Self-adaptive separation and enhancement method for ultrasonic guided wave multi-mode signals

Abstract

The invention provides a self-adaptive separation and enhancement method for ultrasonic guided wave multi-mode signals, and relates to the technical field of nondestructive detection.The method comprises the steps that a baseline state voucher is established and verified, comprehensive baseline reliability is obtained by calculating voucher validity, maturity and stability, and the quality of reference data is ensured; thirdly, calculating mode purity, an energy attenuation mismatch coefficient and a signal structure complexity entropy based on the physical relevance of the real-time signal so as to quantify the intrinsic conformity of the real-time signal; then, constructing a dynamic working condition base line matched with the current working condition, and calculating an environment working condition deviation degree representing a real-time signal deviation degree by applying a mahalanobis distance; and finally, taking the environment working condition deviation degree, the baseline reliability and the like as state input, and outputting an optimal self-adaptive enhancement strategy and a pipe section integrity evaluation result by the reinforcement learning model. Through step-by-step quantitative evaluation and intelligent decision making, dynamic adaptive adjustment of a signal processing strategy is realized.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, specifically to an adaptive separation and enhancement method for ultrasonic guided wave multimode signals. Background Technology

[0002] Ultrasonic guided wave technology, capable of rapidly scanning pipelines tens or even hundreds of meters long from a single excitation point, has become a key technology for online monitoring of long-distance pipelines in the oil, gas, and chemical industries. Current technology excites guided waves of a specific mode within the pipeline wall, causing them to propagate along the pipeline axis. The integrity of the pipeline is then assessed by analyzing the reflected or mode-transformed signals generated after the waves interact with discontinuities in the pipeline structure (such as welds, flanges, corrosion defects, and cracks).

[0003] In practical applications, ideal guided wave detection relies on exciting and propagating a single, pure guided wave mode within the pipeline, such as the torsional mode T(0,1) sensitive to circumferential defects or the longitudinal mode L(0,2) sensitive to axial defects. The entire theoretical basis of signal analysis, including defect localization and quantification, is based on the premise that the current mode propagates at a predictable speed and attenuation characteristics. However, during actual propagation, when this pre-set target mode interacts with inherent structural features on the pipeline, such as welds, elbows, and supports, mode conversion occurs, exciting various stray modes with different propagation speeds and characteristics. These stray modes superimpose with the reflected signals of the target mode, forming extremely complex waveforms. This impure signal has two serious consequences: first, the stray mode itself is misinterpreted as a defect echo, leading to false alarms; second, as coherent noise, it masks weak echoes from real, minute defects, resulting in missed detections.

[0004] Currently, traditional signal processing workflows generally lack effective means to quantitatively assess the degree of aliasing, meaning they cannot a priori determine the inherent "availability" and "reliability" of a signal before analysis. Analyzing a "low-quality" signal distorted due to severe mode impurity significantly reduces the reliability of the results. Such analysis makes it difficult to effectively separate weak target signals from this complex noise background in later stages, resulting in poor separation performance and the continued obscuring of defective features. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive separation and enhancement method for ultrasonic guided wave multimode signals to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An adaptive separation and enhancement method for ultrasonic guided wave multimode signals, the specific steps of which include: S1: establishing and verifying a baseline state certificate based on historical calibration data associated with a specific pipe segment, the baseline state certificate being used to characterize the inherent acoustic characteristics of the pipe segment under ideal working conditions;

[0008] S2: Acquire the real-time multimodal pipeline signal to be evaluated, collected by the sensor;

[0009] S3: Based on the physical correlation within the real-time multimodal pipeline signal, calculate the intrinsic signal conformity index to quantify its deviation from the theoretical physical model;

[0010] S4: Based on the baseline status certificate, combined with real-time environment and operating parameters, construct a dynamic operating condition baseline, and combine the intrinsic signal compliance index to calculate the environmental operating condition deviation degree, which characterizes the degree of deviation between the real-time signal and the dynamic operating condition baseline.

[0011] S5: Based on the environmental operating condition deviation, generate an adaptive enhancement strategy for real-time multimodal pipeline signals and pipeline integrity assessment results.

[0012] Furthermore, verifying the baseline status certificate specifically includes: confirming the timeliness and coverage of the historical calibration data on which the certificate relies to obtain the certificate validity status; after verifying the baseline status certificate, it further includes: calculating the certificate maturity, which characterizes the richness of historical data, based on the time span of the historical calibration data; calculating the certificate stability, which characterizes the reliability, based on the statistical stability of the historical calibration data under different operating conditions; and weightedly integrating the certificate validity status, certificate maturity, and certificate stability to obtain the comprehensive baseline reliability.

[0013] Furthermore, the calculation of the intrinsic signal compliance index specifically includes: calculating the mode purity, which characterizes the distribution of guided wave energy between the preset target mode and spurious modes; calculating the energy attenuation mismatch coefficient, which characterizes the difference between the actual attenuation rate of the signal and the theoretical attenuation model; and calculating the signal structure complexity entropy, which quantifies the disorder of the guided wave signal waveform structure.

[0014] Furthermore, the signal structure complexity entropy includes: based on wavelet packet transform of the guided wave signal, jointly quantifying the uniformity of its energy distribution and the degree of disorder of its instantaneous characteristics in each frequency band; taking mode purity, energy attenuation mismatch coefficient and signal structure complexity entropy as inputs, and performing probabilistic inference fusion through a pre-trained Bayesian network to obtain the posterior probability distribution of the intrinsic signal conformity index.

[0015] Furthermore, the signal structure complexity entropy is obtained by statistically learning from labeled signal samples containing different types of defects or noise in the historical database; wavelet packet energy spectrum analysis is performed on the signal, and Shannon information entropy calculation logic is applied.

[0016] Furthermore, the higher the reliability of the integrated baseline, the more accurate, stable, and reliable the historical acoustic model of the current pipe segment is judged by this method, and the better the prior knowledge of signal processing is. When the intrinsic signal compliance index is lower, it indicates that the guided wave signal currently acquired by this method has more severe mode aliasing, abnormal attenuation, or structural distortion, the signal quality is worse, and the potential interference or defect risk is higher.

[0017] Furthermore, the dynamic operating condition baseline includes statistical distribution parameters of historical normal signal characteristics corresponding to the pipe section and the current environmental operating state; the calculation of the environmental operating condition deviation specifically includes: constructing a current signal state feature vector containing intrinsic signal compliance index and comprehensive baseline reliability; determining the dynamic operating condition baseline as the feature mean vector and feature covariance matrix of historical normal signals matching the current operating condition of the pipe section; and applying the Mahalanobis distance algorithm to calculate the distance between the current signal state feature vector and the dynamic operating condition baseline as the environmental operating condition deviation.

[0018] Furthermore, the generation of the adaptive enhancement strategy specifically includes: taking the environmental condition deviation as a key input, selecting and configuring the optimal combination of signal processing operators; and triggering a preset adaptive diagnostic level based on the evaluation of the environmental condition deviation; defining the triggering of the preset adaptive diagnostic level specifically includes: taking the environmental condition deviation and the comprehensive baseline reliability as state inputs and feeding them into a pre-trained reinforcement learning model; and having the reinforcement learning model output and execute the optimal enhancement strategy with the maximum expected diagnostic accuracy based on the state input.

[0019] Furthermore, the state input of the reinforcement learning model further includes pipeline criticality, which represents the importance or risk level of the current pipeline segment; after executing the optimal reinforcement strategy, the reward value is calculated based on the signal-to-noise ratio improvement score and the signal clarity improvement score of the enhanced signal.

