A Deep Learning-Based Intelligent Discrimination Method and System for Neutron Detector Pulse Signals

Abstract

This invention belongs to the field of nuclear logging data processing technology, specifically involving a deep learning-based intelligent identification method and system for neutron detector pulse signals. The method includes the following steps: acquiring the pulse signal output by the detector; when the rising edge of the pulse signal exceeds a preset noise floor threshold, extracting a time series containing the complete pulse signal and normalizing the time series to obtain a normalized dimensionless voltage series; calculating the absolute value of the first-order difference between adjacent sampling points in the dimensionless voltage series, and weighting the calculation by combining the time distance between the current sampling point and the peak time to obtain the waveform oscillation entropy index, which characterizes the degree of micro-oscillation of the waveform. This invention can effectively eliminate more than 98% of neutron tube arcing interference, significantly purify the high-energy portion of the energy spectrum, and solve the problem of unusable data from aging neutron tubes.

Description

Technical Field

[0001] This invention belongs to the field of nuclear logging data processing technology, specifically relating to a method and system for intelligent identification of neutron detector pulse signals based on deep learning. Background Technology

[0002] Pulsed neutron logging is a core method for evaluating reservoir oil saturation and lithological composition. This technology relies on a downhole neutron generator emitting high-energy fast neutrons into the formation and using a scintillator detector to collect the gamma rays fed back from the formation. During operation, the neutron generator requires a high-voltage electric field of approximately 100 kV to accelerate deuterium ions to bombard a tritium target. In actual logging operations, especially with increased instrument usage time or rising downhole temperatures, the vacuum insulation performance inside the neutron tube decreases, leading to minute sparking phenomena, or micro-discharges, between the high-voltage electrodes. This phenomenon generates interference signals.

[0003] In existing technologies, electromagnetic interference pulses generated by ignition typically have extremely high amplitudes, often falling at the high-energy end of the energy spectrum, such as the oxygen and carbon window regions of inelastic scattering gamma. Traditional discrimination methods based on amplitude thresholds not only fail to eliminate ignition signals but also misclassify them as important high-energy signals, leading to severe distortion in carbon-oxygen ratio calculations. Furthermore, traditional trapezoidal shaping or Gaussian shaping algorithms only focus on the pulse area or amplitude, ignoring the pulse's microscopic texture, making it difficult to distinguish between ignition signals with high-frequency oscillations or irregular spikes and smooth unipolar nuclear detector pulses.

[0004] To address the aforementioned issues, while large convolutional neural networks could be used to identify waveforms, these models consume significant computational resources and storage space, making them unsuitable for real-time operation on low-power FPGAs or DSP chips in the high-temperature environment of downhole drilling. This fails to meet the microsecond-level processing time requirements for single pulses. Therefore, there is an urgent need for an intelligent pulse signal identification method that can accurately eliminate high-energy ignition noise while meeting the requirements for low-power real-time processing in downhole environments. Summary of the Invention

[0005] This invention provides a method and system for intelligent discrimination of neutron detector pulse signals based on deep learning, in order to solve the technical problem in the prior art where ignition signals and high-energy nuclear signals are mixed and difficult to remove by traditional methods.

[0006] In a first aspect, the present invention provides a method for intelligent discrimination of neutron detector pulse signals based on deep learning, comprising the following steps:

[0007] The pulse signal output by the detector is acquired. When the rising edge of the pulse signal exceeds the preset noise floor threshold, the time series containing the complete pulse signal is extracted and normalized to obtain the normalized dimensionless voltage series.

[0008] The absolute value of the first-order difference between adjacent sampling points in the dimensionless voltage sequence is calculated, and weighted by combining the time distance between the current sampling point and the peak moment, to obtain the waveform oscillation entropy index, which characterizes the degree of micro oscillation of the waveform.

[0009] Based on the waveform oscillation entropy exponent and the original peak amplitude of the pulse signal, a nonlinear activation function incorporating an amplitude penalty mechanism is used to calculate the neutron pulse confidence level for judging the validity of the signal.

