阅读说明：本技术 一种基于1D-Unet的低场核磁共振仪器信号校正方法 (Low-field nuclear magnetic resonance instrument signal correction method based on 1D-Unet ) 是由 聂生东 王欣 侯学文 王广利 于 2020-07-06 设计创作，主要内容包括：本发明提供了一种基于1D-Unet的低场核磁共振仪器信号校正方法,包括以下步骤：步骤1,将两台相同型号的低场核磁共振仪器分为目标仪器和待校正仪器,将目标仪器作为主机,待校正仪器作为从机,而后利用主机采集样本的多条第一CPMG原始信号,并利用从机采集样本的多条第二CPMG原始信号；步骤2,对每条第一CPMG原始信号和每条第二CPMG原始信号分别进行归一化预处理操作,得到主机归一化数据和从机归一化数据,并保存主机归一化数据的预处理数据结构体；步骤3,构建1D-Unet网络并通过1D-Unet网络校正从机数据,得到校正后数据；步骤4,通过预处理数据结构体对校正后数据进行反归一化处理,得到与主机数据在同一量纲下的标准数据,完成主机和从机之间的信号校正。(The invention provides a low-field nuclear magnetic resonance instrument signal correction method based on 1D-Unet, which comprises the following steps: step 1, dividing two low-field nuclear magnetic resonance instruments with the same model into a target instrument and an instrument to be corrected, taking the target instrument as a host and the instrument to be corrected as a slave, then acquiring a plurality of first CPMG original signals of a sample by using the host, and acquiring a plurality of second CPMG original signals of the sample by using the slave; step 2, respectively carrying out normalization preprocessing operation on each first CPMG original signal and each second CPMG original signal to obtain host normalization data and slave normalization data, and storing a preprocessing data structure body of the host normalization data; step 3, constructing a 1D-Unet network and correcting slave data through the 1D-Unet network to obtain corrected data; and 4, performing inverse normalization processing on the corrected data through the preprocessed data structure body to obtain standard data in the same dimension as the host data, and completing signal correction between the host and the slave.)

1. A low-field nuclear magnetic resonance instrument signal correction method based on 1D-Unet is characterized by comprising the following steps:

step 1, dividing two low-field nuclear magnetic resonance instruments with the same model into a target instrument and an instrument to be corrected, and taking the target instrument as a host M_oSaid apparatus to be calibrated acting as a slave M_cAnd then using said host M_oCollecting a plurality of first CPMG original signals of a sample, and utilizing the slave M_cCollecting a plurality of second CPMG original signals of the sample;

step 2, respectively carrying out normalization preprocessing operation on each first CPMG original signal and each second CPMG original signal to obtain host normalization data M'_oAnd slave normalized data M'_cAnd storing the host normalized data M'_oIs used to preprocess the data structure S_o；

Step 3, constructing a 1D-Unet network and correcting the slave machine normalized data M 'through the 1D-Unet network'_cCorrected data M 'are obtained'_co；

Step 4, preprocessing the data structure S through the data structure_oTo the corrected data M'_coPerforming inverse normalization processing to obtain data M of the host_oStandard data M under the same dimension_coCompleting the host M_oAnd said slave M_cThe signal correction in between.

2. The signal correction method for 1D-Unet-based low-field NMR instrument according to claim 1, characterized in that:

in step 2, the normalization preprocessing operation is performed on each CPMG original signal, and a data normalization formula is as follows:

in the formula (1), M_i,maxAnd M_i,minRespectively representing the maximum value and the minimum value in the ith CPMG original signal.

3. The signal correction method for 1D-Unet-based low-field NMR instrument according to claim 1, characterized in that:

wherein the preprocessed data structure S_oThe maximum value and the minimum value of n first CPMG original signals are saved and are arranged according to the sample acquisition sequence.

4. The signal correction method for 1D-Unet-based low-field NMR instrument according to claim 1, characterized in that:

wherein, in the step 3, the coding layer of the 1D-Unet network is constructed by Convolition 1D and Maxpooling1D, and is used for completing the extraction of key features of sample data,

the decoding layer of the 1D-Unet network is constructed by Usamling 1D and Copy, and is used for fusing down-sampling features and amplifying the dimensionality of feature signals.

