An electromagnetic detection defect qualitative and quantitative method based on multi-feature fusion and model learning

Abstract

The application discloses a kind of based on multi-feature fusion and model learning's electromagnetic detection defect qualitative and quantitative method, builds multi-modal electromagnetic detection system, acquires the multi-frequency eddy current of measured component / multi-channel magnetic flux leakage signal and lift-off data, after adaptive denoising, interference suppression and normalization preprocessing, extract four categories 44-dimensional multidimensional features, obtain optimal feature subset by three-level screening strategy;Adopt channel attention and cross attention mechanism to realize multi-feature adaptive fusion, build "shared backbone+double branch" dual-task joint learning model, complete transfer pre-training, fine-tuning and lightweight compression;Finally output defect qualitative and quantitative results, support dual-technology result fusion and risk rating.The application realizes the full-depth coverage detection of special equipment metal components from surface to middle-deep defects, qualitative classification accuracy is greater than or equal to 98%, depth quantitative average relative error is less than or equal to 5%, single-frame inference time is less than or equal to 10ms, which can be directly deployed on handheld devices to realize real-time detection on site.

Description

Technical Field

[0001] This invention belongs to the field of nondestructive testing technology for special equipment, and specifically relates to an electromagnetic detection defect identification method. It is applicable to the automated qualitative classification and quantitative parameter inversion of surface, near-surface and medium-deep defects of metal components of special equipment such as pressure vessels, pressure pipelines, hoisting machinery, boilers, and storage tanks. It is compatible with ferromagnetic and non-ferromagnetic conductive materials and covers two core electromagnetic detection technologies: eddy current detection and magnetic flux leakage (MFL). Background Technology

[0002] Electromagnetic nondestructive testing (EMB) is based on the principles of electromagnetic induction and magnetic flux leakage. It identifies defects by detecting the distortion signal of the electromagnetic field in the tested component. It boasts core advantages such as non-contact operation, high detection speed, no need for coupling agents, and the ability to penetrate coatings, making it the mainstream technology system for defect detection of metal components in special equipment. Eddy current testing excels at detecting surface and near-surface defects in conductive materials, while magnetic flux leakage testing has higher sensitivity for deep defects in ferromagnetic materials. The two technologies are highly complementary.

[0003] As my country's special equipment develops towards high-parameter, long-cycle, and high-reliability operation, the operating environment of in-service equipment is becoming increasingly complex, placing more stringent demands on the full-depth coverage, automation level, identification accuracy, and anti-interference capabilities of defect detection. However, existing electromagnetic detection technologies and improvement schemes still have many core pain points in practical engineering applications:

[0004] First, traditional electromagnetic testing heavily relies on manual interpretation of signal curves and spectra by testing personnel, and the accuracy and consistency of the identification results depend entirely on the experience of the personnel. In large-scale, long-distance in-service testing scenarios, testing personnel are highly susceptible to misjudgments and omissions due to fatigue, and the testing efficiency is extremely low, failing to meet the needs of large-scale testing. According to statistics, the efficiency of traditional manual electromagnetic testing is only a few meters per hour. For the testing of long-distance pressure pipelines and the bottom plates of large storage tanks, it often takes weeks or even months, seriously affecting the normal operation cycle of the equipment.

[0005] Secondly, most existing methods use single features for defect characterization, reflecting only a portion of the defect's electromagnetic response characteristics. For complex defects such as stress corrosion cracks, micropitting, incomplete weld penetration, and deep corrosion pits, these methods suffer from low feature discrimination, weak qualitative classification capabilities, and large quantitative inversion errors. In particular, for surface micro-defects with a depth <1mm and deep defects with a depth >5mm, the detection sensitivity is severely insufficient, often failing to detect early defects in a timely manner, posing significant hidden dangers to the safe operation of special equipment.

[0006] Furthermore, the on-site inspection environment for special equipment is complex, commonly characterized by strong electromagnetic interference, vibration and noise, fluctuations in lift-off effect, and material inhomogeneity. Features extracted by existing methods are easily contaminated, exhibit weak generalization ability under different operating conditions, and suffer from poor consistency in test results. Simultaneously, most existing technologies employ either eddy current or magnetic flux leakage detection alone, failing to achieve the complementary advantages of both techniques and making it difficult to cover the full-depth defect detection needs from surface to deep layers. For example, eddy current detection alone cannot effectively detect mid-to-deep defects, while magnetic flux leakage detection alone has limited ability to detect surface micro-defects.

[0007] Furthermore, most existing model-based electromagnetic detection methods can only achieve qualitative classification or quantitative inversion independently. A few dual-task schemes use two independent models to process data separately, failing to leverage the feature correlations between tasks for collaborative optimization. This results in high model redundancy and high computational cost. Moreover, a unified feature fusion and learning framework has not been designed specifically for the characteristics of eddy current and leakage magnetic signals, making it unsuitable for multimodal electromagnetic detection data and difficult to deploy on handheld devices for real-time on-site detection.

[0008] In summary, there is an urgent need in the industry to develop an electromagnetic defect identification method that is compatible with eddy current and magnetic flux leakage detection, has full depth coverage, a high degree of automation, strong robustness to complex working conditions, and can achieve qualitative and quantitative collaborative optimization, in order to solve the core pain points of existing technologies and improve the reliability and efficiency of special equipment inspection. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a qualitative and quantitative method for electromagnetic defect detection based on multi-feature fusion and model learning. This method unifies the signal processing and model learning framework for eddy current and magnetic flux leakage detection, and solves the technical problems of traditional methods, such as single technology, high dependence on manual labor, insufficient single feature representation ability, poor robustness in complex working conditions, and disconnect between qualitative and quantitative methods. This enables automated, high-precision, and highly robust detection of defects in metal components of special equipment, from surface to deep layers.

[0010] A qualitative and quantitative method for electromagnetic detection defects based on multi-feature fusion and model learning is proposed. The electromagnetic detection includes two techniques: eddy current detection and magnetic flux leakage (MFL) detection. The method comprises the following steps: S1 Electromagnetic detection signal acquisition and preprocessing: A multi-modal electromagnetic detection system is built to acquire multi-frequency eddy current detection signals / multi-channel magnetic flux leakage detection signals and lift-off data from the metal components of the tested special equipment. Adaptive denoising, lift-off effect suppression, and signal normalization are sequentially performed on the original signals to construct and segment a defect sample dataset; S2 Multi-dimensional defect feature extraction and screening: For standardized signal samples, four categories of multi-dimensional features are extracted: time-domain statistical features, frequency-domain transform features, intrinsic electromagnetic response features, and physical sensitivity features. The three-level screening strategy obtains the optimal feature subset; S3 Multi-feature Adaptive Fusion: The optimal feature subset is mapped to the feature space, and the feature weights are adaptively allocated through the attention mechanism. After cross-feature association fusion and standardization, a global fusion feature vector is obtained; S4 Dual-task Joint Learning Model Construction and Training: A qualitative classification and quantitative regression dual-task joint learning model with "shared backbone + dual branches" is constructed. Based on the dataset, the model pre-training, fine-tuning, generalization verification and lightweight compression are completed; S5 Defect Identification Result Output and Verification: The signal to be detected is preprocessed, feature extracted and fused and then input into the trained model to output the qualitative classification and quantitative inversion results of defects, and the defect risk rating and result verification are completed.

[0011] Corresponding technical effects:

[0012] 1. A unified eddy current + leakage magnetic flux multimode electromagnetic detection framework was constructed to achieve complementary advantages of the two technologies, covering full-depth defect detection of metal components of special equipment from the surface to the middle and deep layers, and solving the pain point of limited detection depth of single technology;

[0013] 2. It achieves fully automated processing of the entire testing process, completely eliminating reliance on the experience of testing personnel, and improving testing efficiency by more than 80% compared to traditional manual methods, making it suitable for large-volume, long-distance in-service testing scenarios;

[0014] 3. Through multi-dimensional feature extraction and adaptive fusion, the ability to characterize defect features and resist interference under complex working conditions is greatly improved, and the consistency of detection results is ≥95%;

[0015] 4. It adopts a dual-task joint learning architecture to achieve synergistic optimization of qualitative classification and quantitative inversion, balancing detection accuracy and inference speed, and can be directly adapted to the needs of real-time on-site detection.

