Ultrasonic flaw detection method and system for carbon steel bar

Abstract

The invention relates to the technical field of new material detection, and particularly discloses an ultrasonic flaw detection method and system for a carbon steel bar. The method comprises the following steps: acquiring an appearance image and a section label, constructing a probe path, generating a sound field compensation factor, collecting and preprocessing an echo signal to form a multi-modal signal sequence, further extracting time sequence and spatial characteristics of the multi-modal signal sequence to generate fusion representation, and finally realizing defect classification and label generation. And dynamically updating model parameters and path configuration based on a classification result. The method can improve the detection intelligence and robustness under a complex structure, has high adaptability, high precision and high self-optimization capability, and is suitable for intelligent defect identification of industrial grade carbon steel materials.

Description

Technical Field

[0001] The present invention relates to the technical field of new material detection, and in particular to an ultrasonic flaw detection method and system for carbon steel bars. Background Art

[0002] Ultrasonic flaw detection technology, as an important means of non-destructive detection of internal defects in metal materials, has been widely used in fields such as metallurgy, mechanical manufacturing, and material forming. Traditional flaw detection methods mainly rely on fixed probe trajectories and regular sound field propagation models to analyze ultrasonic echoes, which has the following limitations:

[0003] For carbon steel bars with complex cross-sectional shapes or tiny bumps and depressions on the surface, traditional fixed-path scanning cannot flexibly match structural changes, resulting in insufficient sound field coverage in some areas and severe echo distortion.

[0004] Most methods only extract signal features at a single scale, ignoring the deep semantic correlation of ultrasonic signals in the time dimension and spatial distribution, making it difficult to effectively express the evolution characteristics of complex defects.

[0005] Most current mainstream flaw detection systems use a "feedforward" architecture, which is unable to adaptively adjust the detection model or optimize the subsequent path based on the recognition results. This causes the model to easily overfit historical patterns and is difficult to cope with complex working condition changes.

[0006] To address the above problems, existing studies have attempted to introduce technologies such as deep learning models, time series analysis methods, and graph structure optimization mechanisms for improvement. However, there is still a lack of a unified closed-loop solution that combines regional residual dynamic modeling, model parameter evolution optimization, and path planning feedback reconstruction. Summary of the Invention

[0007] The present invention provides an ultrasonic flaw detection method and system for carbon steel bars, so as to solve the problem of how to realize regional residual-driven adaptive model parameter update and path reconstruction feedback based on the multimodal ultrasonic echo signal and defect identification label of the carbon steel bar, thereby improving the intelligence and robustness of ultrasonic detection under complex geometric structures.

[0008] In order to solve the above technical problems, the present invention provides an ultrasonic flaw detection method for carbon steel bars, comprising:

[0009] Obtain the external structural image and cross-sectional geometry labels of the carbon steel bar and construct the initial scanning path configuration of the probe;

[0010] generating a sound field compensation factor based on the initial scanning path configuration and coupling medium property information;

[0011] Based on the sound field compensation factor and the initial scanning path configuration, collecting ultrasonic echo signals and performing preprocessing to generate a multimodal signal sequence;

[0012] Extracting temporal features and spatial features of the multimodal signal sequence to generate a feature fusion representation;

[0013] Perform defect classification and label generation based on the feature fusion representation, and output defect identification labels;

[0014] The defect identification label and the feature fusion representation are obtained, model parameters are updated, and a detection parameter configuration is generated.

[0015] Furthermore, the obtaining of the external structural image and cross-sectional geometric labels of the carbon steel bar and the construction of the initial scanning path configuration of the probe include:

[0016] Obtain the shape structure image and cross-sectional geometry labels of carbon steel bars, perform edge feature extraction and geometric contour recognition, and generate a geometric information graph;

[0017] Based on the geometric information graph, extracting cross-sectional shape parameters and boundary coordinate information, and constructing an initial scanning path structure of the probe;

[0018] Node compensation and speed adjustment are performed on the initial scanning path structure of the probe to generate an initial scanning path configuration.

[0019] Furthermore, generating the sound field compensation factor includes:

[0020] Acquiring the initial scanning path configuration and water immersion coupling medium property information, performing an acoustic wave penetration simulation calculation, and obtaining a path coupling coefficient;

[0021] Calculating the acoustic field imbalance at each scanning point based on the path coupling coefficient and the distribution characteristics of the nanoparticles in the coupling medium;

[0022] The sound field imbalance is minimized and adjusted to generate a sound field compensation factor for signal deconstruction and filtering configuration.

[0023] Furthermore, the collecting of ultrasonic echo signals and preprocessing to generate a multimodal signal sequence includes:

[0024] Based on the sound field compensation factor and the initial scanning path configuration, synchronously performing ultrasonic probe scanning and collecting echo signal data;

[0025] Performing filtering and envelope extraction on the echo signal data to remove high-frequency noise interference and generate a preprocessing signal sequence;

[0026] The preprocessed signal sequence is aligned and mapped with the cross-sectional geometric labels to generate a multimodal signal sequence.