[0020] The reinforcement learning model is updated online using reward values ​​to optimize its subsequent policy selection. The smaller the environmental condition deviation, the more closely the current signal state matches its normal state under the current environmental conditions, the healthier the signal, and the lower the risk of anomalies. The optimal enhancement policy is a specific combination of actions selected by the reinforcement learning model from a set of discrete actions that includes various filters, compensation algorithms, and reconstruction models.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] This invention improves the reliability of signal analysis data sources by establishing a multi-dimensional evaluation system encompassing credential validity status (V), credential maturity (M), and credential stability (S), and calculating a comprehensive baseline reliability (Rc). This method abandons the blind reliance on baseline data quality found in traditional techniques, introducing an automated quality control threshold of "verification before use." By quantitatively reviewing the timeliness, operational condition coverage, and consistency of baseline data, it ensures that only high-quality, highly relevant baseline credentials can be used for subsequent differential comparison and analysis. This quality control mechanism from the source effectively avoids the "garbage in, garbage out" problem caused by using outdated, unrepresentative, or unstable baselines, laying a solid and reliable data foundation for the accuracy of the entire detection process.

[0023] This invention further enhances the accuracy and depth of signal diagnosis by deconstructing real-time signals across multiple physical dimensions. It utilizes two-dimensional Fourier transform to calculate mode purity, compares theoretical models to calculate energy attenuation mismatch coefficients, and applies wavelet packet energy entropy to calculate signal structural complexity. Bayesian networks are then employed for probabilistic fusion. Compared to traditional methods that only focus on signal amplitude changes, this invention comprehensively and quantitatively characterizes the "health" of a signal from three orthogonal physical dimensions: mode composition, energy propagation patterns, and waveform structural orderliness. This deep signal insight allows the method not only to determine whether a signal is abnormal but also to reveal the underlying physical causes of the abnormality (e.g., whether it is dominated by mode mixing or abnormal attenuation).

[0024] This invention constructs a dynamic baseline that matches real-time operating conditions and applies the Mahalanobis distance algorithm to calculate the environmental operating condition deviation D. env This deviation is then used as a key state input to drive the reinforcement learning model to select the optimal reinforcement strategy, thereby improving the adaptive capability and intelligence level of the entire signal processing process. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall method flow of the present invention;

[0026] Figure 2 This is a schematic diagram of the energy distribution spectrum S(f,k) of the multimode signal fk in the S311 pipeline in the specification of this invention. Detailed Implementation

[0027] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0028] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0029] Example 1:

[0030] Please see Figure 1 This invention provides a technical solution: an adaptive separation and enhancement method for multi-mode ultrasonic guided wave signals, the specific steps of which include:

[0031] S1: Establish and validate baseline status credentials based on historical calibration data associated with specific pipeline segments. Baseline status credentials are used to characterize the inherent acoustic properties of the pipeline segment under ideal operating conditions.

[0032] Verifying the baseline status credential specifically includes: confirming the timeliness and coverage of the historical calibration data upon which the credential relies, thus obtaining the credential validity status; after verifying the baseline status credential, further steps include: calculating the credential maturity, representing the richness of historical data, based on the time span of the historical calibration data; calculating the credential stability, representing its reliability, based on the statistical stability of the historical calibration data under different operating conditions; and weightedly integrating the credential validity status, credential maturity, and credential stability to obtain the comprehensive baseline reliability. A preset baseline availability threshold, T, is used. hresholdRc The preferred value is set to 0.70.

[0033] It should be noted that the availability threshold T hresholdRc The calibration method is as follows: higher T hresholdRc A value (e.g., 0.90) performs rigorous screening, ensuring that only high-quality baseline credentials are used for subsequent analysis. This effectively reduces the risk of false alarms due to outdated or unstable baselines (i.e., misclassifying normal signals as anomalous). However, the trade-off is that baselines of acceptable quality may be frequently rejected, leading to unnecessary recalibration requests and reducing the online availability of the method.

[0034] Lower T hresholdRcA value (e.g., 0.60) is more lenient, accommodating more baseline credentials and ensuring the continuity of the detection task. However, this increases the risk of missed detections due to the adoption of poor-quality baselines, which may mask real defects (i.e., failure to identify real abnormal signals). This embodiment prepares a simulated dataset of "baseline credential heating detection signals" containing various quality states. By combining parameters in S11, S12, and S13 (such as time span, number of collection days, signal standard deviation, etc.), virtual baseline credentials with different comprehensive baseline reliability Rc (ranging from 0.5 to 0.95) are generated. Some credentials are marked as "high-quality and usable," while others are marked as "low-quality and unusable." The simulated detection signals of known states are partly "defect-free" pure signals and partly "defective" signals superimposed with typical weak defect features. In offline data analysis, a baseline availability threshold T is set. hresholdRc The search range is [0.60, 0.85], with a step size of 0.01. For each T... hresholdRc Candidate values: Iterate through all baseline vouchers in the dataset, and select the candidate values ​​based on Rc>T. hresholdRc The rules determine whether to "accept" or "reject" these baseline credentials. For each "accepted" baseline credential, it is used to analyze (e.g., perform differential processing) all "no-defect" and "defective" test signals. The statistical analysis results are used to count the number of false alarms, i.e., the number of times a "low-quality unusable" baseline was "accepted," resulting in a "no-defect" signal being misjudged as abnormal. The count of missed alarms is calculated as the sum of the number of times a "high-quality usable" baseline was incorrectly "rejected" and the number of times a "low-quality unusable" baseline was "accepted," resulting in a "defective" signal not being identified. Each T... hresholdRc The total false alarm rate and total false negative rate under candidate values. Finally, the parameter value that minimizes the weighted sum of the false alarm rate and false negative rate is selected as T. hresholdRc The preferred value is determined by the above calibration process. In this embodiment, T is finally determined through the above calibration procedure. hresholdRc= 0.70 can ensure the availability of the method while keeping the risks of false positives and false negatives to an acceptable minimum. If the overall baseline reliability is greater than or equal to the baseline availability threshold, the current baseline status certificate is determined to be reliable and available; if step S1 is executed successfully, it proceeds normally to step S2 to start acquiring real-time signals for analysis. If the overall baseline reliability is less than the baseline availability threshold, the current baseline status certificate is determined to be of insufficient quality and unavailable, triggering the first warning instruction to be sent to the operation and maintenance platform. The first warning instruction includes: insufficient baseline quality, recalibration is recommended;

[0035] This section elaborates and deepens step S1. It aims to address the key question: is the reliability of the baseline state credentials upon which it relies?

[0036] The calculation of the intrinsic signal compliance index specifically includes: calculating the mode purity, which characterizes the distribution of guided wave energy between the preset target mode and spurious modes; calculating the energy attenuation mismatch coefficient, which characterizes the difference between the actual attenuation rate of the signal and the theoretical attenuation model; and calculating the signal structure complexity entropy, which quantifies the disorder of the guided wave signal waveform structure.

[0037] In this specific embodiment, the application scenario is set as a key section of a long-distance natural gas pipeline located in Northwest China, traversing the Gobi Desert. The current region exhibits typical environmental complexity: diurnal temperature variations can exceed 30°C, and the internal gas pressure fluctuates periodically between 4.5 MPa and 5.5 MPa. The monitoring object is a 100-meter-long DN800 pipeline section. This section includes two standard circumferential welds (located 30 meters and 65 meters from the excitation end, respectively) and a large-curvature elbow (located at 80 meters), all inherent structural features. A ring transducer array consisting of 16 piezoelectric probes is installed at the beginning of the pipeline section to excite and receive the T(0,1) torsional guided wave mode, with the center operating frequency set to 64 kHz. The T(0,1) mode was chosen because it has low energy leakage during propagation in liquid-filled pipelines and is sensitive to circumferential defects.

[0038] During the initial commissioning of the pipeline or after a comprehensive overhaul, continuous calibration data acquisition is conducted for 30 days after confirming that the current 100-meter pipe section is free of macroscopic defects such as corrosion and cracks using an internal detector (set to PIG). This method automatically acquires one set of data per hour, obtaining approximately 720 sets of high signal-to-noise ratio baseline signals. Each set of data not only contains a complete guided wave A-scan signal but also simultaneously records the corresponding environmental and operating parameters: pipeline external wall temperature, denoted as T, and pipeline internal pressure, denoted as P. Each set of data is stored in a historical database and marked as a baseline dataset.