[0016] 5. The output results include qualitative classification of defects, quantitative parameters, and hazard rating, providing a direct basis for safety evaluation and maintenance decisions for special equipment.

[0017] In S1, the multimodal electromagnetic detection system includes an eddy current detection subsystem and a magnetic flux leakage detection subsystem. The eddy current detection subsystem includes a multi-frequency eddy current detector, an absolute + differential combined eddy current probe, a three-dimensional displacement scanning mechanism, a laser displacement sensor, and a data acquisition module. The excitation frequency setting range is 10Hz to 10MHz, and the sampling frequency is not less than 10 times the highest excitation frequency. It synchronously acquires eddy current time-domain signals, impedance plane real / imaginary part signals, and probe lift-off data at different excitation frequencies. The magnetic flux leakage detection subsystem includes an excitation power supply, a permanent magnet / electromagnetic magnetization device, a Hall sensor array, a three-dimensional displacement scanning mechanism, a laser displacement sensor, and a data acquisition module. The magnetization intensity setting range is 5000A / m to 50000A / m, and the sampling frequency is not less than 10kHz. It synchronously acquires magnetic flux leakage normal / tangential time-domain signals and probe lift-off data from different channels.

[0018] Corresponding technical effects:

[0019] 1. It integrates eddy current and magnetic flux leakage dual detection subsystems, which can flexibly select single technology or dual technology fusion detection according to the material of the component being tested and the depth of the defect, making it applicable to a wider range of scenarios;

[0020] 2. It adopts a combination of absolute and differential eddy current probes and Hall sensor arrays, which simultaneously takes into account the sensitivity and spatial resolution of defect detection;

[0021] 3. Synchronously acquire probe lift-off data to provide core data support for subsequent lift-off effect suppression and signal amplitude correction, thereby improving the accuracy of signal preprocessing.

[0022] The signal preprocessing in S1 specifically includes: S101 Adaptive denoising: using the db4 wavelet basis to perform 4-6 level wavelet decomposition on the original time-domain signal, and using an adaptive soft thresholding function to threshold the high-frequency detail coefficients to reconstruct a denoised signal; S102 Interference suppression: for eddy current signals, using the synchronously acquired lift-off data as the reference component, principal component analysis (PCA) orthogonal decomposition is used to separate and remove the lift-off interference component; for magnetic leakage signals, PCA orthogonal decomposition is used to separate and remove the magnetization inhomogeneity and background magnetic field interference components, and the amplitude of the magnetic leakage signal is corrected by combining the lift-off data; S103 Signal normalization: using the minimum-maximum normalization method, all signal amplitudes are linearly normalized to the [0,1] interval; S104 Dataset construction and enhancement: using the defect location as the center to extract a time-domain window to obtain standardized signal samples, labeling the samples with qualitative and quantitative labels, and performing data enhancement through noise addition, amplitude scaling, time offset, mirror flipping, and channel permutation, and dividing the dataset into training set, validation set, and test set in a 7:2:1 ratio.

[0023] Corresponding technical effects:

[0024] 1. The db4 wavelet basis with adaptive soft thresholding is used to effectively eliminate strong electromagnetic interference and vibration noise on site, while retaining the effective characteristics of the defect signal;

[0025] 2. Differentiated PCA orthogonal decomposition method is adopted to address the problems of eddy current lift-off fluctuation, non-uniform leakage magnetic magnetization, and background magnetic field interference, respectively, based on the characteristics of eddy current and leakage magnetic field signals.

[0026] 3. Minimum-maximum normalization eliminates signal amplitude deviations caused by equipment gain, material inhomogeneity, and component wall thickness differences, ensuring data consistency;

[0027] 4. Five data augmentation methods were used to address the problem of insufficient samples of defects in special equipment. The dataset was divided into a 7:2:1 ratio to provide a high-quality, standardized sample base for model training.

[0028] In S2, the four categories of multi-dimensional features extracted are: time-domain statistical features, including peak value, peak-to-peak value, mean, variance, root mean square, kurtosis, skewness, waveform factor, peak factor, impulse factor, margin factor, and signal gradient extrema, totaling 12 dimensions; frequency-domain transform features, extracted from the time-domain signal after Fast Fourier Transform (FFT), Wavelet Packet Transform (WPT), and Hilbert-Huang Transform (HHT), including fundamental amplitude, harmonic amplitude, dominant frequency, frequency band energy, frequency band energy percentage, energy entropy, marginal spectral peak value, instantaneous frequency mean, and instantaneous amplitude variance, totaling 17 dimensions; and intrinsic electromagnetic response features. Eddy current detection corresponds to intrinsic impedance plane characteristics, including impedance amplitude, phase angle, real extrema, imaginary extrema, major-minor axis ratio of the impedance trajectory, trajectory enclosed area, average trajectory curvature, and phase change rate. Magnetic leakage detection corresponds to intrinsic magnetic circuit characteristics, including leakage magnetic normal / tangential amplitude ratio, magnetic field gradient, magnetic flux change rate, leakage magnetic signal peak-to-valley spacing, hysteresis loop slope, and leakage magnetic trajectory enclosed area, totaling 8 dimensions. Physical sensitivity characteristics include signal amplitude ratio under different excitation frequencies / magnetization intensities, phase difference / magnetic field phase difference, lift-off sensitivity coefficient, conductivity change rate, permeability change characteristics, and magnetization depth coefficient, totaling 7 dimensions.

[0029] Corresponding technical effects:

[0030] 1. A 44-dimensional feature system with four major categories that is compatible with both eddy current and leakage magnetic flux technologies was designed to comprehensively cover the time-domain fluctuations, frequency-domain distributions, electromagnetic intrinsic responses and physical properties of defects, thus solving the problems of incomplete characterization and low distinguishability of single features.

[0031] 2. Extract specific intrinsic features (eddy current impedance plane features, leakage magnetic circuit intrinsic features) for different detection technologies, directly correlate them with the type and size of defects, and significantly improve the accuracy of qualitative classification and quantitative inversion;

[0032] 3. Physically sensitive characteristics directly reflect the physical properties of defects, further enhancing the detection sensitivity for surface micro-defects and deep defects.

[0033] The three-level screening strategy of S2 is as follows: S201 Linear correlation screening: Calculate the Pearson correlation coefficient between each feature and the qualitative and quantitative labels, and remove redundant features with a correlation coefficient lower than 0.3; S202 Nonlinear correlation screening: Calculate the maximum information coefficient (MIC) between each feature and the label, and remove weakly correlated features with a maximum information coefficient lower than 0.2; S203 Optimal subset screening: Use the sequence forward selection (SFS) algorithm, with model recognition accuracy as the evaluation index, to screen and obtain the optimal feature subset.

[0034] Corresponding technical effects:

[0035] 1. A three-level progressive screening strategy of "linear correlation - non-linear correlation - optimal subset" is adopted to remove redundant and weakly correlated features layer by layer, which can reduce the feature dimension by 45%-50%;

[0036] 2. While ensuring the ability to represent defect features, significantly reduce the computational load of the model, reduce the risk of model overfitting, and improve inference speed;

[0037] 3. Use model recognition accuracy as the final evaluation metric to select the optimal feature subset, ensuring that the selected features contribute to defect recognition to the maximum extent.