5. The signal correction method for 1D-Unet-based low-field NMR instrument according to claim 1, characterized in that:

wherein, in the step 3, the slave machine normalization data M'_cAs an input layer of the 1D-Unet network, normalizing the host normalized data M'_oAs an output layer of the 1D-Unet network, and the data of the input layer and the data of the output layer correspond to each other one by one according to the sample acquisition sequence,

the loss function of the 1D-Unet network is a relative residual equation, such as equation (2),

in the formula (2), M'_o,iAnd M'_co,iRespectively represent the ith piece of host normalized data M'_oAnd the ith piece of corrected data M'_coAnd | | · | | represents the norm of L2, and the average value of the cumulative sum of the relative residuals of all samples is the loss rate of the 1D-Unet network.

Technical Field

The invention belongs to the technical field of deep learning and nuclear magnetic resonance signal processing, and particularly relates to a low-field nuclear magnetic resonance instrument signal correction method based on 1D-Unet.

Background

The low-field nuclear magnetic resonance technology is a new technology which is rapidly developed in recent years and is used for observing and analyzing physical parameters of samples, which marks that nuclear magnetic resonance is moved from high-end molecular chemical structure research and medical examination services to wider industrial and agricultural fields, such as food industry, agriculture, mining industry, chemical industry and the like, and plays an increasingly important role in application and scientific research of related fields. The low-field nuclear magnetic resonance technology mainly relates to hydrogen nuclei in oil, water and high molecular polymers: after acquiring the resonance signal generated by the hydrogen nuclei of the sample, the sample characteristics or distribution are analyzed and measured through signal processing, so that the research and analysis of the carrier or environment (such as pore size) are indirectly carried out. The technology can deeply observe and analyze the internal information of the substance without damaging the sample, and has the characteristics of rapidness, accuracy, one machine with multiple parameters, no radiation, no environmental pollution and the like.

The low-field nuclear magnetic resonance signal is weak, the signal-to-noise ratio is low, and the distribution of the static magnetic field and the radio frequency field directly determines the magnitude of the nuclear magnetic resonance signal, so the design of a magnet and a radio frequency coil is the core of a low-field nuclear magnetic resonance instrument. In practical application, the difference of the field intensity of the magnet and the difference of the winding of the radio frequency coil also causes the difference of the amplitude of the nuclear magnetic resonance signal received by the signal receiver, so that the experimental result obtained by using a low-field nuclear magnetic resonance instrument of a certain type may not completely reproduce the same experiment in another type of instrument with the same type. Similarly, such signal differences occur in near infrared devices, and therefore, algorithms such as Direct normalization (DS), piecewise Direct normalization (PDS), and Orthogonal Signal Correction (OSC) are proposed to solve the problem of signal correction between devices of the same model. However, since the corrected signal obtained by the above method is dimensionless after the data preprocessing (normalization or normalization), a method capable of correcting the low-field nuclear magnetic resonance signal and restoring the corrected signal to the original dimension is required.

Disclosure of Invention

The present invention is made to solve the above problems, and an object of the present invention is to provide a method for correcting a signal of a low-field nuclear magnetic resonance instrument based on 1D-Unet.

The invention provides a low-field nuclear magnetic resonance instrument signal correction method based on 1D-Unet, which is characterized by comprising the following steps: step 1, dividing two low-field nuclear magnetic resonance instruments with the same model into a target instrument and an instrument to be corrected, and taking the target instrument as a host M_oWith the apparatus to be calibrated as slave M_cAnd then using the host M_oCollecting a plurality of first CPMG original signals of the sample and utilizing the slave M_cCollecting a plurality of second CPMG original signals of the sample;

step 2, respectively carrying out normalization preprocessing operation on each first CPMG original signal and each second CPMG original signal to obtain host normalization data M'_oAnd slave normalized data M'_cAnd storing host normalized data M'_oIs used to preprocess the data structure S_o；

Step 3, constructing a 1D-Unet network and correcting the normalized data M 'of the slave machine through the 1D-Unet network'_cCorrected data M 'are obtained'_co；

Step 4, preprocessing the data structure S_oTo corrected data M'_coPerforming inverse normalization processing to obtain host data M_oStandard data M under the same dimension_coTo complete the host M_oAnd a slave M_cThe signal correction in between.

The signal correction method of the 1D-Unet-based low-field nuclear magnetic resonance instrument can also have the following characteristics: in step 2, normalization preprocessing operation is performed on each CPMG original signal, and a data normalization formula is as follows:

in the formula (1), M_i,maxAnd M_i,minRespectively representing the maximum value and the minimum value in the ith CPMG original signal.

In the inventionThe signal correction method for the 1D-Unet-based low-field nuclear magnetic resonance instrument can also have the following characteristics: wherein the data structure S is preprocessed_oThe maximum value and the minimum value of n first CPMG original signals are stored in the CPMG and are arranged according to the sample acquisition sequence.