[0027] Furthermore, the step of extracting temporal features and spatial features of the multimodal signal sequence and generating a feature fusion representation includes:

[0028] Based on the multimodal signal sequence, extracting temporal variation features using a long-term and short-term dependency modeling method;

[0029] Jointly modeling the temporal variation characteristics and the cross-sectional position relationship to extract spatial attention features;

[0030] Gated fusion processing is performed on the temporal change feature and the spatial attention feature to generate a feature fusion representation.

[0031] Furthermore, the defect classification and label generation includes:

[0032] The feature fusion representation is input into the classification model to perform a discrimination operation on the categories of cracks, inclusions and pore defects;

[0033] Conduct credibility assessment and grade classification processing on the discriminant operation results to generate preliminary classification labels;

[0034] The preliminary classification labels are subjected to position association processing and structured packaging to generate defect identification labels.

[0035] Furthermore, the updating of model parameters and generating detection parameter configuration includes:

[0036] The defect identification label and the feature fusion representation are obtained, and the regional residual value and the reconstruction error are calculated.

[0037] Furthermore, the flaw detection method further includes:

[0038] Based on the reconstruction error and historical comparison sample data, the internal parameters of the classification model and feature extraction module are updated.

[0039] Furthermore, the flaw detection method further includes:

[0040] The updated model parameters are fed back into the path planning parameter structure to generate the detection parameter configuration for subsequent scan optimization.

[0041] Furthermore, an ultrasonic flaw detection system for carbon steel bars is applied to any of the above-mentioned flaw detection methods, comprising:

[0042] Image and label processing module, used to obtain the shape structure image and cross-sectional geometric labels of the carbon steel bar and construct the initial scanning path configuration of the probe;

[0043] An acoustic field optimization module, used to generate an acoustic field compensation factor based on the initial scanning path configuration and coupling medium property information;

[0044] A signal acquisition and processing module is used to acquire and pre-process ultrasonic echo signals based on the sound field compensation factor and the initial scanning path configuration to generate a multimodal signal sequence;

[0045] Feature extraction module, used to extract temporal and spatial features of multimodal signal sequences and generate feature fusion representation;

[0046] Defect recognition module, which is used to classify defects and generate labels based on feature fusion representation, and output defect recognition labels;

[0047] The model update module is used to obtain defect identification labels and feature fusion representations, update model parameters and generate detection parameter configurations.

[0048] The following are its main beneficial effects:

[0049] (1) This paper quantifies the multimodal signal residuals in the defect area and constructs an adaptive weight update mechanism. Compared with traditional static models, this method can autonomously adjust the parameters of the classification model and feature extractor based on the real-time error evolution during the detection process, thereby improving the accuracy and robustness of defect recognition.

[0050] (2) This invention maps identification tags and historical evolution errors to the path planning layer, dynamically correcting the scanning path through the path potential field. Compared with the traditional fixed path mode, it significantly improves the detection coverage and sensitivity of complex cross-sections or areas with uneven materials.

[0051] (3) The present invention constructs a complete closed-loop system from multimodal signal sequence, spatial feature fusion, model parameter evolution to path feedback optimization, so that the detection system has human-like learning ability and can continuously evolve and optimize under different carbon steel bar specifications or manufacturing batches, and is widely applicable to the complex and changing environment of industrial sites. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic flow chart of an ultrasonic flaw detection method for carbon steel bars provided in an embodiment of the present application;

[0053] Figure 2 This is a structural block diagram of an ultrasonic flaw detection system for carbon steel bars provided in an embodiment of the present application. DETAILED DESCRIPTION

[0054] Example 1: Reference Figure 1 This embodiment provides a flow chart of an ultrasonic flaw detection method for carbon steel bars. The flow chart may include at least steps S100-S600:

[0055] S100, obtaining an external structural image and a cross-sectional geometric label of a carbon steel bar, and constructing an initial scanning path configuration of the probe;

[0056] S200, generating a sound field compensation factor based on the initial scanning path configuration and coupling medium property information;

[0057] S300, based on the sound field compensation factor and the initial scanning path configuration, collecting ultrasonic echo signals and performing preprocessing to generate a multimodal signal sequence;

[0058] S400, extracting temporal features and spatial features of the multimodal signal sequence to generate a feature fusion representation;

[0059] S500, performing defect classification and label generation based on the feature fusion representation, and outputting a defect identification label;

[0060] S600: Obtain the defect identification label and the feature fusion representation, update model parameters and generate a detection parameter configuration.

[0061] Step S100 at least includes steps S110-S130:

[0062] S110. Obtain an external structural image and a cross-sectional geometric label of the carbon steel bar, perform edge feature extraction and geometric contour recognition, and generate a geometric information graph.

[0063] First, the full surface appearance image of the carbon steel bar is collected by an industrial camera, a structured light module or a laser scanning device. The appearance structure image is a set of continuously constructed two-dimensional image frames, which can be converted into an image matrix with separable grayscale, edges and contours.

[0064] Furthermore, the cross-sectional geometry labels of the carbon steel bars registered in the database are obtained. The cross-sectional geometry labels are a set of predefined parameters associated with the material structure, including whether the bar is a circular cross-section, a hexagonal cross-section, an elliptical cross-section or other specific cross-sectional categories, and correspondingly store the number of edge corner points, side length distribution, edge type mark and center offset measurement.