[0038] The multi-feature adaptive fusion of S3 specifically includes: S301 Feature space mapping: The four categories of selected features are embedded through fully connected layers and mapped to the feature space of the same dimension to obtain multiple sets of feature vectors with unified dimension; S302 Channel attention adaptive weighting: The squeeze-excitement (SE) channel attention module is used to adaptively allocate channel weights to multiple sets of feature vectors, giving high weights to feature channels that contribute greatly to defect identification and low weights to interference feature channels; S303 Cross-feature association fusion: The cross-attention module is used to capture the complementary associations between features of different categories and features of different detection technologies, calculate the attention weight matrix between features, and generate the fused global fusion feature vector; S304 Feature standardization: The fused global fusion feature vector is batch normalized.

[0039] Corresponding technical effects:

[0040] 1. Eliminating dimensional differences between features of different categories through feature space mapping lays the foundation for multi-feature fusion;

[0041] 2. An SE channel attention module is used to achieve adaptive allocation of feature weights, which automatically strengthens features that contribute greatly to defect identification and suppresses interfering features, thus solving the problem of poor generalization of traditional fixed-weight fusion.

[0042] 3. The cross-attention module can capture complementary correlations between features of different categories and features of different detection techniques, fully explore the complementary information of eddy current and leakage magnetic field signals, and further improve the characterization ability and anti-interference of fused features;

[0043] 4. Batch normalization eliminates differences in feature distribution, improving model convergence speed and training stability.

[0044] In S4, the dual-task joint learning model adopts a lightweight architecture of "shared backbone + dual branches," specifically including: Shared backbone network: The input is a globally fused feature vector, using two fully connected layers as shared feature extraction layers, coupled with ReLU activation function and Dropout layer, with a Dropout rate of 0.2–0.4, outputting deep features shared by both tasks; Qualitative classification branch: The input is shared deep features, consisting of a classification head composed of two fully connected layers + Softmax activation function, outputting the probability distribution of defect types, with the branch loss function using cross-entropy loss function; Quantitative regression branch: The input is shared deep features, consisting of a regression head composed of two fully connected layers + Linear activation function, outputting the geometric parameters of defect length, depth, and width, with the branch loss function using weighted mean square error (MSE) loss function, assigning higher weights to surface micro-defects with depth <1mm and deep defects with depth >5mm; Joint loss function: The total model loss is the weighted sum of classification loss and regression loss, as shown in the formula... ,in For classifying losses, To regress the loss, , These are dynamically adjustable weighting coefficients, with initial values ​​of 0.4 and 0.6 respectively.

[0045] Corresponding technical effects:

[0046] 1. Adopting a lightweight architecture of "shared backbone + dual branches", feature reuse is achieved through the shared backbone network, which greatly reduces model redundancy and reduces the number of parameters and computation.

[0047] 2. The dual-branch structure can simultaneously output qualitative defect classification results and quantitative parameters such as length, depth, and width, solving the problem of disconnect between qualitative and quantitative methods in traditional methods;

[0048] 3. The quantitative regression branch uses a weighted MSE loss function, which assigns higher weights to surface micro-defects with a depth of <1mm and deep defects with a depth of >5mm, significantly improving the quantitative detection accuracy across the entire depth range.

[0049] 4. The dynamic weighted joint loss function enables collaborative optimization of classification and regression tasks. The weights of the two tasks can be adjusted according to actual detection needs to adapt to different application scenarios.

[0050] The model training and optimization in S4 specifically includes: S401 Transfer learning pre-training: The model is pre-trained on a publicly available large-scale electromagnetic detection defect dataset (including eddy current and magnetic leakage datasets) to obtain pre-training weights; S402 Model fine-tuning: The pre-trained model is fine-tuned using the training set. The Adam optimizer is used, with an initial learning rate of 1e-4. A cosine decay strategy is adopted, with the learning rate decreasing to 0.5 every 20 training epochs. The batch size is 16-64, and the maximum number of training epochs is 200; S403 Overfitting suppression: An early stopping mechanism is adopted, stopping training early when the validation set loss does not decrease for 10 consecutive training epochs, combined with L2 regularization; S404 Generalization verification: The five-fold cross-validation method is used to verify the model's generalization ability; S405 Model lightweight compression: The trained model is quantized with 8-bit integers and pruned to make the single-frame signal inference time ≤10ms.

[0051] Corresponding technical effects:

[0052] 1. Transfer learning pre-training utilizes prior knowledge from publicly available large-scale datasets to effectively address the model overfitting problem in small-sample scenarios of special equipment defects;

[0053] 2. The Adam optimizer and cosine decay learning rate strategy ensure fast and stable convergence of the model, while the early stopping mechanism and L2 regularization further suppress overfitting;

[0054] 3. Five-fold cross-validation fully verifies the model's generalization ability, ensuring stable detection performance under different working conditions;

[0055] 4.8-bit integer quantization and channel pruning enable lightweight model compression, reducing model size by more than 70% and achieving a single-frame signal inference time of ≤10ms. This allows for direct deployment on handheld electromagnetic detection equipment, meeting real-time on-site detection requirements.

[0056] S5 specifically includes: S501 Signal processing: Preprocessing, feature extraction, and feature fusion are performed sequentially on the eddy current / leakage magnetic field signals collected on-site to obtain a global fused feature vector; S502 Model inference and post-processing: The feature vector to be detected is input into the trained dual-task joint learning model, which outputs the probability distribution of defect types and geometric parameters. The sliding window non-maximum suppression (NMS) algorithm is used to remove duplicate identification results, and continuous signal segments of the same defect are merged to obtain the final qualitative and quantitative results; S503 Multi-technology result fusion: If eddy current and leakage magnetic field signals are collected simultaneously, the identification results of the two technologies are fused using DS evidence theory to obtain a comprehensive qualitative and quantitative conclusion; S504 Defect hazard rating: Based on the special equipment safety standards, the hazard rating is performed according to the type, size, and location of the defect; S505 Result verification and model iteration: The identification results are verified through metallographic dissection, ultrasonic testing, and penetrant testing. Newly added defect samples after verification are added to the dataset for model iteration and optimization.

[0057] Corresponding technical effects:

[0058] 1. A standardized signal processing procedure ensures consistency between on-site data acquisition and training data, avoiding recognition errors caused by differences in data distribution;

[0059] 2. The sliding window non-maximum suppression algorithm removes duplicate recognition results and merges continuous signal segments with the same defect, thereby improving the accuracy and completeness of the recognition results;

[0060] 3. By adopting the DS evidence theory and integrating the results of eddy current and magnetic flux leakage technologies, the overall defect detection rate can be increased to 100%, and the qualitative and quantitative errors can be reduced by more than 30%.

[0061] 4. Conduct defect risk assessments in accordance with national special equipment safety standards to provide direct and actionable basis for equipment operation, maintenance, and repair decisions;

[0062] 5. The result verification and model iteration mechanism can continuously supplement new defect samples, thereby continuously improving the detection performance and scenario adaptability of the model.

[0063] For carbon steel special equipment components, the preferred excitation frequencies for eddy current testing are 100Hz, 1kHz, 10kHz, and 100kHz, and the preferred magnetization intensities for flux leakage testing are 10000A / m, 20000A / m, and 30000A / m. For stainless steel / aluminum alloy non-ferromagnetic components, the preferred excitation frequencies for eddy current testing are 1kHz, 10kHz, 100kHz, and 1MHz, and the preferred magnetization intensities for flux leakage testing are 30000A / m, 40000A / m, and 50000A / m.

[0064] Corresponding technical effects:

[0065] 1. Based on the electromagnetic properties of carbon steel and stainless steel / aluminum alloy, the combination of eddy current detection excitation frequency and leakage magnetic flux detection magnetization intensity is optimized respectively to take into account the detection sensitivity of defects of different materials and different depths.

[0066] 2. Targeted parameter settings can further improve the defect detection rate and identification accuracy, and enhance the adaptability of the method to special equipment components of different materials.

