CN-121994936-A - Method for detecting internal defects of steel plate

CN121994936ACN 121994936 ACN121994936 ACN 121994936ACN-121994936-A

Abstract

The invention provides a steel plate internal defect detection method, which belongs to the technical field of industrial intelligent detection based on machine learning, and comprises the steps of deploying a multi-channel ultrasonic phased array probe on a steel plate detection platform, acquiring an ultrasonic full matrix, acquiring related data, inputting the data into a multi-mode signal preprocessing module, generating a structured multi-channel space-time signal data body according to the geometric relationship of an acoustic path, extracting local time domain characteristics of channels through a lightweight signal interaction network layer, generating an inter-channel interaction weight matrix by combining a physical perception attention module of acoustic attenuation compensation, dynamically weighting and summing, and adding initial weight bias to specific signals to obtain a fusion signal field with enhanced signal-to-noise ratio. And then inputting the three-dimensional acoustic scattering intensity distribution pattern into a deep defect feature extraction module, and finally inputting a defect detection model, wherein the three types of sub-networks are used for respectively predicting the probability distribution of the defect type, the three-dimensional space position coordinates and the equivalent size. The invention improves the depth of defect characterization and intelligent recognition accuracy.

Inventors

JING YUQIANG
WANG DAWEI
SONG YANQIANG
HAN GANG
ZHANG XIAOBING
ZHAO DONGDONG
HAN QI
SUN HAO

Assignees

日照市昱岚新材料有限公司

Dates

Publication Date: 20260508
Application Date: 20260127

Claims (10)

