Steel plate internal defect detection method

Abstract

The invention provides a steel plate internal defect detection method, and belongs to the technical field of industrial intelligent detection based on machine learning. The method comprises the following steps: deploying a multi-channel ultrasonic phased array probe on a steel plate detection platform, collecting an ultrasonic full matrix, obtaining related data, inputting the data into a multi-modal signal preprocessing module, generating a structured multi-channel space-time signal data body according to a sound wave path geometrical relationship, extracting channel local time domain characteristics through a lightweight signal interaction network layer, and carrying out multi-modal signal detection on the channel local time domain characteristics. A physical perception attention module combined with sound path distance attenuation compensation generates an inter-channel interaction weight matrix, dynamic weighted summation is carried out, initial weight bias is added to a specific signal, and a signal-to-noise ratio enhanced fusion signal field is obtained. And inputting the three-dimensional acoustic scattering intensity distribution diagram into a deep defect feature extraction module, extracting a three-dimensional acoustic scattering intensity distribution diagram, finally inputting the three-dimensional acoustic scattering intensity distribution diagram into a defect detection model, and respectively predicting defect type probability distribution, three-dimensional space position coordinates and an equivalent size by three types of sub-networks. According to the invention, the depth of defect characterization and the intelligent identification precision are improved.

Description

Technical Field

[0001] This invention belongs to the field of industrial intelligent inspection technology based on machine learning, and particularly relates to a method for detecting internal defects in steel plates. Background Technology

[0002] As a key structural material in modern industry and infrastructure construction, the internal quality of steel plates directly affects the safety and service life of various important equipment and projects. During manufacturing or service, defects such as cracks, porosity, and slag inclusions may occur within the steel plates. These defects are often highly concealed, and if not detected in time, they will gradually expand under load and environmental conditions, eventually leading to component failure or even serious safety accidents. Traditional ultrasonic testing methods face limitations in defect qualitative, quantitative, and feature analysis. As industrial testing technology evolves towards intelligence and precision, developing a testing method that deeply integrates physical mechanisms and data-driven approaches is crucial. This method could intelligently determine and invert the type, location, and size of defects, improving the objectivity and accuracy of testing and providing reliable data for structural health assessment and life prediction. However, current research has the following limitations: 1. The depth and intelligence of defect characterization are limited. Existing technologies mostly rely on macroscopic signal responses based on single physical principles for discrimination, lacking the ability to perform refined analysis and learning of the deep physical waveform characteristics generated by the interaction between defects and ultrasonic waves, making it difficult to accurately identify complex and diverse defects; 2. Insufficient multi-source information fusion and physical guidance. Existing methods typically treat signal processing, feature extraction, and defect inversion as relatively independent steps, failing to deeply embed multimodal information and physical prior knowledge from the detection process into the feature learning and decision-making model, thus limiting the robustness and interpretability of the model in complex industrial scenarios; 3. The system has a weak ability to handle uncertainties and difficult-to-detect samples. Existing systems mostly focus on detecting routine defects, lack a dynamic evaluation mechanism for the reliability of the detection results themselves, and also lack the ability to actively learn and self-optimize models for high-uncertainty samples or boundary cases, resulting in performance bottlenecks.

[0003] Chinese invention patent application CN202511545230.0 discloses a fully automated online detection system for internal defects in steel plates based on ultrasonic phased arrays, relating to the field of non-destructive testing technology. It includes an intelligent analysis module, communicatively connected to the online detection module, used to run an artificial intelligence rating model to process and initially evaluate the original ultrasonic testing data, outputting an initial evaluation result including the spatial location of the defects; and a high-precision verification module, also communicatively connected to the intelligent analysis module, used to extract corresponding sample steel plates based on the spatial location of the defects in the initial evaluation result, and to verify and measure them using high-precision detection methods to generate true defect data. However, this solution requires extracting sample steel plates based on the initial evaluation result for high-precision verification. This physical sampling method is destructive testing, and the verification process is significantly decoupled from the online detection process in time, making it impossible to utilize high-precision physical property constraints in real-time and non-destructively to improve the model's instantaneous prediction accuracy during the detection process.

[0004] Chinese invention patent application CN201380068387.8 provides a device and method for detecting internal defects in steel plates. The device includes: a total defect detection unit, which detects all defects, including surface defects and internal defects, based on the intensity of leakage magnetic flux measured by generating magnetic flux in the moving direction of the steel plate; a surface defect detection unit, which detects surface defects in a defined detection area, including all defects detected by the total defect detection unit, based on the intensity of leakage magnetic flux measured by generating magnetic flux along the thickness direction of the steel plate; and a data processing unit, which, for the detection area, subtracts the surface defects detected by the surface defect detection unit from all defects detected by the total defect detection unit, thereby detecting only internal defects present in the detection area. However, the data processing part of this scheme only performs simple subtraction to remove surface defects. This logic belongs to low-dimensional signal post-processing and fails to deeply embed complex geometric and physical relationships such as probe pose and sound beam path into the feature extraction network, resulting in poor robustness of the model in complex industrial noise environments. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a method for detecting internal defects in steel plates. In this method, a core dataset integrating precise physical parameters and waveform information is constructed, and a signal preprocessing module based on spatiotemporal alignment and interactive weighted fusion is designed. Then, deformable convolution and physical field inversion networks are used for deep feature extraction. Finally, the accurate identification of defect types and quantitative evaluation of attributes are achieved through physical simulation twin metrics and joint constraint regression.

[0006] The present invention provides a method for detecting internal defects in steel plates, comprising the following steps: S1, deploy a multi-channel ultrasonic phased array probe on the steel plate inspection platform to collect the full ultrasonic matrix, thereby obtaining the original radio frequency signal sequence, probe spatial position coordinates, sound beam incident angle, and transmit-receive transducer pair index matrix; S2 inputs the S1 data into the multimodal signal preprocessing module. Using the probe's spatial position coordinates, the sound beam incident angle, and the geometric relationship of the acoustic path defined by the transmit-receive transducer index matrix, a structured multi-channel spatiotemporal signal data volume is generated. The convolutional layer in the lightweight signal interaction network layer extracts the channel local temporal feature vector of each channel. The physical perception attention module based on sound path attenuation compensation generates a channel-level statistical description and learns and outputs the inter-channel interaction weight matrix. The inter-channel interaction weight matrix is ​​used to dynamically weight and sum the multi-channel spatiotemporal signal data volume, and initial weight biases are assigned to signals of specific state categories to obtain a fused signal field with enhanced signal-to-noise ratio. S3 inputs the signal-to-noise ratio enhanced fused signal field into the deep defect feature extraction module to extract a three-dimensional acoustic scattering intensity distribution map; S4. Input the acoustic scattering intensity distribution map into the defect detection model. The defect type classification subnetwork, spatial location regression subnetwork, and size regression subnetwork predict the probability distribution of the defect type, the three-dimensional spatial location coordinates of the defect, and the equivalent size of the defect, respectively.