[0067] In the adaptive weighting step of channel attention, a CBAM hybrid attention module can be used, which combines channel attention and spatial attention to allocate feature weights.

[0068] Corresponding technical effects: Optimize feature weights from both channel and spatial dimensions to further enhance effective features, suppress interfering features, and improve the robustness of the model under strong interference conditions.

[0069] An LSTM layer can be added to the shared backbone network to extract the temporal correlation features of electromagnetic signals.

[0070] Corresponding technical effects: Effectively captures the correlation characteristics of continuous signals, improving the identification accuracy and characterization ability of continuous defects (such as long cracks and uniform corrosion bands).

[0071] Compared with the prior art, the present invention has the following significant advantages:

[0072] 1. Multimodal technology integration for full-depth defect coverage. This invention unifies the technical framework of eddy current detection and magnetic flux leakage detection, allowing for flexible selection of single or dual-technology integration based on detection needs. Eddy current detection is responsible for surface and near-surface defects (0-3mm), while magnetic flux leakage detection is responsible for mid-to-deep defects (3-20mm), achieving full-depth defect detection of metal components in special equipment from surface to depth, and solving the problem of limited detection depth of single technologies.

[0073] 2. Comprehensive multi-dimensional feature representation significantly improves defect identification accuracy. This invention extracts four major categories of multi-dimensional features compatible with eddy currents and leakage magnetic flux: time-domain statistics, frequency-domain transformation, intrinsic electromagnetic response, and physical sensitivity, comprehensively covering the electromagnetic response characteristics of defects. Testing shows that the qualitative classification accuracy of this invention is ≥98%, the average relative error for quantitative defect depth is ≤5%, and the average relative error for quantitative length and width is ≤3%, far exceeding traditional single electromagnetic detection methods.

[0074] 3. Adaptive feature fusion with strong robustness in complex working conditions. This invention adopts a multi-feature adaptive fusion method based on an attention mechanism, which can automatically allocate weights to different feature channels and different detection technology features. It effectively suppresses the influence of factors such as strong electromagnetic interference, lift-off effect, material inhomogeneity, and magnetization fluctuations in the field. It can adapt to the field detection conditions of special equipment with different materials, different wall thicknesses, and different interference environments, and the consistency of detection results is ≥95%.

[0075] 4. Dual-task joint learning for qualitative and quantitative collaborative optimization. This invention constructs a unified dual-task joint learning model, which achieves feature reuse through a shared backbone network, significantly reducing model redundancy; and achieves collaborative optimization of qualitative classification and quantitative inversion through a dynamically weighted joint loss function, improving detection efficiency by more than 80% compared to traditional manual methods.

[0076] 5. Lightweight design, adapted for real-time on-site detection. This invention solves the problem of insufficient defect samples for special equipment through transfer learning and contrastive learning strategies; through model quantization and pruning, the inference time for a single frame signal is ≤10ms, which can be directly deployed on handheld electromagnetic detection equipment to meet the real-time detection needs of in-service sites.

[0077] 6. High versatility and wide applicability. This invention is applicable to the defect detection of metal components in various special equipment such as pressure vessels, pressure pipelines, boilers, hoisting machinery, and storage tanks. It is compatible with both ferromagnetic and non-ferromagnetic conductive materials, and the detection process and results fully comply with the requirements of national electromagnetic testing standards such as GB / T7735, GB / T28703, and GB / T27698. Attached Figure Description

[0078] Figure 1 This is a flowchart illustrating the overall process of a qualitative and quantitative method for electromagnetic defect detection based on multi-feature fusion and model learning.

[0079] Figure 2 A schematic diagram of a multi-dimensional defect feature extraction and adaptive fusion framework;

[0080] Figure 3 This is a schematic diagram of the network architecture of a dual-task joint learning model;

[0081] Figure 4 This is a confusion matrix diagram for qualitative defect identification in Example 1;

[0082] Figure 5 This is a graph showing the error distribution of quantitative defect depth detection in Example 1.

[0083] Figure 6 This is a histogram showing the relative error distribution of defect depth detection in Example 1. Detailed Implementation

[0084] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0085] Reference Figures 1 to 6 This invention provides a qualitative and quantitative method for electromagnetic detection defects based on multi-feature fusion and model learning. The core process includes five main steps: electromagnetic detection signal acquisition and preprocessing, multi-dimensional defect feature extraction and screening, multi-feature adaptive fusion, dual-task joint learning model construction and training, and defect identification result output and verification. The specific implementation process is as follows:

[0086] Step 1: Electromagnetic detection signal acquisition and preprocessing

[0087] The core objective of this step is to acquire high-quality electromagnetic detection signals and construct a standardized defect dataset to provide a data foundation for subsequent feature extraction and model training. The specific implementation process is as follows:

[0088] 1. Detection System Setup: A multi-modal electromagnetic detection system is built, which can switch between eddy current detection subsystems or magnetic flux leakage detection subsystems according to detection requirements, or integrate both subsystems simultaneously to achieve dual-technology fusion detection. The system hardware includes a multi-frequency digital eddy current detector / excitation power supply, an absolute + differential combined eddy current probe / permanent magnetization device + Hall sensor array, a three-dimensional displacement scanning mechanism, a laser displacement sensor, and a high-speed data acquisition module.

[0089] 2. Data Acquisition Parameter Settings: For the material and wall thickness of the metal components of the special equipment under test, set the corresponding excitation frequency (eddy current) or magnetization intensity (magnetic leakage), with a frequency range of 10Hz-10MHz and a magnetization intensity range of 5000A / m-50000A / m; set the sampling frequency to be no less than 10 times the highest excitation frequency (eddy current) or no less than 10kHz (magnetic leakage); set the scanning speed of the three-dimensional displacement scanning mechanism to ensure that the probe scans the measured area at a uniform speed.

[0090] 3. Signal Acquisition: The detection probe is driven by a three-dimensional displacement scanning mechanism to perform uniform scanning of the detection area of ​​the component under test. Electromagnetic time domain signals, impedance plane real / imaginary part signals (eddy currents) or leakage magnetic normal / tangential signals (leakage magnetics) of different excitation frequencies / different channels are acquired simultaneously. At the same time, the lift-off fluctuation data of the probe is acquired simultaneously through a laser displacement sensor.

[0091] 4. Signal preprocessing: (1) Adaptive denoising: The original time-domain signal was decomposed into 4 to 6 layers using the db4 wavelet basis to obtain low-frequency approximation coefficients and high-frequency detail coefficients; the high-frequency detail coefficients were thresholded using an adaptive soft threshold function to eliminate electromagnetic interference and vibration noise in the signal; the processed coefficients were reconstructed using wavelets to obtain the denoised electromagnetic signal. (2) Interference suppression: For eddy current signals, the lifted data collected synchronously was used as the reference component, and the PCA orthogonal decomposition method was used to separate and remove the lifted interference component; for leakage magnetic signals, the PCA orthogonal decomposition method was used to separate and remove the magnetization inhomogeneity and background magnetic field interference components, and the leakage magnetic signal amplitude was corrected by combining the lifted data. (3) Signal normalization: The minimum-maximum normalization method was used to linearly normalize all preprocessed signal amplitudes to the [0,1] interval to eliminate signal amplitude deviations caused by equipment gain, material inhomogeneity, and component wall thickness differences. (4) Dataset construction and partitioning: Standardized signal samples are obtained by extracting time-domain windows centered on the defect location. Each sample is labeled with qualitative labels (no defects, transverse cracks, longitudinal cracks, pitting, uniform corrosion, incomplete penetration, slag inclusions, etc.) and quantitative labels (length, depth, width). Data augmentation is performed by adding noise, amplitude scaling, time offset, mirror flipping, and channel replacement. Finally, the dataset is randomly divided into training set, validation set and test set in a ratio of 7:2:1.

