A method for processing and identifying radionuclide radiation signals of a pulse element logging tool

Abstract

The present invention discloses a method for processing and identifying radionuclide radiation signals of a pulsed element logging instrument. The method comprises the following steps: radioactive nuclides generate gamma photons through a series of processes, which are absorbed by a detector and finally converted into electrical signals to obtain gamma ray signals; amplification and filtering are then performed on the gamma ray signals, and the processed waveforms are sampled and digitized using an ADC. The digitized waveforms are then sent to a digital signal processor for peak search and other operations, thereby obtaining charge information of the signal to be measured. The charge information is then statistically analyzed to form a radionuclide energy spectrum, and finally the element type in the gamma energy spectrum is identified using a multi-resolution wavelet convolutional neural network model.

Description

Technical Field

[0001] The present invention belongs to the technical field of nuclide signal identification, and more particularly, relates to a method for processing and identifying nuclide radiation signals of a pulse element logging tool. Background Art

[0002] In the field of oil and gas extraction, detecting formation information in areas where oil and gas resources are to be developed is essential to ensure production safety and maximize economic benefits. This has led to the development of numerous logging methods, such as spontaneous potential logging, gamma logging, and imaging logging. Among these, controlled neutron and X-ray sources have gradually gained attention with advances in nuclear instrumentation and industrial development, and have been widely applied in the logging field, providing a key technical means for the exploration and development of oil, gas, and other mineral resources. Controlled neutron source logging technology determines formation porosity, density, oil and gas saturation, and elemental content by measuring the neutrons and gamma rays produced by the interaction of neutrons with formation materials. Corresponding logging techniques include controlled neutron porosity logging, neutron gamma density logging, pulsed neutron oil and gas saturation logging, and controlled neutron formation element logging.

[0003] Rocks are composed of various minerals and fluids. Controlled neutron elemental logging can measure the elemental content of the surrounding formation, thereby revealing properties such as mineralogy, lithology, and formation structure. Fast neutrons emitted by a controlled neutron source react with formation materials through inelastic fast neutron scattering and thermal neutron capture. Different nuclides emit characteristic gamma rays of varying energies during these nuclear reactions with neutrons. By recording spectra such as inelastic gamma energy spectra, capture gamma energy spectra, fast neutron time spectra, and thermal neutron decay time spectra, spectral analysis methods can be used to obtain elemental information for reservoir evaluation.

[0004] However, the biggest challenge facing domestic controlled neutron element logging tools is their outdated circuit technology. The instrument's measurement accuracy and speed lag significantly behind those of its international counterparts, and the data analysis algorithms for energy and time spectra no longer meet current measurement requirements. Existing gamma-ray signal acquisition solutions are mostly based on analog circuits. The signal acquisition and processing schemes are relatively simple, insufficient to address the severe baseline drift of the electrical signals. They are also susceptible to noise and distortion, resulting in low signal quality and low signal amplitude detection accuracy. The acquired radioactive element energy spectra have weak peaks and high noise, making it difficult to distinguish the nuclides within the energy spectrum. Conventional spectroscopic techniques struggle to identify and classify weak peaks in gamma and X-ray spectra. Peak matching techniques face difficulties with complex overlapping peaks for lower-resolution spectra. Furthermore, due to limitations associated with NaI(TI) scintillation detectors, few algorithms address spectral identification of complex overlapping peaks. Summary of the Invention

[0005] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method for processing and identifying radionuclide radiation signals of a pulse element logging tool to quickly and accurately identify radioactive elements.

[0006] To achieve the above-mentioned object of the invention, the present invention provides a method for processing and identifying radionuclide radiation signals suitable for pulse element logging, which is characterized by comprising the following steps:

[0007] (1) The detector collects the gamma ray signal generated by the reaction between neutrons and elements on the well wall, and then adds labels to the elements in the gamma ray signal;

[0008] (2) Using a voltage stabilizing diode to scale the amplitude of the gamma electric signal to x volts with a scaling parameter scale, and then adding a DC bias DC_bias to adjust the signal to [-x / 2 volts, x / 2 volts] to obtain a front-end processing signal Sig_preprocess;

[0009] (3) Use ADC to collect the front-end processing signal Sig_preprocess to obtain the original digital signal Sig_digital, and then perform low-pass filtering on the original digital signal Sig_digital to filter out high-frequency noise and obtain the signal to be corrected Sig_denoised;

[0010] (4) Collect a data segment of length N_sample from the signal to be corrected Sig_denoised as a baseline correction reference signal Sig_sample;