[0007] Preferably, the specific process of generating the structured multi-channel spatiotemporal signal data volume is as follows: By utilizing the spatial coordinates of the probe, the incident angle of the sound beam, and the geometric relationship of the sound wave path defined by the index matrix of the transmitter-receiver transducer, the sound wave path represented by each A-scan signal is mapped to a unified, predefined virtual three-dimensional grid coordinate system inside the steel plate through ray tracing and time delay calculation. Each grid point is assigned a time-corrected signal amplitude from different sound wave paths, thereby generating a structured multi-channel spatiotemporal signal data volume, in which each channel corresponds to a unique sound wave path sampled in three-dimensional space.

[0008] Preferably, the convolutional layer in the lightweight signal interaction network layer first scans the temporal signal of each channel through a convolutional layer containing 32 one-dimensional convolutional kernels to extract the local temporal feature vector of each channel. Subsequently, the feature vector is fed into a physical perception attention module based on sound path attenuation compensation. First, a global average pooling layer is used to generate channel-level statistical descriptions. Then, two fully connected layers are used to learn the basic weights and combine them with the physical prior weight vector to output an inter-channel interaction weight matrix. Each element of the inter-channel interaction weight matrix quantifies the correlation strength of any two channel signals in representing defect scattering information. Finally, the inter-channel interaction weight matrix is ​​used to dynamically weight and sum the original multi-channel spatiotemporal signal data volume, and channels with signal state categories of high signal-to-noise ratio and strong defect scattering signals are given higher initial weight biases to obtain a fused signal field with enhanced signal-to-noise ratio where the number of channels is fused to one.

[0009] Preferably, the signal state category determination method is as follows: The signal-to-noise ratio (SNR) of the entire signal sequence and the energy attenuation coefficient of the signal within the theoretical scattering time window of the defect are calculated. The start and end times of this window are calculated by subtracting and adding a preset signal pulse half-width time, centered on the theoretical transit time obtained by dividing the sum of the distance from the transmitting transducer to the three-dimensional spatial position of the defect and the distance from the receiving transducer to the three-dimensional spatial position of the defect by the medium sound speed. The SNR calculation selects the first 50 sampling points of the original RF signal sequence as a pure noise sampling segment, and defines the root mean square value of the signal amplitude in this segment as the reference noise level. The energy attenuation coefficient is defined as the ratio of the average energy of the second half of the signal to the average energy of the first half of the signal within this time window. If the SNR of a signal sample is greater than 20 dB and its energy attenuation coefficient is less than 0.5, the signal state category of the sample is determined to be a high SNR, strong defect scattering signal. If the above conditions are not met, it is marked as a medium-low SNR, weak or complex scattering signal.

[0010] Preferably, the deep defect feature extraction module first utilizes an adaptive defect scattering feature extraction network composed of anisotropic deformable convolutional modules guided by the direction of acoustic beam propagation to extract primary, intermediate, and high-level scattering feature maps layer by layer, outputting a high-level abstract defect feature tensor; secondly, it constructs a data-driven acoustic scattering intensity map inversion network, whose encoder part reuses the adaptive defect scattering feature extraction network; finally, its decoder part maps the feature tensor into a three-dimensional data-driven acoustic scattering intensity distribution map through deconvolutional and convolutional layers.

[0011] Preferably, the adaptive extraction network for defect scattering features consists of a four-layer anisotropic deformable convolutional module guided by the direction of sound beam propagation. The first layer uses a 3×3×3 convolutional kernel to perform initial sampling in the input signal field. A parallel anisotropic offset learning network takes the input feature map and the recorded incident angle of the sound beam as conditional variables. First, the input features pass through a 3×3×3 convolutional layer to obtain intermediate spatial geometric features. Then, these features are concatenated with the incident angle vector. Finally, a 1×1×1 convolutional layer outputs a constrained offset field, utilizing the ultrasonic waves in the medium... The directionality of propagation constrains the sampling range of the convolution kernel, outputting a primary scattering feature map containing local waveform fluctuations. The second layer uses a 5×5×5 convolution kernel, taking the primary scattering feature map as input, to learn more complex spatial deformations, enabling the convolution kernel to adaptively focus on signal abrupt change regions, outputting a mid-level scattering feature map that characterizes the mid-scale scattering pattern. The third layer uses a 7×7×7 convolution kernel to extract a high-level scattering feature map reflecting the macroscopic contour of the defect. The fourth layer uses a 9×9×9 convolution kernel to integrate the multi-scale features of the first three layers, outputting a high-level abstract defect feature tensor containing fine geometric and orientation information of the defect.

[0012] Preferably, the data-driven acoustic scattering intensity map inversion network is an encoder-decoder structure; the encoder part reuses the trained defect scattering feature adaptive extraction network to encode the signal-to-noise ratio enhanced fused signal field into a high-level abstract defect feature tensor; the decoder part consists of three deconvolutional layers and two convolutional layers connected in series: the first deconvolutional layer upsamples the high-level abstract defect feature tensor by 2x in the spatial dimension to obtain a low-resolution initial inversion feature map; the second and third deconvolutional layers upsample by 2x in sequence to gradually restore spatial details and obtain a high-resolution inversion feature map; finally, two 3x3 convolutional layers are used for channel adjustment and detail optimization to output a three-dimensional data-driven acoustic scattering intensity distribution map.

[0013] Preferably, the defect detection model employs a physically guided defect attribute regression head, comprising three parallel sub-networks: the defect type classification sub-network first maps the input feature tensor to a 256-dimensional type hidden feature vector through a fully connected layer, then maps the type hidden feature vector to 3-dimensionality through another fully connected layer, and finally normalizes it through a Softmax layer, outputting the probability distribution of the defect belonging to the type of crack, pore, or inclusion; the spatial location regression sub-network consists of two fully connected layers, the first layer maps the input feature tensor to a 128-dimensional intermediate location feature, and the second layer directly maps the intermediate location feature and outputs the three-dimensional spatial location coordinates of the defect; the size regression sub-network has the same structure as the location sub-network and outputs the equivalent size of the defect.

[0014] Preferably, the defect detection model constructs a physically constrained twin network during training. This twin network contains two branches with identical structures and shared weights: a query branch and a reference branch. The input to the reference branch is simulation data generated using a finite element ultrasonic propagation model. This model is based on the elastic wave equation, setting the four sides of the computational domain as absorbing boundary conditions to simulate an infinitely large plate environment, and the upper and lower surfaces as free stress boundaries. It simulates the propagation and scattering of ultrasonic waves by setting steel plate material parameters and pre-defined defect geometric models based on the real physical properties of various defects. The defect geometric model is a simplified three-dimensional geometry built in the simulation software based on the labeled defect type, three-dimensional spatial location, and equivalent size of the defect. For crack types, a rectangular shape with tiny openings is used. The model is structured as follows: for pore types, an ellipsoidal model is used; for inclusion types, a polyhedral model with irregular surfaces is used. By assigning the three-dimensional spatial location and equivalent size parameters of the defects to the geometry, a radio frequency signal sequence and its synchronization parameters covering various defect scenarios are generated. The input of the query branch is the measured signal of the steel plate to be tested. Both branches are processed by an adaptive extraction network of defect scattering features, and each outputs a high-level abstract defect feature tensor. The physical information-constrained Siamese network is trained through a Siamese metric space learning module. The Siamese metric space learning module calculates the Euclidean distance between the feature tensors from the two branches and uses a triplet loss function. The training objective is to make the distance between measured features and simulated features with the same real physical properties of defects smaller than the distance between them and features with different properties.