[0092] Step 2: Multi-dimensional defect feature extraction and screening

[0093] The core objective of this step is to comprehensively extract multi-dimensional features compatible with eddy currents and leakage flux, and to select the optimal feature subset with high discriminative power and low redundancy. The specific implementation process is as follows:

[0094] 1. Four categories of multi-dimensional feature extraction: For each standardized signal sample, four categories of features are extracted: time-domain statistical features, frequency-domain transformation features, electromagnetic response intrinsic features, and physical sensitivity features. (1) Time-domain statistical features: Peak value, peak-to-peak value, mean, variance, root mean square, kurtosis, skewness, waveform factor, peak factor, impulse factor, margin factor, and signal gradient extremum are extracted, totaling 12 dimensions, to characterize the time-domain fluctuation law of the signal caused by the defect. (2) Frequency-domain transformation features: The time-domain signal is transformed by FFT, WPT, and HHT to extract the fundamental amplitude, harmonic amplitude, main frequency, frequency band energy, frequency band energy ratio, energy entropy, marginal spectrum peak value, instantaneous frequency mean, and instantaneous amplitude variance, totaling 17 dimensions, to characterize the frequency domain distribution law of the signal and the corresponding depth information of the defect. (3) Intrinsic characteristics of electromagnetic response: Eddy current detection extracts 8-dimensional impedance plane characteristics such as impedance amplitude and phase angle; leakage magnetic field detection extracts 8-dimensional magnetic circuit intrinsic characteristics such as leakage magnetic field normal / tangential amplitude ratio and magnetic field gradient, which directly correspond to the type and size of the defect. (4) Physical sensitivity characteristics: Extracts 7 dimensions of signal amplitude ratio, phase difference, lift-off sensitivity coefficient, conductivity change rate, permeability change characteristics, and magnetization depth coefficient under different excitation frequencies / magnetization intensities, which are directly related to the physical nature of the defect.

[0095] 2. Three-level feature selection: (1) Linear correlation selection: Calculate the Pearson correlation coefficient between each feature and the label, and remove redundant features with a correlation coefficient lower than 0.3. (2) Nonlinear correlation selection: Calculate the maximum information coefficient (MIC) between each feature and the label, and remove weakly correlated features with a MIC value lower than 0.2. (3) Optimal subset selection: Use the Sequence Forward Selection (SFS) algorithm and use the model recognition accuracy as the evaluation index to select the optimal feature subset.

[0096] Step 3: Adaptive fusion of multiple features

[0097] The core objective of this step is to achieve adaptive weighted fusion of multiple features to generate a globally fused feature vector with high representational power. The specific implementation process is as follows:

[0098] 1. Feature Space Mapping: The four categories of features selected are embedded through fully connected layers and mapped to a feature space of the same dimension, resulting in multiple sets of feature vectors with uniform dimension.

[0099] 2. Adaptive weighted channel attention: The SE channel attention module is used to adaptively allocate channel weights to multiple sets of feature vectors, giving high weights to feature channels that contribute significantly to defect identification and low weights to interference feature channels.

[0100] 3. Cross-feature association fusion: A cross-attention module is used to capture the complementary associations between different categories of features and between eddy current and leakage magnetic field features, calculate the attention weight matrix between features, and generate a fused global feature vector.

[0101] 4. Feature standardization: Batch normalization is performed on the fused global feature vectors to eliminate differences in feature distribution and improve the convergence speed and stability of the model.

[0102] Step 4: Construction and Training of the Dual-Task Joint Learning Model

[0103] The core objective of this step is to construct a unified dual-task joint learning model, complete the model training optimization and lightweight compression, and the specific implementation process is as follows:

[0104] 1. Dual-task joint model architecture: A lightweight architecture of "shared backbone + dual branches" is adopted: (1) Shared backbone network: The input is the global fusion feature vector. Two fully connected layers are used as shared feature extraction layers, with ReLU activation function and Dropout layer (Dropout rate 0.2~0.4) to output the deep features shared by the two tasks. (2) Qualitative classification branch: It consists of two fully connected layers + Softmax activation function, outputting the probability distribution of defect types. The loss function adopts the cross-entropy loss function. (3) Quantitative regression branch: It consists of two fully connected layers + Linear activation function, outputting three geometric parameters: defect length, depth, and width. The loss function adopts the weighted MSE loss function, giving higher weights to micro defects with depth <1mm and deep defects with depth >5mm. (4) Joint loss function: The initial values ​​of α and β are 0.4 and 0.6, respectively, and are dynamically adjusted during training.

[0105] 2. Model Training and Optimization: (1) Transfer Learning Pre-training: First, pre-training for 100 epochs on a publicly available large-scale electromagnetic detection defect dataset to obtain pre-training weights. (2) Model Fine-tuning: Fine-tuning is performed using the target training set. The Adam optimizer has an initial learning rate of 1e-4, which is reduced to 0.5 every 20 epochs. The batch size is 16-64, and the maximum training epoch is 200. (3) Overfitting Suppression: An early stopping mechanism is adopted (the validation set loss stops if it does not decrease for 10 consecutive epochs), combined with L2 regularization. (4) Generalization Validation: The five-fold cross-validation method is used to verify the generalization ability of the model. (5) Model Lightweight Compression: 8-bit integer quantization and channel pruning are performed to ensure that the inference time of a single frame signal is ≤10ms.

[0106] Step 5: Output and Verification of Defect Identification Results

[0107] The core objective of this step is to achieve automated defect identification of the signal to be detected, output the final result, and complete verification and optimization. The specific implementation process is as follows:

[0108] 1. Signal processing: The electromagnetic signals to be detected collected on site are preprocessed, feature extracted and fused sequentially to obtain the global feature vector to be detected.

[0109] 2. Model Inference and Post-processing: Input the feature vector to be detected into the trained model, and output the probability distribution of defect type and geometric parameters; use the sliding window non-maximum suppression algorithm to remove duplicate identification results and merge continuous signal segments of the same defect.

[0110] 3. Fusion of results from multiple technologies: If eddy current and leakage magnetic field signals are collected simultaneously, the identification results of the two technologies are fused using the DS evidence theory to obtain a comprehensive qualitative and quantitative conclusion.

[0111] 4. Defect risk rating: Based on the special equipment safety standards such as GB150 "Pressure Vessels" and GB50235 "Code for Construction of Industrial Metal Piping Engineering", the risk rating is carried out according to the type, size and location of the defect.

[0112] 5. Result Validation and Model Iteration: The identification results are validated through metallographic dissection, ultrasonic testing, and penetrant testing. Newly added defect samples after validation are added to the dataset for model iteration and optimization.

[0113] Example 1: Application of eddy current testing in weld seams of 20# carbon steel pressure pipelines

[0114] The testing object in this embodiment is a 20# carbon steel pressure pipeline commonly used in special equipment. The pipeline specification is Φ108mm×6mm. Different types and sizes of artificial defects were prefabricated in the weld area, specifically including: transverse cracks (length 220mm, depth 0.55mm), longitudinal cracks (length 220mm, depth 0.55mm), pitting corrosion (diameter 0.53mm, depth 0.53mm), and uniform corrosion (depth 0.5~4mm), totaling 120 sets of defect samples, combined with 30 sets of defect-free samples, for a total of 150 sets of original samples.