[0011] (5) Using the baseline correction reference signal Sig_sample as a window, perform sliding dot product processing on the correction signal Sig_denoised to obtain the cross-correlation signal Sig_patten;

[0012] (6) Use the peak detection method to determine the peak point of the cross-correlation signal Sig_patten, then use the cubic spline interpolation method to connect the peak points to obtain the baseline signal baseline, and finally subtract the baseline signal baseline from the signal to be corrected Sig_denoised to obtain the corrected signal Sig_corrected;

[0013] (7) Use the peak detection method to detect the pulse amplitude of the corrected signal Sig_corrected and save the record;

[0014] (8) Repeat steps (1), (4), and (5), with the signal pulse amplitude as the x-axis and the amplitude count rate as the y-axis, and obtain a gamma energy spectrum by counting the amplitude information in a time window of length T;

[0015] (9) By changing the type of well wall elements and repeating the above steps, multiple gamma energy spectra are obtained;

[0016] (10) Construct a multi-resolution wavelet convolutional neural network model;

[0017] The multi-resolution wavelet convolutional neural network model includes an input layer, a wavelet transform module, a CNN feature extraction module, a fully connected layer, and an output layer;

[0018] The input layer receives the normalized gamma energy spectrum, and the wavelet transform module and CNN feature extraction module are used to extract the energy spectrum features of the gamma energy spectrum. After the energy spectrum features are classified by the fully connected layer, the types of each element in the gamma energy spectrum are output through the output layer.

[0019] Finally, the training is repeated and the loss function value is calculated by the binary cross entropy of the multi-label task until the wavelet convolutional neural network model converges;

[0020] (11) Identify the well wall elements;

[0021] The detector collects the gamma-ray signal generated by the reaction between neutrons and the elements on the well wall in real time, obtains the gamma energy spectrum according to the method described in steps (1)-(8), and then inputs the gamma energy spectrum into the converged wavelet convolutional neural network model after normalization processing, thereby outputting the element type in the gamma energy spectrum.

[0022] The object of the invention of the present invention is achieved like this:

[0023] The present invention provides a method for processing and identifying radionuclide radiation signals of a pulsed element logging tool. The method comprises the following steps: radioactive nuclides generate gamma photons through a series of processes, which are absorbed by a detector and finally converted into electrical signals to obtain gamma ray signals; amplification and filtering are then performed on the gamma ray signals; an ADC is used to sample and digitize the processed waveforms; the digitized waveforms are then sent to a digital signal processor for peak search and other operations, thereby obtaining charge information of the signal to be measured; statistics are then taken of the charge information to form a radionuclide energy spectrum; and finally, a multi-resolution wavelet convolutional neural network model is used to identify the element type in the gamma energy spectrum.

[0024] At the same time, the method for processing and identifying radionuclide radiation signals of a pulse element logging tool of the present invention also has the following beneficial effects:

[0025] (1) The present invention adopts digital signal processing, which significantly improves the resolution of gamma spectrum, significantly reduces noise, and makes the peak of weak elements obvious, which shows that the system has good accuracy and robustness;

[0026] (2) The present invention implements baseline correction with the assistance of a priori signals, improves the baseline offset problem, and further ensures the accuracy of baseline correction.

[0027] (3) The present invention combines the advantages of wavelet transform and CNN feature extraction to design a neural network, and adopts an end-to-end approach to achieve fast and accurate nuclide identification, which can meet the real-time and accuracy requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a flow chart of a method for processing and identifying radionuclide radiation signals of a pulse element logging tool of the present invention;

[0029] Figure 2 This is the circuit schematic diagram of the detector;

[0030] Figure 3 This is a schematic diagram of the multi-resolution wavelet convolutional neural network model;

[0031] Figure 4 It is a schematic diagram of the structure after the wavelet transform module and the CNN feature extraction module are integrated;

[0032] Figure 5 It is a structural diagram of a single CNN feature extraction module. DETAILED DESCRIPTION

[0033] The following describes the specific embodiments of the present invention in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, such descriptions will be omitted here.

[0034] Example

[0035] Figure 1 The present invention is a flow chart of a method for processing and identifying radionuclide radiation signals of a pulse element logging tool.