[0039] When this method needs to evaluate newly acquired real-time signals, it first performs step S1 to quantify the reliability of historical baseline credentials, including the following steps:

[0040] S11. Calculate the validity status of the voucher, denoted as V. The specific steps include:

[0041] S111. Obtain the current date of this method and the database creation date of historical calibration data, and calculate the time span between the two, which is recorded as the first time span;

[0042] S112. Obtain the current date of this method and the date of the most recent baseline fast verification operation, and calculate the time span between the two, which is denoted as the second time span;

[0043] S113. Compare the first time span with the preset database timeliness threshold (set to 36 months) to obtain the first comparison result;

[0044] S114. Compare the second time span with the preset recent verification time threshold (set to 6 months) to obtain the second comparison result;

[0045] S115. When the first comparison result is less than or equal to and the second comparison result is less than or equal to, the validity status of the voucher is assigned to the valid status value (set to 1); otherwise, it is assigned to the invalid status value (set to 0).

[0046] The specific principle and purpose of calculating the validity status of the credentials is that the longer the pipeline has been in service, the more likely even minor changes in material properties and stress states may cause the earliest baseline data to become outdated. Step S11 aims to confirm the timeliness of the baseline data. During implementation, two rules are set: 1) The comprehensive calibration collection date (i.e., the database establishment date) should be within the last 36 months. 2) There should be at least one rapid baseline verification operation triggered by maintenance personnel within the last 6 months (the current operation only collects a set of data when the operating conditions are confirmed to be stable and compares it with the historical baseline; if the deviation is less than a threshold, it passes).

[0047] An example of calculating credential validity status includes: Setting the current date to October 1, 2024. The database creation date is found to be May 15, 2023 (within 36 months). The most recent baseline quick verification date is found to be August 10, 2024 (within 6 months). Both rules are met, so the credential validity status V=1 (valid). If either condition is not met, V=0, and an alert to update the baseline is issued to the operator.

[0048] S12. Calculate the voucher maturity level, denoted as M. The specific steps include:

[0049] S121. Obtain the difference between the total number of days of historical calibration data collection and the preset experience days threshold (set to 20 days), and multiply the current difference by the negative value of the preset gain coefficient (set to 0.1) to obtain the exponential term;

[0050] S122. Perform the natural exponential function operation on the exponential term to obtain the exponential result;

[0051] S123. Add the exponent result to the value 1 to obtain the denominator term;

[0052] S124. Perform the division operation between the value 1 and the denominator to obtain the basic maturity factor;

[0053] S125. Obtain the coverage rate of historical calibration data to the preset normal operating condition range, and multiply the coverage rate by the basic maturity factor. The resulting product is the certificate maturity.

[0054] The principle and purpose of calculating credential maturity is that the sample size and the range of operating conditions covered by baseline data determine its representativeness. The richer the sample, the more mature the baseline model, and the more accurate the description of various normal operating conditions. Specifically, the implementation process involves using a non-linear function to quantify maturity, comprehensively considering the total number of days T for data collection. days And coverage Crange.

[0055] M=1 / [1+exp(-k×(T days- T threshold ))]×Crange;

[0056] Where k is the gain coefficient (set to k=0.1), T threshold The threshold for the number of experience days (set T) threshold =20 days). Crange is the operating condition coverage rate, defined as the percentage of area in historical data (temperature, pressure) that covers the preset normal operating condition range (setting, temperature -10~40℃, pressure 4.5~5.5MPa).

[0057] Example of voucher maturity calculation: The data collection period is Tdays = 30 days. Analysis shows that historical data covers 85% of the preset normal operating condition range, i.e., Crange = 0.85. Substituting into the formula: M = 1 / [1 + exp(-0.1 × (30 - 20))] × 0.85 = 1 / (1 + e -1 )×0.85≈0.731×0.85=0.621. Therefore, the maturity of the voucher M=0.621.

[0058] S13. Calculate the stability of the voucher, denoted as S. The specific steps include:

[0059] S131. Select multiple sets of baseline signals from historical calibration data under specific operating conditions (set temperature 25±1℃, pressure 5.0±0.1MPa);

[0060] S132. Extract the peak amplitude of the preset inherent structural feature echo (set the reflection wave packet of the first circumferential weld) in each set of baseline signals;

[0061] S133. Calculate the statistical mean and statistical standard deviation of all extracted peak amplitudes;

[0062] S134. Divide the statistical standard deviation by the statistical mean to obtain the relative standard deviation, denoted as RSD;

[0063] S135. Multiply the relative standard deviation by a preset penalty factor (set to 5) to obtain the penalty term;

[0064] S136. Perform the subtraction operation between the value 1 and the penalty term to obtain the intermediate value of stability;

[0065] S137. Compare the intermediate value of stability with the value 0, and select the larger value as the stability of the certificate.

[0066] The principle and purpose of baseline stability: Even under similar operating conditions, if the baseline signal itself fluctuates greatly, it indicates the presence of an uncertain interference source or instability of the method. Step S13 aims to quantify the consistency of the baseline data itself. The implementation process involves automatically filtering all (N=50 sets) baseline signals from the historical database under the most common operating conditions (set, temperature 25±1℃, pressure 5.0±0.1MPa). The peak amplitude A of the reflected wave packet from the first circumferential weld in these signals is extracted as a stable structural characteristic echo. The relative standard deviation (RSD) of these N amplitudes is calculated. RSD = σA / μA (standard deviation / mean); σA represents the peak amplitude values ​​A1 to A2 of these 50 sets. 50 The sample standard deviation; μA is the peak amplitude values ​​from A1 to A1 in these 50 groups. 50 The arithmetic mean of the RSD; stability S is defined as S=max(0,1-α×RSD), where α is a penalty factor (set to α=5) to amplify the effect of RSD.

[0067] An example of calculating document stability is as follows: 50 baseline signals are selected, and the echo amplitude of the first weld seam is extracted. The calculated mean μA = 0.92V, and the standard deviation σA = 0.04V. Substituting into the formula, RSD = 0.04 / 0.92 ≈ 0.0435. S = max(0, 1 - 5 × 0.0435) = max(0, 1 - 0.2175) = 0.7825. Therefore, the document stability S = 0.7825.

[0068] S14. The comprehensive baseline reliability, denoted as Rc, is calculated using a weighted summation method. The document validity status V is multiplied by a first preset weight to obtain the first weighted term; the document maturity M is multiplied by a second preset weight to obtain the second weighted term; the document stability S is multiplied by a third preset weight to obtain the third weighted term; the first, second, and third weighted terms are then added together to obtain the comprehensive baseline reliability. The weights are set empirically, with validity typically used as the threshold and given the highest weight.

[0069] Rc = wv × V + wm × M + ws × S; Set weights: wv = 0.5, wm = 0.2, ws = 0.3 (the sum of the weights is 1). Substituting the calculated values: Rc = 0.5 × 1 + 0.2 × 0.621 + 0.3 × 0.7825 = 0.5 + 0.1242 + 0.23475 = 0.85895; finally, the overall baseline reliability of the current pipe section is Rc ≈ 0.86.

[0070] Through the above steps, this method has completed a comprehensive verification of the baseline state credentials and obtained a quantified reliability score of 0.86. This is a high value, indicating that the method can highly rely on the current historical baseline data. This Rc value will serve as a key trust parameter in subsequent steps, used to adjust the weights of the dynamic operating condition baseline or as one of the input states to the reinforcement learning model, thereby achieving true adaptability throughout the evaluation and enhancement process.