The signal correction method of the 1D-Unet-based low-field nuclear magnetic resonance instrument can also have the following characteristics: in step 3, an encoding layer of the 1D-Unet network is constructed through Convolition 1D and Maxpooling1D and used for completing extraction of key features of sample data, and a decoding layer of the 1D-Unet network is constructed through Usamling 1D and Copy and used for fusing down-sampling features and amplifying dimensions of feature signals.

The signal correction method of the 1D-Unet-based low-field nuclear magnetic resonance instrument can also have the following characteristics: wherein, in step 3, the slave machine is normalized to data M'_cAs an input layer of the 1D-Unet network, host normalized data M'_oAs the output layer of the 1D-Unet network, the data of the input layer and the data of the output layer are in one-to-one correspondence according to the sample acquisition sequence, the loss function of the 1D-Unet network is a relative residual error formula, such as formula (2),

in the formula (2), M'_o,iAnd M'_co,iRespectively represent ith host normalized data M'_oAnd the ith piece of corrected data M'_coAnd | | · | | represents the norm of L2, and the average value of the cumulative sum of all samples relative to the residual errors is the loss rate of the 1D-Unet network.

Action and Effect of the invention

According to the 1D-Unet-based low-field nuclear magnetic resonance instrument signal correction method, the corrected data is obtained by processing the slave machine data through constructing the 1D-Unet network, and feature screening is not needed to be carried out on the data in the processing process, so that all data features can be directly used for signal correction, and real data distribution can be reflected; and standard data obtained by performing inverse normalization processing on the corrected data of the slave computer and the host computer data are in the same dimension, so that the processing method of the host computer data can be completely reproduced. Therefore, the 1D-Unet-based low-field nuclear magnetic resonance instrument signal correction method is high in calculation accuracy, fast in time, good in robustness and capable of performing reliable and stable signal correction.

Drawings

FIG. 1 is a flow chart of a 1D-Unet based low-field NMR instrument signal correction method in an embodiment of the invention;

FIG. 2 is a network framework diagram of a 1D-Unet network in an embodiment of the invention;

FIG. 3 is CuSO in an embodiment of the present invention₄A comparison curve graph of host data, slave data of the aqueous solution, slave data corrected by a low-field nuclear magnetic resonance instrument signal correction method based on 1D-Unet and slave data corrected by a DS algorithm;

FIG. 4 is an enlarged view of a portion of the curve of FIG. 3;

FIG. 5 is a graph comparing host data, slave data corrected by a signal correction method of a 1D-Unet-based low-field NMR instrument, and slave data corrected by a DS algorithm of edible oil according to an embodiment of the present invention;

fig. 6 is an enlarged view of a portion of the curve of fig. 5.

Detailed Description

In order to make the technical means and functions of the present invention easy to understand, the present invention is specifically described below with reference to the embodiments and the accompanying drawings.

FIG. 1 is a flow chart of a 1D-Unet based low-field NMR instrument signal correction method in an embodiment of the invention.

As shown in fig. 1, the method for correcting a signal of a low-field nmr instrument based on 1D-Unet of the present embodiment includes the following steps:

step 1, dividing two low-field nuclear magnetic resonance instruments with the same model into a target instrument and an instrument to be corrected, and taking the target instrument as a host M_oWith the apparatus to be calibrated as slave M_cAnd then using the host M_oCollecting a sampleAnd a plurality of first CPMG original signals, and utilizes the slave M_cA plurality of second CPMG raw signals of the sample are collected.

Step 2, respectively carrying out normalization preprocessing operation on each first CPMG original signal and each second CPMG original signal to obtain host normalization data M'_oAnd slave normalized data M'_cAnd storing host normalized data M'_oIs used to preprocess the data structure S_o。

In step 2, each CPMG original signal is subjected to normalization preprocessing operation, and a data normalization formula is as follows:

in the formula (1), M_i,maxAnd M_i,minRespectively representing the maximum value and the minimum value in the ith CPMG original signal.

Preprocessing a data structure S_oThe maximum value and the minimum value of n first CPMG original signals are stored in the CPMG and are arranged according to the sample acquisition sequence.

Step 3, constructing a 1D-Unet network and correcting the normalized data M 'of the slave machine through the 1D-Unet network'_cCorrected data M 'are obtained'_co。

Fig. 2 is a network framework diagram of a 1D-Unet network in an embodiment of the invention.

