Split type steel pipe surface detection system and method

Abstract

The application discloses a split type steel pipe surface detection system and method, and relates to the technical field of steel pipe surface detection.The system comprises a steel pipe conveying assembly, a quick-change base, a composite detection unit and a fusion operation module; the quick-change base is fixed to a steel pipe conveying assembly detection station, is provided with a mechanical quick-lock structure and an electrical blind plug type quick-change interface; the composite detection unit can be split type disassembled through the quick-change interface, integrates a multi-spectral vision sensor and an eddy current array sensor, and is used for collecting multi-spectral images and eddy current electromagnetic signals of a steel pipe surface; the fusion operation module pre-stores calibration reference parameters, performs multi-level composite processing on multi-modal data to obtain fusion feature data sets, and inputs a pre-trained convolutional neural network to complete defect recognition.The application can realize quick disassembly, maintenance and assembly of the detection unit, combines multi-modal multi-level fusion and CNN determination, takes into account detection accuracy and real-time performance, and is suitable for online surface defect detection of a steel pipe production line.

Description

Technical Field

[0001] This invention relates to the field of steel pipe surface inspection technology, specifically to a split-type steel pipe surface inspection system and method. Background Technology

[0002] As a core basic component in fields such as petrochemicals, power transmission, and construction engineering, steel pipes are subject to surface and near-surface defects that directly affect the operational safety and service life of equipment. Currently, the mainstream industrial production lines use either single multispectral visual inspection or single eddy current inspection to achieve online flaw detection on the surface of steel pipes. Single multispectral visual inspection relies on optical imaging principles and can only collect the morphological information of the outermost layer of the pipe. It is severely affected by environmental interference such as oxide scale, surface water stains, and oil stains generated during the rolling process. The spectral reflection characteristics of oxide scale are highly similar to those of microcracks, which can easily lead to false detections. Water stains and oil stains can cover up minute surface cracks. At the same time, it is completely unable to detect hidden defects such as delamination and slag inclusions within a certain depth range below the surface. Such defects are prone to fatigue fracture under subsequent high pressure and heavy load conditions, causing major safety accidents. Single eddy current inspection is based on the principle of electromagnetic induction and can penetrate the surface to detect near-surface damage. However, it has extremely poor ability to distinguish surface morphological defects. Harmless pseudo-defects such as pitting, scratches, and peeling that do not affect the performance of use will produce electromagnetic response signals similar to real cracks, resulting in a high false alarm rate of the equipment and a significant increase in the cost of manual re-inspection. Moreover, the results of eddy current inspection lack intuitive morphological information and cannot accurately determine the specific shape, size, and hazard level of defects.

[0003] Currently available multispectral eddy current composite flaw detection equipment generally adopts an integrated, fixed structure. The detection unit is rigidly connected to the steel pipe conveyor frame by welding or bolts. When a sensor malfunctions and needs replacement, calibration, or maintenance, the entire detection frame must be disassembled and the machine shut down. Each maintenance requires a long downtime, severely impacting the continuous operation efficiency of the production line. At the algorithm level, traditional multimodal feature fusion schemes often employ a multi-classifier integrated architecture. Each base classifier operates independently and lacks an effective feature-level deep fusion mechanism. This fails to fully leverage the shape recognition advantages of multispectral vision and the depth detection advantages of eddy current detection, resulting in a trade-off between accuracy and robustness in defect identification, and failing to meet the high-precision detection requirements under complex industrial conditions. Summary of the Invention

[0004] The present invention provides a split-type steel pipe surface inspection system and method to solve the technical problems of insufficient single-modal detection accuracy, cumbersome maintenance and production change of integrated equipment, and insufficient fusion depth of multi-modal fusion algorithms, which makes it difficult to balance the accuracy and robustness of defect identification.

[0005] According to one aspect of the present invention, a split-type steel pipe surface inspection system is provided, comprising a quick-change base, a composite inspection unit, and a fusion computing module;

[0006] The quick-change base is fixed to the inspection station of the steel pipe conveying assembly and is equipped with a mechanical quick-lock structure and an electrical blind-plug quick-change interface;

[0007] The composite detection unit is detached and assembled from the quick-change base via an electrical blind-plug quick-change interface, and integrates a multispectral visual sensor and an eddy current array sensor to acquire multispectral images and eddy current electromagnetic signals on the surface of the steel pipe.

[0008] The fusion computing module pre-stores calibration reference parameters and performs multi-level composite processing on multispectral images and eddy current electromagnetic signals before inputting them into a pre-trained convolutional neural network to identify surface defects in steel pipes, thereby obtaining surface detection results for steel pipes.

[0009] Preferably, the multi-stage composite processing method includes the following steps:

[0010] Tensor decomposition and band information entropy weighted transformation are performed on multispectral images to generate multiscale spectral feature tensors. Then, the multiscale spectral feature tensors are fitted with a hybrid probability model and combined with a graph cut optimization algorithm to split them into local catastrophe feature tensors and matrix steady-state feature tensors.