[0071] This embodiment introduces a baseline credential quantitative evaluation mechanism based on "verification before use." Its core idea is to conduct a multi-dimensional and methodological review and scoring of the "credibility" of the baseline data itself, serving as a reference, before performing any real-time signal analysis. This fundamentally solves the root problem in existing technologies where reliance on a baseline of unknown quality or that has already failed leads to unreliable subsequent analysis results. By decomposing the abstract "baseline quality" into calculable indicators (V, M, S) and merging them into a single comprehensive reliability Rc, this method provides a clear and quantitative decision-making basis for whether to adopt the current baseline, eliminating the subjectivity and uncertainty of human judgment and ensuring the accuracy of the analysis results from the source. Based on a preset availability threshold, the method can automatically determine whether the baseline credential is qualified. When the baseline quality is insufficient, it does not blindly continue erroneous analysis but proactively triggers early warning instructions, prompting maintenance personnel to perform necessary maintenance or data recalibration. This proactive quality control mechanism identifies risks at the beginning of the analysis, improving the overall intelligence level of the monitoring method. By ensuring that the analysis is performed against a high-quality, highly relevant baseline, false difference signals caused by mismatch between the baseline and real-time operating conditions are effectively avoided, directly reducing the false alarm rate. At the same time, a stable and complete baseline provides a cleaner background for capturing weak, genuine defect signals, indirectly reducing the risk of missed detections.

[0072] Example 2

[0073] S2: Acquire the real-time multimodal pipeline signal to be evaluated, collected by the sensor;

[0074] S1 and S2 are the data preparation stages. S3 and S4 are the core analysis stages, providing quantitative evidence for the decision in S5. S5 is the final decision output stage, and its adaptive nature depends entirely on the deviation calculated in S4. This is a progressive process with a clear causal relationship.

[0075] This section elaborates and deepens step S1. It aims to address a key issue: the reliability of the baseline state credentials used. The innovation lies in employing an intelligent fusion method based on probabilistic reasoning (Bayesian networks) instead of a simple weighted summation. This makes the fusion process more scientific, capable of handling nonlinear relationships and uncertainties between indicators, resulting in a more reliable conformity index.

[0076] The higher the reliability of the integrated baseline, the more accurate, stable and reliable the historical acoustic model of the current pipe section is judged by this method, and the better the prior knowledge of signal processing is. The lower the intrinsic signal compliance index, the more severe the mode mixing, abnormal attenuation or structural distortion of the currently acquired guided wave signal is judged by this method, the worse the signal quality is, and the higher the potential interference or defect risk is.

[0077] S3: Based on the physical correlation within the real-time multimodal pipeline signal, calculate the intrinsic signal conformity index to quantify its deviation from the theoretical physical model;

[0078] The calculation of the intrinsic signal compliance index specifically includes: calculating the mode purity, which characterizes the distribution of guided wave energy between the preset target mode and spurious modes; calculating the energy attenuation mismatch coefficient, which characterizes the difference between the actual attenuation rate of the signal and the theoretical attenuation model; and calculating the signal structure complexity entropy, which quantifies the disorder of the guided wave signal waveform structure.

[0079] Symmetric mode S0, antisymmetric mode A0, or longitudinal mode L(0,m), torsional mode T(0,m), and bending mode F(n,m) in the pipe can coexist and propagate simultaneously.

[0080] The signal structure complexity entropy includes: based on wavelet packet transform of the guided wave signal, jointly quantifying the uniformity of its energy distribution and the degree of disorder of its instantaneous characteristics in each frequency band; taking mode purity, energy attenuation mismatch coefficient and signal structure complexity entropy as inputs, and performing probabilistic inference fusion through a pre-trained Bayesian network to obtain the posterior probability distribution of the intrinsic signal conformity index.

[0081] Wavelet packet transform is a powerful time-frequency analysis tool that can finely characterize the structure of a signal at different frequencies, making it very suitable for calculating complexity entropy. 2) Bayesian networks are probabilistic graphical models for handling uncertain problems. They are well-suited for fusing indicators (including patterns, energy, and structure) from different sources with varying meanings, and providing a more reliable comprehensive evaluation result with probabilistic significance.

[0082] The signal structure complexity entropy is obtained by statistical learning of labeled signal samples containing different types of defects or noise in historical databases; wavelet packet energy spectrum analysis is performed on the signal, and Shannon information entropy calculation logic is applied.

[0083] In this embodiment, the intrinsic conformity assessment of the signal is performed, following the scenario and conclusion of Embodiment 1. In Embodiment 1, this method confirmed that the comprehensive baseline reliability Rc of the historical baseline certificate was 0.86, which is higher than the availability threshold of 0.70, so the process normally proceeds to S2. A scenario is set where, at a certain moment (e.g., 15:00 on October 2, 2024), the pipeline operation center needs to conduct a routine health assessment on the current critical pipe section. At this moment, the pipe outer wall temperature T is 28℃, and the internal pressure P is 5.2MPa. The current array consists of 16 piezoelectric probes, which are equidistantly distributed on the circumference of the pipe. During the acquisition process, the voltage time-domain waveform received by each probe over a period of time is recorded sequentially, forming a real-time multimodal pipeline signal, denoted as S. raw (t,θ); is called the real-time multimodal pipeline signal because, in addition to the desired T(0,1) target mode, the signal also contains L(0,2) longitudinal mode and F(1,3) bending mode with a certain amount of energy as the main spurious modes. S raw (t,θ) represents the real-time multimodal pipeline signal to be evaluated. From a data structure perspective, the real-time multimodal pipeline signal S... raw (t,θ) represents a two-dimensional data matrix, which consists of the original signals in the time and spatial domains. Here, t represents the time axis, corresponding to the sampling time point of each probe; θ represents the spatial axis, corresponding to the angular position of each of the 16 probes on the circumference of the pipe.

[0084] The signal structure complexity entropy includes: based on wavelet packet transform of the guided wave signal, jointly quantifying the uniformity of its energy distribution and the degree of disorder of its instantaneous characteristics in each frequency band; taking mode purity, energy attenuation mismatch coefficient and signal structure complexity entropy as inputs, and performing probabilistic inference fusion through a pre-trained Bayesian network to obtain the posterior probability distribution of the intrinsic signal conformity index.

[0085] Wavelet packet transform is a powerful time-frequency analysis tool that can finely characterize the structure of a signal at different frequencies, making it very suitable for calculating complexity entropy. 2) Bayesian networks are probabilistic graphical models for handling uncertain problems. They are well-suited for fusing indicators (including patterns, energy, and structure) from different sources with varying meanings, and providing a more reliable comprehensive evaluation result with probabilistic significance.

[0086] The signal structure complexity entropy is obtained by statistical learning of labeled signal samples containing different types of defects or noise in historical databases; wavelet packet energy spectrum analysis is performed on the signal, and Shannon information entropy calculation logic is applied.

[0087] In acquiring real-time multimodal pipeline signal S raw After (t,θ), S3 is executed immediately, which quantifies the deviation of the current real-time signal from the ideal physical model by calculating three sub-indices.

[0088] S31. Calculate the purity of the model, denoted as... mode The specific steps include:

[0089] S311. Using two-dimensional Fourier transform (2D-FFT) technology, the real-time multimodal pipeline signal S raw (t,θ) is transformed from the spatiotemporal domain to the frequency and wavenumber domains to obtain the energy distribution spectrum S(f,k). On the fk spectrum, different guided wave modes will fall on their respective theoretical dispersion curves. f represents the frequency, and k represents the wavenumber.

[0090] Please refer to Figure 2 The vertical axis (Y-axis) represents frequency (f), measured in Hertz (Hz). It indicates the speed of wave oscillation. The horizontal axis (X-axis) represents wave number (k), measured in meters (m⁻¹). It indicates the spatial density of the wave (inversely proportional to wavelength). Color / brightness in the graph represents energy intensity. The brighter or redder the color at a point (f,k), the stronger the signal energy at the given combination of frequency and wave number.

[0091] S312. Extract the energy distribution spectrum, identify the theoretical dispersion curve of the target mode, and one or more main non-target modes' theoretical dispersion curves. Non-target modes include longitudinal modes or bending modes. As mentioned above, in addition to the desired T(0,1) target mode, the signal also contains L(0,2) longitudinal mode and F(1,3) bending mode with a certain amount of energy as main spurious modes. Figure 2 The system identifies at least three curves: one is the dispersion curve of the target mode T(0,1), and the other two are the dispersion curves of the expected main non-target modes (spurious modes) L(0,2) and F(1,3).