[0036] In this embodiment, if Figure 1 As shown, the present invention provides a method for processing and identifying radionuclide radiation signals of a pulse element logging tool, comprising the following steps:

[0037] S1. The detector collects the gamma ray signal generated by the reaction between neutrons and elements on the well wall, and then adds labels to the elements in the gamma ray signal;

[0038] In this embodiment, the detector is a key component of the pulse element logging tool and the only non-high temperature component, such as Figure 2 As shown, the detector consists of a probe (including a scintillation crystal, a photomultiplier tube) and a photomultiplier tube high-voltage circuit, which is used to convert the energy of gamma rays into an electrical signal that can be measured.

[0039] In this embodiment, the FPGA in the pulse element logging instrument uses the A3P1000-PQG208I chip of Microsemi's ProASIC3 series to implement gamma ray signal acquisition and processing. The type of each radioactive element in the gamma ray signal is marked by adding labels to facilitate the subsequent training of the network model.

[0040] S2. Use a voltage stabilizing diode to scale the gamma signal to 2V using the scaling parameter scale, then add a DC bias DC_bias to adjust the signal to [-1V, 1V] to obtain the front-end processing signal Sig_preprocess.

[0041] S3, using ADC to collect the front-end processing signal Sig_preprocess to obtain the original digital signal Sig_digital, and then perform low-pass filtering on the original digital signal Sig_digital to filter out noise above 1kHz to obtain the signal to be corrected Sig_denoised;

[0042] In this embodiment, the ADC is a 12-bit high-speed, low-power successive approximation AD7482BSTZ produced by Analog Devices.

[0043] S4, collecting a data segment with a length of N_sample from the signal to be corrected Sig_denoised as a baseline correction reference signal Sig_sample;

[0044] S5. Using the baseline correction reference signal Sig_sample as a window, perform sliding dot product processing on the correction signal Sig_denoised to obtain a cross-correlation signal Sig_patten.

[0045] S6. Use the peak detection method to determine the peak point of the cross-correlation signal Sig_patten, then use the cubic spline interpolation method to connect the peak points to obtain the baseline signal baseline, and finally subtract the baseline signal baseline from the signal to be corrected Sig_denoised to obtain the corrected signal Sig_corrected; in this embodiment, cubic spline interpolation can divide the known data into several segments, construct a cubic function for each segment, and ensure that the connection between the piecewise functions has the properties of zero-order continuity, first-order derivative continuity, and second-order derivative continuity.

[0046] S7, using a peak detection method to detect the pulse amplitude of the corrected signal Sig_corrected, and save the record;

[0047] S8. Repeat steps S1, S4, and S5, with the signal pulse amplitude as the x-axis and the amplitude count rate as the y-axis, and obtain a gamma energy spectrum by counting the amplitude information within a time window of length T;

[0048] S9, by changing the type of well wall elements, repeating the above steps to obtain multiple gamma energy spectra;

[0049] S10. Construct a multi-resolution wavelet convolutional neural network model;

[0050] like Figure 3 As shown, the multi-resolution wavelet convolutional neural network model includes an input layer, a wavelet transform module, a CNN feature extraction module, a fully connected layer and an output layer. In this embodiment, the wavelet transform module and the CNN feature extraction module are integrated into one model. On the one hand, the wavelet transform makes up for the part of the information lost by CNN as a limited multi-resolution analysis. On the other hand, CNN can better extract features at different resolutions. Integrating them as a whole can more effectively learn and extract data features.

[0051] The input layer receives the normalized gamma energy spectrum, and the wavelet transform module and CNN feature extraction module are used to extract the energy spectrum features of the gamma energy spectrum. After the energy spectrum features are classified by the fully connected layer, the types of each element in the gamma energy spectrum are output through the output layer.

[0052] In this embodiment, if Figure 4 As shown, the wavelet transform module adopts four-level wavelet decomposition, where each level of wavelet decomposition decomposes the received input into high-frequency features and low-frequency features. The received inputs corresponding to the 1st to 4th level wavelet decompositions are gamma energy spectrum, first-level low-frequency features, second-level low-frequency features, and third-level low-frequency features respectively.