[0015] Preferably, in actual deployment, five Siamese networks with physical information constraints are initialized and trained in parallel. The networks have the same structure but random initial weights, forming a multi-model ensemble system. For the measured signal of the same detection area, all five Siamese networks with physical information constraints run independently and output five sets of defect attribute prediction results. The statistical variance of the five sets of prediction results on each attribute is calculated. For the predicted three-dimensional spatial coordinates, if the variance of the X, Y, and Z coordinate components is less than 0.1 mm, the detection result is considered to have high reliability. If the variance of any coordinate component exceeds this threshold, the detection area is marked as a high uncertainty area and a reliability warning is triggered.

[0016] The innovative points and beneficial effects of this invention include: (1) Enhancing the depth of defect characterization and the accuracy of intelligent identification: Existing technologies mostly rely on macroscopic signal responses based on single physical principles such as ultrasound or magnetic leakage for discrimination, which makes it difficult to analyze the deep and microscopic scattering patterns generated by the interaction between defects and sound waves. This invention adaptively extracts the microscopic waveform features and spatial patterns of defect scattering through deformable convolutional networks, and combines them with a metric space constructed by a physical simulation twin network, enabling the model to learn and distinguish the essential scattering feature differences of different types of defects such as cracks, pores, and inclusions. This achieves a deep and intelligent mapping from signal to physical properties, significantly improving the accuracy of identification and classification of complex defects. (2) Achieving deep fusion of multi-source physical information and enhancing the physical interpretability of the model: Existing methods typically separate signal processing, feature extraction, and defect inversion, lacking in-depth utilization of geometric and physical relationships such as precise probe pose and acoustic beam path. In the preprocessing stage, this invention maps multi-channel signals to a unified spatial grid through the geometric relationship of acoustic wave paths, achieving signal alignment in physical space; in the feature extraction and decision-making stage, an acoustic scattering intensity map is generated through a physical field inversion network, which is used to constrain the physical spatial consistency of attribute regression. This approach of deeply embedding physical mechanisms into the data-driven model not only improves the robustness of the model in complex noise environments but also enhances the physical interpretability of the entire detection process; (3) Strengthening the quantitative management of detection uncertainty and the system's adaptive evolution capability: Existing systems lack an evaluation mechanism for the reliability of detection results themselves, making it difficult to handle difficult-to-detect samples with ambiguous boundaries or low signal-to-noise ratios. This invention achieves objective evaluation and early warning of the reliability of a single detection by using multi-model integration and parallel inference to dynamically calculate the statistical variance of the prediction results. More importantly, the system can automatically archive high-uncertainty cases and perform targeted fine-tuning of the model through subsequent re-verification and annotation, forming a closed-loop online self-optimization mechanism. This enables the system to continuously learn from failed or difficult cases, constantly expand the stable detection range, and achieve self-improvement in detection performance and reliability. Attached Figure Description

[0017] Figure 1 This is a flowchart of the overall method of the present invention.

[0018] Figure 2 This is a structural diagram of the deep defect feature extraction module based on adaptive convolution and physical field inversion of the present invention.

[0019] Figure 3 This is a structural diagram of the defect detection model based on physical simulation twin metric and joint constraint regression of the present invention.

[0020] Figure 4 This is a diagram of the internal structure of the steel plate in an embodiment of the present invention.

[0021] Figure 5The figure shows the experimental results of the regression performance analysis of the physical properties of defects under complex noise environment in an embodiment of the present invention.

[0022] Figure 6 The figure shows the experimental results of the online self-optimization effect analysis of the system based on uncertainty mining in the embodiments of the present invention. Detailed Implementation

[0023] The present invention proposes a method for detecting internal defects in steel plates, the overall process of which is as follows: Figure 1 As shown: S1. First, a multi-channel ultrasonic phased array probe is deployed on the steel plate inspection platform. For steel plate samples containing artificial pre-made defects and natural defects, the core data ultrasonic full matrix is ​​collected to capture the original radio frequency signal sequence. Second, the probe spatial position coordinates, sound beam incident angle, and transmit-receive transducer pair index matrix corresponding to each set of signals are recorded simultaneously and accurately. Third, the true three-dimensional morphology of the defects is obtained using X-ray computed tomography, and the defect type, three-dimensional spatial location, and equivalent size of the defects are labeled accordingly. All the above data are then encapsulated to construct an ultrasonic core dataset for deep feature learning.

[0024] S2 inputs the data obtained from S1 into a multimodal signal preprocessing module based on spatiotemporal parameter alignment and interactive weighted fusion. First, the signal is preprocessed. Utilizing the geometric relationship of the acoustic wave path and based on the assumption of a homogeneous isotropic medium velocity model, the acoustic wave path is mapped to a virtual three-dimensional grid coordinate system with a grid step size less than one-quarter of the ultrasonic wavelength through ray tracing and time delay calculation, generating a structured multi-channel spatiotemporal signal data volume. Second, the convolutional layer in the lightweight signal interaction network layer extracts the channel local temporal feature vector of each channel. Third, a physical perception attention module based on acoustic path attenuation compensation generates channel-level statistical descriptions and learns and outputs the inter-channel interaction weight matrix. Finally, the inter-channel interaction weight matrix is ​​used to dynamically weight and sum the multi-channel spatiotemporal signal data volume, and initial weight biases are assigned to signals of specific state categories to obtain a fused signal field with enhanced signal-to-noise ratio.

[0025] S3 inputs the signal-to-noise ratio enhanced fused signal field into the deep defect feature extraction module based on adaptive convolution and physical field inversion. First, it utilizes a defect scattering feature adaptive extraction network composed of anisotropic deformable convolution modules guided by the direction of sound beam propagation to extract primary, intermediate, and high-level scattering feature maps layer by layer, outputting a high-level abstract defect feature tensor. Second, it constructs a data-driven acoustic scattering intensity map inversion network, whose encoder part reuses the defect scattering feature adaptive extraction network. Finally, its decoder part maps the feature tensor into a three-dimensional data-driven acoustic scattering intensity distribution map through deconvolution and convolution layers, and uses the defect contour approximation constructed from the real physical properties of the defects marked in S1 as a soft supervision reference for training.