[0010] The confidence level of the neutron pulse is compared with a preset passing threshold. If the confidence level of the neutron pulse is greater than or equal to the passing threshold, the pulse signal is determined to be a valid nuclear signal and included in the energy spectrum. If the confidence level of the neutron pulse is less than the passing threshold, the pulse signal is determined to be arcing noise and discarded.

[0011] Furthermore, the time series is normalized to satisfy the expression:

[0012]

[0013] In the formula, This is the normalized dimensionless voltage sequence. For the first Voltage values ​​at each sampling point This represents the maximum voltage value extracted from the time series.

[0014] Furthermore, the waveform oscillation entropy exponent satisfies the expression:

[0015]

[0016] In the formula, The waveform oscillation entropy exponent. The total length of the extracted waveform sequence. This is the normalized dimensionless voltage sequence. For sampling point index, The logarithmic base protection constant, This is the distance weighting coefficient. This is the index of the sampling point corresponding to the peak time.

[0017] Furthermore, the logarithmic base protection constant The value of is greater than or equal to the natural constant. Distance weighting coefficient Weights are used to amplify oscillation characteristics far from the peak region.

[0018] Furthermore, the neutron pulse confidence level satisfies the expression:

[0019]

[0020] In the formula, The confidence level of the neutron pulse. The waveform oscillation entropy exponent. To reference the oscillation entropy threshold, The power index of oscillation sensitivity. This is the magnitude penalty coefficient. This represents the maximum voltage value of the pulse signal. This is the maximum allowable voltage value of the measurement range.

[0021] Furthermore, the reference oscillation entropy threshold is determined based on the average value of the waveform oscillation entropy exponent of the single-pulse signal generated by the standard radiation source; the oscillation sensitivity power exponent is used to control the steepness of the discrimination function.

[0022] Furthermore, the neutron pulse confidence level is compared with a preset passage threshold, including:

[0023] Set the threshold to 0.5;

[0024] when At that time, retain the maximum voltage value of the pulse signal. And accumulate it to the channel address corresponding to the multichannel energy spectrum;

[0025] when When the pulse signal is interrupted, the value of the ignition counter is incremented by 1.

[0026] Furthermore, The calculation process is implemented in the FPGA through a pre-calculated and stored lookup table, and real-time calculation is performed according to... and The difference can be obtained directly by looking up the table.

[0027] Furthermore, the method is applied to low-power FPGA or DSP chips in high-temperature downhole environments.

[0028] Secondly, the present invention provides a deep learning-based intelligent discrimination system for neutron detector pulse signals, including a memory and a processor. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned deep learning-based intelligent discrimination method for neutron detector pulse signals is implemented.

[0029] The beneficial effects are as follows: This invention utilizes a lightweight pulse-gated network based on waveform oscillation entropy characteristics, without relying on complex black-box deep networks. Instead, it embeds physical feature operators into the front end to construct an oscillation entropy sensing layer. This method can effectively eliminate over 98% of neutron tube arcing interference, significantly purifying the high-energy portion of the energy spectrum and solving the problem of unusable data from aging neutron tubes. Simultaneously, the core algorithm involves only simple computations, processing a single pulse in less than 1 microsecond, allowing it to easily run on low-power downhole chips, meeting the real-time requirements of high-count-rate logging, and reducing the variance of the processed curve by over 60%, effectively extending the service life of the neutron tube. Attached Figure Description

[0030] Figure 1 This is a flowchart of a deep learning-based intelligent discrimination method for neutron detector pulse signals.

[0031] Figure 2 A comparison of the microscopic waveform oscillation characteristics of nuclear signals and ignition noise.

[0032] Figure 3 This is a comparison chart of the high-energy gamma spectrum quality before and after intelligent identification. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] An embodiment of the intelligent discrimination method for neutron detector pulse signals based on deep learning provided by this invention:

[0035] like Figure 1 As shown, the intelligent discrimination method for neutron detector pulse signals based on deep learning includes the following steps:

[0036] S1: Acquire the pulse signal output by the detector. When the rising edge of the pulse signal exceeds the preset noise floor threshold, extract the time series containing the complete pulse signal and normalize the time series to obtain the normalized dimensionless voltage series.