[0053] The first-level high-frequency features are passed through a 1*1 convolution layer and a downsampling layer to generate the first-level residual connection; the first-level high-frequency features are spliced ​​with the first-level low-frequency features in the form of row splicing and pass through the first-level feature extraction module to generate the first-level convolution output; the second-level wavelet high-frequency features and the second-level wavelet low-frequency features are spliced ​​in the form of row splicing and pass through a 1*1 convolution layer and a 1*5 convolution layer with 64 convolution kernels to generate the first-level wavelet connection and the first-level wavelet convolution output; the first-level residual connection, the first-level convolution output, the first-level wavelet connection and The first-level wavelet convolution output is spliced ​​in the form of column splicing to form the first-level feature; the first-level feature is passed through a 1*1 convolution layer and a downsampling layer to generate the second-level residual connection; the first-level feature is passed through the second-level feature extraction module to generate the second-level convolution output; the third-level wavelet high-frequency feature and the third-level wavelet low-frequency feature are spliced ​​in the form of row splicing and then passed through a 1*1 convolution layer and 2 1*5 convolution layers with 32 and 64 convolution kernels respectively to generate the second-level wavelet connection and the second-level wavelet convolution output; the second-level residual connection and the second-level convolution are respectively generated. The output, the second-level wavelet connection and the second-level wavelet convolution output are spliced ​​in the form of column splicing to form the second-level features; the second-level features are passed through a 1*1 convolution layer and a downsampling layer to generate the third-level residual connection; the second-level features are passed through the third-level feature extraction module to generate the third-level convolution output; the fourth-level wavelet high-frequency features and the fourth-level wavelet low-frequency features are spliced ​​in the form of row splicing and then passed through a 1*1 convolution layer and 3 1*5 convolution layers with 32, 64 and 128 convolution kernels respectively to generate the third-level wavelet connection and the third-level wavelet convolution output. The third-level residual connection, the third-level convolution output, the third-level wavelet connection, and the third-level wavelet convolution output are concatenated in columns to form the third-level features. The third-level features are passed through a 1*1 convolution layer and a downsampling layer to produce a fourth-level residual connection. The third-level features are passed through the fourth-level feature extraction module to produce a fourth-level convolution output. The fourth-level residual connection and the fourth-level convolution output are concatenated in columns to form the fourth-level features. The fourth-level features are extracted by the wavelet transform module and the feature extraction module and input into the fully connected layer for classification. Among them, the 1*1 convolution layer connects the features at different wavelet resolutions to form a residual connection to prevent the network gradient from disappearing.

[0054] like Figure 5 As shown in Figure 1, the CNN feature extraction module consists of two convolutional layers and one downsampling layer, where the convolution kernel length of each convolutional layer is 5. The CNN feature extraction module uses downsampling to reduce the feature dimension while doubling the number of filters, which not only reduces the feature dimension of the output layer but also alleviates the problem of feature disappearance.

[0055] Finally, the training is repeated and the loss function value is calculated by the binary cross entropy of the multi-label task until the wavelet convolutional neural network model converges;

[0056] S11, identify the well wall elements;

[0057] The detector collects the gamma-ray signal generated by the reaction between neutrons and the elements on the well wall in real time, obtains the gamma energy spectrum according to the method described in steps S1-S8, and then inputs the gamma energy spectrum into the converged wavelet convolutional neural network model after normalization processing, thereby outputting the element type in the gamma energy spectrum.

[0058] Although the above describes the illustrative specific embodiments of the present invention to facilitate understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concepts of the present invention are protected.

Claims

1. A method for processing and identifying radionuclide radiation signals of a pulse element logging tool, characterized in that: The following steps are involved: (1) The detector collects the gamma ray signal generated by the reaction between neutrons and elements on the well wall, and then adds labels to the elements in the gamma ray signal; (2) Using a voltage stabilizing diode to scale the amplitude of the gamma electric signal to x volts with a scaling parameter scale, and then adding a DC bias DC_bias to adjust the signal to [-x / 2 volts, x / 2 volts] to obtain a front-end processing signal Sig_preprocess; (3) Use ADC to collect the front-end processing signal Sig_preprocess to obtain the original digital signal Sig_digital, and then perform low-pass filtering on the original digital signal Sig_digital to filter out high-frequency noise and obtain the signal to be corrected Sig_denoised; (4) Collect a data segment of length N_sample from the signal to be corrected Sig_denoised as a baseline correction reference signal Sig_sample; (5) Using the baseline correction reference signal Sig_sample as a window, perform sliding dot product processing on the correction signal Sig_denoised to obtain the cross-correlation signal Sig_patten; (6) Use the peak detection method to determine the peak point of the cross-correlation signal Sig_patten, then use the cubic spline interpolation method to connect the peak points to obtain the baseline signal baseline, and finally subtract the baseline signal baseline from the signal to be corrected Sig_denoised to obtain the corrected signal Sig_corrected; (7) Use the peak detection method to detect the pulse amplitude of the corrected signal Sig_corrected and save the record; (8) Repeat steps (1), (4), and (5), with the signal pulse amplitude as the x-axis and the amplitude count rate as the y-axis, and obtain a gamma energy spectrum by counting the amplitude information in a time window of length T; (9) By changing the type of well wall elements and repeating the above steps, multiple gamma energy spectra are obtained; (10) Construct a multi-resolution wavelet convolutional neural network model; The multi-resolution wavelet convolutional neural network model includes an input layer, a wavelet transform module, a CNN feature extraction module, a fully connected layer, and an output layer; The input layer receives the normalized gamma energy spectrum, and the wavelet transform module and CNN feature extraction module are used to extract the energy spectrum features of the gamma energy spectrum. After the energy spectrum features are classified by the fully connected layer, the types of each element in the gamma energy spectrum are output through the output layer. Finally, the training is repeated and the loss function value is calculated by the binary cross entropy of the multi-label task until the wavelet convolutional neural network model converges; (11) Identify well wall elements; The detector collects the gamma-ray signal generated by the reaction between neutrons and the elements on the well wall in real time, obtains the gamma energy spectrum according to the method described in steps (1)-(8), and then inputs the gamma energy spectrum into the converged wavelet convolutional neural network model after normalization processing, thereby outputting the element type in the gamma energy spectrum.