[0115] The specific implementation steps of this embodiment are as follows:

[0116] 1. Electromagnetic detection signal acquisition and preprocessing: (1) Acquisition parameter settings: The eddy current detection subsystem of the multi-mode electromagnetic detection system is used, with an absolute + differential combination eddy current probe. The excitation frequency is set to four groups: 100Hz, 1kHz, 10kHz, and 100kHz. The sampling frequency is set to 1MHz, and the scanning speed is 10mm / s. Simultaneously, a laser displacement sensor is used to acquire probe lift-off data, and the lift-off fluctuation range is controlled within 0 to 0.5mm. (2) Signal acquisition: The three-dimensional displacement scanning mechanism is used to drive the probe to perform uniform spiral scanning on the pipe weld area. The eddy current time domain signal, impedance plane real part and imaginary part signal are acquired at each excitation frequency. 10 sets of repeated signals are acquired for each sample. (3) Signal preprocessing: The time domain signal is decomposed into 5 layers using the db4 wavelet basis and adaptive soft thresholding is used for noise reduction; PCA orthogonal decomposition is used to remove the lifting interference; minimum-maximum normalization is performed to the [0,1] interval; a signal window of 2000 sampling points is extracted with the defect location as the center, and qualitative and quantitative labels are marked. The dataset is expanded to 1500 groups through data augmentation and divided into 1050 training sets, 300 validation sets, and 150 test sets in a 7:2:1 ratio.

[0117] 2. Multi-dimensional defect feature extraction and screening: Four categories of 44-dimensional features are extracted for each signal sample. Through a three-level screening strategy, 3-dimensional redundant features with correlation coefficient <0.3 are first removed, then 2-dimensional weakly correlated features with MIC <0.2 are removed, and finally the SFS algorithm is used to screen out the 24-dimensional optimal feature subset, reducing the feature dimension by 45%.

[0118] 3. Multi-feature adaptive fusion: The four categories of features selected are mapped to a 64-dimensional unified feature space. The SE channel attention module is used for weight allocation. The complementary correlation between features is captured by cross-attention modules to generate a 128-dimensional global fusion feature vector and perform batch normalization.

[0119] 4. Construction and training of dual-task joint learning model: A "shared backbone + dual branches" architecture is built. The shared backbone is set with two 64-dimensional fully connected layers and a dropout rate of 0.3. The quantitative regression branch assigns double weight to micro-defects with a depth of <1mm. The joint loss function is initially set as total loss = 0.4 × classification loss + 0.6 × regression loss.

[0120] First, the model is pre-trained for 100 epochs on a publicly available eddy current detection dataset. Then, it is fine-tuned using the training set in this embodiment. The Adam optimizer has an initial learning rate of 1e-4, a batch size of 32, a maximum training epoch of 200, and an early stopping mechanism. After training, the model is compressed by 8-bit quantization, resulting in a 75% reduction in model size and a single-frame signal inference time of ≤8ms.

[0121] 5. Defect identification results output and verification: After testing, the overall accuracy of defect classification on the test set was 98.67%, the recall rate of each type of defect was ≥97%, and the false positive rate of no-defect samples was ≤1%; the average relative error of defect depth was 4.21%, and the average relative error of length and width was 2.15%.

[0122] The compressed model was deployed on a handheld electromagnetic detector to conduct on-site inspection of the welds of the No. 20 steel pressure pipeline in a chemical plant. A total of 100 welds were inspected, and the 12 defects identified were verified by metallographic dissection and ultrasonic testing. The consistency of the results was ≥97%, and the detection efficiency was improved by 85% compared with traditional manual interpretation.

[0123] Example 2: Application of eddy current testing in S30408 ​​austenitic stainless steel pressure vessel cylinder

[0124] The testing object in this embodiment is the S30408 ​​austenitic stainless steel pressure vessel cylinder commonly used in special equipment. The cylinder has a specification of Φ500mm×8mm and is a non-ferromagnetic material. Different types and sizes of artificial defects were pre-fabricated in the cylinder base material and the circumferential weld area, specifically including: circumferential cracks (length 15mm, depth 0.4mm), axial cracks (length 20mm, depth 0.6mm), intergranular corrosion (depth 0.3~2.5mm), pitting corrosion (diameter 0.4mm, depth 0.4mm), and weld incomplete fusion (length 25mm, depth 0.8mm), totaling 100 sets of defect samples, combined with 40 sets of defect-free samples, for a total of 140 sets of original samples.

[0125] The specific implementation steps of this embodiment are as follows:

[0126] 1. Electromagnetic detection signal acquisition and preprocessing: (1) Acquisition parameter settings: Eddy current detection subsystem is used. For non-ferromagnetic stainless steel materials, the excitation frequency is set to four groups: 1kHz, 10kHz, 100kHz, and 1MHz. The sampling frequency is set to 10MHz, and the scanning speed is 8mm / s. Simultaneous acquisition of lift-off data is performed, and the lift-off fluctuation range is controlled within 0 to 0.6mm. (2) Signal acquisition: A three-dimensional displacement scanning mechanism is used to drive the probe to perform uniform circumferential scanning on the pressure vessel cylinder base material and the circumferential weld area. 12 sets of repetitive signals are acquired for each sample. (3) Signal preprocessing: The time domain signal is decomposed into 6 layers for noise reduction using the db4 wavelet basis. The PCA orthogonal decomposition method is used to remove lift-off interference. Minimum-maximum normalization is performed. A signal window of 1500 sampling points is extracted, and the dataset is expanded to 1680 groups through data augmentation and divided into 7:2:1 groups.

[0127] 2. Multi-dimensional defect feature extraction and screening: 44-dimensional features in four categories are extracted. A three-level screening strategy is used to remove redundant features in four dimensions and weakly correlated features in two dimensions. Finally, the optimal feature subset of 22 dimensions is selected, reducing the feature dimension by about 50%.

[0128] 3. Multi-feature adaptive fusion: The filtered features are mapped to a 64-dimensional unified feature space, and fused using the SE channel attention module and cross-attention module to generate a 128-dimensional global fusion feature vector.

[0129] 4. Dual-task joint learning model construction and training: The shared backbone dropout rate is set to 0.25; the quantitative regression branch assigns 2.5 times the weight to micro-defects with a depth <1mm; it is first pre-trained on a public eddy current detection dataset and then fine-tuned; the model volume is compressed by 72% after quantization, and the single-frame inference time is ≤7ms.

[0130] 5. Defect identification results output and verification: The overall accuracy of defect classification on the test set was 98.21%, the recall rate of each type of defect was ≥96%, and the false positive rate of no-defect samples was ≤1.2%; the average relative error of defect depth was 4.52%, and the average relative error of length and width was 2.38%.

[0131] On-site inspection of the S30408 ​​stainless steel reactor body in service at a pharmaceutical factory was conducted. A total of 8 circumferential welds and 20㎡ of base material area were inspected. The 9 intergranular corrosion and microcrack defects identified were verified by penetrant testing and metallographic sampling. The consistency of the results was ≥96%, and the inspection efficiency was improved by 82%.

[0132] Example 3: Application of magnetic flux leakage detection for corrosion defects in Q235B steel bottom plates of storage tanks

[0133] The testing object in this embodiment is Q235B carbon steel used for the bottom plate of crude oil storage tanks. The steel plate is 12mm thick and is a ferromagnetic material. The bottom plate surface has a 100μm thick anti-corrosion coating. It is subject to complex working conditions such as localized rust and electromagnetic interference from the soil. The samples include artificially prefabricated corrosion pits of different sizes and naturally occurring uniform corrosion defects in in-service storage tanks. Specifically, they include: pitting pits (diameter 0.8-5mm, depth 0.5-8mm), uniform corrosion (depth 1-10mm), and localized corrosion pit groups, totaling 90 defect samples, plus 40 defect-free samples, for a total of 130 original samples.