[0011] Wavelet packet band decomposition and depth mapping feature separation are performed sequentially on the eddy current electromagnetic signal to obtain the matrix electromagnetic feature tensor and the defect disturbance feature tensor.

[0012] The local mutation feature tensor and the defect perturbation feature tensor are fused across modes by combining a spatial deep attention mechanism with a graph convolutional network to generate the first composite defect feature tensor.

[0013] The matrix steady-state feature tensor and the matrix electromagnetic feature tensor are mapped by a kernel method and fused with Bayesian posterior probability to generate the second matrix background feature tensor.

[0014] Based on the calibration benchmark parameters, feature-level weighted fusion, decision-level evidence fusion, and feedback-based weight iterative optimization are sequentially performed on the first composite defect feature tensor and the substrate background feature tensor until the preset convergence condition is met, resulting in a fused feature dataset.

[0015] Preferably, the method for obtaining the multi-scale spectral feature tensor is as follows:

[0016] Tucker tensor decomposition was performed on the multispectral image to obtain the core tensor and factor matrix;

[0017] Perform multi-level two-dimensional discrete wavelet transform on each spatial slice of the core tensor to obtain wavelet coefficients at different scales, including approximation coefficients and detail coefficients;

[0018] The information entropy of each spectral band is calculated, and the wavelet coefficients at different scales are weighted and fused by combining the spectral band weight matrix in the calibration reference parameters to generate a multi-scale spectral feature tensor containing multiple spatial scales.

[0019] Preferably, the method for obtaining the local mutation feature tensor and the matrix steady-state feature tensor is as follows:

[0020] The global feature map corresponding to each spatial scale of the multi-scale spectral feature tensor is divided into multiple spatial sub-blocks. A two-component Gaussian mixture model is fitted to all feature pixels in each spatial sub-block to obtain the assignment probability of each feature pixel in the spatial sub-block to the first Gaussian component and the second Gaussian component. The latent attributes of the feature pixels are marked according to the assignment probability.

[0021] A Markov random field based on 8 neighborhoods is constructed for the global feature region formed by stitching together all spatial sub-blocks at a single spatial scale. An energy function consisting of data terms and smoothing terms is set, and the energy function is minimized by the α-expanded graph cut algorithm to complete the classification attribute division of all feature pixels at this spatial scale and obtain the initial segmentation result corresponding to this spatial scale.

[0022] The initial segmentation results corresponding to each spatial scale are fused by cross-scale voting. After removing isolated noise pixels through morphological post-processing, the features of the defect region and the features of the normal matrix region are extracted respectively, and the local mutation feature tensor and the matrix steady-state feature tensor are generated accordingly.

[0023] Preferably, the method for obtaining the matrix electromagnetic characteristic tensor and the defect perturbation characteristic tensor is as follows:

[0024] Multi-level wavelet packet decomposition is performed on the time-domain eddy current signal of each coil in the eddy current electromagnetic signal to obtain the wavelet packet coefficients corresponding to several frequency bands, and the total energy of each frequency band is calculated.

[0025] Call the frequency band division matrix in the calibration reference parameters to divide all frequency bands corresponding to each coil into two categories: the base frequency band and the defect frequency band;

[0026] Based on the eddy current skin depth principle, the detection depth corresponding to each frequency band is determined, and the detection depth range is divided into multiple depth intervals in advance. The total energy of the matrix frequency band and the total energy of the defect frequency band of each coil in each depth interval are statistically analyzed.

[0027] The total energy data of all coils in the eddy current array sensor in two frequency bands at various depth ranges are summarized and integrated to form the matrix electromagnetic characteristic tensor and the defect disturbance characteristic tensor corresponding to the eddy current array.

[0028] Preferably, the method for obtaining the first composite defect feature tensor is as follows:

[0029] Calculate the space-depth attention weight matrix of the local mutation feature tensor and the space-depth attention weight matrix of the defect perturbation feature tensor respectively;

[0030] Establish the mapping relationship between the physical spatial coordinates of the eddy current array coil and the pixel coordinates of the multispectral image, and map the defect perturbation feature tensor to the pixel coordinate system of the multispectral image;

[0031] Construct a graph structure with pixels of the multispectral image as nodes, where edge weights are determined by the spatial distance between pixels and the attention weights.

[0032] A two-layer graph convolutional network is used to fuse the mapped defect perturbation feature tensor and the local mutation feature tensor, and the weights are adjusted by combining attention weights to generate the first composite defect feature tensor.

[0033] Preferably, the method for obtaining the background feature tensor of the second substrate is as follows:

[0034] By selecting the radial basis function as the kernel function, the matrix steady-state feature tensor and the matrix electromagnetic feature tensor are mapped to the high-dimensional feature space to obtain the visual matrix features and electromagnetic matrix features;

[0035] The covariance matrices of visual matrix features and electromagnetic matrix features are solved separately. Whitening is then performed based on the covariance matrices to eliminate the correlation between visual matrix features and electromagnetic matrix features.