[0026] S4. Construct a defect detection model based on physical simulation twin metric and joint constraint regression. First, establish a twin network with physical information constraints. The reference branch inputs simulation data generated based on the elastic wave equation and defect geometry model, and the query branch inputs measured signals. Both use a defect scattering feature adaptive extraction network to extract high-level abstract defect feature tensors. Second, use the twin metric space learning module to calculate the Euclidean distance between feature tensors and use a triplet loss function. Third, design a physical-guided defect attribute regression head. Through defect type classification subnetwork, spatial location regression subnetwork, and size regression subnetwork, predict the probability distribution of defect type, the three-dimensional spatial location coordinates of the defect, and the equivalent size of the defect, respectively. Finally, combine regression loss and physical consistency loss for joint constraint. The physical consistency loss calculates the volume crossover ratio between the predicted bounding box and the significantly highlighted areas in the data-driven acoustic scattering intensity distribution map.

[0027] S5, during system deployment, runs a system reliability enhancement platform based on integrated uncertainty quantification and closed-loop feedback. First, it initializes and trains multiple twin networks with physical information constraints in parallel to form a multi-model integrated system, which runs independently and outputs multiple sets of defect attribute prediction results. Second, the multi-model integrated reliability assessment unit calculates the statistical variance of the prediction results, marking areas with variance exceeding a threshold as high-uncertainty areas and triggering reliability warnings. Third, it saves the data from high-uncertainty areas to an incremental sample database and uses higher-precision detection methods to supplement or confirm the labeling of the true physical attributes of the defects. Finally, it uses samples from the incremental sample database to fine-tune all parameters in the multi-model integrated system, achieving continuous self-improvement in detection performance and result reliability.

[0028] The specific implementation of the present invention will be further described below with reference to specific embodiments.

[0029] S1. Construction of the core ultrasonic dataset for deep feature learning First, a multi-channel ultrasonic phased array probe and a high-precision synchronous acquisition system are deployed on a steel plate inspection platform. For steel plate samples containing artificially created defects and natural defects, the core data ultrasonic full matrix is ​​acquired to capture the original radio frequency signal sequence. Second, the probe spatial position coordinates, beam incident angle, and transmit-receive transducer pair index matrix corresponding to each set of signals are recorded synchronously and accurately. Third, the true three-dimensional morphology of the defects is obtained using X-ray computed tomography, and the defect type, three-dimensional spatial location, and equivalent size of the defects are labeled accordingly. All the above data are then encapsulated to construct an ultrasonic core dataset for deep feature learning.

[0030] S1-1 Core Signal and Synchronization Parameter Acquisition: A multi-channel ultrasonic phased array probe and a high-precision synchronous acquisition system are deployed on the steel plate inspection platform. For steel plate samples containing artificially created defects and natural defects, the core data is acquired through ultrasonic full-matrix capture of the original radio frequency signal sequence. This sequence is a collection of original A-scan radio frequency signals obtained by sequentially exciting and receiving all transmitter-receiver pairs, completely preserving the physical waveform information of the interaction between the sound wave and the defect. The probe's spatial position coordinates, beam incidence angle, and a transmitter-receiver pair index matrix are synchronously and accurately recorded for each set of ultrasonic full-matrix capture of the original radio frequency signal sequence. Each row of the transmitter-receiver pair index matrix uniquely identifies a specific transmitter-receiver pair.

[0031] S1-2 Refined Defect Physical Attribute Labeling: X-ray computed tomography (CT) was used to scan and reconstruct the steel plate sample in three dimensions to obtain the true three-dimensional morphology of the defects. Based on the CT reconstruction results, the corresponding true physical attributes of the defects were labeled for each set of original radio frequency signal sequences acquired in S1-1 using the full-matrix ultrasonic wave capture method. These attributes included: defect type, three-dimensional spatial location of the defect, and equivalent size of the defect. The defect type was classified into three categories based on the actual situation: cracks, porosity, and inclusions. The three-dimensional spatial location of the defect was represented by coordinate values ​​in the length, width, and depth directions, and the equivalent size of the defect was represented by dimension values ​​in the length, width, and depth directions. This labeling process established a precise correspondence between the original ultrasonic wave signals and the physical attributes of the defects.

[0032] The original radio frequency signal sequence captured by the full ultrasonic matrix, the probe spatial position coordinates, the incident angle of the sound beam, the transmit-receive transducer pair index matrix, and the real physical properties of defects labeled in S1-2 are encapsulated together to construct an ultrasonic core dataset for deep feature learning.

[0033] S2. Multimodal signal preprocessing module based on spatiotemporal parameter alignment and interactive weighted fusion The data obtained from S1 is input into a multimodal signal preprocessing module based on spatiotemporal parameter alignment and interactive weighted fusion. First, the signal is preprocessed. Utilizing the geometric relationship of the acoustic wave path and based on the assumption of a homogeneous isotropic medium velocity model, the acoustic wave path is mapped to a virtual three-dimensional grid coordinate system with a grid step size less than one-quarter of the ultrasonic wavelength through ray tracing and time delay calculation, generating a structured multi-channel spatiotemporal signal data volume. Second, the convolutional layer in the lightweight signal interaction network layer extracts the channel local temporal feature vector of each channel. Third, a physical perception attention module based on sound path attenuation compensation generates channel-level statistical descriptions and learns and outputs the inter-channel interaction weight matrix. Finally, the inter-channel interaction weight matrix is ​​used to dynamically weight and sum the multi-channel spatiotemporal signal data volume, and initial weight biases are assigned to signals of specific state categories to obtain a fused signal field with enhanced signal-to-noise ratio.

[0034] S2-1 Multi-channel Signal Spatiotemporal Alignment Based on Virtual Mesh Mapping: This method preprocesses the original RF signal sequences from the full-matrix ultrasonic acquisition in the core ultrasonic dataset, including bandpass filtering and gain compensation to suppress noise. Innovatively, it utilizes the probe's spatial coordinates, beam incident angle, and the geometric relationship of the acoustic path defined by the transmit-receive transducer index matrix. Through ray tracing and time delay calculation, it maps the acoustic path represented by each Amplitude Scan signal onto a unified, predefined virtual three-dimensional mesh coordinate system within the steel plate. This process assigns time-corrected signal amplitudes from different acoustic paths to each mesh point, generating a structured multi-channel spatiotemporal signal data volume. Each channel corresponds to a unique acoustic path sampled in three-dimensional space, achieving precise alignment of multi-view signals in both spatial and temporal dimensions.