[0092] S313. The target mode energy is calculated by integrating the energy distributed along the theoretical dispersion curve of the target mode on the energy distribution spectrum; the respective energy integration regions are delineated on the energy distribution spectrum. The energy of each non-target mode, including longitudinal mode energy and bending mode energy, is calculated by integrating the energy distributed along the theoretical dispersion curve of each non-target mode on the energy distribution spectrum; these are then summed to obtain the total non-target mode energy, denoted as E. spurious ; E spurious1 Represents the longitudinal mode energy, corresponding to the L(0,2) longitudinal energy, E spurious2 This represents the bending mode energy, corresponding to the F(1,3) bending mode energy;

[0093] S314. Calculate the target mode energy E target and target mode energy E target Total non-target mode energy E spurious The ratio of the sum to obtain the pattern purity P. mode ; ;

[0094] S32. Calculate the energy attenuation mismatch coefficient, denoted as Catt. The specific steps include:

[0095] The principle and purpose of calculating the energy attenuation mismatch coefficient is that, in a uniform pipe without newly formed defects, the energy of a specific guided wave mode will follow a predictable exponential attenuation law as the propagation distance increases. This law is primarily determined by the pipe material, geometry, and the guided wave mode itself. This step aims to quantify the consistency of the signal in the energy propagation dimension by comparing the actual attenuation characteristics of the real-time signal with the theoretical attenuation model. Any significant deviation (i.e., mismatch) may indicate anomalies such as diffuse corrosion, coating peeling, or changes in the medium along the pipe.

[0096] S321, from the real-time multimodal pipeline signal S raw In (t,θ), a phased array synthesis algorithm is used to extract a pure one-dimensional target mode signal. Based on the physical characteristics of the target mode T(0,1) (its wave structure is axially symmetric in the circumferential direction, i.e., phase consistent), the phased array synthesis algorithm is used to analyze the signals S received by 16 probes. raw (t,θ), where i=1-16, are processed. The 16 time-domain signals are directly added and averaged at each time point t to obtain the synthesized one-dimensional signal S. synth (t), mathematically represented as ;θ iThe angle of the i-th probe on the spatial axis is represented by . Signals received from probes 1 to 16 are averaged. During the synthesis process, the axisymmetric T(0,1) mode signal is effectively enhanced (due to their in-phase superposition), while non-target modes such as the non-axisymmetric bending mode F(1,3) are suppressed (due to their phase difference, they cancel each other out). The final result is a one-dimensional time-domain signal with a higher signal-to-noise ratio, primarily containing T(0,1) mode information.

[0097] S322, From the synthesized one-dimensional signal S synth In (t), the echoes of two inherent structural features are located and identified based on the signal flight time. In this embodiment, these two features are the first circumferential weld reflection wave packet located at a distance of d1=30 meters from the excitation end, and the second circumferential weld reflection wave packet located at d2=65 meters.

[0098] Perform a Hilbert transform on each wave packet to obtain its envelope, and extract the peak amplitude from the envelope. Record the peak amplitude of the first weld echo as A1; record the peak amplitude of the second weld echo as A2.

[0099] S323. Calculate the actual attenuation coefficient α between the two echoes according to the Lambert-Beer law. actual The unit is Np / m, nappers per meter, and the specific formula is as follows: The 2 in the denominator is because the signal traveled a round trip distance.

[0100] S324. Based on the current operating conditions (pipe material, geometry, internal medium, T(0,1) mode, center frequency 64kHz, temperature 28℃, pressure 5.2MPa), retrieve the corresponding theoretical attenuation coefficient α from the pre-built multiphysics coupling attenuation database stored within the method. theory .

[0101] S325 (Mismatch Coefficient Calculation): Calculate the normalized difference between the actual and theoretical values ​​to obtain the energy attenuation mismatch coefficient C. att C att It is designed as a positive indicator; the closer it is to 1, the better the match.

[0102] Data Example: From the synthesized signal S synth In (t), the echo amplitude of the first weld was measured to be A1 = 0.58V, and the echo amplitude of the second weld was measured to be A2 = 0.41V. α actual =(ln(0.58)-ln(0.41)) / (2×(65-30))=0.345 / 70≈0.00493Np / m. Furthermore, the theoretical attenuation coefficient α under the current operating conditions is obtained by querying the database. theory=0.00450 Np / m. Substitute into the formula to calculate C. att =max(0,1-|0.00493-0.00450| / 0.00450)=max(0,1-0.0956)≈0.904.

[0103] S33. Calculate the signal structure complexity entropy, denoted as E. comp The specific steps include:

[0104] S331, Using the enhanced one-dimensional composite signal S generated in step S321 synth (t). For the synthesized signal S synth (t) Perform 4-level wavelet packet decomposition. The db4 wavelet basis decomposition is selected. The current decomposition process will synthesize the one-dimensional signal S. synth The frequency band of (t) is divided layer by layer, eventually yielding 2 4 =16 non-overlapping sub-bands. Synthetic signal S synth (t) is decomposed into 16 corresponding wavelet packet coefficient sequences, which represent the components of the signal in these 16 sub-bands.

[0105] S332. For the j-th sub-band (j=1-16), calculate its energy E. j The energy calculation method is as follows: square each coefficient in the wavelet packet coefficient sequence of the current sub-band, and then sum all the squared values. ;

[0106] h is the time series index in the coefficient sequence; c jh Represents the dimensional synthesized signal S synth (t) is the amplitude at the j-th sub-band and the h-th time point, in volts;

[0107] S333, Calculate the total energy (j ranges from 1 to 16). Then, calculate the proportion of energy p of each sub-band to the total energy. j Calculate the proportion of the j-th sub-band; ;

[0108] S334. Calculate the entropy value H(p) of the current energy distribution according to Shannon's information entropy formula. Furthermore, 0 × log2(0) = 0;

[0109] S335. Divide the calculated entropy value H(p) by its maximum possible value log2(16) = 4 (the maximum entropy is taken when the energy is completely uniformly distributed across the 16 subbands). This scales its range to [0,1]. Then, invert the normalized entropy value H(p) to obtain the signal structure complexity entropy E. comp ; Since a higher entropy value indicates a more chaotic signal (higher non-uniformity), to make it a positive indicator (a higher value indicates a better structure and lower complexity), the normalized entropy value is subtracted from 1. The principle and purpose of calculating signal structural complexity entropy is that the energy of an ideal guided wave signal composed of a single mode should be highly concentrated near the center frequency, and the signal waveform structure should be relatively simple and ordered. Conversely, when a signal contains multiple modes, multipath reflections, or is subject to broadband noise interference, its energy will disperse over a wider frequency band, causing the signal waveform to become complex, chaotic, and unpredictable in the time domain. This step uses wavelet packet energy entropy to quantify the disorder of this structure. Data example: For the synthesized signal S... synth Wavelet packet decomposition was performed on (t) to calculate the energy proportions of the 16 sub-bands p1,...,p16. The calculated Shannon entropy value is H(p) = 2.85. comp =max(0,1-2.85 / 4)=max(0,1-0.7125)=0.2875. This lower E comp The value of 0.2875 intuitively indicates that the waveform structure of the current real-time signal is relatively complex and disordered, which is quite different from the ideal simple waveform.

[0110] The principle of this embodiment lies in constructing a multi-dimensional, physically driven intrinsic quality assessment framework for real-time guided wave signals, namely, calculating the intrinsic signal conformity index. This principle no longer views the signal in isolation, but rather deconstructs and quantifies it jointly from three complementary physical perspectives: First, it quantifies mode purity directly in the frequency and wavenumber domains using a two-dimensional Fourier transform, fundamentally measuring the degree of deviation of the signal from the ideal physical model of "single-mode propagation." Second, it calculates the energy attenuation mismatch coefficient by comparing actual and theoretical attenuation coefficients, verifying whether the signal follows established physical laws of energy dissipation during propagation, thus possessing the ability to detect diffusion anomalies. Third, it calculates the signal structural complexity entropy using wavelet packet energy entropy, quantifying the disorder and chaos of the signal waveform from an information theory perspective, reflecting the comprehensive distortion effect after the superposition of multiple interference factors. The core innovation of this invention lies in the fact that it does not use simple linear weighting to integrate these three indicators, but instead introduces a pre-trained Bayesian network for probabilistic inference fusion. This nonlinear fusion method can more scientifically handle the complex dependencies and uncertainties between various indicators, and output a compliance index with clear probabilistic meaning, which has a much higher credibility than traditional scoring mechanisms.