[0036] Based on Bayes' theorem, calculate the posterior probability distribution of visual matrix features relative to electromagnetic matrix features, and the posterior probability distribution of electromagnetic matrix features relative to visual matrix features.

[0037] Based on the posterior probability distribution, the visual matrix features and electromagnetic matrix features after whitening are weighted and fused to generate the second matrix background feature tensor.

[0038] Preferably, the method for obtaining the fused feature dataset is as follows:

[0039] Using the initial fusion weight coefficients in the calibration reference parameters, the first composite defect feature tensor and the second substrate background feature tensor are fused at the feature level to obtain the initial fused features;

[0040] The initial fused features are input into multiple base classifiers to obtain the prediction results and uncertainty estimates of each base classifier;

[0041] The DS evidence theory is used to perform decision-level fusion of the prediction results and uncertainty estimates of multiple base classifiers to obtain the global prediction result and global uncertainty.

[0042] Calculate the iteration error. If the iteration error meets the preset iteration convergence threshold, stop the iteration and use the initial fused features as the fused feature dataset.

[0043] If the iteration error does not meet the preset iteration convergence threshold, the initial fusion weight coefficients are updated according to the iteration error gradient, and feature-level weighted fusion and decision-level fusion are performed again until the iteration error meets the preset iteration convergence threshold or the number of iterations reaches the preset maximum number of iterations.

[0044] According to another aspect of the present invention, a method for inspecting the surface of a split-type steel pipe is provided, for use in a split-type steel pipe surface inspection system, comprising:

[0045] Multispectral images and eddy current electromagnetic signals on the surface of steel pipes are acquired by a composite detection unit that integrates a multispectral visual sensor and an eddy current array sensor.

[0046] The calibration reference parameters are preset, and the multispectral image and eddy current electromagnetic signal are processed in multiple stages and then input into a pre-trained convolutional neural network to identify defects on the surface of the steel pipe, so as to obtain the surface inspection results of the steel pipe.

[0047] Due to the adoption of the above technical solution, the technical progress achieved by this invention compared to the prior art is as follows:

[0048] 1. The composite detection unit and quick-change base of this invention adopt a split quick-installation structure, which can be quickly disassembled and assembled with the help of mechanical quick lock and electrical blind plug interface. Equipment maintenance and steel pipe production change debugging do not require disassembling the whole machine frame, effectively shortening downtime and adapting to the needs of flexible production lines with multiple varieties.

[0049] 2. This invention employs multispectral and eddy current dual-mode collaborative acquisition and detection. It relies on tensor decomposition, graph cut optimization, and wavelet packet frequency band decomposition to separate matrix and defect features, overcoming the inherent defects of single vision or single eddy current detection. It can identify surface morphological defects of pipe materials and detect hidden subcutaneous damage, reducing the probability of false detection and missed detection.

[0050] 3. This invention completes the cross-modal fusion of matrix and defect features in two ways. After weight iteration and convergence, a fused feature dataset is generated and uniformly connected to a single convolutional neural network for judgment. It abandons the multi-classifier stacking fusion mode and deeply combines the technical advantages of the two detection methods to improve the accuracy and stability of defect identification under complex working conditions. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the overall structure of the split-type steel pipe surface inspection system of the present invention.

[0052] Figure 2 This is a flowchart of the multi-stage composite treatment method for steel pipe surface inspection according to the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Examples, such as Figure 1 As shown, the system in this embodiment adopts a split-type disassembly and assembly architecture, consisting of four parts: steel pipe conveying assembly, quick-change base, composite detection unit, and fusion computing module.

[0055] The steel pipe conveying assembly in this embodiment includes a steel pipe idler roller group, a frequency conversion feed drive mechanism, and a pipe diameter adaptive centering fixture, which is fixedly deployed on the steel pipe production line frame, allowing for complete assembly and disassembly without disassembly. It is also equipped with a high-precision incremental encoder, which uses the encoder trigger timing to achieve synchronous hardware sampling by multispectral vision sensors and eddy current array sensors. Furthermore, the outer layer of the steel pipe idler roller group is covered with a polyurethane buffer pad, which can prevent new scratches from being generated during the conveying of the steel pipe outer wall and thus avoid interference with detection.

[0056] In this embodiment, the quick-change base is made of cast aluminum and fixedly installed at the testing station of the steel pipe conveying assembly. It integrates an eccentric locking mechanical quick-lock structure and an aviation plug-type electrical blind-plug quick-change interface. The mechanical locking effective stroke of the eccentric locking mechanical quick-lock structure is 12mm. In the locked state, the circumferential movement error of the composite testing unit is less than 0.03mm. The aviation plug-type electrical blind-plug quick-change interface integrates three types of interfaces: power supply terminal, gigabit network communication terminal, and analog signal acquisition terminal. The total assembly and disassembly time of the composite testing unit is ≤3min, realizing tool-free split-type quick assembly and disassembly.