[0037] Specifically, the system first utilizes the high-speed ADC circuit of the downhole instrument to monitor the detector output in real time. When the rising edge of the detected pulse signal exceeds a preset noise floor threshold, for example, a noise floor threshold of 20mV, the FPGA logic unit automatically locks and extracts a time series containing the complete pulse signal. Assuming the number of extracted sampling points is... , record the The voltage value at each sampling point is .

[0038] To eliminate the interference of amplitude magnitude on subsequent waveform morphology analysis, the extracted time series needs to be normalized. First, the series is traversed to obtain the maximum voltage value. ,like If valid, meaning the noise level is greater than the noise floor threshold, then normalization is performed according to the following formula:

[0039]

[0040] In the formula, This is a normalized dimensionless voltage sequence, with values ​​ranging from 0 to 1.

[0041] For example, suppose the voltage sequence of a captured pulse signal is [0.1, 0.5, 1.5, 2.0, 1.2, 0.4, ...]. At this time... Then the normalized sequence It becomes [0.05, 0.25, 0.75, 1.0, 0.6, 0.2, ...]. Through this processing, regardless of whether the original signal is strong or weak, its waveform characteristics are unified to the same scale, which facilitates subsequent processing.

[0042] By normalizing the acquired pulse signals, the influence of signal amplitude differences on waveform morphology analysis can be eliminated, ensuring that subsequent feature extraction focuses on the microscopic texture and shape features of the waveform, thus laying the foundation for accurate identification of ignition signals.

[0043] S2 calculates the absolute value of the first-order difference between adjacent sampling points in the dimensionless voltage sequence, and performs a weighted calculation based on the time distance between the current sampling point and the peak moment to obtain the waveform oscillation entropy index, which characterizes the degree of micro-oscillation of the waveform.

[0044] Specifically, the essential difference between the ignition signal and the actual nuclear signal lies in the following: the nuclear signal is determined by crystal light emission and RC circuit discharge, resulting in a smooth waveform; the ignition signal is determined by plasma fluctuations and distributed inductance-capacitance resonance, resulting in a rough waveform accompanied by oscillations. To accurately capture these microscopic oscillations, this invention designs a waveform oscillation entropy index. .

[0045] The waveform oscillation entropy index uses a difference operator to extract the roughness of the waveform and a logarithmic weighting function to amplify the sensitivity to non-peak regions. The formula for calculating the waveform oscillation entropy index is as follows:

[0046]

[0047] In the formula, The total length of the extracted waveform sequence is, for example, 32. It is the absolute value of the first-order difference between adjacent sampling points, representing the local rate of change of the waveform; This is the logarithmic base protection constant, exemplarily taking the value of... ,make sure The argument is always greater than or equal to This ensures that the logarithmic calculation result is always greater than or equal to 1, preventing the weight from being negative or zero; This is the distance weighting coefficient, exemplarily set to a value of 0.5; Index of the current sampling point Distance from peak time The time distance.

[0048] Assumption , ;

[0049] For a smooth nuclear signal, near the peak ( ), Very small, weight Meanwhile, due to the smoothness of the waveform, It is also relatively small, and the product of the two is relatively small.

[0050] For a firing signal, suppose an oscillation occurs at the tail, for example, 10 points away from the peak, i.e. At this time, the weight term is Due to oscillations, the difference value It will increase significantly, for example, from 0.05 during smoothing to 0.3, at which point the contribution value is... This is much larger than the smoothed signal. The final accumulated value is... The values ​​will be significantly higher than the actual signal. In FPGAs, logarithmic terms can be pre-calculated and stored as lookup tables, allowing for direct lookup during real-time calculations without consuming significant resources for logarithmic operations.

[0051] By calculating the waveform oscillation entropy exponent and utilizing the differential operator and distance weighting mechanism, the oscillation characteristics in the non-peak region can be effectively amplified, thereby accurately quantifying the micro-roughness of the waveform and achieving keen capture of the ignition signal characteristics.

[0052] S3, based on the waveform oscillation entropy exponent and the original peak amplitude of the pulse signal, uses a nonlinear activation function that incorporates an amplitude penalty mechanism to calculate the neutron pulse confidence level used to judge the validity of the signal.