[0035] S2-2 Signal-Level Cross-Perspective Fusion Based on Lightweight Interactive Network: A lightweight signal interactive network layer is designed to directly fuse the multi-channel spatiotemporal signal data volume output from S2-1. The lightweight signal interactive network layer first scans the temporal signal of each channel through a convolutional layer containing 32 one-dimensional convolutional kernels, extracting the local temporal feature vector of each channel. Subsequently, these feature vectors are fed into a physics-aware attention module based on sound path attenuation compensation. This module no longer relies purely on data-learned weights but introduces a physical prior weight vector determined by the sound wave propagation distance calculated in S2-1. This explicitly encodes the spherical wave diffusion attenuation law from physics into the attention mechanism, forcing the network to focus on channels with shorter sound paths and higher theoretical energy, thus making the attention mechanism physically interpretable. The physics-aware attention module first generates channel-level statistical descriptions through a global average pooling layer. Then, it learns basic weights through two fully connected layers and combines these with a physics-prior weight vector to output an inter-channel interaction weight matrix. Each element of this matrix quantifies the correlation strength between any two channel signals in representing defect scattering information. Finally, the original multi-channel spatiotemporal signal data volume is dynamically weighted and summed using the inter-channel interaction weight matrix. Channels with a signal state category of "high signal-to-noise ratio, strong defect scattering signal" are assigned higher initial weight biases, resulting in a new, signal-to-noise ratio-enhanced fused signal field with a unified channel count. This process strengthens the common components related to potential defect scattering in all channel signals while suppressing random noise and irrelevant structural echoes.

[0036] The signal state category is determined by calculating the signal-to-noise ratio (SNR) of the entire signal sequence and the energy attenuation coefficient of the signal within the theoretical scattering time window of the defect. The start and end times of this window are calculated by subtracting and adding a preset half-width signal pulse time, centered on the theoretical transit time obtained by dividing the sum of the distances from the transmitting transducer to the three-dimensional spatial location of the defect and the distances from the receiving transducer to the three-dimensional spatial location of the defect by the speed of sound in the medium. Specifically, the original radio frequency (RF) signal sequence acquired by S1-1 is quantitatively analyzed, and two objective indicators are calculated for each signal sample: the SNR of the entire signal sequence and the energy attenuation coefficient of the signal within the theoretical scattering time window of the defect. For the SNR calculation, the first 50 sampling points of the original RF signal sequence captured by the ultrasonic full matrix are selected as the pure noise sampling segment, and the root mean square (RMS) value of the signal amplitude in this segment is defined as the reference noise level. The energy attenuation coefficient is defined as the ratio of the average energy of the second half of the signal to the average energy of the first half of the signal within this time window. If a signal sample has a signal-to-noise ratio (SNR) greater than 20 dB and an energy attenuation coefficient less than 0.5, it is determined that the sample comes from a clear measurement with a high SNR and dominated by defect scattering signals, and its signal state category is marked as a high SNR, strong defect scattering signal; if the above conditions are not met, it is marked as a medium-low SNR, weak or complex scattering signal.

[0037] S3. Deep Defect Feature Extraction Module Based on Adaptive Convolution and Physical Field Inversion The signal-to-noise ratio enhanced fused signal field is input into a deep defect feature extraction module based on adaptive convolution and physical field inversion, with the structure as follows: Figure 2 As shown, firstly, an adaptive extraction network for defect scattering features, composed of anisotropic deformable convolutional modules guided by the direction of sound beam propagation, is used to extract primary, intermediate, and high-level scattering feature maps layer by layer, outputting a high-level abstract defect feature tensor. Secondly, a data-driven acoustic scattering intensity map inversion network is constructed, with its encoder part reusing the adaptive extraction network for defect scattering features. Finally, its decoder part maps the feature tensor to a three-dimensional data-driven acoustic scattering intensity distribution map through deconvolutional and convolutional layers, and uses the defect contour approximation constructed from the real physical properties of the defect labeled in S1 as a soft supervision reference for training.

[0038] S3-1 Adaptive Extraction of Defect Scattering Features Based on Anisotropic Deformable Convolution: The signal-to-noise ratio enhanced fused signal field output from S2-2 is input into the adaptive extraction network for defect scattering features. The main body of the adaptive extraction network for defect scattering features consists of a four-layer anisotropic deformable convolution module guided by the direction of sound beam propagation. The first layer uses a 3×3×3 convolutional kernel for initial sampling in the input signal field. A parallel anisotropic offset learning network takes the input feature map and the incident beam angle recorded in S1 as conditional variables. This sub-network first passes the input features through a 3×3×3 convolutional layer to obtain intermediate spatial geometric features. Then, this feature is concatenated with the incident angle vector, and finally, a 1×1×1 convolutional layer outputs a constrained offset field. This restricts the offset of the convolutional kernel to deform primarily along the direction of sound beam propagation and its normal, rather than undergoing free deformation. Utilizing the directionality of ultrasound propagation in the medium, the sampling range of the convolutional kernel is constrained, making it more consistent with the physical laws of sound wave scattering and reducing interference from invalid background noise. The output is a primary scattering feature map containing local waveform fluctuations. The second layer uses a 5×5×5 convolutional kernel, taking the primary scattering feature map as input, to further learn more complex spatial deformations. This allows the convolutional kernel to adaptively focus on regions of signal abrupt change, outputting a mid-level scattering feature map that characterizes mesoscale scattering patterns. The third layer uses a 7×7×7 convolutional kernel to extract a high-level scattering feature map that reflects the macroscopic contour of the defect. The fourth layer uses a 9×9×9 convolutional kernel to integrate the multi-scale features of the first three layers and output a high-level abstract defect feature tensor containing fine geometric and orientation information of the defect.

[0039] S3-2 Implicit Physics Field Inversion Based on Encoder-Decoder Structure: A data-driven acoustic scattering intensity map inversion network is constructed, which is an encoder-decoder structure. The encoder part reuses the defect scattering feature adaptive extraction network trained in S3-1 to encode the signal-to-noise ratio enhanced fused signal field into a high-level abstract defect feature tensor. The decoder part consists of three deconvolutional layers and two convolutional layers connected in series: the first deconvolutional layer upsamples the high-level abstract defect feature tensor by 2x in the spatial dimension to obtain a low-resolution initial inversion feature map; the second and third deconvolutional layers upsample by 2x sequentially to gradually restore spatial details and obtain a high-resolution inversion feature map; finally, two 3x3 convolutional layers are used for channel adjustment and detail optimization to output a three-dimensional data-driven acoustic scattering intensity distribution map. The training of the defect scattering feature adaptive extraction network uses the defect contour approximation constructed based on the real physical properties of the defect obtained by CT in S1-2 as a soft supervision reference. The specific method for constructing the defect contour approximation is as follows: First, initialize a three-dimensional mesh tensor with the same resolution as the output distribution map. Then, with the three-dimensional spatial coordinates of the defect marked in S1-2 as the center, use the length, width and depth of the equivalent size of the defect to define the covariance matrix of the three-dimensional Gaussian function, generate a three-dimensional Gaussian heat map, and map and fill it into the mesh tensor, thereby forming a soft mask that can reflect the spatial occupancy probability of the defect, so that the output distribution map approximates the real defect scatterer distribution in spatial morphology, realizing the end-to-end mapping from the original signal to the approximate physical imaging.