[0057] The composite detection unit in this embodiment integrates a multispectral vision sensor and an eddy current array sensor, and is detached from the quick-connect base using an aviation-type blind-plug quick-connect interface. The multispectral vision sensor is a 4-channel industrial multispectral camera with spectral center wavelengths of 450nm, 550nm, 650nm, and 850nm, and a resolution of 2048×2048 pixels, used to output multispectral images. H=W=2048, S=4; a ring-shaped diffused lighting source is also provided to eliminate the shadows cast by the curved surface of the steel pipe; the eddy current array sensor uses a 32-channel differential wound coil array with a coil excitation frequency range of 50kHz-1MHz to output eddy current electromagnetic signals. Number of coil channels Number of sampling points per time Meanwhile, the eddy current array sensor is equipped with a pre-signal signal conditioning circuit to filter, amplify, and perform analog-to-digital conversion on the original eddy current electromagnetic signal before transmitting it to the back-end fusion computing module.

[0058] In this embodiment, the fusion computing module adopts a collaborative computing architecture combining FPGA and industrial control computer. The FPGA hardware accelerates the real-time decomposition and feature preprocessing of eddy current electromagnetic signal wavelet packets, while the industrial control computer is responsible for multispectral tensor decomposition, graph cut optimization, cross-modal feature fusion, iterative weight optimization, and CNN defect inference. Furthermore, the fusion computing module locally stores a pre-calibrated set of calibration reference parameters, which specifically includes a spectral band weight matrix. Eddy current frequency band division matrix Initial fusion weight coefficients Iterative convergence threshold And the maximum number of iterations of the algorithm The calibration reference parameter set was calibrated and solidified after being calibrated by offline experiments on the defect-free standard 20# steel pipe.

[0059] In this embodiment, the system is deployed on a 20# carbon steel seamless steel pipe production line. The outer diameter of the steel pipe being tested is φ60mm-φ168mm. The conveying speed of the steel pipe on the production line is adjustable from 0.5-8m / s. This system can accurately identify four types of high-frequency industrial defects in steel pipes: surface cracks, subcutaneous delamination, rolling pits, and surface peeling. The system acquires multispectral images and eddy current electromagnetic signals of the steel pipes on the production line through a composite detection unit. The fusion calculation module performs multi-level composite processing on the acquired multispectral images and eddy current electromagnetic signals to obtain a fused feature dataset.

[0060] Please refer to the multi-level composite processing described in this embodiment. Figure 2 As shown, the specific steps are as follows:

[0061] S1. Perform tensor decomposition and band information entropy weighted transformation on the multispectral image to generate a multi-scale spectral feature tensor. Then, use a hybrid probability model fitting and graph cut optimization algorithm to decompose the multi-scale spectral feature tensor into a local mutation feature tensor and a matrix steady-state feature tensor.

[0062] In this embodiment, S1 specifically refers to: firstly, performing Tucker tensor decomposition and multi-level two-dimensional discrete wavelet transform to process the multispectral image. According to the tensor decomposition formula of Tucker tensor decomposition: The multispectral image is decomposed, where, Represents the core tensor, , This is the factor matrix corresponding to the spatial dimensions of the multispectral image. The spectral dimension factor matrix of the multispectral image; the core tensor obtained from the decomposition. Each spatial slice is uniformly subjected to a three-level two-dimensional discrete wavelet transform, and each level of wavelet transform can separate approximate coefficients. With detail coefficient Where 's' represents three spatial scales, from 1 to 3. Spatial scale is also known as scale. The four spectral bands are numbered 1-4; then the information entropy of each individual spectral band is calculated. The formula for calculating the information entropy is: In the formula, For the first The statistical distribution probability of pixel grayscale values ​​within each spectral band is used to retrieve the pre-stored spectral band weight matrix within the calibration reference parameters. The spectral band weight matrix is ​​a fixed parameter matrix obtained through offline calibration of a defect-free standard 20# carbon steel pipe using multispectral imaging. It is used to differentiate the weighting coefficients based on the sensitivity of different bands to defects and the steel pipe matrix, suppressing the weights of low-sensitivity spectral bands severely affected by light and oil contamination, and amplifying the proportion of band features with high defect identification. The spectral band weights and information entropy are used to adjust the wavelet coefficients (approximate coefficients) at each scale. With detail coefficient Perform weighted calculations The weighted wavelet coefficients are reassembled and spliced ​​to generate a multi-scale spectral feature tensor containing information at three spatial scales. , =512, =512, =3 represents three spatial scales. This multi-scale spectral feature tensor preserves the morphology and spectral variation information of the steel pipe surface at different imaging scales.