[0065] Specifically, the outer structure image and the cross-sectional geometric label are used as input, and edge enhancement processing is performed on the edges of the carbon steel bar in the image space. An edge detection algorithm is used to extract the edge response area in the image to form an edge response map. A contour recognition operation is then performed on the edge response map to eliminate background interference and image artifact information, and a closed contour is output.

[0066] The contour recognition result is matched and verified with the cross-section geometric label, and the reliability of image recognition is judged based on the contour closure degree and the structure fitting deviation. The recognition results that do not meet the error threshold are repeatedly collected and reconstructed.

[0067] Furthermore, a geometric information graph including boundary vectors, centroid positions, side length scales, and contour types is generated to provide structural input for constructing the probe path structure in S120.

[0068] S120 . Based on the geometric information graph, extract cross-sectional shape parameters and boundary coordinate information, and construct an initial scanning path structure of the probe.

[0069] In S120 , based on the geometric information graph generated in S110 , the cross-sectional shape parameters and boundary coordinate information recorded in the graph are parsed, and a path construction operation is performed.

[0070] Specifically, the following three elements are extracted from the geometric information graph: the coordinates of the cross-section center point, the coordinates of each edge node, and the boundary type identifier. The edge node coordinates are used to construct the path starting point set and path transition point set, the center point is used for probe path optimization and staggered scanning settings, and the boundary type identifier is used to match the steering control rules during path generation.

[0071] Furthermore, different forms of probe initial scanning path structures are constructed according to the cross-section type:

[0072] For circular cross-sections, an equal-angle segmentation algorithm is performed to form closed-loop spiral paths with equal radius;

[0073] For hexagonal cross-sections, six directional vector path segments are constructed to form a broken line trajectory that combines edge alignment with corner tangent transitions;

[0074] For ellipses and non-standard cross-sections, an adaptive path generation strategy based on boundary curvature continuity and centroid aggregation is implemented. The generated path includes equidistant point cloud strips, edge fitting segments, and corner compensation areas.

[0075] The path structure is output in the form of a vector path graph, which contains the path node sequence, the distance between adjacent nodes, and the geometric label index corresponding to each node.

[0076] The above path structure is used as input for node compensation and speed adjustment processing in S130 to ensure structural continuity and consistency of scanning dynamic control.

[0077] S130 , performing node compensation and speed adjustment processing on the probe initial scanning path structure to generate an initial scanning path configuration.

[0078] After receiving the initial probe scanning path structure output by S120, S130 performs node-level structural optimization. First, based on the degree of angle mutation between path segments, the distance between adjacent nodes, and the boundary category identification, it determines the path segments that need to be interpolated or the corner transition areas that need to be corrected.

[0079] Specifically, cubic interpolation operations are performed on node segments in the path that do not meet the continuity criteria, and velocity buffers are marked at the interpolation points. Transition curve fitting segments are set for critical corner nodes to construct smooth curve trajectories and suppress trajectory jitter caused by high-speed switching.

[0080] Furthermore, each path node is assigned a scanning speed weight value, which is determined based on factors such as the acoustic impedance of the area where the node is located, the complexity of the geometric features, and the estimated signal reflection intensity. The speed weight is used in the calculation of the path coupling coefficient in step S210.

[0081] Finally, the output is the initial scanning path configuration containing information such as node position sequence, path segment number, speed weight set, and vector direction between nodes.

[0082] The initial scanning path configuration serves as a necessary input for generating the sound field compensation factor in S200 and participates in the joint modeling of the path and medium. The velocity weights and distribution density of each node in the path directly influence the calculation model of the path coupling coefficient and are further transmitted to the sound field imbalance calculation logic in S220.

[0083] Step S200 at least includes steps S210-S230:

[0084] S210 , obtaining the initial scanning path configuration and water immersion coupling medium property information, performing acoustic wave penetration simulation calculation, and obtaining a path coupling coefficient.

[0085] Specifically, when executing S210, the initial scanning path configuration outputted by S130 is first called, which includes a path node sequence, a path segment structure, a speed weight set, and path length information. The path node is a set of three-dimensional points in space with a clear position and order relationship.

[0086] Simultaneously, the system acquires the properties of the water-immersion coupling medium corresponding to the path configuration. These properties include medium type, water temperature, water pressure, density, acoustic impedance, and the type and concentration of suspended nanoparticles. This property information is set during system initialization or periodically collected and transmitted by an external detection module.

[0087] After the path configuration is matched with the medium properties, a sound wave penetration simulation is performed. This simulation calculates the energy attenuation, phase delay factor, and reflection probability of the sound wave during its propagation through the medium for each node connected to the path segment.

[0088] The overall acoustic impedance matching and effective penetration depth at the segment level are calculated by combining the angular rate of change and velocity weight of each path segment. Each calculation result is mapped to a coupling response value corresponding to the path segment, which in turn forms a complete list of path coupling coefficients.

[0089] The path coupling coefficient is used as a structured output and input to step S220 for imbalance modeling.

[0090] S220 , calculating the sound field imbalance at each scanning point based on the path coupling coefficient and the distribution characteristics of the nanoparticles in the coupling medium.

[0091] In step S220, the path coupling coefficient obtained in step S210 is used as a basic input, and the nanoparticle distribution characteristics of the coupling medium are called to construct a multi-point sound field balance analysis model.