[0040] S4. Defect detection model based on physical simulation twin metric and joint constraint regression A defect detection model based on physical simulation twin metrics and joint constraint regression is constructed, and the model structure is as follows: Figure 3 As shown, firstly, a physical information-constrained Siamese network is established. The reference branch input is simulation data generated based on the elastic wave equation and the defect geometric model, and the query branch input is the measured signal. Both are output as high-level abstract defect feature tensors through an adaptive extraction network of defect scattering features. Secondly, the Euclidean distance between feature tensors is calculated using a Siamese metric space learning module, and a triplet loss function is applied. Thirdly, a physical-guided defect attribute regression head is designed, which predicts the defect type probability distribution, the three-dimensional spatial coordinates of the defect, and the equivalent size of the defect through a defect type classification sub-network, a spatial location regression sub-network, and a size regression sub-network, respectively. Finally, joint constraints are applied by combining regression loss and physical consistency loss, where the physical consistency loss calculates the volume crossover ratio between the predicted bounding box and the significantly highlighted areas in the data-driven acoustic scattering intensity distribution map.

[0041] S4-1 Construction of a Physical Simulation-Driven Twin Metric Space: This involves constructing a physically constrained twin network. This network comprises two branches with identical structures and shared weights: a query branch and a reference branch. The reference branch receives simulation data generated using a finite element ultrasonic propagation model. This model is based on the elastic wave equation, simulating an infinitely large plate environment by setting the four sides of the computational domain as absorbing boundary conditions and the upper and lower surfaces as free stress boundaries. The propagation and scattering of ultrasonic waves are simulated by setting steel plate material parameters and pre-defined defect geometric models based on the real physical properties of various defects in S1-2. The defect geometric model is a simplified three-dimensional geometry built in the simulation software based on the defect type, three-dimensional spatial location, and equivalent size labeled in S1-2: for crack types, a cuboid model with tiny openings is used; for porosity types, an ellipsoid model is used; and for inclusion types, a polyhedron model with irregular surfaces is used. By assigning the three-dimensional spatial location and equivalent size parameters of the defects from S1-2 to these geometries, a simulated ultrasonic full-matrix acquisition signal sequence and its synchronization parameters covering various defect scenarios can be generated. The input to the query branch is the measured signal of the steel plate under test. Both branches are processed sequentially by the defect scattering feature adaptive extraction network described in S3, and each ultimately outputs a high-level abstract defect feature tensor. The physically constrained Siamese network is trained through a Siamese metric space learning module, which calculates the Euclidean distance between the feature tensors from the two branches and uses a triplet loss function. The training objective is to ensure that the distance between measured features and simulated features with the same real physical properties of defects is less than the distance between them and features with different properties.

[0042] S4-2 Physically Guided Defect Attribute Direct Regression and Consistency Constraints: A physically guided defect attribute regression head is designed, taking the high-level abstract defect feature tensor output by the query branch as input. The regression head contains three parallel sub-networks: The first is a defect type classification sub-network. This sub-network first maps the input feature tensor to a 256-dimensional type hidden feature vector through a fully connected layer, then maps this type hidden feature vector to 3D through another fully connected layer, and finally normalizes it through a Softmax layer, outputting the probability distribution of the defect type (crack, porosity, inclusion). The second is a spatial location regression sub-network, consisting of two fully connected layers. The first layer maps the input feature tensor to a 128-dimensional intermediate location feature, and the second layer directly maps and outputs the 3D spatial location coordinates of the defect. The third is a size regression sub-network, with the same structure as the location sub-network, outputting the equivalent size of the defect. The total loss function of the physics-guided defect attribute regression head consists of two parts: the first part is the regression loss, which is a weighted sum of cross-entropy loss and mean squared error loss, directly supervising the approximation of the predicted value to the true physical attributes of the defect labeled in S1-2; the second part is the physical consistency loss, which constructs a three-dimensional bounding box from the predicted defect location and size, and calculates the volume cross-union ratio between it and the significantly bright regions after binarization in the acoustic scattering intensity distribution map driven by the data output of S3-2. During binarization, 50% of the global maximum intensity value in the current distribution map is selected as the segmentation threshold, and the consistency between the numerical prediction and the imaging results in physical space is constrained by maximizing this ratio.

[0043] S5, System Reliability Enhancement Platform Based on Integrated Uncertainty Quantification and Closed-Loop Feedback In the system deployment phase, a system reliability enhancement platform based on integrated uncertainty quantification and closed-loop feedback is run. First, multiple twin networks constrained by physical information are initialized and trained in parallel to form a multi-model integrated system, which runs independently and outputs multiple sets of defect attribute prediction results. Second, the multi-model integrated reliability assessment unit calculates the statistical variance of the prediction results, marks areas with variance exceeding a threshold as high uncertainty areas, and triggers reliability warnings. Third, the data of high uncertainty areas are saved to an incremental sample database, and higher-precision detection methods are used to supplement or confirm the labeling of their true physical attributes of defects. Finally, samples from the incremental sample database are used to fine-tune all parameters in the multi-model integrated system to achieve continuous self-improvement of detection performance and result reliability.

[0044] S5-1 Multi-Model Integration and Dynamic Reliability Assessment for Detection: In actual deployment, five Siamese networks constrained by physical information as described in S4 are initialized and trained in parallel. These networks have the same structure but randomized initial weights, forming a multi-model integration system. For measured signals in the same detection area, all five Siamese networks constrained by physical information operate independently, outputting five sets of defect attribute prediction results. The multi-model integration reliability assessment unit calculates the statistical variance of these five sets of prediction results for each attribute. For the predicted three-dimensional spatial coordinates, if the variance of each of the X, Y, and Z coordinate components is less than 0.1 mm, the detection result is considered to have high reliability; if the variance of any coordinate component exceeds this threshold, the detection area is marked as a high-uncertainty area, and a reliability warning is triggered, prompting the inspection personnel to review the data.

[0045] S5-2 Online Self-Optimization Mechanism Based on Uncertainty Sample Mining: An online self-optimization mechanism based on uncertainty feedback is established. The system automatically packages and saves the original ultrasonic full-matrix capture radio frequency signal sequences, along with all acquisition parameters and intermediate processing results, corresponding to the high-uncertainty regions marked in S5-1, into an incremental sample database. A batch of samples is periodically extracted from the incremental sample database. For these samples, higher-precision detection methods are prioritized for re-verification to supplement or confirm the true physical attribute annotations of the defects. Subsequently, these newly annotated, challenging samples are used to perform a round of targeted fine-tuning of the parameters in the deployed multi-model ensemble system. The specific fine-tuning strategy is to freeze all weights of the deep defect feature extraction module described in S3, update only the fully connected layer parameters in the physically guided defect attribute regression head described in S4-2, and set the learning rate to 10% of the initial training learning rate. This closed-loop process enables the system to continuously learn from difficult-to-distinguish cases, gradually expanding its stable detection range and reducing prediction uncertainty.

[0046] Experimental verification and analysis: To verify the effectiveness of the proposed method for constructing an ultrasonic core dataset and detecting defects based on deep feature learning, this experiment uses ultrasonic full-matrix capture data of various types of steel plate defects as the research object. The focus is on evaluating the accuracy of physical property regression and the performance of the online adaptive optimization mechanism under complex noise environments. Experimental metrics include mean absolute error of defect size, signal-to-noise ratio robustness, and the improvement in online detection accuracy. Figure 4 The internal structure of the electro-galvanized steel sheet is shown, displaying cross-sectional views of the zinc (Zn) and steel (Steel) layers, with the interface clearly visible and irregular areas marked in red.