[0063] After constructing the multi-scale spectral feature tensor, the process proceeds to Gaussian mixture model fitting and Markov random field graph segmentation. In this embodiment, the global feature map corresponding to each spatial scale is uniformly divided into several continuous spatial sub-blocks of size 32×32 pixels. A two-component Gaussian mixture probability model is constructed for all feature pixels within any sub-block. In the formula, , These are the weighting coefficients for the two types of Gaussian components, respectively. Given a normal Gaussian distribution function, after solving the parameters of the two-component Gaussian mixture probability model through maximum likelihood estimation, the pixel attributes are marked according to the probability values ​​of each feature pixel belonging to the two types of Gaussian components. Feature pixels that are likely to belong to the first Gaussian component are marked as normal matrix pixels, and feature pixels that belong to the second Gaussian component are marked as defect mutation pixels.

[0064] A Markov random field with an 8-neighborhood is constructed based on the complete global feature region formed by stitching together all spatial sub-blocks at a single spatial scale. The energy function of this Markov random field, consisting of data terms and a smoothing term, is set as follows: In the formula, For each data item, i represents the category to which feature pixel i belongs. The negative logarithmic probability of (0 = normal matrix pixels, 1 = defective aberration pixels), The smoothing term represents the penalty for adjacent feature pixels of different categories. The α-dilated graph cut algorithm iteratively minimizes the overall energy function value to obtain the initial segmentation result at a single spatial scale. After performing cross-scale voting fusion processing on the initial segmentation results corresponding to the three spatial scales, isolated noise pixels are removed using opening operation morphological filtering. Based on this, defect region features are extracted and combined into a local mutation feature tensor. The combination of features of the normal matrix region forms the matrix steady-state feature tensor. .

[0065] S2. Wavelet packet band decomposition and depth mapping feature separation are performed sequentially on the eddy current electromagnetic signal to obtain the matrix electromagnetic feature tensor and the defect disturbance feature tensor.

[0066] In this embodiment, S2 specifically refers to: firstly, processing the 32-channel eddy current electromagnetic signals... A 5-level dB4 wavelet packet decomposition process was performed. The single-channel time-domain eddy current signal was decomposed into 32 non-overlapping frequency bands, with each channel corresponding to a coil. The wavelet packet coefficients corresponding to each frequency band were extracted. Where p ranges from 1 to 32, representing the coil number; k is the frequency band number from 1 to 32; and n is the discrete sampling point number within a single frequency band. The energy calculation formula is as follows: The total energy of each frequency band is calculated one by one. The value of the total energy directly reflects the strength of the eddy current electromagnetic signal response in the corresponding frequency band. At the same time, the frequency band division matrix pre-stored in the calibration reference parameters is retrieved. The frequency band division matrix is ​​a binary label matrix generated through eddy current calibration of artificially defective steel pipes at different depth standards. It is used to quickly distinguish whether the signals in each frequency band belong to the inherent electromagnetic response of the steel pipe matrix or the electromagnetic disturbance signal caused by defects, and automatically classify the matrix frequency band and the defect frequency band into two major categories. According to the frequency band division matrix, all frequency bands are divided into a matrix frequency band set. Defect frequency band set This enables accurate classification of the frequency bands corresponding to the matrix and defects.

[0067] The calculation formula based on the eddy current skin depth principle is as follows: Complete the detection depth conversion for each frequency band, where, The center angular frequency of the k-th frequency band ( (The center frequency of the corresponding frequency band) The permeability of free space, The relative permeability of 20# carbon steel. To determine the conductivity of the steel pipe material; the detection depth corresponding to the entire frequency band is uniformly divided into 5 continuous depth intervals, and the interval width is... Determined by offline standard sample calibration; the total energy of the matrix frequency band and the total energy of the defect frequency band for each coil in the five depth ranges are statistically analyzed. The formula for summarizing the total energy of the matrix frequency band is: The formula for summing the total energy of the defect frequency band is: After traversing all 32 coils, the data is integrated and summarized to generate the matrix electromagnetic characteristic tensor. With the defect perturbation characteristic tensor .

[0068] S3. The local mutation feature tensor and the defect perturbation feature tensor are fused across modes by combining a spatial deep attention mechanism with a graph convolutional network to generate the first composite defect feature tensor.

[0069] In this embodiment, S3 specifically involves: firstly solving the spatial-depth attention weight matrix of the local mutation feature tensor and the defect perturbation feature tensor, and then setting two sets of learnable attention weight matrices. , Attention weights are obtained by operating on the SoftMax activation function. and This attention weight and It can adaptively quantify the effective contribution ratio of local mutation features and defect disturbance features at different spatial locations and subcutaneous depths, and automatically suppress invalid feature interference caused by acquisition noise.