1. The method for detecting the internal defects of the steel plate is characterized by comprising the following steps of: S1, deploying a multi-channel ultrasonic phased array probe on a steel plate detection platform, and collecting an ultrasonic full matrix, so as to obtain an original radio frequency signal sequence, a probe space position coordinate, an acoustic beam incidence angle and a transmitting-receiving transducer pair index matrix; S2, inputting S1 data into a multi-mode signal preprocessing module, generating a structured multi-channel space-time signal data body by utilizing a probe space position coordinate, an acoustic beam incidence angle and an acoustic path geometric relation defined by a transmitting-receiving transducer on an index matrix, extracting a channel local time domain feature vector of each channel by utilizing a convolution layer in a lightweight signal interaction network layer, generating channel level statistical description by a physical perception attention module based on acoustic attenuation compensation, and learning and outputting an inter-channel interaction weight matrix; s3, inputting the fusion signal field with the enhanced signal-to-noise ratio into a deep defect feature extraction module, and extracting a three-dimensional acoustic scattering intensity distribution map; S4, inputting the acoustic scattering intensity distribution map into a defect detection model, and respectively predicting defect type probability distribution, three-dimensional space position coordinates of defects and defect equivalent sizes through a defect type classification sub-network, a space position regression sub-network and a size regression sub-network.
2. The method for detecting the internal defects of the steel plate according to claim 1, wherein the specific process for generating the structured multi-channel space-time signal data body is as follows: By utilizing the spatial position coordinates of the probe, the incidence angle of the sound beam and the geometrical relationship of the sound wave path defined by the transmitting-receiving transducer pair index matrix, the sound wave path represented by each A-scan signal is mapped into a unified and predefined virtual three-dimensional grid coordinate system inside the steel plate through ray tracing and time delay calculation, and signal amplitude values from different sound wave paths after time correction are distributed to each grid point, so that a structured multichannel space-time signal data body is generated, wherein each channel corresponds to a unique sampling of the sound wave path in the three-dimensional space.
3. The method for detecting the internal defects of the steel plate according to claim 2, wherein the convolution layer in the lightweight signal interaction network layer scans time domain signals of each channel through the convolution layer containing 32 one-dimensional convolution kernels to extract channel local time domain feature vectors of each channel, then the feature vectors are sent into a physical perception attention module based on acoustic attenuation compensation, channel level statistical description is firstly generated through a global average pooling layer, basic weight is obtained through two layers of full-connection layer learning, and an inter-channel interaction weight matrix is output by combining physical priori weight vectors, each element of the inter-channel interaction weight matrix quantifies the correlation strength of any two channel signals in terms of representing defect scattering information, and finally the inter-channel interaction weight matrix is used for carrying out dynamic weighted summation on original multi-channel space-time signal data bodies and giving higher initial weight bias to channels with signal state categories of high signal to noise ratio and strong defect scattering signals to obtain a fused signal-noise ratio enhanced fusion signal field.
4. A method for detecting the internal defects of a steel plate according to claim 3 is characterized in that the signal state type judgment mode is that the signal-to-noise ratio of a signal full sequence and the energy attenuation coefficient of a signal in a defect theoretical scattering time window are calculated, the starting time and the ending time of the window are calculated by taking the sum of the distance from a transmitting transducer to the defect three-dimensional space position plus the distance from a receiving transducer to the defect three-dimensional space position divided by the theoretical transit time obtained by dividing the sound velocity of a medium as the center, the first 50 sampling points of the original radio frequency signal sequence are selected as pure noise sampling segments when the signal-to-noise ratio is calculated, the root mean square value of the signal amplitude of the segments is defined as a reference noise level, the energy attenuation coefficient is defined as the ratio of the average energy of the second half signal to the average energy of the first half signal in the time window, if the signal-to-noise ratio of a certain signal sample is larger than 20 dB and the energy attenuation coefficient is smaller than 0.5, the sample signal state type is judged to be marked as a high signal-to-noise ratio and a strong defect signal, and if the conditions are not met, the signal-to be marked as a medium-low signal-noise ratio or a complex signal.
5. A method for detecting the internal defects of steel plates according to claim 1, wherein the deep defect feature extraction module firstly utilizes a defect scattering feature self-adaptive extraction network formed by an anisotropic deformable convolution module guided based on the propagation direction of sound beams to extract a primary scattering feature map, a middle scattering feature map and a high-level scattering feature map layer by layer and output a high-level abstract defect feature tensor, secondly constructs a data-driven acoustic scattering intensity map inversion network, an encoder part of which multiplexes the defect scattering feature self-adaptive extraction network, and finally a decoder part of which maps the feature tensor into a three-dimensional data-driven acoustic scattering intensity distribution map through a deconvolution layer and a convolution layer.
6. The method for detecting the internal defects of the steel plate according to claim 5, wherein the defect scattering characteristic self-adaptive extraction network main body is composed of four layers of anisotropic deformable convolution modules guided based on the propagation direction of sound beams; the first layer uses a convolution kernel of 3 x 3 to perform initial sampling in an input signal field, and inputs an input feature map and a recorded sound beam incidence angle as a condition variable through a parallel anisotropic offset learning network, firstly, the input feature is enabled to pass through a convolution layer of 3 x 3 to obtain space geometrical intermediate features, then the features are spliced with incidence angle vectors, finally, a limited offset field is output through a convolution layer of 1 x 1, the sampling range of the convolution kernel is restrained by utilizing the directionality of ultrasonic wave propagation in a medium, and a primary scattering feature map containing local waveform fluctuation features is output; the second layer uses a convolution kernel of 5 multiplied by 5, takes a primary scattering feature map as input, learns more complex spatial deformation, enables the convolution kernel to be adaptively focused on a signal mutation area, and outputs a middle-level scattering feature map capable of representing a middle-scale scattering mode; the third layer uses a convolution kernel of 7 multiplied by 7 to extract a high-level scattering feature map reflecting the macroscopic outline of the defect, the fourth layer uses a convolution kernel of 9 multiplied by 9 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.
7. The method for detecting the internal defects of the steel plate according to claim 6, wherein the data-driven acoustic scattering intensity map inversion network is of an encoder-decoder structure, the encoder part multiplexes a trained defect scattering characteristic self-adaptive extraction network, a signal-to-noise ratio enhanced fusion signal field is encoded into a high-level abstract defect characteristic tensor, the decoder part is formed by connecting three deconvolution layers and two convolution layers in series, the first deconvolution layer carries out 2-time upsampling on the high-level abstract defect characteristic tensor in a space dimension to obtain an initial inversion characteristic map with low resolution, the second deconvolution layer and the third deconvolution layer carry out 2-time upsampling in sequence to gradually recover space details to obtain a high-resolution inversion characteristic map, and finally, channel adjustment and detail optimization are carried out through the two 3x3 convolution layers to output a three-dimensional data-driven acoustic scattering intensity map.
8. The method for detecting the internal defects of the steel plate according to claim 1, wherein the defect detection model adopts a physically guided defect attribute regression head and comprises three parallel sub-networks, wherein the defect type classification sub-network firstly maps an input characteristic tensor into a 256-dimensional type hidden characteristic vector through a full connection layer, then maps the type hidden characteristic vector into 3 dimensions through a full connection layer, finally normalizes through a Softmax layer and outputs the type probability distribution of defects belonging to cracks, air holes and slag inclusion, the spatial position regression sub-network consists of two full connection layers, the first layer maps the input characteristic tensor into a 128-dimensional position intermediate characteristic, the second layer directly maps the position intermediate characteristic and outputs the three-dimensional spatial position coordinates of the defects, and the size regression sub-network is structurally the same-position sub-network and outputs the equivalent sizes of the defects.
9. The method for detecting the internal defects of the steel plate according to claim 8, wherein the defect detection model is used for constructing a physical information constraint twin network during training, and the physical information constraint twin network comprises two branches with the same structure and shared weight, namely a query branch and a reference branch; the input of the reference branch is simulation data generated by utilizing a finite element ultrasonic propagation model, the finite element ultrasonic propagation model is based on an elastic wave equation, the peripheral side surface of a calculation domain is set to be an absorption boundary condition so as to simulate an infinite plate environment, the upper surface and the lower surface are free stress boundaries, the propagation and scattering processes of ultrasonic waves in the finite element ultrasonic propagation model are simulated by setting steel plate material parameters and a defect geometric model preset according to real physical properties of various defects, the defect geometric model is a simplified three-dimensional geometric body established in simulation software according to the marked defect type, the three-dimensional space position of the defect and the equivalent defect size, the simplified three-dimensional geometric body is established in simulation software, for the crack type, a cuboid model with a tiny opening is adopted, for the air hole type, an ellipsoid model is adopted, for the slag inclusion type, a radio frequency signal sequence covering various defect scenes and synchronous parameters thereof are generated, the input of the query branch is an actual measurement signal of the steel plate to be tested, the two branches are processed through a defect scattering characteristic self-adaption extraction network, finally, a high-level abstract defect characteristic tensor is output, the physical information constraint is used for the calculation of a twin-dimensional space training module is used for measuring the twin-space tensor distance between two branches through a twin-space learning module, and a triplet loss function is adopted, and the training target is that the distance between the measured characteristic and the simulation characteristic with the same defect real physical attribute is smaller than the distance between the measured characteristic and the simulation characteristic with different attribute characteristics.
10. The method for detecting the internal defects of the steel plate according to claim 9, wherein in actual deployment, five twin networks constrained by physical information are initialized and trained in parallel, the network structures are the same but initial weights are random to form a multi-model integrated system, all the twin networks constrained by the physical information independently operate for actual measurement signals of the same detection area, five groups of defect attribute prediction results are output, statistical variances of the five groups of prediction results on all the attributes are calculated, if variances of X, Y, Z three coordinate components of the three dimensional spatial position coordinates are lower than 0.1 millimeter, the detection result is judged to have high reliability, and if the variances of any coordinate component exceed the threshold, the detection area is marked as a high uncertainty area, and reliability warning is triggered.