[0134] The specific implementation steps of this embodiment are as follows:

[0135] 1. Electromagnetic detection signal acquisition and preprocessing: (1) Acquisition parameter settings: The leakage magnetic field detection subsystem of the multi-mode electromagnetic detection system is adopted, using a permanent magnet magnetization device and a 16-channel Hall sensor array. The magnetization intensity is set to three groups: 10000A / m, 20000A / m, and 30000A / m. The sampling frequency is set to 20kHz, and the scanning speed is 5mm / s. Simultaneous acquisition of lift-off data is performed, and the lift-off fluctuation range is controlled within 0 to 1.2mm. (2) Signal acquisition: The three-dimensional displacement scanning mechanism drives the leakage magnetic field probe to perform grid scanning on the detection area of ​​the tank bottom plate. Simultaneous acquisition of 16 channels of leakage magnetic field normal and tangential time domain signals is performed. 12 sets of repetitive signals are acquired for each sample. (3) Signal preprocessing: The time domain signal is decomposed into five layers for noise reduction using the db4 wavelet basis; PCA orthogonal decomposition method is used to remove magnetization inhomogeneity and background magnetic field interference, and the signal amplitude is corrected by combining the extracted data; minimum-maximum normalization is performed; a signal window of 3000 sampling points is extracted, and data enhancement is performed by adding noise, amplitude scaling and channel replacement, expanding the dataset to 1820 groups, which are divided into 7:2:1.

[0136] 2. Multi-dimensional defect feature extraction and screening: Four categories of 44-dimensional features are extracted for each signal sample. A three-level screening strategy is used to remove 6 redundant features and 3 weakly correlated features. Finally, the optimal feature subset of 23 dimensions is selected, reducing the feature dimension by about 48%.

[0137] 3. Multi-feature adaptive fusion: The filtered features are mapped to a 64-dimensional unified feature space. The CBAM hybrid attention module is used to allocate channel and spatial weights. The complementary correlation between leakage magnetic features is captured by cross-attention modules to generate a 128-dimensional global fusion feature vector.

[0138] 4. Dual-task joint learning model construction and training: A shared backbone with one LSTM layer is added to extract temporal correlation features, and the dropout rate is set to 0.3; the quantitative regression branch assigns 2.5 times the weight to micro-corrosion pits with a depth of <1mm and deep corrosion with a depth of >5mm; it is first pre-trained on a publicly available magnetic flux leakage detection dataset for tank bottom plates, and then fine-tuned; the model volume is compressed by 76% after quantization, and the single-frame inference time is ≤8ms.

[0139] 5. Defect identification results output and verification: The overall accuracy of defect classification on the test set was 97.80%, the recall rate of each type of defect was ≥96%, and the false positive rate of no-defect samples was ≤1.3%; the average relative error of defect depth was 4.68%, and the average relative error of length and width was 2.45%.

[0140] On-site inspection of the bottom plate of a 100,000 cubic meter crude oil storage tank at an oil depot covered an area of ​​500 square meters. The 15 deep corrosion defects identified were verified by ultrasonic thickness measurement and excavation, with a consistency of ≥95% and an 83% improvement in inspection efficiency.

[0141] Example 4: Application of Eddy Current + Magnetic Flux Leakage Fusion Detection for Full-Depth Defects in 20# Carbon Steel Pressure Pipelines

[0142] The test object in this embodiment is a 20# carbon steel pressure pipe with a pipe specification of Φ219mm×10mm. Surface microcracks (depth 0.4~1mm) and deep corrosion pits (depth 3~8mm) were prefabricated, totaling 80 defect samples, along with 30 defect-free samples, for a total of 110 original samples.

[0143] This embodiment employs both eddy current detection and magnetic flux leakage detection technologies. Eddy current is used to detect surface microcracks, while magnetic flux leakage is used to detect deep corrosion pits. The identification results of the two technologies are fused using the DS evidence theory. The specific implementation steps are as follows:

[0144] 1. Eddy current and leakage magnetic field signals were collected according to the methods of Example 1 and Example 3 respectively, and a fusion dataset was constructed. The dataset was expanded to 1650 sets through data augmentation.

[0145] 2. Extract multi-dimensional features of eddy current and leakage flux respectively, select the optimal feature subsets of each, and then fuse the features of the two technologies through cross-attention modules to generate a global fusion feature vector.

[0146] 3. A dual-task joint learning model is used for training, and the joint loss function is initially set as total loss = 0.45 × classification loss + 0.55 × regression loss.

[0147] 4. During model reasoning, the identification results of eddy current and leakage magnetic field are first output separately, and then the DS evidence theory is used to fuse them to obtain the final qualitative and quantitative conclusions.

[0148] Tests showed that the overall defect detection rate of the fusion detection reached 100%, the qualitative classification accuracy rate was 99.09%, and the average relative error of defect depth quantification was 4.12%. Compared with single eddy current detection, the detection rate of deep defects was improved by 15%; compared with single magnetic flux leakage detection, the detection rate of surface micro-defects was improved by 12%, and the overall quantitative error was reduced by more than 30%, realizing high-precision full-depth detection of pressure pipelines from the surface to the depth.

Claims

1. A qualitative and quantitative method for electromagnetic defect detection based on multi-feature fusion and model learning, characterized in that, The electromagnetic detection includes two techniques: eddy current detection and magnetic flux leakage (MFL) detection. The method includes the following steps: S1 Electromagnetic detection signal acquisition and preprocessing: A multi-mode electromagnetic detection system is built to acquire multi-frequency eddy current detection signals / multi-channel magnetic flux leakage detection signals and lift-off data of the metal components of the special equipment under test. Adaptive denoising, lift-off effect suppression, and signal normalization are sequentially performed on the original signals to construct and divide the defect sample dataset; S2 Multi-dimensional defect feature extraction and screening: For standardized signal samples, four categories of multi-dimensional features are extracted: time-domain statistical features, frequency-domain transformation features, intrinsic electromagnetic response features, and physical sensitivity features. A three-level screening strategy is used to obtain the optimal feature subset. S3 Multi-feature Adaptive Fusion: The optimal feature subset is mapped to the feature space, and the feature weights are adaptively allocated through an attention mechanism. After cross-feature association fusion and standardization, a global fused feature vector is obtained. S4 Dual-Task Joint Learning Model Construction and Training: A dual-task joint learning model of qualitative classification and quantitative regression with a "shared backbone + dual branches" is constructed. Based on the dataset, the model is pre-trained, fine-tuned, generalized, validated, and lightweight compressed. S5 Defect Identification Result Output and Validation: The signal to be detected is pre-processed, feature extracted, and fused, and then input into the trained model. The model outputs the qualitative classification and quantitative inversion results of defects, and completes the defect risk rating and result validation. In S1, the multimodal electromagnetic detection system includes an eddy current detection subsystem and a magnetic flux leakage detection subsystem. The eddy current detection subsystem includes a multi-frequency eddy current detector, an absolute + differential combined eddy current probe, a three-dimensional displacement scanning mechanism, a laser displacement sensor, and a data acquisition module. The excitation frequency setting range is 10Hz to 10MHz, and the sampling frequency is not less than 10 times the highest excitation frequency. It synchronously acquires eddy current time-domain signals, impedance plane real / imaginary part signals, and probe lift-off data at different excitation frequencies. The magnetic flux leakage detection subsystem includes an excitation power supply, a permanent magnet / electromagnetic magnetization device, a Hall sensor array, a three-dimensional displacement scanning mechanism, a laser displacement sensor, and a data acquisition module. The magnetization intensity setting range is 5000A / m to 50000A / m, and the sampling frequency is not less than 10kHz. It synchronously acquires magnetic flux leakage normal / tangential time-domain signals and probe lift-off data from different channels.

2. The qualitative and quantitative method for electromagnetic defect detection based on multi-feature fusion and model learning according to claim 1, characterized in that, The signal preprocessing in S1 specifically includes: S101 Adaptive denoising: using the db4 wavelet basis to perform 4-6 level wavelet decomposition on the original time-domain signal, and using an adaptive soft thresholding function to threshold the high-frequency detail coefficients to reconstruct a denoised signal; S102 Interference suppression: for eddy current signals, using the synchronously acquired lift-off data as the reference component, principal component analysis (PCA) orthogonal decomposition is used to separate and remove the lift-off interference component; for magnetic leakage signals, PCA orthogonal decomposition is used to separate and remove the magnetization inhomogeneity and background magnetic field interference components, and the amplitude of the magnetic leakage signal is corrected by combining the lift-off data; S103 Signal normalization: using the minimum-maximum normalization method, all signal amplitudes are linearly normalized to the [0,1] interval; S104 Dataset construction and enhancement: using the defect location as the center to extract a time-domain window to obtain standardized signal samples, labeling the samples with qualitative and quantitative labels, and performing data enhancement through noise addition, amplitude scaling, time offset, mirror flipping, and channel permutation, and dividing the dataset into training set, validation set, and test set in a 7:2:1 ratio.