As shown in fig. 2, in step 3, the coding layer of the 1D-Unet network is constructed by constraint 1D and maxporoling 1D to complete extraction of key features of sample data, and the decoding layer of the 1D-Unet network is constructed by upsampling 1D and Copy to fuse downsampled features and amplify dimensions of feature signals.

In this embodiment, the constraint 1D layer is used to complete the extraction of key features of sample data, the size of the Convolution kernel is 1 × 3, the filling mode is same as same,

the Maxpooling layer is used for reducing the characteristic number of the convolutional layer and further reducing the operation parameter so as to accelerate the calculation speed, the Maxpooling layer has 5 pooling layers, the size of the Maxpooling window is 2, 5 and 5 respectively,

the Copy layer is used for cascading the shallow characteristic and the deep characteristic, so that the maximum stratification of the corrected signal reflects the original signal information,

the Upsampling layer is used for completing deconvolution of signals through the Upsampling layer, so that the signals are gradually decoded and amplified to input dimension, and the window sizes of the Upsampling layer are respectively 5, 2 and 2.

In step 3, normalizing the slave machine data M'_cAs an input layer of the 1D-Unet network, host normalized data M'_oAs an output layer of the 1D-Unet network, and the data of the input layer and the data of the output layer correspond to each other one by one according to the sample acquisition sequence,

the loss function of the 1D-Unet network is a relative residual equation, such as equation (2),

in the formula (2), M'_o,iAnd M'_co,iRespectively represent ith host normalized data M'_oAnd the ith piece of corrected data M'_coAnd | | · | | represents the norm of L2, and the average value of the cumulative sum of all samples relative to the residual errors is the loss rate of the 1D-Unet network.

Step 4, preprocessing the data structure S_oTo corrected data M'_coPerforming reverse normalization to obtain data M 'of the host computer'_oStandard data M under the same dimension_coTo complete the host M_oAnd a slave M_cThe signal correction in between.

CuSO is also selected in this embodiment₄The aqueous solution and the edible oil are subjected to simulation experiments and compared with a DS algorithm, and the experimental process is as follows:

preparing 6 CuSO with different concentrations₄Aqueous solution (2mmol/L,3mmol/L,4mmol/L,5mmol/L,6mmol/L,7mmol/L), host M_oAnd the slave M_cRespectively collecting CPMG original signals under each concentration, collecting 50 times for each concentration to obtain 300 samples, wherein 210 training sets and 90 test sets adopt the corrected signals of the test sets to evaluate the results,

host M_oAnd the slave M_cThe CPMG original signals of 52 brands of edible oil are respectively collected, 468 samples are totally collected, 364 training sets and 104 testing sets are adopted, and the results are evaluated by adopting the signals corrected by the testing sets.

FIG. 3 is CuSO in an embodiment of the present invention₄Fig. 4 is a partial curve enlarged view of fig. 3, fig. 5 is a comparative curve view of the master data, the slave data corrected by the low-field nmr signal correction method based on 1D-Unet, and the slave data corrected by the DS algorithm of the edible oil according to the embodiment of the present invention, and fig. 6 is a partial curve enlarged view of fig. 5.

As shown in FIGS. 3-6, CuSO₄The host data and the slave data of the aqueous solution and the edible oil have obvious signal amplitude difference, the slave data is almost negligible in difference with the host data after being corrected by a DS algorithm and a 1D-Unet-based low-field nuclear magnetic resonance instrument signal correction method, but the comparison analysis shows that the signal corrected by the 1D-Unet-based low-field nuclear magnetic resonance instrument signal correction method is almost coincident with the original signal of the host, and the signal corrected by the DS algorithm still has tiny amplitude difference with the original signal of the host.

Table 1 is CuSO₄The relative residual values before and after the correction of the aqueous solution signal are shown in Table 2.

TABLE 1 CuSO₄Relative residual value before and after aqueous solution signal correction

	Before correction	DS algorithm correction	Correction of the method of the invention
				Relative residual error	0.0132	0.0043	0.0019

TABLE 2 relative residual values before and after correction of edible oil signals

	Before correction	DS algorithm correction	Correction of the method of the invention
				Relative residual error	0.0212	0.0212	0.0051

As shown in tables 1 and 2, the relative residual value between the slave data and the master data after being corrected by the 1D-Unet-based low-field nuclear magnetic resonance instrument signal correction method of the present invention is the smallest, so that the 1D-Unet-based low-field nuclear magnetic resonance instrument signal correction method of the present invention has a good correction capability in the low-field nuclear magnetic resonance instrument signal correction.