[0053] Specifically, after calculating the waveform oscillation entropy exponent, the system needs to convert it into a probability value between 0 and 1 to determine whether the signal is a valid neutron / gamma pulse. This invention designs a nonlinear activation function incorporating an amplitude penalty mechanism to determine the neutron pulse confidence level. The calculation formula is as follows:

[0054]

[0055] In the formula, To reference the oscillation entropy threshold, it is usually obtained in a ground-based laboratory using a standard radioactive source; The oscillation sensitivity power exponent controls the steepness of the discrimination function; This is the magnitude penalty coefficient; This represents the maximum voltage value of the pulse signal. This is the maximum voltage value allowed by the ADC range.

[0056] Assumption , , , ;

[0057] Scenario A: A small-amplitude true signal. , ;

[0058] Substituting into the formula: the oscillation term in the denominator is The penalty range is as follows: ;

[0059] but If the result is greater than 0.5, it is considered valid.

[0060] Situation B: A strong ignition signal. , ;

[0061] Substituting into the formula: the oscillation term in the denominator is The penalty range is as follows: .

[0062] but The result is much less than 0.5, so it is deemed invalid.

[0063] It can be seen that the magnitude penalty item Large-amplitude signals were rigorously examined. If the signal amplitude was very large, even a slight oscillation would cause the denominator to increase rapidly, resulting in the neutron pulse confidence level approaching 0, thus leading to its rejection.

[0064] By introducing a nonlinear activation function that incorporates an amplitude penalty mechanism, the effectiveness of the pulse can be comprehensively evaluated based on the degree of oscillation and the signal amplitude. In particular, it enables more stringent screening of large-amplitude signals, effectively preventing high-energy ignition noise from entering the energy spectrum and improving data quality.

[0065] S4. Compare the neutron pulse confidence level with a preset passing threshold. If the neutron pulse confidence level is greater than or equal to the passing threshold, the pulse signal is determined to be a valid nuclear signal and included in the energy spectrum. If the neutron pulse confidence level is less than the passing threshold, the pulse signal is determined to be arcing noise and discarded.

[0066] Specifically, the FPGA executes the final gating logic based on the neutron pulse confidence calculated in step S3, typically setting the pass threshold to 0.5.

[0067] when When the system determines that the pulse signal is a genuine nuclear signal, it retains the amplitude value of the pulse signal, i.e., the maximum voltage value. And add it to the channel address corresponding to the multichannel energy spectrum.

[0068] when When the pulse signal is identified as ignition noise, the system discards the data and excludes it from the energy spectrum. Simultaneously, the ignition counter inside the FPGA is incremented by 1 to monitor the health status of the neutron tube in real time.

[0069] Reference Figure 2 The diagram shows a comparison of the microscopic waveform oscillation characteristics of nuclear signals and ignition noise. The solid blue line represents the real nuclear pulse, which has a smooth waveform and a low calculated waveform oscillation entropy index, such as 0.42, and is therefore considered a valid signal. The dashed red line represents the ignition interference pulse, which has obvious high-frequency sawtooth oscillations at the beginning of the rising edge and the tail of the falling edge. The calculated waveform oscillation entropy index is extremely high, such as 4.85, and is therefore judged as noise and discarded by the algorithm.

[0070] Reference Figure 3 The image shows a comparison of the quality of the high-energy gamma spectrum before and after intelligent identification. The blue-filled area represents the original collected spectrum, where there are a large number of false signals caused by arcing at the high-energy end, forming unnatural bulges. The red solid line represents the purified spectrum after processing by this invention. The abnormal counts at the high-energy end are removed, clearly revealing the true oxygen peak, which highly overlaps with the standard reference spectrum represented by the green dashed line.

[0071] By setting a clear gating threshold to judge the calculated confidence level, it is possible to quickly classify and process pulse signals, completely eliminate ignition noise while retaining the effective core signal, and monitor the equipment status in real time, thereby significantly improving the accuracy and reliability of logging data.

[0072] An embodiment of the intelligent discrimination system for neutron detector pulse signals based on deep learning provided by this invention:

[0073] The deep learning-based intelligent discrimination system for neutron detector pulse signals includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the aforementioned deep learning-based intelligent discrimination method for neutron detector pulse signals.