3. The qualitative and quantitative method for electromagnetic defect detection based on multi-feature fusion and model learning according to claim 1, characterized in that, In S2, the four categories of multi-dimensional features extracted are: time-domain statistical features, including peak value, peak-to-peak value, mean, variance, root mean square, kurtosis, skewness, waveform factor, peak factor, impulse factor, margin factor, and signal gradient extrema, totaling 12 dimensions; Frequency domain transformation features are extracted from the time-domain signal after Fast Fourier Transform (FFT), Wavelet Packet Transform (WPT), and Hilbert-Huang Transform (HHT), including fundamental amplitude, harmonic amplitude, dominant frequency, band energy, band energy ratio, energy entropy, marginal spectral peak value, instantaneous frequency mean, and instantaneous amplitude variance, totaling 17 dimensions; electromagnetic response intrinsic features, corresponding to impedance plane intrinsic features for eddy current detection, include impedance amplitude, phase angle, real part extrema, imaginary part extrema, impedance trajectory major axis to minor axis ratio, trajectory enclosed area, trajectory average curvature, and phase change rate; leakage magnetic field detection corresponds to magnetic circuit intrinsic features, including leakage magnetic field normal / tangential amplitude ratio, magnetic field gradient, magnetic flux change rate, leakage magnetic field signal peak-to-valley spacing, hysteresis loop slope, and leakage magnetic field trajectory enclosed area, totaling 8 dimensions; physical Sensitive features include the signal amplitude ratio, phase difference / magnetic field phase difference, lift-off sensitivity coefficient, conductivity change rate, permeability change characteristics, and magnetization depth coefficient under different excitation frequencies / magnetization intensities, totaling 7 dimensions.

4. The qualitative and quantitative method for electromagnetic defect detection based on multi-feature fusion and model learning according to claim 1, characterized in that, The three-level screening strategy of S2 is as follows: S201 Linear correlation screening: Calculate the Pearson correlation coefficient between each feature and the qualitative and quantitative labels, and remove redundant features with a correlation coefficient lower than 0.3; S202 Nonlinear correlation screening: Calculate the maximum information coefficient (MIC) between each feature and the label, and remove weakly correlated features with a maximum information coefficient lower than 0.2; S203 Optimal subset screening: Use the sequence forward selection (SFS) algorithm, with model recognition accuracy as the evaluation index, to screen and obtain the optimal feature subset.

5. The qualitative and quantitative method for electromagnetic defect detection based on multi-feature fusion and model learning according to claim 1, characterized in that, The multi-feature adaptive fusion of S3 specifically includes: S301 Feature space mapping: The four categories of selected features are embedded through fully connected layers and mapped to the feature space of the same dimension to obtain multiple sets of feature vectors with unified dimension; S302 Channel attention adaptive weighting: The squeeze-excitement (SE) channel attention module is used to adaptively allocate channel weights to multiple sets of feature vectors, giving high weights to feature channels that contribute greatly to defect identification and low weights to interference feature channels; S303 Cross-feature association fusion: The cross-attention module is used to capture the complementary associations between features of different categories and features of different detection technologies, calculate the attention weight matrix between features, and generate the fused global fusion feature vector; S304 Feature standardization: The fused global fusion feature vector is batch normalized.

6. The qualitative and quantitative method for electromagnetic detection defects based on multi-feature fusion and model learning according to claim 1, characterized in that, In S4, the dual-task joint learning model adopts a lightweight architecture of "shared backbone + dual branches," specifically including: Shared backbone network: The input is a globally fused feature vector, using two fully connected layers as shared feature extraction layers, coupled with ReLU activation function and Dropout layer, with a Dropout rate of 0.2–0.4, outputting deep features shared by both tasks; Qualitative classification branch: The input is shared deep features, consisting of a classification head composed of two fully connected layers + Softmax activation function, outputting the probability distribution of defect types, with the branch loss function using cross-entropy loss function; Quantitative regression branch: The input is shared deep features, consisting of a regression head composed of two fully connected layers + Linear activation function, outputting the geometric parameters of defect length, depth, and width, with the branch loss function using weighted mean square error (MSE) loss function, assigning higher weights to surface micro-defects with depth <1mm and deep defects with depth >5mm; Joint loss function: The total model loss is the weighted sum of classification loss and regression loss, as shown in the formula... ,in For classifying losses, To regress the loss, , These are dynamically adjustable weighting coefficients, with initial values ​​of 0.4 and 0.6 respectively.

7. The qualitative and quantitative method for electromagnetic defect detection based on multi-feature fusion and model learning according to claim 1, characterized in that, The model training and optimization in S4 specifically includes: S401 Transfer learning pre-training: The model is pre-trained on a publicly available large-scale electromagnetic detection defect dataset (including eddy current and magnetic leakage datasets) to obtain pre-training weights; S402 Model fine-tuning: The pre-trained model is fine-tuned using the training set. The Adam optimizer is used, with an initial learning rate of 1e-4. A cosine decay strategy is adopted, with the learning rate decreasing to 0.5 every 20 training epochs. The batch size is 16-64, and the maximum number of training epochs is 200; S403 Overfitting suppression: An early stopping mechanism is adopted, stopping training early when the validation set loss does not decrease for 10 consecutive training epochs, combined with L2 regularization; S404 Generalization verification: The five-fold cross-validation method is used to verify the model's generalization ability; S405 Model lightweight compression: The trained model is quantized with 8-bit integers and pruned to make the single-frame signal inference time ≤10ms.

8. The qualitative and quantitative method for electromagnetic detection defects based on multi-feature fusion and model learning according to claim 1, characterized in that, S5 specifically includes: S501 Signal processing: Preprocessing, feature extraction, and feature fusion are performed sequentially on the eddy current / leakage magnetic field signals collected on-site to obtain a global fused feature vector; S502 Model inference and post-processing: The feature vector to be detected is input into the trained dual-task joint learning model, which outputs the probability distribution of defect types and geometric parameters. The sliding window non-maximum suppression (NMS) algorithm is used to remove duplicate identification results, and continuous signal segments of the same defect are merged to obtain the final qualitative and quantitative results; S503 Multi-technology result fusion: If eddy current and leakage magnetic field signals are collected simultaneously, the identification results of the two technologies are fused using DS evidence theory to obtain a comprehensive qualitative and quantitative conclusion; S504 Defect hazard rating: Based on the special equipment safety standards, the hazard rating is performed according to the type, size, and location of the defect; S505 Result verification and model iteration: The identification results are verified through metallographic dissection, ultrasonic testing, and penetrant testing. Newly added defect samples after verification are added to the dataset for model iteration and optimization.

9. The qualitative and quantitative method for electromagnetic detection defects based on multi-feature fusion and model learning according to claim 2, characterized in that: • For carbon steel special equipment components, the preferred excitation frequencies for eddy current testing are 100Hz, 1kHz, 10kHz, and 100kHz, and the preferred magnetization intensities for leakage flux testing are 10000A / m, 20000A / m, and 30000A / m. • For stainless steel / aluminum alloy non-ferromagnetic components, the preferred excitation frequencies for eddy current testing are 1kHz, 10kHz, 100kHz, and 1MHz, and the preferred magnetization intensities for leakage flux testing are 30000A / m, 40000A / m, and 50000A / m.

Images