[0111] The beneficial effect of this embodiment is that the method no longer processes a signal of unknown quality "blindly," but can a priori and accurately "diagnose" the specific condition of the current signal in terms of mode aliasing, energy attenuation, and structural distortion. This deep signal insight enables subsequent signal separation and enhancement strategies to truly achieve adaptive adjustment "tailored to the specific needs," thereby effectively improving the accuracy of defect identification and the overall reliability of the method under complex working conditions.

[0112] Example 3

[0113] S4: Based on the baseline status certificate, combined with real-time environment and operating parameters, construct a dynamic operating condition baseline, and combine the intrinsic signal compliance index to calculate the environmental operating condition deviation degree, which characterizes the degree of deviation between the real-time signal and the dynamic operating condition baseline.

[0114] S5: Based on the environmental operating condition deviation, generate an adaptive enhancement strategy for real-time multimodal pipeline signals and pipeline integrity assessment results.

[0115] The dynamic operating condition baseline includes statistical distribution parameters of historical normal signal characteristics corresponding to the pipe section and the current environmental operating state. The calculation of the environmental operating condition deviation specifically includes: constructing a current signal state feature vector containing intrinsic signal compliance index and comprehensive baseline reliability; determining the dynamic operating condition baseline as the feature mean vector and feature covariance matrix of historical normal signals that match the current operating condition of the pipe section; and applying the Mahalanobis distance algorithm to calculate the distance between the current signal state feature vector and the dynamic operating condition baseline as the environmental operating condition deviation.

[0116] The specific steps for generating an adaptive enhancement strategy include: using environmental condition deviation as a key input, selecting and configuring the optimal combination of signal processing operators; and triggering a preset adaptive diagnostic level based on the evaluation of the environmental condition deviation. Defining the triggering of the preset adaptive diagnostic level specifically involves: using both the environmental condition deviation and the overall baseline reliability as state inputs, feeding them into a pre-trained reinforcement learning model; and having the reinforcement learning model output and execute the optimal enhancement strategy with the highest expected diagnostic accuracy based on the state inputs.

[0117] The state input of the reinforcement learning model further includes pipeline criticality, which represents the importance or risk level of the current pipeline segment; after executing the optimal reinforcement strategy, the reward value is calculated based on the improvement in the signal-to-noise ratio of the enhanced signal or the matching degree with known defect samples.

[0118] The reinforcement learning model is updated online using reward values ​​to optimize its subsequent policy selection. The smaller the environmental condition deviation, the more closely the current signal state matches its normal state under the current environmental conditions, the healthier the signal, and the lower the risk of anomalies. The optimal enhancement policy is a specific combination of actions selected by the reinforcement learning model from a set of discrete actions that includes various filters, compensation algorithms, and reconstruction models.

[0119] In Example 2, S3 has calculated the intrinsic signal compliance index: Pmode=0.841, Catt=0.904, Ecomp=0.2875; and the overall baseline reliability Rc=0.86 is known.

[0120] S4. Calculate the environmental operating condition deviation, denoted as D. env The specific steps include:

[0121] S41. Construct the current signal state feature vector, denoted as V. current The three intrinsic signal conformity indices calculated in S3 are combined with the comprehensive baseline reliability Rc calculated in S1 to form a four-dimensional feature vector, which comprehensively characterizes the current signal's overall state. The current signal state feature vector is denoted as V. current =[P mode C att E comp Based on the previous calculation results, the eigenvector is: V current =[0.841,0.904,0.2875,0.86];

[0122] S42. Set up a historical normal signal feature database. The historical normal signal feature database stores a large number of signals that have been identified as "healthy" in historical periods, along with their corresponding operating parameters and signal state feature vectors. Specific steps include:

[0123] S421. Operating Condition Matching: Set the screening conditions as temperature T within the range of 28±2℃ and pressure P within the range of 5.2±0.3MPa. Retrieve all historical health signal feature vectors that match this operating condition range from the database and summarize them into a historical health feature vector set.

[0124] S422. Perform statistical analysis on the selected N sets of historical health feature vectors and calculate their mean vector μ. baseline The covariance matrix Σ baseline These two together constitute the dynamic baseline under the current operating conditions, describing the central tendency and fluctuation range of normal signal characteristics.

[0125] S43. Calculate the Mahalanobis distance as the environmental operating condition deviation D. env Specifically, the Mahalanobis distance algorithm is applied to calculate the current signal state feature vector V.current To the dynamic operating condition baseline (by μ) baseline and Σ baseline Mahalanobis distance is the distance between features (defined by a multidimensional normal distribution). It takes into account the correlation between features and is a scale-independent distance metric. Its calculation formula is:

[0126]

[0127] in, It is the inverse of the covariance matrix, and T denotes the vector transpose. D env The larger the value, the further the current signal state deviates from the normal distribution.

[0128] The principle and purpose of environmental condition deviation measurement are as follows: the "health" of a signal is relative and must be compared with the normal state under the current operating conditions. This step aims to construct a "dynamic operating condition baseline" that perfectly matches the real-time operating conditions (temperature 28℃, pressure 5.2MPa), and then accurately quantify the degree of deviation between the current signal state and the current baseline through multidimensional statistical distance (Mahaviron distance). This deviation measure D... env It is a comprehensive indicator that takes into account both the intrinsic quality of the signal and the reliability of historical baselines, and can more robustly reflect the true health status of the signal.

[0129] Data Example: After database filtering and statistical calculation, the dynamic baseline parameters under the current operating conditions are as follows:

[0130] Mean vector μ baseline =[0.92,0.95,0.85,0.90] (representing the characteristics of an ideal health signal);

[0131] covariance matrix Σ baseline A 4x4 matrix;

[0132] V current μ baseline and Σ baseline Substituting into the formula, we obtain the environmental condition deviation D. env =2.58.

[0133] S44. Set the first and second grade thresholds. The preferred value for the first grade threshold is 1.5. When the environmental condition deviation is <1.5, it indicates that the current signal state is highly matched with the dynamic condition baseline, and the comprehensive feature vector V of the signal is... current It falls within the high confidence interval of the normal statistical distribution; this indicates that all the characteristic indicators of the current signal are in line with its expected performance under the current temperature, pressure and other operating conditions, and no abnormal signs are found, so the first label is generated;

[0134] The preferred value for the second grading threshold is set to 3.0; when 1.5 ≤ Denv A value <3.0 indicates a moderate deviation between the current signal state and the dynamic operating condition baseline, and the signal's comprehensive feature vector V... current This exceeds the normal fluctuation range, indicating a potential risk, which could be the formation of early micro-defects, slight drift in sensor performance, or mild noise interference in the environment, resulting in a second label indicating a moderate deviation.

[0135] When D env A value ≥3.0 indicates a significant deviation between the current signal state and the dynamic operating condition baseline, and the signal's comprehensive feature vector V current It has fallen into the low probability region of the normal statistical distribution, indicating the existence of a high probability of an abnormal event and the risk of external interference sources. The third label generated is severe deviation.

[0136] It should be noted that the first and second level thresholds together constitute the core of risk assessment in this method. Their settings are mutually restrictive: if the threshold is too low, the method will be overly sensitive, easily misreporting normal operating condition fluctuations as risks; if the threshold is too high, it may be slow to respond to early minor defects, resulting in missed detections. Therefore, the optimal values ​​of these two parameters are not determined in isolation, but are obtained through joint calibration experiments using the following method: Constructing a calibration dataset: In this embodiment, an offline simulation signal dataset containing three pre-labeled categories is prepared, and the corresponding environmental operating condition deviation D has been calculated for each group of signals. env :

[0137] The "Normal State" dataset contains defect-free signals collected under various normal operating conditions.