[0092] Nanoparticle distribution characteristics include particle size range, spatial distribution density, particle activity parameters, and particle projections on the scanned surface. Based on a predefined particle transport model within the coupling medium, the system calculates the particle participation ratio at each scan point—the ratio of the number of particles per unit volume around each scan point to the global average particle concentration.

[0093] Combining the acoustic resistance and propagation intensity of the corresponding scanning point in the path coupling coefficient, a penetration response comparison function is established for different points to reflect the impact of particle concentration differences on the balance of sound wave propagation.

[0094] Furthermore, the response values ​​corresponding to each node in the entire scanning path are normalized and calculated to obtain the sound field response deviation, and a sound field imbalance surface of the entire path is formed in the spatial dimension.

[0095] The sound field imbalance is the core input of the S230 sound field adjustment strategy and is the key basis for judging whether the flaw detection signal has a consistent deconstruction basis.

[0096] S230: Minimize the sound field imbalance and generate a sound field compensation factor for signal deconstruction and filtering configuration.

[0097] In step S230 , based on the sound field imbalance surface calculated in step S220 , adjustment model construction and compensation factor generation operations are performed.

[0098] First, a compensation priority table is created based on the degree of sound field imbalance at each scanning point. For areas with higher imbalance, the system assigns higher weights to participate in compensation adjustments.

[0099] Then, a dynamic point weighting algorithm is executed to reconstruct the original coupling coefficient under the particle response weight, and the scanning speed and node angle changes are adjusted synchronously to generate a set of sound field adjustment coefficients corresponding to each node in the scanning path.

[0100] The adjustment coefficient is mapped into a sound field compensation factor in the system. The compensation factor is a structured data unit that records the filter enhancement value, reflection area magnification factor, and sampling density multiple of each path point.

[0101] The resulting sound field compensation factor will serve as a key control parameter for S300 signal acquisition and filtering processing, controlling the echo signal sampling rate, enhancement filter selection, and dynamic frequency domain window size.

[0102] Step S300 at least includes steps S310-S330:

[0103] S310 , based on the sound field compensation factor and the initial scanning path configuration, synchronously perform ultrasonic probe scanning and collect echo signal data.

[0104] Specifically, the system schedules the seven-axis scanning mechanism to execute path trajectory control, guiding the ultrasonic probe to continuously move in three-dimensional space according to the node sequence and path segment sequence in the initial scanning path configuration. The probe's scanning speed and transition parameters at path corners are controlled in real time based on the speed weights and node compensation results configured in S130.

[0105] During synchronization, the acoustic wave transmission module adjusts the pulse emission intensity, frequency interval, and transmission angle based on the velocity state and corresponding acoustic field compensation factor at each node. The value of the acoustic field compensation factor affects the acoustic wave penetration power and angular accuracy, ensuring stable reflection sampling capabilities at boundary transition areas and acoustic impedance mutation points along the scanning path.

[0106] After the probe receives the ultrasonic echo signal from the interior of the carbon steel bar, the system caches and stores the echo data for each time window, recording the signal amplitude, echo delay, scan node index, and path segment label. All signal data is labeled according to the acquisition sequence, path segment number, and spatial position index to form the original echo signal data set.

[0107] S320: Perform filtering processing and envelope extraction on the echo signal data to remove high-frequency noise interference and generate a preprocessing signal sequence.

[0108] In S320, the system calls the original echo signal data set collected in S310 and performs customized filtering using the filtering parameters matched with the sound field compensation factor, including: intermediate frequency bandpass filtering, abnormal peak reduction filtering and periodic interference suppression algorithm.

[0109] The filtering process uses a dynamic window width adjustment mechanism. The filter window size, reconstruction amplitude threshold, and sampling point compression ratio are all determined by the sound field compensation factor generated in S230. The system automatically switches to an enhanced filtering strategy based on the degree of sound field imbalance in the area where different scanning points are located, achieving fine-grained analysis of boundary signals in complex cross-sectional areas.

[0110] After filtering, the system performs envelope extraction on the enhanced time-domain echo signal sequence. Envelope extraction is used to restore the changing trend of reflection intensity, remove high-frequency carrier information, and enhance the edge characteristics of reflectors.

[0111] Finally, the filtered signal is uniformly packaged with its corresponding probe scanning time, path segment number, and node index to generate a structured preprocessed signal sequence, which serves as the input for S330 mapping to generate multimodal data.

[0112] S330 , aligning and mapping the preprocessed signal sequence with the cross-sectional geometric labels to generate a multimodal signal sequence.

[0113] In S330 , the cross-sectional geometry tag extracted in S110 is first called. The tag contains information such as cross-sectional shape type, number of boundary points, cross-sectional symmetry, and structure identification number, and is used to complete the spatial mapping between the preprocessed signal and the geometric model.

[0114] For each data frame in the preprocessed signal sequence, the system locates its spatially corresponding position in the cross-sectional profile diagram based on its path segment index and node number, completing the signal physical structural alignment. The spatial reference information used for this alignment operation is the probe initial scanning path structure constructed in S120.