[0074] The deep learning-based intelligent discrimination system for neutron detector pulse signals also includes other components well known to those skilled in the art, such as communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

[0075] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device. Any application or module described in this invention can be implemented using computer-readable / executable instructions stored or otherwise maintained by such a computer-readable medium.

[0076] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent discrimination of neutron detector pulse signals based on deep learning, characterized in that, Includes the following steps: The pulse signal output by the detector is acquired. When the rising edge of the pulse signal exceeds the preset noise floor threshold, the time series containing the complete pulse signal is extracted and normalized to obtain the normalized dimensionless voltage series. The absolute value of the first-order difference between adjacent sampling points in the dimensionless voltage sequence is calculated, and weighted by combining the time distance between the current sampling point and the peak moment, to obtain the waveform oscillation entropy index, which characterizes the degree of micro oscillation of the waveform. Based on the waveform oscillation entropy exponent and the original peak amplitude of the pulse signal, a nonlinear activation function incorporating an amplitude penalty mechanism is used to calculate the neutron pulse confidence score for judging signal validity, satisfying the expression: , The confidence level of the neutron pulse. The waveform oscillation entropy exponent. To reference the oscillation entropy threshold, The power index of oscillation sensitivity. This is the magnitude penalty coefficient. This represents the maximum voltage value of the pulse signal. This is the maximum allowable voltage value of the measurement range; The confidence level of the neutron pulse is compared with a preset passing threshold. If the confidence level of the neutron pulse is greater than or equal to the passing threshold, the pulse signal is determined to be a valid nuclear signal and is included in the energy spectrum. If the confidence level of the neutron pulse is less than the passing threshold, the pulse signal is determined to be ignition noise and discarded. The neutron pulse confidence level is compared with a preset passage threshold, including setting the passage threshold to 0.5; when At that time, retain the maximum voltage value of the pulse signal. And accumulate it to the channel address corresponding to the multichannel energy spectrum; when When the pulse signal is interrupted, the value of the ignition counter is incremented by 1.

2. The intelligent discrimination method for neutron detector pulse signals based on deep learning according to claim 1, characterized in that, The time series is normalized to satisfy the expression: In the formula, This is the normalized dimensionless voltage sequence. For the first Voltage values ​​at each sampling point This represents the maximum voltage value extracted from the time series.

3. The intelligent discrimination method for neutron detector pulse signals based on deep learning according to claim 1, characterized in that, The waveform oscillation entropy exponent satisfies the expression: In the formula, The waveform oscillation entropy exponent. The total length of the extracted waveform sequence. This is the normalized dimensionless voltage sequence. For sampling point index, The logarithmic base protection constant, This is the distance weighting coefficient. This is the index of the sampling point corresponding to the peak time.

4. The intelligent discrimination method for neutron detector pulse signals based on deep learning according to claim 3, characterized in that, Logarithmic base protection constant The value of is greater than or equal to the natural constant. Distance weighting coefficient Weights are used to amplify oscillation characteristics far from the peak region.

5. The intelligent discrimination method for neutron detector pulse signals based on deep learning according to claim 1, characterized in that, The reference oscillation entropy threshold is determined based on the average value of the waveform oscillation entropy exponent of the single-pulse signal generated by the standard radiation source; the oscillation sensitivity power exponent is used to control the steepness of the discrimination function.

6. The intelligent discrimination method for neutron detector pulse signals based on deep learning according to claim 3, characterized in that, The calculation process is implemented in the FPGA through a pre-calculated and stored lookup table, and real-time calculation is performed according to... and The difference can be obtained directly by looking up the table.

7. The intelligent discrimination method for neutron detector pulse signals based on deep learning according to claim 1, characterized in that, The method is applied to low-power FPGA or DSP chips in high-temperature downhole environments.

8. A deep learning-based intelligent discrimination system for neutron detector pulse signals, characterized in that, It includes a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the intelligent discrimination method for neutron detector pulse signals based on deep learning as described in any one of claims 1-7 is implemented.

Images