[0070] Based on the previously calibrated geometric mapping parameters, a one-to-one mapping relationship is established between the physical installation coordinates of the 32-channel eddy current array coils in the eddy current array sensor and the pixel coordinates of the multispectral image. This is used to map the defect perturbation feature tensor with a dimension of 32×5. Mapping to a unified pixel coordinate system of multispectral images enables spatial alignment and matching of visual and electromagnetic features.

[0071] After spatial alignment and matching are completed, a pixel topology graph structure is constructed, using pixels from the multispectral image as independent nodes in the graph structure, and the weights of the edges connecting the nodes are defined. `dist` represents the Euclidean distance between two pixel nodes, from which the adjacency matrix is ​​obtained. The diagonal elements of the adjacency matrix are filled with 0 by default, indicating that the nodes initially have no self-connections. A two-layer concatenated graph convolutional network is constructed. Before performing graph convolution, the adjacency matrix is ​​first normalized to obtain a normalized adjacency matrix. The convolution operation formula for graph convolutional networks is: , For the weight parameters of the graph convolutional network, The ReLU activation function is used to introduce a nonlinear transformation; the attention weights obtained from the solution are then... and The weighted features are superimposed onto the graph convolution output features, the weight ratios of features in different regions are adjusted, and finally the first composite defect feature tensor is generated. The first composite defect feature tensor fuses optical morphology abrupt change information with eddy current subcutaneous defect perturbation information.

[0072] S4. The matrix steady-state feature tensor and the matrix electromagnetic feature tensor are mapped by kernel method and fused with Bayesian posterior probability to generate the second matrix background feature tensor.

[0073] In this embodiment, S4 specifically refers to: First, the radial basis function (RBF) is selected as the high-dimensional mapping kernel function, and the expression of the kernel function is: In the formula, For kernel width hyperparameter, Let i be the i-th set of basis feature vectors in the high-dimensional feature space. The j-th matrix feature vector in the high-dimensional feature space is determined through offline calibration using standard defect-free steel pipe samples; the steady-state feature tensor of the matrix is ​​then analyzed. Matrix electromagnetic characteristic tensor By performing kernel space mapping, the original low-dimensional matrix steady-state feature tensor and matrix electromagnetic feature tensor are mapped to a high-dimensional feature space to obtain high-dimensional visual matrix features. Characteristics of electromagnetic matrix And solve for the covariance matrices of visual matrix features and electromagnetic matrix features respectively. , Feature whitening transformation is carried out using the covariance matrix. In the formula, This is the new feature matrix after whitening and removal of dimensional redundancy. The original high-dimensional matrix feature matrix before whitening treatment is used to eliminate data redundancy and dimensional correlation between visual matrix features and electromagnetic matrix features, and to remove feature coupling interference caused by uniform physical and chemical fluctuations of steel pipe substrate.

[0074] Based on Bayes' theorem, a probabilistic fusion of dual-modal matrix features is carried out, according to... The bidirectional posterior probability distributions of visual matrix features relative to electromagnetic matrix features and electromagnetic matrix features relative to visual matrix features are calculated separately. These posterior probability distributions characterize the matching confidence of the two matrix features under the same steel pipe substrate condition. Using the values ​​of the bidirectional posterior probability distributions as weighting coefficients, a weighted fusion operation is performed on the two types of matrix features after whitening treatment, ultimately generating the second substrate background feature tensor. The second substrate background feature tensor fully characterizes the combined spectral and electromagnetic background features of the defect-free region of the steel pipe substrate.

[0075] S5. Combining the calibration benchmark parameters, perform feature-level weighted fusion, decision-level evidence fusion, and feedback-based weight iterative optimization on the first composite defect feature tensor and the base material background feature tensor in sequence until the preset convergence condition is met, and obtain the fused feature dataset.

[0076] In this embodiment, S5 specifically involves: first, retrieving the initial fusion weight coefficients stored in the calibration benchmark parameters. ,include and The weights satisfy ,according to Feature-level weighted fusion is completed to obtain the first round of initial fused features, among which, For defect feature weights, The base features are weighted; the initial fusion features are simultaneously input into the three base classifiers: SVM, BP neural network, and K-nearest neighbor. Each base classifier independently completes feature prediction and outputs the prediction result, i.e. the defect classification result. At the same time, the uncertainty estimate corresponding to the single classifier is calculated based on the classification confidence interval to quantify the reliability of the single classifier's recognition result.