[0115] Furthermore, the system classifies each signal frame according to the path segment and binds them to the corresponding geometric tags to form a composite data unit with section type, node relative coordinates and reflection intensity tags.

[0116] The system normalizes and distributes the data units within all path segments to construct a multimodal signal sequence that has both temporal dimension and spatial structure labels, providing complete data features for entering the S400 feature extraction module.

[0117] Step S400 at least includes steps S410-S430:

[0118] S410 , extracting temporal variation features based on the multimodal signal sequence using a long-term and short-term dependency modeling method.

[0119] In S410 , the multimodal signal sequence is first reconstructed on a time axis according to the scanning path segment numbers, ensuring that all signal sequences have monotonically increasing time tags in the structured input.

[0120] The system adopts a time series dependency model with bidirectional correlation modeling capabilities, inputs reflection amplitude information and time labels point by point, and constructs a time series feature set based on local delay patterns and global attenuation trends.

[0121] Specifically, the system calculates the amplitude change rate of the signal's rising edge, peak duration, and falling edge, and combines this with node scan speed tags to construct a dynamic time-domain signature for each reflection event along its path. This signature demonstrates the reflection stability and signal morphology consistency across different structural regions.

[0122] The system uniformly encodes the time domain feature set formed in each signal curve, and adds path segment identification information after encoding to generate a time series change feature output including time domain pattern, position sequence number and response amplitude.

[0123] The output provides a basic temporal feature expression for the spatial modeling step in S420 and is one of the key inputs for the spatiotemporal fusion processing.

[0124] S420: Jointly model the temporal variation characteristics and the cross-sectional position relationship to extract spatial attention features.

[0125] In S420 , the system combines the temporal variation features extracted in S410 with the geometric structure information retained in S330 to construct a spatial attention model based on structure index and feature heat.

[0126] First, the mapping points of each reflection event in the cross-sectional geometric label are extracted, and a local aggregation operation is performed on the mapping area to count the concentration and difference of the signal time domain features in each area.

[0127] Furthermore, the system uses the cross-section edge type, boundary curvature and regional density as geometric prior parameters, combined with the temporal feature response intensity, to generate a spatial position perception weight map.

[0128] The system calculates the consistency between the signal activity level and the structural distribution in each structural area, establishes an attention distribution function, and weights the spatial importance of different scanning nodes in the temporal features.

[0129] The output spatial attention features contain the node spatial position index, the importance score of the corresponding position, and the mapping feature density, which is the second main input for fusion processing.

[0130] S430. Perform gated fusion processing on the temporal change feature and the spatial attention feature to generate a feature fusion representation.

[0131] In S430, the system takes the temporal change features extracted in S410 and the spatial attention features constructed in S420 as joint inputs and performs fusion through the set gating mechanism.

[0132] The gating mechanism includes a position matching module, a feature comparison module, and a weighted fusion module. The system first uses the node index as the connection basis to match the temporal features with the spatial features one by one.

[0133] The system then calculates the response difference between each temporal feature and its corresponding spatial attention weight to control the fusion ratio. For nodes with small differences, the system implements a high-weight retention strategy to strengthen the joint feature expression; for nodes with large differences, it implements feature compensation and information reconstruction mechanisms to balance the contribution of the two types of features to the fusion result.

[0134] The final output feature fusion is represented as a structured tensor, which contains the time domain dynamic structure, spatial geometric response and fusion feature label, and is the input basis for the S500 module to perform defect classification and label generation.

[0135] Step 500 at least includes steps S510-S530:

[0136] S510 , inputting the feature fusion representation into a classification model to perform a discrimination operation on the categories of cracks, inclusions and pore defects.

[0137] In S510, the feature fusion representation generated by S430 is first obtained, in which the fusion expression data of the temporal variation feature and the spatial attention feature corresponding to the multimodal signal sequence is embedded.

[0138] Specifically, the system loads the feature fusion representation as a structured tensor and feeds it into a pre-set classification model in batches, following the order of scan path nodes. This classification model is a pre-trained defect type discrimination model, built using a deep discriminant architecture capable of accepting high-dimensional features and trained with overfitting constraints using comparison test block data.

[0139] The classification model outputs the probability response value of each feature fusion unit according to the structural information in the unit, which corresponds to crack, inclusion and pore defects respectively.

[0140] Furthermore, the system performs a first-order label mapping operation based on the discriminant response value, and uses the maximum response category output by each node as its initial defect type, providing a classification basis for the credibility analysis and grade grading of S520.

[0141] The classification output is recorded in the form of a structured list, including path node index, discriminant type label, category response value and feature fusion segment position identifier.

[0142] S520: Perform credibility assessment and grade classification processing on the discriminant operation results to generate preliminary classification labels.

[0143] In S520 , the system uses the discriminant type label and response value outputted in S510 as input and performs a credibility calculation process based on the response distribution density and the confidence interval of the historical samples.

[0144] Specifically, the discriminant response values ​​of all nodes are first regularized, and the corresponding credibility template library is constructed based on the standard distribution of each type of defect in historical samples.

[0145] Then, the actual response value is matched in the template library to obtain a combination of matching rate and response stability, and the credibility score is calculated comprehensively.

[0146] According to the scoring results, each node is assigned to a different confidence level interval, such as high confidence, medium confidence and low confidence, and the level discrimination result is output in combination with the type label.