2. The method for processing and identifying radionuclide radiation signals of a pulse element logging tool according to claim 1, characterized in that: The wavelet transform module adopts four-level wavelet decomposition, wherein each level of wavelet decomposition decomposes the received input into high-frequency features and low-frequency features, and the received inputs corresponding to the 1st to 4th level wavelet decompositions are gamma energy spectrum, first-level low-frequency features, second-level low-frequency features, and third-level low-frequency features respectively; The first-level high-frequency features are passed through a 1*1 convolution layer and a downsampling layer to generate the first-level residual connection; the first-level high-frequency features are spliced ​​with the first-level low-frequency features in the form of row splicing and pass through the first-level feature extraction module to generate the first-level convolution output; the second-level wavelet high-frequency features and the second-level wavelet low-frequency features are spliced ​​in the form of row splicing and pass through a 1*1 convolution layer and a 1*5 convolution layer with 64 convolution kernels to generate the first-level wavelet connection and the first-level wavelet convolution output; the first-level residual connection, the first-level convolution output, the first-level wavelet connection and The first-level wavelet convolution output is spliced ​​in the form of column splicing to form the first-level feature; the first-level feature is passed through a 1*1 convolution layer and a downsampling layer to generate the second-level residual connection; the first-level feature is passed through the second-level feature extraction module to generate the second-level convolution output; the third-level wavelet high-frequency feature and the third-level wavelet low-frequency feature are spliced ​​in the form of row splicing and then passed through a 1*1 convolution layer and 2 1*5 convolution layers with 32 and 64 convolution kernels respectively to generate the second-level wavelet connection and the second-level wavelet convolution output; the second-level residual connection and the second-level convolution are respectively generated. The output, the second-level wavelet connection and the second-level wavelet convolution output are spliced ​​in the form of column splicing to form the second-level features; the second-level features are passed through a 1*1 convolution layer and a downsampling layer to generate the third-level residual connection; the second-level features are passed through the third-level feature extraction module to generate the third-level convolution output; the fourth-level wavelet high-frequency features and the fourth-level wavelet low-frequency features are spliced ​​in the form of row splicing and then passed through a 1*1 convolution layer and 3 1*5 convolution layers with 32, 64 and 128 convolution kernels respectively to generate the third-level wavelet connection and the third-level wavelet convolution output. The third-level residual connection, the third-level convolution output, the third-level wavelet connection and the third-level wavelet convolution output are spliced ​​in the form of column splicing to form the third-level features; the third-level features are passed through a 1*1 convolution layer and a downsampling layer to generate the fourth-level residual connection; the third-level features are passed through the fourth-level feature extraction module to generate the fourth-level convolution output; the fourth-level residual connection and the fourth-level convolution output are spliced ​​in the form of column splicing to form the fourth-level features; the fourth-level features are used as the features extracted by the wavelet transform module and the feature extraction module and input into the fully connected layer for classification.

3. The method for processing and identifying radionuclide radiation signals of a pulse element logging tool according to claim 1, characterized in that: The CNN feature extraction module consists of two convolutional layers and one downsampling layer, wherein the convolution kernel length of each convolutional layer is 5.

Images