[0047] 1. Regression performance analysis of physical properties of defects under complex noise environment The experiment first selected test samples with different signal-to-noise ratio levels based on the constructed ultrasonic core dataset to evaluate the accuracy of the defect detection model based on physical simulation twin metric and joint constraint regression in predicting the equivalent size of defects. To visually demonstrate the advantages of the multimodal signal preprocessing module S2 based on spatiotemporal parameter alignment and interactive weighted fusion and the deep defect feature extraction module S3 based on adaptive convolution and physical field inversion in this invention in terms of noise resistance, the experiment introduced a standard three-dimensional convolutional network and a traditional full-focus imaging method as comparative benchmarks. Figure 5 As shown, the horizontal axis represents the signal-to-noise ratio level of the input signal, the vertical axis represents the mean absolute error of the defect size, and the three broken lines represent the performance change trends of different methods.

[0048] Experimental results show that as the signal-to-noise ratio (SNR) decreases, the error of traditional all-focusing imaging methods increases significantly, indicating that they heavily rely on signal quality and their beam focusing capability is greatly weakened under noise interference. While standard 3D convolutional networks possess some feature extraction capabilities, they still exhibit significant performance degradation in the low SNR range, making it difficult to extract effective information from weak echoes submerged in noise. In contrast, the method of this invention maintains the lowest mean absolute error across the entire SNR range. Particularly under low SNR conditions, the error curve is flat, demonstrating that the spatiotemporal alignment of multi-channel signals based on virtual mesh mapping in S2 and the interactive weighted fusion in S2-2 significantly suppress random noise. Simultaneously, the 3D deformable convolutional module in S3-1 effectively extracts weak defect scattering features. Furthermore, the physical consistency loss introduced in S4-2 constrains the geometric correspondence between the regression results and physical imaging, further ensuring the accuracy of size prediction under complex acoustic field interference.

[0049] 2. Analysis of the Online Self-Optimization Effect of the System Based on Uncertainty Mining To verify the actual performance of the system reliability enhancement platform based on integrated uncertainty quantification and closed-loop feedback described in S5, an experiment simulated the online operation scenario after system deployment. Through multiple iterations, the experiment compared and analyzed three different strategies: the uncertainty sample mining strategy proposed in this invention, the random sample sampling strategy, and the static model benchmark without updates. Figure 6 As shown, the horizontal axis represents the number of online optimization iterations, the vertical axis represents the defect detection accuracy, and the lines of different colors reflect the trajectory of system performance evolution over time.

[0050] Analysis shows that the accuracy of the static model baseline remains at the initial level, unable to adapt to newly emerging complex defect samples, and even fluctuates slightly due to minor drifts in field conditions. While the random sample sampling strategy gradually improves performance by increasing training data, the convergence speed is slow and the improvement is limited, indicating that a large number of redundant simple samples do not provide sufficient information gain. In stark contrast, the strategy proposed in this invention exhibits a rapid increase followed by a stable high level. This is because the multi-model ensemble system in S5-1 can accurately identify high-uncertainty regions, and the S5-2 mechanism selectively filters these difficult sample examples located at the decision boundary, storing them in the incremental sample database for fine-tuning. This mechanism allows the model to prioritize learning features in existing knowledge blind spots, efficiently correcting parameter biases in the physically constrained Siamese network, achieving continuous self-improvement in detection performance and result reliability, and verifying the reliability and superiority of the closed-loop feedback mechanism in long-term operation.

[0051] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0052] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for detecting internal defects in steel plates, characterized in that, The process includes the following: S1, deploy a multi-channel ultrasonic phased array probe on the steel plate inspection platform to collect the full ultrasonic matrix, thereby obtaining the original radio frequency signal sequence, probe spatial position coordinates, sound beam incident angle, and transmit-receive transducer pair index matrix; S2 inputs the S1 data into the multimodal signal preprocessing module. Using the probe's spatial position coordinates, the sound beam incident angle, and the geometric relationship of the acoustic path defined by the transmit-receive transducer index matrix, a structured multi-channel spatiotemporal signal data volume is generated. The convolutional layer in the lightweight signal interaction network layer extracts the channel local temporal feature vector of each channel. The physical perception attention module based on sound path attenuation compensation generates a channel-level statistical description and learns and outputs the inter-channel interaction weight matrix. The inter-channel interaction weight matrix is ​​used to dynamically weight and sum the multi-channel spatiotemporal signal data volume, and initial weight biases are assigned to signals of specific state categories to obtain a fused signal field with enhanced signal-to-noise ratio. S3 inputs the signal-to-noise ratio enhanced fused signal field into the deep defect feature extraction module to extract a three-dimensional acoustic scattering intensity distribution map; S4. Input the acoustic scattering intensity distribution map into the defect detection model. The defect type probability distribution, the three-dimensional spatial coordinates of the defect, and the equivalent size of the defect are predicted by the defect type classification subnetwork, spatial location regression subnetwork, and size regression subnetwork, respectively.

2. The method for detecting internal defects in steel plates as described in claim 1, characterized in that: The specific process for generating the structured multi-channel spatiotemporal signal data volume is as follows: By utilizing the spatial coordinates of the probe, the incident angle of the sound beam, and the geometric relationship of the sound wave path defined by the index matrix of the transmitter-receiver transducer, the sound wave path represented by each A-scan signal is mapped to a unified, predefined virtual three-dimensional grid coordinate system inside the steel plate through ray tracing and time delay calculation. Each grid point is assigned a time-corrected signal amplitude from different sound wave paths, thereby generating a structured multi-channel spatiotemporal signal data volume, in which each channel corresponds to a unique sound wave path sampled in three-dimensional space.

3. The method for detecting internal defects in steel plates as described in claim 2, characterized in that: The convolutional layer in the lightweight signal interaction network layer first scans the temporal signal of each channel through a convolutional layer containing 32 one-dimensional convolutional kernels, and extracts the channel local temporal feature vector of each channel. Subsequently The feature vectors are fed into a physical perception attention module based on sound path attenuation compensation. First, a global average pooling layer is used to generate channel-level statistical descriptions. Then, two fully connected layers are used to learn the basic weights and combine them with the physical prior weight vector to output an inter-channel interaction weight matrix. Each element of the inter-channel interaction weight matrix quantifies the correlation strength between any two channel signals in representing defect scattering information. Finally, the inter-channel interaction weight matrix is ​​used to dynamically weight and sum the original multi-channel spatiotemporal signal data volume, and channels with high signal-to-noise ratio and strong defect scattering signals are given higher initial weight biases, resulting in a fused signal field with enhanced signal-to-noise ratio where the number of channels is fused to one.