[0147] The system further maps the trustworthy levels to defect grade labels, such as level 1 cracks, level 2 inclusions, or level 3 pores, to form a structured preliminary classification label record set.

[0148] The format of the preliminary classification label includes defect type, confidence level, node index, response stability identifier, and source feature fragment number, providing a reference for position reconstruction and structural encapsulation.

[0149] S530 , performing position association processing and structural packaging on the preliminary classification labels to generate defect identification labels.

[0150] In S530 , the system uses the preliminary classification labels generated in S520 as input, combines the cross-sectional geometry label information and path segment index data saved in S330 , and performs spatial positioning and structural association of the defect labels.

[0151] Specifically, first, according to the node index of each tag, the corresponding position point is searched in the geometric tag structure, and the path segment information and regional geometric properties of the point are obtained, including the cross-sectional shape, boundary type and acoustic mapping point.

[0152] Next, the system constructs a defect location calibration structure, marking each defect with its precise physical coordinates, path segment number, and scanning direction identifier, and forming a "type-level-location" ternary label.

[0153] The system organizes all ternary tags in the order of path segments, packages them into a structured tag data format, and constitutes the final defect identification tag output result.

[0154] The defect identification labels will be input into the S600 module and combined with the model feedback mechanism for reconstruction parameter update and scan path optimization.

[0155] Step S600 at least includes steps S610-S630:

[0156] S610 obtains the defect identification label and the feature fusion representation, and calculates the regional residual value and the reconstruction error.

[0157] Specifically, based on the defect identification label and the feature fusion representation, a model is constructed to calculate the regional residual value and reconstruction error. The specific steps are as follows:

[0158] ①Construct feature evolution state field:

[0159] The probability state evolution of each defect area in the feature fusion space is defined as:

[0160]

[0161] in:

[0162] ψ i (x, t): time-space state function of the i-th defect region;

[0163] α ik (t): weight coefficient of the kth feature in the i-th region in the feature fusion representation;

[0164] φ k (x): spatial feature basis represented by basis function;

[0165] x: corresponds to the spatial position parameter in the image or signal space;

[0166] K: total number of features.

[0167] This formula combines the characteristic coefficients generated by the LSTM-Transformer parallel mechanism to achieve temporal coupling modeling of spatial structure changes.

[0168] ②Calculate the residual offset term:

[0169] The regional residual is defined based on the standardized one-dimensional Schrödinger equation as:

[0170]

[0171] in:

[0172] The residual signal amplitude of region i;

[0173] V i (x): Potential disturbance field of the defect area, generated based on the defect identification label mapping;

[0174] ψ i (x, t) is the time-space state function of the i-th defect area

[0175] Laplace spatial diffusion term, reflecting the change of local structural gradient;

[0176] x: is the spatial position variable.

[0177] This equation is used to calculate the interference response caused by regional structural changes on signal conduction.

[0178] ③Weighted processing of state entropy weight:

[0179] Apply information entropy weight adjustment to all residual terms:

[0180]

[0181] in:

[0182] Shannon entropy of region i;

[0183] p ij : The normalized probability of the jth feature category in the i-th region;

[0184] M: total number of all feature categories;

[0185] λ: entropy adjustment coefficient;

[0186] The entropy-adjusted residual value.

[0187] By introducing information entropy, the uncertainty of multi-category signal responses is weightedly suppressed to achieve optimization of defect area identification accuracy.

[0188] S620: Based on the reconstruction error and historical comparison sample data, update the internal parameters of the classification model and feature extraction module.

[0189] In this step, the system further uses the regional residual value generated in S610 and the historical sample error distribution to construct an update criterion, and realizes the network weight update based on the entropy extreme value drive.

[0190] ① Define the Shannon entropy calculation model for the reference residual distribution of historical samples:

[0191]

[0192] in, is the historical distribution entropy, which represents the stability of historical judgment; p krepresents the normalized frequency of the k-th type of defect sample error in history; K represents the number of historical sample types. This entropy value measures the degree of deviation between the current reconstruction error and the historical error.

[0193] ② Design the residual feedback network structure and update the gated attention unit weight parameters:

[0194]

[0195] in:

[0196] W (t) : Classification model parameter matrix at time t;

[0197] W (t+1) : updated weight;

[0198] P: transfer matrix based on feature entropy correlation;

[0199] α: Damping factor, the typical value is 0.85;

[0200] Feature attention maps output by the parallel LSTM and Transformer models.

[0201] ③ While optimizing parameter feedback, define a residual-guided parameter correction operator to improve robustness:

[0202]

[0203] in:

[0204] μ(n): Möbius function, used to weight parameter perturbations at different historical scales;

[0205] The value of the model's i-th parameter in the n-th iteration of history;

[0206] Δθ i : The offset value of the i-th parameter that needs to be adjusted in this round;

[0207] The gradient response of the weighted residual with respect to the parameters.

[0208] By constructing a multi-scale perturbation function, the robustness and generalization ability of parameter updates are enhanced.

[0209]

[0210] in:

[0211] Θ (t) : represents the set of all model parameters at the current time t;

[0212] Δθi : The updated offset of the i-th parameter, which comes from the weight perturbation result derived from the Möbius inversion;

[0213] Θ new : The final model parameter vector group after integrating reconstruction error feedback, historical sample entropy weight adjustment and structure tensor field feedback.