Description

Method for detecting internal defects of steel plate Technical Field The invention belongs to the technical field of industrial intelligent detection based on machine learning, and particularly relates to a method for detecting internal defects of a steel plate. Background The steel plate is used as a key structural material in modern industry and infrastructure construction, and the internal quality of the steel plate is directly related to the safety and service life of various important equipment and engineering. In the manufacturing or service process, defects such as cracks, air holes, slag inclusion and the like can be generated in the steel plate, the defects are often strong in concealment, if the defects cannot be detected in time, the defects can be gradually expanded under the action of load and environment, and finally, component failure is caused and even serious safety accidents are caused. The traditional ultrasonic detection method faces limitation in defect qualitative, quantitative and characteristic analysis, and along with the evolution of industrial detection technology to intelligent and refined directions, a detection method of depth fusion physical mechanism and data driving is developed, intelligent judgment and inversion of defect types, positions and sizes can be realized, objectivity and accuracy of detection can be improved, and reliable basis can be provided for structural health assessment and life prediction. However, the existing developments have the following limitations: 1. the depth and the degree of intelligence of defect characterization are limited. In the prior art, the macroscopic signal response based on a single physical principle is judged, the capability of carrying out fine analysis and learning on deep physical waveform characteristics generated by the interaction of defects and ultrasonic waves is lacking, and the method is difficult to adapt to the precise identification of complex and various defects; 2. The multi-source information fusion and the physical guidance are insufficient. The existing method generally regards signal processing, feature extraction and defect inversion as relatively independent steps, and multi-mode information and physical priori knowledge in the detection process cannot be deeply embedded into a feature learning and decision-making model, so that the robustness and the interpretation of the model in a complex industrial scene are limited; 3. The uncertainty and the difficult-to-detect sample are weak in processing capacity. The existing system focuses on the detection of conventional defects in multiple ways, lacks a dynamic evaluation mechanism for the reliability of the detection result, lacks the capability of active learning and model self-optimization for high-uncertainty samples or boundary cases, and has a bottleneck in system performance. The invention patent of China with the application number of CN202511545230.0 discloses a full-automatic steel plate internal defect online detection system based on an ultrasonic phased array, which relates to the technical field of nondestructive detection and comprises an intelligent analysis module, a high-precision verification module and a high-precision detection module, wherein the intelligent analysis module is in communication connection with the online detection module and is used for operating an artificial intelligent rating model to process and initially evaluate original ultrasonic detection data and output an initial evaluation result containing a defect space position, and the high-precision verification module is in communication connection with the intelligent analysis module and is used for intercepting a corresponding sample steel plate according to the defect space position in the initial evaluation result and verifying and measuring the sample steel plate by a high-precision detection means so as to generate defect truth value data. However, the scheme needs to intercept the sample steel plate according to the initial evaluation result for high-precision verification, the physical sampling mode belongs to destructive detection, and obvious time decoupling exists between the verification process and an online detection link, so that the real-time prediction precision of the model can not be improved by utilizing the high-precision physical attribute constraint in real time and without damage in the detection process. The Chinese patent application No. CN201380068387.8 provides a device and a method for detecting the internal defects of a steel plate. The steel sheet internal defect detection device includes an all-defect detection unit that detects all defects including surface defects existing on the surface of the steel sheet and internal defects existing inside the steel sheet based on the intensity of leakage magnetic flux measured by generating magnetic flux in the moving direction of the steel sheet, a surface-defect detection unit that detects surface defects