurement sequence. Solving the inversion objective function yields the hydrogen nucleus content of organic pollutants, the average relaxation time, the longitudinal relaxation time, and the transverse relaxation time. This process generates the characteristic spectrum. Characteristic spectrum; Labeling after passing through an unsupervised SVM classification network The feature spectrum is solved in a semi-supervised process to obtain the classification result. The classification result is then added to the labeled sample set to continue training the unsupervised SVM classification network, thus obtaining the semi-supervised SVM classification network. A semi-supervised SVM classification network was used to classify the field NMR responses. The characteristic spectrum is used to adaptively distinguish the types of pollution.

2. The method for interpreting organic pollution using nuclear magnetic resonance based on multi-parameter semi-supervised classification according to claim 1, characterized in that, Acquire the responses generated by the measurement sequence, including: Nuclear magnetic resonance (FID) signals were acquired by sequentially transmitting DC pulses, adiabatic full-wave pulses, and adiabatic half-wave pulses; and the data were collected at intervals after excitation. Repeatedly transmit adiabatic half-wave pulses and acquire inverted recovery FID signals; then at intervals of time... , emits adiabatic full-wave pulses, and at intervals of Repeatedly emit adiabatic full-wave pulses and collect spin echo signals.

3. The method for interpreting organic pollution based on multi-parameter semi-supervised classification according to claim 1, characterized in that, The inversion objective function is constructed for each response generated by the measurement sequence, including: Construct the objective function for inverting nuclear magnetic resonance FID signals; The objective function for inverting nuclear magnetic resonance (FID) signals is expressed as an iterative format; Solve the objective function for the inversion of the nuclear magnetic resonance (FID) signal to obtain the spatial distribution information of the hydrogen nucleus content and average relaxation time of organic pollutants; Using the hydrogen nucleus content and average relaxation time of organic pollutants as inputs, an inversion objective function for the reverse recovery FID signal is established, and the spatial distribution of longitudinal relaxation time is obtained by solving the problem. Using the longitudinal relaxation time as input, an inversion objective function for the spin echo signal is established and converted into an iterative scheme for solution, yielding the spatial distribution of the transverse relaxation time. Combined organic pollution hydrogen nucleus content, longitudinal relaxation time, and transverse relaxation time generation Characteristic spectrum.

4. The method for interpreting organic pollution using nuclear magnetic resonance based on multi-parameter semi-supervised classification according to claim 3, characterized in that, The objective function for inverting nuclear magnetic resonance (NMR) FID signals is: ， in, The hydrogen nucleus content of organic pollutants, The average relaxation time, This is an FID signal from nuclear magnetic resonance imaging. For adiabatic half-wave pulse excitation pulse moment, Indicates different locations underground. For regularization parameters, The smoothness matrix, For the FID response sensitivity kernel function of organic pollution detection, The objective function for inverting nuclear magnetic resonance FID signals; The iterative form of the objective function for inverting nuclear magnetic resonance (FID) signals is: ， in, Let Jacobian matrix be the Jacobian matrix of the sensitivity kernel function corresponding to the nuclear magnetic resonance FID signal to be solved. Represents matrix transpose. This represents the current iteration number. It is the identity matrix. To update the step size, The simulated FID signal response is updated iteratively.

5. The method for interpreting organic pollution using nuclear magnetic resonance based on multi-parameter semi-supervised classification according to claim 4, characterized in that, An inversion objective function for recovering the inverted FID signal is established, and the longitudinal relaxation time is obtained by solving for it. The inversion objective function for recovering the inverted FID signal is as follows: ， in, For longitudinal relaxation time, It is an inverted recovery of the FID signal. The interval time, To determine the sensitivity kernel function for the reverse recovery FID response to organic pollution detection, the longitudinal relaxation time is iteratively updated, and the sensitivity kernel function for the reverse recovery FID response to organic pollution detection is recalculated. The spatial distribution of the longitudinal relaxation time is successively approximated and finally solved. The objective function for inverting and recovering the FID signal is denoted as .

6. The method for interpreting organic pollution using nuclear magnetic resonance based on multi-parameter semi-supervised classification according to claim 5, characterized in that, Establish the objective function for inverting the spin echo signal: ， in, For the lateral relaxation time, It is a spin echo signal. The excitation pulse distance for the adiabatic full-wave, A kernel function for detecting spin echo response sensitivity of organic pollutants. The objective function for inverting the spin echo signal is denoted as .

7. The method for interpreting organic pollution using nuclear magnetic resonance based on multi-parameter semi-supervised classification according to claim 1, characterized in that, The labeled sample set includes typical contaminated samples from the laboratory. Characteristic spectra and corresponding category labels and known contaminated samples from field sites Feature spectrum and corresponding category labels.

8. The method for interpreting organic pollution using nuclear magnetic resonance based on multi-parameter semi-supervised classification according to claim 1, characterized in that, Marked - The characteristic spectrum is solved through a semi-supervised process, including: The predicted pollution category label is obtained by solving an optimization problem through a semi-supervised process, where the optimization problem is: ， in, Let be the normal vector of the hyperplane. It is the hyperplane bias term. For the first One slack variable, , and All are labeled samples Feature spectrum and unlabeled samples Compromise parameters between characteristic spectra This represents the number of uncontaminated samples from field sites. To label the number of samples, To predict pollution category labels.

Images