[0214] S630: Feedback the updated model parameters to the path planning parameter structure to generate a detection parameter configuration for subsequent scanning optimization.

[0215] This step constructs the field potential change map and completes the feedback path generation.

[0216] ①Introduce the Riemann curvature tensor to define the intrinsic correlation between the manifold points in the path structure:

[0217]

[0218] in:

[0219] Manifold tensor component, reflecting the local geometric distortion of the path;

[0220] The contact coefficient is constructed based on the local curvature of the scan path;

[0221] and Partial derivative of a quantity along the μth or νth coordinate direction;

[0222] Indicates the connection term in the μth dimension direction changes;

[0223] Indicates the connection term in the νth dimension direction changes;

[0224] and These two terms are combinations of connection terms in different dimensions, which model the composite connection mode in the path field and constitute the nonlinear part of the curvature change;

[0225] It is used to reconstruct the degree of distortion under the influence of geometric tension of the current path segment and provide geometric field support for dynamic path adjustment.

[0226] ②Build an optimal fallback mechanism for path weights:

[0227] d j =min(d j ,d i +w ij (Θ new ))

[0228] in:

[0229] d j : The shortest path cost from the starting point to node j;

[0230] d i : The minimum cost to the adjacent node i;

[0231] w ij Θ new ): path loss function between nodes under current parameters;

[0232] Θ new : Model weight updated in S620.

[0233] ③ represents the path energy evolution trend:

[0234]

[0235] in:

[0236] Ψ: wave function of the scanning path field;

[0237] Normalized Planck constant;

[0238] m: equivalent acoustic mass of the fitting material;

[0239] Laplace diffusion of path fields;

[0240] V scan (x,Θ new ): The path potential field generated in combination with the updated parameters.

[0241] ④ Comprehensive path strategy function

[0242]

[0243] in:

[0244] Π scan (x): final path configuration;

[0245] It means selecting the path that minimizes the target functional from the set of all paths Γ(x); Γ(x): the set of all feasible paths from the current point to the target area;

[0246] γ: parameterized expression of a specific path (e.g., expressed in arc length or coordinates);

[0247] argmin: select the path with the minimum cost;

[0248] ∫ γ)dx: represents the integration of the local cost of each point on the path γ;

[0249] The cost of geometric adjustments caused by Riemann tension;

[0250] Update the parameter-guided cost function;

[0251] |Ψ(x,t)| 2 : energy density at the path point;

[0252] β: regulatory factor.

[0253] in:

[0254] The specific calculation expression is:

[0255]

[0256] Where: g μν (x) is the metric tensor of the path point.

[0257] for:

[0258]

[0259] where ξ i (x): represents the expression strength of the i-th feature at point x.

[0260] Finally used to generate the final path configuration π scan It is used by subsequent scheduling control modules to achieve self-adjustment and optimization of the path structure.

[0261] This section constructs a dynamic optimization method for the flaw detection model, which can realize high-dimensional dynamic analysis of defect identification feedback information, adaptive parameter adjustment and iterative optimization of path strategy, significantly enhancing the system's generalization ability and detection robustness under complex working conditions, and has significant algorithm innovation value and industrial promotion potential.

[0262] Example 2: Figure 2 FIG. 1 is a block diagram showing a structural diagram of an ultrasonic flaw detection system for carbon steel bars according to an embodiment of the present invention. Figure 2 As shown, the structure may include:

[0263] The image and label processing module 10 is used to obtain the external structural image and cross-sectional geometric labels of the carbon steel bar and construct the initial scanning path configuration of the probe. This module performs the following functions in sequence:

[0264] Acquire and process image data to identify the outer contour and cross-sectional edges of the steel bar;

[0265] Extract cross-sectional geometric parameters and boundary coordinate information;

[0266] A preliminary scanning path structure is constructed and node compensation and velocity calibration are performed in combination with geometric features to generate an optimized initial scanning path configuration.

[0267] The sound field optimization module 20 is used to generate sound field compensation factors based on the initial scanning path configuration and the properties of the water immersion coupling medium to enhance the accuracy of subsequent signal acquisition. Specific functions include:

[0268] Perform simulation calculations on the acoustic wave penetration path to obtain the path coupling coefficient;

[0269] Based on the nanoparticle distribution model in the coupling medium, the acoustic field imbalance of each path node is calculated;

[0270] The sound field imbalance is minimized through dynamic adjustment strategies, generating sound field compensation factors that can be used for deconstruction and filtering.

[0271] The signal acquisition and processing module 30 is used to acquire and pre-process ultrasonic echo signals based on the sound field compensation factor and the scanning path to generate a multimodal signal sequence. This module mainly includes:

[0272] Control the ultrasonic probe to scan synchronously along a preset path and collect signals;

[0273] Perform filtering and envelope extraction operations to effectively remove high-frequency interference;

[0274] The processed echo signal is aligned with the geometric label information to form a multimodal input sequence that can be used for feature extraction.