4. The method for detecting internal defects in steel plates as described in claim 3, characterized in that: The signal state category determination method is as follows: calculate the signal-to-noise ratio (SNR) of the entire signal sequence and the energy attenuation coefficient of the signal within the theoretical scattering time window of the defect. The start and end times of this window are calculated by subtracting and adding the preset half-width signal pulse time, respectively, from the theoretical transit time obtained by dividing the sum of the distance from the transmitting transducer to the three-dimensional spatial position of the defect and the distance from the receiving transducer to the three-dimensional spatial position of the defect by the medium sound speed. When calculating the SNR, the first 50 sampling points of the original radio frequency signal sequence are selected as the pure noise sampling segment, and the root mean square value of the signal amplitude of this segment is defined as the reference noise level. The energy attenuation coefficient is defined as the ratio of the average energy of the second half of the signal to the average energy of the first half of the signal within this time window. If the signal-to-noise ratio of a signal sample is greater than 20 dB and its energy attenuation coefficient is less than 0.5, the signal state category of the sample is determined to be a high signal-to-noise ratio, strong defect scattering signal; if the above conditions are not met, it is marked as a medium-low signal-to-noise ratio, weak or complex scattering signal.

5. The method for detecting internal defects in steel plates as described in claim 1, characterized in that: The deep defect feature extraction module first uses an adaptive extraction network for defect scattering features, which is composed of anisotropic deformable convolutional modules guided by the direction of sound beam propagation, to extract primary scattering feature maps, intermediate scattering feature maps and high-level scattering feature maps layer by layer, and outputs a high-level abstract defect feature tensor. Secondly, a data-driven acoustic scattering intensity map inversion network is constructed, in which the encoder part reuses the defect scattering feature adaptive extraction network; Finally, its decoder section maps the feature tensors into a three-dimensional data-driven acoustic scattering intensity distribution map through deconvolutional and convolutional layers.

6. The method for detecting internal defects in steel plates as described in claim 5, characterized in that: The adaptive extraction network for defect scattering features consists of a four-layer anisotropic deformable convolutional module guided by the direction of sound beam propagation. The first layer uses a 3×3×3 convolutional kernel to perform initial sampling in the input signal field. The input feature map and the recorded incident angle of the sound beam are input as conditional variables through a parallel anisotropic offset learning network. First, the input features are passed through a 3×3×3 convolutional layer to obtain intermediate spatial geometric features. Then, these features are concatenated with the incident angle vector. Finally, a 1×1×1 convolutional layer is passed to output a constrained offset field. The directionality of ultrasonic waves propagating in the medium is used to constrain the sampling range of the convolutional kernel, and the output is a primary scattering feature map containing local waveform fluctuation features. The second layer uses a 5×5×5 convolutional kernel. Taking the primary scattering feature map as input, it learns more complex spatial deformations, enabling the convolutional kernel to adaptively focus on signal change regions and output an intermediate scattering feature map that can characterize the mesoscale scattering mode. The third layer uses a 7×7×7 convolution kernel to extract a high-level scattering feature map that reflects the macroscopic contour of the defect; The fourth layer uses a 9×9×9 convolution kernel to integrate the multi-scale features of the first three layers and outputs a high-level abstract defect feature tensor containing fine geometric and orientation information of the defects.

7. The method for detecting internal defects in steel plates as described in claim 6, characterized in that: The data-driven acoustic scattering intensity map inversion network is an encoder-decoder structure; the encoder part reuses the trained defect scattering feature adaptive extraction network to encode the signal-to-noise ratio enhanced fused signal field into a high-level abstract defect feature tensor. The decoder consists of three deconvolutional layers and two convolutional layers connected in series: the first deconvolutional layer upsamples the high-level abstract defect feature tensor by 2x in the spatial dimension to obtain a low-resolution initial inversion feature map; the second and third deconvolutional layers upsample by 2x in sequence to gradually restore spatial details and obtain a high-resolution inversion feature map; finally, two 3x3 convolutional layers are used for channel adjustment and detail optimization to output a three-dimensional data-driven acoustic scattering intensity distribution map.

8. The method for detecting internal defects in steel plates as described in claim 1, characterized in that: The defect detection model employs a physically guided defect attribute regression head, comprising three parallel sub-networks: the defect type classification sub-network first maps the input feature tensor to a 256-dimensional type hidden feature vector through a fully connected layer, then maps this type hidden feature vector to 3D through another fully connected layer, and finally normalizes it through a Softmax layer, outputting the probability distribution of the defect type (crack, porosity, inclusion); the spatial location regression sub-network consists of two fully connected layers, the first layer maps the input feature tensor to a 128-dimensional intermediate location feature, and the second layer directly maps the intermediate location feature and outputs the 3D spatial location coordinates of the defect; the size regression sub-network has the same structure as the location sub-network and outputs the equivalent size of the defect.

9. The method for detecting internal defects in steel plates as described in claim 8, characterized in that: The defect detection model constructs a physically constrained Siamese network during training. The physically constrained Siamese network contains two branches with identical structures and shared weights, namely the query branch and the reference branch. The input to the reference branch is the simulation data generated using the finite element ultrasonic propagation model. The finite element ultrasonic propagation model is based on the elastic wave equation. The four sides of the computational domain are set as absorbing boundary conditions to simulate the environment of an infinitely large plate, and the upper and lower surfaces are free stress boundaries. The propagation and scattering process of ultrasonic waves in the plate is simulated by setting the steel plate material parameters and the defect geometry model pre-set according to the real physical properties of various defects. The defect geometry model is a simplified three-dimensional geometry built in simulation software based on the labeled defect type, three-dimensional spatial location, and equivalent size of the defect: for crack types, a cuboid model with tiny openings is used; for porosity types, an ellipsoid model is used; and for inclusion types, a polyhedron model with irregular surfaces is used. By assigning the three-dimensional spatial location and equivalent size parameters of the defect to the geometry, a radio frequency signal sequence and its synchronization parameters covering various defect scenarios are generated. The input of the query branch is the measured signal of the steel plate under test. Both branches are processed by an adaptive defect scattering feature extraction network, and each ultimately outputs a high-level abstract defect feature tensor. The physical information-constrained Siamese network is trained through a Siamese metric space learning module, which calculates the Euclidean distance between the feature tensors from the two branches and uses a triplet loss function. The training objective is to make the distance between measured features and simulated features with the same real physical properties of defects smaller than the distance between them and features with different properties.

10. The method for detecting internal defects in steel plates as described in claim 9, characterized in that: In actual deployment, five Siamese networks with physical information constraints are initialized and trained in parallel. The networks have the same structure but random initial weights, forming a multi-model ensemble system. For the measured signal of the same detection area, all five Siamese networks with physical information constraints run independently and output five sets of defect attribute prediction results. The statistical variance of the five sets of prediction results on each attribute is calculated. For the predicted three-dimensional spatial coordinates, if the variance of the X, Y, and Z coordinate components is less than 0.1 mm, the detection result is considered to have high reliability. If the variance of any coordinate component exceeds this threshold, the detection area is marked as a high uncertainty area and a reliability warning is triggered.

Images