[0077] The DS evidence theory is introduced to complete the decision-level fusion of multiple classifiers, and the basic trust assignment function is constructed based on the output results of each base classifier. By fusing the prediction results and uncertainty estimates of all classifiers through the classic evidence combination rule of DS, the global prediction result and global uncertainty are output; the L2 norm error calculation formula is used: Solve for the iteration error. Global prediction labels representing the global prediction results The standard sample's authentic label is a numerical label representing the defect type (0: no defect, 1: surface crack, 2: subcutaneous delamination, 3: pitting / peeling) determined by manual metallographic sectioning. It is generated by manually measuring the defect location and type after cutting the steel pipe sample under test using a metallographic microscope. This is for iteration error; this embodiment presets an iteration convergence threshold. Maximum number of iterations ,when If the iteration convergence is determined at a certain time, the current initial fused features are directly used as the final fused feature dataset. ;like If the current iteration number is less than the maximum iteration number, then update the initial fusion weight coefficients in reverse along the negative gradient of the error. The entire process, including weighted fusion of repeated features, multi-classifier prediction, and DS evidence-based decision fusion, continues until the error meets the convergence threshold or the number of iterations reaches the upper limit. After the iteration terminates, the converged fused feature dataset is output. .

[0078] This embodiment is based on the obtained fusion feature dataset. The input format is standardized to the network standard, and a pre-trained CNN (Convolutional Neural Network) is used to determine defects. The hierarchical structure of this CNN is as follows: input layer, two convolutional layers (3×3 kernels, 16 and 32 channels respectively), global average pooling layer, fully connected layer, and classification output layer. This pre-trained CNN was trained offline using tens of thousands of manually labeled steel pipe surface defect samples, and can output four categories of judgment results: no defect, surface crack, subcutaneous delamination, and pitting / peeling, simultaneously outputting the circumferential location of the defect on the steel pipe surface and the damage depth range. This embodiment will fuse the feature dataset. After being standardized into the network's standard input format, the data is input into a pre-trained CNN for defect classification and recognition, yielding the final surface inspection results of the steel pipe.

[0079] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them; modifications to the technical solutions described in the foregoing embodiments, or equivalent substitutions of some of the technical features, do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A split-type steel pipe surface inspection system, characterized in that, Includes a quick-change base, a composite detection unit, and a fusion computing module; The quick-change base is fixed to the inspection station of the steel pipe conveying assembly and is equipped with a mechanical quick-lock structure and an electrical blind-plug quick-change interface; The composite detection unit is detached and assembled from the quick-change base via an electrical blind-plug quick-change interface, and integrates a multispectral visual sensor and an eddy current array sensor to acquire multispectral images and eddy current electromagnetic signals on the surface of the steel pipe. The fusion computing module pre-stores calibration reference parameters and performs multi-level composite processing on multispectral images and eddy current electromagnetic signals before inputting them into a pre-trained convolutional neural network to identify surface defects in steel pipes, thereby obtaining surface detection results for steel pipes.

2. The split-type steel pipe surface inspection system according to claim 1, characterized in that, The multi-stage composite processing method includes the following steps: Tensor decomposition and band information entropy weighted transformation are performed on multispectral images to generate multiscale spectral feature tensors. Then, the multiscale spectral feature tensors are fitted with a hybrid probability model and combined with a graph cut optimization algorithm to split them into local catastrophe feature tensors and matrix steady-state feature tensors. Wavelet packet band decomposition and depth mapping feature separation are performed sequentially on the eddy current electromagnetic signal to obtain the matrix electromagnetic feature tensor and the defect disturbance feature tensor. The local mutation feature tensor and the defect perturbation feature tensor are fused across modes by combining a spatial deep attention mechanism with a graph convolutional network to generate the first composite defect feature tensor. The matrix steady-state feature tensor and the matrix electromagnetic feature tensor are mapped by a kernel method and fused with Bayesian posterior probability to generate the second matrix background feature tensor. Based on the calibration benchmark parameters, feature-level weighted fusion, decision-level evidence fusion, and feedback-based weight iterative optimization are sequentially performed on the first composite defect feature tensor and the substrate background feature tensor until the preset convergence condition is met, resulting in a fused feature dataset.

3. The split-type steel pipe surface inspection system according to claim 2, characterized in that, The method for obtaining the multi-scale spectral feature tensor is as follows: Tucker tensor decomposition was performed on the multispectral image to obtain the core tensor and factor matrix; Perform multi-level two-dimensional discrete wavelet transform on each spatial slice of the core tensor to obtain wavelet coefficients at different scales, including approximation coefficients and detail coefficients; The information entropy of each spectral band is calculated, and the wavelet coefficients at different scales are weighted and fused by combining the spectral band weight matrix in the calibration reference parameters to generate a multi-scale spectral feature tensor containing multiple spatial scales.

4. The split-type steel pipe surface inspection system according to claim 3, characterized in that, The method for obtaining the local mutation feature tensor and the matrix steady-state feature tensor is as follows: The global feature map corresponding to each spatial scale of the multi-scale spectral feature tensor is divided into multiple spatial sub-blocks. A two-component Gaussian mixture model is fitted to all feature pixels in each spatial sub-block to obtain the assignment probability of each feature pixel in the spatial sub-block to the first Gaussian component and the second Gaussian component. The latent attributes of the feature pixels are marked according to the assignment probability. A Markov random field based on 8 neighborhoods is constructed for the global feature region formed by stitching together all spatial sub-blocks at a single spatial scale. An energy function consisting of data terms and smoothing terms is set, and the energy function is minimized by the α-expanded graph cut algorithm to complete the classification attribute division of all feature pixels at this spatial scale and obtain the initial segmentation result corresponding to this spatial scale. The initial segmentation results corresponding to each spatial scale are fused by cross-scale voting. After removing isolated noise pixels through morphological post-processing, the features of the defect region and the features of the normal matrix region are extracted respectively, and the local mutation feature tensor and the matrix steady-state feature tensor are generated accordingly.