[0138] The "Early Minor Defects" dataset was obtained through finite element simulation or by machining minute artificial defects (such as grooves less than 5% of the wall thickness) on test pipes. env The value usually shows a slight increase.

[0139] The "Significant Structural Anomalies" dataset corresponds to clearly defined, noteworthy defect signals, with Denv values ​​significantly deviating from the normal range. The first grading threshold is calibrated to maximize the correct detection rate of "early, minor defects" while controlling the false alarm rate of "normal state" signals to a preset, acceptable low level (e.g., below 2%). This is achieved through D... env Receiver operation characteristic analysis is performed on the data distribution to find the optimal balance point that satisfies the current constraints. After calibration, the preferred value for the current threshold is 1.5.

[0140] The second grading threshold was calibrated as follows: Based on the established first threshold (1.5), a rigid objective was to ensure a detection rate of over 99.5% for "significant structural anomalies," while minimizing the proportion of "early minor defects" mistakenly upgraded to high-risk. By analyzing the distribution of data with Denv values ​​exceeding 1.5, the dividing point that most effectively distinguishes between "early" and "significant" defects was identified. After calibration, the optimal value for the current threshold was determined to be 3.0.

[0141] S5. Generate adaptive enhancement strategies and pipe segment integrity assessments; based on the comprehensive deviation D calculated in the previous step. env This embodiment uses a pre-trained reinforcement learning model to complete the decision-making process.

[0142] First, historical and simulation databases are collected to train the reinforcement learning model. The historical database includes real detection signals accumulated over many years with clear labels ("healthy", "XX mm corrosion defect", "XX type crack", "sensor failure") and their corresponding operating parameters. The simulation database includes a large number of simulation signals generated using techniques such as finite element analysis (FEA) under various operating conditions, defect types, and sizes. Then, with the help of experts in the field, the action space in the reinforcement learning model is predefined to guide the model to prioritize learning tasks related to integrity assessment before being applied to step S5, which specifically includes:

[0143] S51. The built-in pre-trained reinforcement learning model defines the input information upon which the model's decisions are based. In this embodiment, the state vector SRL includes three dimensions:

[0144] 1. Environmental operating condition deviation D env : Core input, reflecting the degree of signal abnormality.

[0145] 2. Overall baseline reliability Rc: Reflects the reliability of historical data.

[0146] 3. Pipeline Criticality A crit The preset static parameters characterize the importance of the current pipe segment, including: 1 represents the first level of normal area, 2 represents the second level of sensitive area, 3 represents the third level of important area, 4 represents the fourth level of critical area, and 5 represents the fifth level of high-consequence crossing area. The specific identification is as follows:

[0147] Acrit=1 indicates that the pipeline section is located in a sparsely populated remote area, open farmland, or wasteland. There are no important ecological protection targets or large-scale infrastructure in the area. Acrit=2 indicates that the pipeline section is adjacent to a small town, a general water source protection area (such as a non-drinking grade river), or crosses a minor transportation route (such as a rural road). Acrit=3 indicates that the pipeline section is located in the suburbs of an city, an industrial park, or crosses an important transportation artery (such as a national highway or railway). Acrit=4 indicates that the pipeline section is located in a densely populated residential area, a commercial area, or near schools, hospitals, or other high-density population places, or crosses a large drinking water source or an important river. Acrit=5 indicates that the pipeline section is located in the city center (CBD), a major transportation hub (such as under a subway), or crosses an irreplaceable key ecological protection area or a national first-level drinking water source protection area.

[0148] The predefined action space defines the set of discrete operations for selecting the reinforcement learning model, including:

[0149] B0 Strategy: Maintain the status quo; applicable to D. env In extremely rare cases.

[0150] Strategy B1: Applicable to high-frequency noise interference.

[0151] B2 Strategy: Matched Filtering - Enhances weak defect signals of specific shapes.

[0152] B3 Strategy: Mode Compensation Algorithm - Corrects signal distortion caused by mode conversion.

[0153] The reward function R is used to evaluate the performance of a policy across a discrete set of operations during training and online updates. R = w1 × SNRScore + w2 × DScore, where SNRScore represents the signal-to-noise ratio improvement score and DScore represents the signal clarity improvement score; and w1 and w2 are weighting coefficients.

[0154] The signal-to-noise ratio improvement score is obtained by dividing the ratio of the signal power to the noise power of the processed signal by the ratio of the signal power to the noise power of the unprocessed signal. The quotient reflects the actual gain of the signal-to-noise ratio.

[0155] The signal clarity improvement score is obtained by determining the ratio of the increment of the signal-to-noise ratio improvement score relative to the preset lower limit of improvement to the range defined by the preset upper limit of improvement and the preset lower limit of improvement. The ratio is used to linearly map the signal-to-noise ratio improvement score to a standardized score range.

[0156] S52. Construct the State Vector SRL, setting the current segment criticality Acrit=4. Combine the current state information into the SRL as the state input for the reinforcement learning model: SRL=[D env ,Rc,Acrit =[2.58,0.86,4]. The state vector SRL is input into the pre-trained reinforcement learning model. The reinforcement learning model outputs a Q-value for each possible policy (B0-B3), where the Q-value represents the predicted reward. The policy with the largest Q-value is selected as the optimal reinforcement policy. Assuming the model outputs Q-values ​​of Q(SRL,B0)=1.2, Q(SRL,B1)=3.5, Q(SRL,B2)=5.8, and Q(SRL,B3)=4.1, then selecting the largest Q-value means that the reinforcement learning model selects policy B2: matched filtering is performed, and the method automatically calls the matched filtering module to process the original signal S. raw (t,θ) or the synthesized signal S synth (t) is filtered to generate the enhanced signal S. enhanced (t).

[0157] S53, Based on D env The second label, with a moderate deviation of 2.58, and the reinforcement learning model executing the B2 strategy, automatically generated the first execution result, which included a signal state showing a moderate deviation from the normal operating baseline, indicating a potential anomaly risk. Matched filtering enhancement processing using the B2 strategy was performed to improve the detectability of weak defect signals. The pipe section integrity conclusion was that no clear, severe defect response was found, shortening the next inspection cycle to 3 months.

[0158] S54. After executing strategy B2 in S52, the method quantifies the signal enhancement effect to calculate the reward value R. Assume that the signal-to-noise ratio (SNR) improvement after strategy B2 is 4 dB. Based on the reward function and its component calculation method defined in S51, calculate the SNR improvement score (SNRScore): the current value is the quotient of the linear SNR of the signal before and after processing. A 4 dB improvement corresponds to 10... (4 / 10) A 10x improvement in linear signal-to-noise ratio, therefore SNRScore = 10. 0.4 ≈2.51.

[0159] Calculate the signal clarity improvement score DScore, mapping a 4dB improvement value to a standard score range. Assuming the preset lower limit of improvement is 0.5dB, the upper limit is 8.0dB, and the score range is [0,10], then DScore = 10 × (4 - 0.5) / (8.0 - 0.5) ≈ 4.67.

[0160] Based on the reward function R=w1×SNRScore+w2×DScore defined in S51, and substituting the preset weights w1=0.8 and w2=0.2, the final reward value of this policy execution is obtained as: R=0.8×2.51+0.2×4.67=2.008+0.934=2.942.

[0161] S55. Constructing Experience Tuples and Performing Online Model Updates. The method encapsulates the entire decision-making and evaluation process into an experience tuple (S, B, R, S'), including the current state S (i.e., SRL=[2.58, 0.86, 4]), the action B (i.e., policy B2), the reward R (i.e., 2.942), and the subsequent new state S' (i.e., the new state vector observed after the action). The constructed experience tuple is stored in a pre-defined experience replay memory. When the pre-defined update conditions are met, the method randomly samples a batch of experience tuples from the experience replay memory. The network parameters of the reinforcement learning model are updated using gradient descent, with the goal of minimizing the temporal difference error between the predicted Q-value and the target Q-value, continuously learning from real-world application data, and self-optimizing its decision-making strategy.

[0162] The technical principle of this embodiment is to construct a dynamic operating condition baseline (defined by the mean vector and covariance matrix) that matches the real-time operating conditions, and to use the Mahalanobis distance algorithm to uniformly map the multi-dimensional features obtained in the previous steps (including the intrinsic quality of the signal and the reliability of the historical baseline) into a comprehensive environmental operating condition deviation D. env The core principle is that when evaluating signal anomalies, not only are the deviations of each feature considered, but also the statistical correlation between features, thus providing a more robust and accurate quantification of the degree of "anomaly".

[0163] A further innovation lies in the fact that this embodiment does not employ rigid rules to respond to this deviation, but instead treats it as a core state, inputting it along with baseline reliability and segment criticality into a pre-trained reinforcement learning decision model. Based on strategies learned from massive amounts of historical and simulation data, the reinforcement learning decision model can dynamically and objectively (i.e., maximizing expected diagnostic accuracy) select the optimal combination of signal enhancement strategies. This process transcends the logic of traditional single-decision approaches, achieving true context awareness and intelligent decision-making.

[0164] The beneficial effects of this embodiment are that it achieves a high degree of automation and continuous self-optimization of the detection process. The method not only automatically matches and executes the most appropriate enhancement strategies and diagnostic conclusions based on the real-time anomaly level of the signal, the reliability of historical data, and the risk level of the pipeline segment, but also, through its built-in reward feedback and online update mechanism, allows its decision-making capabilities to continuously evolve and improve over time. This closed loop from "diagnosis" to "action" to "learning" transforms the entire detection method from a static analysis tool into an intelligent method capable of adapting to the environment, accumulating experience, and continuously improving its performance.

[0165] It should be noted that all calculation formulas in this application employ regression analysis, including but not limited to machine learning algorithms, to deeply analyze the collected parameters and identify their natural trends and interrelationships. Specialized software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Then, cross-validation and other methods are used to objectively evaluate the model performance, and continuous feedback and optimization are combined to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their effectiveness and accuracy. In all calculation formulas in this application, the parameters in each formula undergo dimensionless processing within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless processing techniques include, but are not limited to, min-max-normalization and Z-score standardization.

[0166] The algorithm of this invention is implemented as a Python script. Before executing the core logic, the program first executes a data loading module (e.g., using the widely used pandas library in Python), configured to read the aforementioned spreadsheet file and load its contents into the program's working memory (e.g., a DataFrame data structure). Subsequent algorithm steps will directly query and retrieve the required configuration parameters from this memory data structure.

[0167] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An adaptive separation and enhancement method of ultrasonic guided wave multi-modal signals, characterized in that, The specific steps include: S1: Establish and validate baseline status credentials based on historical calibration data associated with specific pipeline segments. Baseline status credentials are used to characterize the inherent acoustic properties of the pipeline segment under ideal operating conditions. S2: Acquire the real-time multimodal pipeline signal to be evaluated, collected by the sensor; S3: Based on the physical correlation within the real-time multimodal pipeline signal, calculate the intrinsic signal conformity index to quantify its deviation from the theoretical physical model; S4: Based on the baseline status certificate, combined with real-time environment and operating parameters, construct a dynamic operating condition baseline, and combine the intrinsic signal compliance index to calculate the environmental operating condition deviation degree, which characterizes the degree of deviation between the real-time signal and the dynamic operating condition baseline. S5: Based on the environmental operating condition deviation, generate an adaptive enhancement strategy for real-time multimodal pipeline signals and pipeline integrity assessment results.

2. The adaptive separation and enhancement method for ultrasonic guided wave multi-mode signals according to claim 1, characterized in that: Verifying the baseline status certificate specifically includes: confirming the timeliness and coverage of the historical calibration data on which the certificate relies to obtain the certificate validity status; after verifying the baseline status certificate, it further includes: calculating the certificate maturity, which characterizes the richness of historical data, based on the time span of the historical calibration data; calculating the certificate stability, which characterizes the reliability, based on the statistical stability of the historical calibration data under different operating conditions; and weightedly integrating the certificate validity status, certificate maturity, and certificate stability to obtain the comprehensive baseline reliability.

3. The adaptive separation and enhancement method for ultrasonic guided wave multi-mode signals according to claim 1, characterized in that: The calculation of the intrinsic signal compliance index specifically includes: calculating the mode purity, which characterizes the distribution of guided wave energy between the preset target mode and spurious modes; calculating the energy attenuation mismatch coefficient, which characterizes the difference between the actual attenuation rate of the signal and the theoretical attenuation model; and calculating the signal structure complexity entropy, which quantifies the disorder of the guided wave signal waveform structure.

4. The adaptive separation and enhancement method for ultrasonic guided wave multi-mode signals according to claim 1, characterized in that: The signal structure complexity entropy includes: based on wavelet packet transform of the guided wave signal, jointly quantifying the uniformity of its energy distribution and the degree of disorder of its instantaneous characteristics in each frequency band; taking mode purity, energy attenuation mismatch coefficient and signal structure complexity entropy as inputs, and performing probabilistic inference fusion through a pre-trained Bayesian network to obtain the posterior probability distribution of the intrinsic signal conformity index.

5. The adaptive separation and enhancement method for ultrasonic guided wave multi-mode signals according to claim 4, characterized in that: Historical databases are collected by statistical learning of labeled signal samples containing different types of defects or noise. The signal structure complexity entropy is obtained by performing wavelet packet energy spectrum analysis on the signal and applying Shannon information entropy calculation logic.

6. The adaptive separation and enhancement method for ultrasonic guided wave multi-mode signals according to claim 5, characterized in that: The higher the reliability of the integrated baseline, the more accurate, stable and reliable the historical acoustic model of the current pipe section is judged by this method, and the better the prior knowledge of signal processing is. The lower the intrinsic signal compliance index, the more severe the mode mixing, abnormal attenuation or structural distortion of the currently acquired guided wave signal is judged by this method, the worse the signal quality is, and the higher the potential interference or defect risk is.

7. The adaptive separation and enhancement method for ultrasonic guided wave multimode signals according to claim 6, characterized in that: The dynamic operating condition baseline includes statistical distribution parameters of historical normal signal characteristics corresponding to the operating status of the pipeline section and the current environment; The calculation of environmental condition deviation specifically includes: constructing a current signal state feature vector that includes intrinsic signal compliance index and comprehensive baseline reliability; The dynamic operating condition baseline is determined as the characteristic mean vector and characteristic covariance matrix of historical normal signals that match the current operating conditions of the pipe section. The Mahalanobis distance algorithm is applied to calculate the distance between the current signal state characteristic vector and the dynamic operating condition baseline, which is used as the environmental operating condition deviation.

8. The adaptive separation and enhancement method for ultrasonic guided wave multimode signals according to claim 7, characterized in that: The specific steps for generating an adaptive enhancement strategy include: using environmental condition deviation as a key input, selecting and configuring the optimal combination of signal processing operators; and triggering a preset adaptive diagnostic level based on the evaluation of the environmental condition deviation. Defining the triggering of the preset adaptive diagnostic level specifically involves: using both the environmental condition deviation and the overall baseline reliability as state inputs, feeding them into a pre-trained reinforcement learning model; and having the reinforcement learning model output and execute the optimal enhancement strategy with the highest expected diagnostic accuracy based on the state inputs.

9. The adaptive separation and enhancement method for ultrasonic guided wave multimode signals according to claim 1, characterized in that: The state input to the reinforcement learning model further includes pipeline criticality, which represents the importance or risk level of the current pipeline segment; After implementing the optimal enhancement strategy, the reward value is calculated based on the signal-to-noise ratio improvement score and the signal clarity improvement score of the enhanced signal. The reinforcement learning model is updated online using reward values ​​to optimize its subsequent policy selection. The smaller the environmental condition deviation, the more closely the current signal state matches its normal state under the current environmental conditions, the healthier the signal, and the lower the risk of anomalies. The optimal enhancement policy is a specific combination of actions selected by the reinforcement learning model from a set of discrete actions that includes various filters, compensation algorithms, and reconstruction models.

Images