5. The split-type steel pipe surface inspection system according to claim 2, characterized in that, The method for obtaining the matrix electromagnetic characteristic tensor and the defect perturbation characteristic tensor is as follows: Multi-level wavelet packet decomposition is performed on the time-domain eddy current signal of each coil in the eddy current electromagnetic signal to obtain the wavelet packet coefficients corresponding to several frequency bands, and the total energy of each frequency band is calculated. Call the frequency band division matrix in the calibration reference parameters to divide all frequency bands corresponding to each coil into two categories: the base frequency band and the defect frequency band; Based on the eddy current skin depth principle, the detection depth corresponding to each frequency band is determined, and the detection depth range is divided into multiple depth intervals in advance. The total energy of the matrix frequency band and the total energy of the defect frequency band of each coil in each depth interval are statistically analyzed. The total energy data of all coils in the eddy current array sensor in two frequency bands at various depth ranges are summarized and integrated to form the matrix electromagnetic characteristic tensor and the defect disturbance characteristic tensor corresponding to the eddy current array.

6. The split-type steel pipe surface inspection system according to claim 2, characterized in that, The method for obtaining the first composite defect feature tensor is as follows: Calculate the space-depth attention weight matrix of the local mutation feature tensor and the space-depth attention weight matrix of the defect perturbation feature tensor respectively; Establish the mapping relationship between the physical spatial coordinates of the eddy current array coil and the pixel coordinates of the multispectral image, and map the defect perturbation feature tensor to the pixel coordinate system of the multispectral image; Construct a graph structure with pixels of the multispectral image as nodes, where edge weights are determined by the spatial distance between pixels and the attention weights. A two-layer graph convolutional network is used to fuse the mapped defect perturbation feature tensor and the local mutation feature tensor, and the weights are adjusted by combining attention weights to generate the first composite defect feature tensor.

7. The split-type steel pipe surface inspection system according to claim 2, characterized in that, The method for obtaining the background feature tensor of the second substrate is as follows: By selecting the radial basis function as the kernel function, the matrix steady-state feature tensor and the matrix electromagnetic feature tensor are mapped to the high-dimensional feature space to obtain the visual matrix features and electromagnetic matrix features; The covariance matrices of visual matrix features and electromagnetic matrix features are solved separately. Whitening is then performed based on the covariance matrices to eliminate the correlation between visual matrix features and electromagnetic matrix features. Based on Bayes' theorem, calculate the posterior probability distribution of visual matrix features relative to electromagnetic matrix features, and the posterior probability distribution of electromagnetic matrix features relative to visual matrix features. Based on the posterior probability distribution, the visual matrix features and electromagnetic matrix features after whitening are weighted and fused to generate the second matrix background feature tensor.

8. The split-type steel pipe surface inspection system according to claim 2, characterized in that, The method for obtaining the fused feature dataset is as follows: Using the initial fusion weight coefficients in the calibration reference parameters, the first composite defect feature tensor and the second substrate background feature tensor are fused at the feature level to obtain the initial fused features; The initial fused features are input into multiple base classifiers to obtain the prediction results and uncertainty estimates of each base classifier; The DS evidence theory is used to perform decision-level fusion of the prediction results and uncertainty estimates of multiple base classifiers to obtain the global prediction result and global uncertainty. Calculate the iteration error. If the iteration error meets the preset iteration convergence threshold, stop the iteration and use the initial fused features as the fused feature dataset. If the iteration error does not meet the preset iteration convergence threshold, the initial fusion weight coefficients are updated according to the iteration error gradient, and feature-level weighted fusion and decision-level fusion are performed again until the iteration error meets the preset iteration convergence threshold or the number of iterations reaches the preset maximum number of iterations.

9. A method for inspecting the surface of a split-type steel pipe, the method being used to implement the split-type steel pipe surface inspection system according to any one of claims 1-9, characterized in that, Includes the following steps: Multispectral images and eddy current electromagnetic signals on the surface of steel pipes are acquired by a composite detection unit that integrates a multispectral visual sensor and an eddy current array sensor. The calibration reference parameters are preset, and the multispectral image and eddy current electromagnetic signal are processed in multiple stages and then input into a pre-trained convolutional neural network to identify surface defects of steel pipes, so as to obtain the surface inspection results of steel pipes.

Images