[0275] The feature extraction module 40 is used to extract discriminative temporal and spatial features from the multimodal signal sequence and construct a feature fusion representation. This module has the following functions:

[0276] Extract time series feature sequences based on long-term and short-term dependency models;

[0277] Jointly consider the cross-sectional position relationship and explore the importance of spatial dimensions through the attention mechanism;

[0278] The gating mechanism is used to fuse temporal and spatial features to generate a fusion vector set with strong representation capabilities.

[0279] The defect recognition module 50 is used to classify defects based on feature fusion representation and generate structured label information. Its core functions include:

[0280] The fused features are input into a multi-class classification model to identify typical defect categories such as cracks, inclusions, and pores;

[0281] Evaluate the confidence of the classification results and perform grading based on the threshold strategy;

[0282] The preliminary label information is structurally encapsulated and the defect label results with position index are output.

[0283] The model update module 60 is used to obtain defect identification labels and feature fusion representations, perform model adaptive optimization based on feedback errors, and generate detection parameter configurations. It mainly implements the following functions:

[0284] Construct the state field and residual function based on the Schrödinger equation, and combine it with Shannon entropy to achieve error weighted compensation;

[0285] The PageRank and LSTM-Transformer parallel model is used to perform dynamic weight optimization, while the Möbius inversion mechanism is introduced to enhance parameter robustness.

[0286] Combining the Riemannian geometric curvature tensor with the Bellman-Ford path selection strategy, feedback is used to form the optimized ultrasonic scanning path configuration, realizing closed-loop regulation of the system.

[0287] An ultrasonic flaw detection system for carbon steel bars provided by an embodiment of the present invention has the following beneficial effects by constructing a complete closed-loop structure from image acquisition, path planning, signal compensation, feature extraction, defect identification, to parameter updating:

[0288] The path structure is generated by combining image recognition and geometric modeling to improve the geometric accuracy of the initial scanning configuration;

[0289] By simulating sound wave propagation and optimizing compensation factors, the stability and effectiveness of signal acquisition are significantly enhanced;

[0290] Introducing LSTM and attention mechanisms to integrate signal timing and spatial information into modeling, improving defect recognition accuracy;

[0291] An original parameter update and path feedback mechanism has been constructed to achieve the system's robustness, adaptability and optimal path dynamic scheduling capabilities.

[0292] This system is particularly suitable for high-sensitivity and high-precision ultrasonic non-destructive testing of carbon steel bars with complex structures, high deformation or irregular defect distribution, and has significant engineering promotion and application value.

Claims

1. An ultrasonic flaw detection method for carbon steel bars, characterized in that: The following steps are involved: Obtain the external structural image and cross-sectional geometry labels of the carbon steel bar and construct the initial scanning path configuration of the probe; generating a sound field compensation factor based on the initial scanning path configuration and coupling medium property information; Based on the sound field compensation factor and the initial scanning path configuration, collecting ultrasonic echo signals and performing preprocessing to generate a multimodal signal sequence; Extracting temporal features and spatial features of the multimodal signal sequence to generate a feature fusion representation; Perform defect classification and label generation based on the feature fusion representation, and output defect identification labels; The defect identification label and the feature fusion representation are obtained, model parameters are updated, and a detection parameter configuration is generated.

2. The flaw detection method according to claim 1, characterized in that: The step of obtaining the appearance structure image and cross-sectional geometry labels of the carbon steel bar and constructing the initial scanning path configuration of the probe includes: Obtain the shape structure image and cross-sectional geometry labels of carbon steel bars, perform edge feature extraction and geometric contour recognition, and generate a geometric information graph; Based on the geometric information graph, extracting cross-sectional shape parameters and boundary coordinate information, and constructing an initial scanning path structure of the probe; Node compensation and speed adjustment are performed on the initial scanning path structure of the probe to generate an initial scanning path configuration.

3. The flaw detection method according to claim 1, characterized in that: Generating the sound field compensation factor includes: Acquiring the initial scanning path configuration and water immersion coupling medium property information, performing an acoustic wave penetration simulation calculation, and obtaining a path coupling coefficient; Calculating the acoustic field imbalance at each scanning point based on the path coupling coefficient and the distribution characteristics of the nanoparticles in the coupling medium; The sound field imbalance is minimized and adjusted to generate a sound field compensation factor for signal deconstruction and filtering configuration.

4. The flaw detection method according to claim 1, characterized in that: The ultrasonic echo signal is collected and pre-processed to generate a multimodal signal sequence, including: Based on the sound field compensation factor and the initial scanning path configuration, synchronously performing ultrasonic probe scanning and collecting echo signal data; Performing filtering and envelope extraction on the echo signal data to remove high-frequency noise interference and generate a preprocessing signal sequence; The preprocessed signal sequence is aligned and mapped with the cross-sectional geometric labels to generate a multimodal signal sequence.

5. The flaw detection method according to claim 1, characterized in that: The step of extracting temporal features and spatial features of a multimodal signal sequence and generating a feature fusion representation includes: Based on the multimodal signal sequence, extracting temporal variation features using a long-term and short-term dependency modeling method; Jointly modeling the temporal variation characteristics and the cross-sectional position relationship to extract spatial attention features; Gated fusion processing is performed on the temporal change feature and the spatial attention feature to generate a feature fusion representation.

6. The flaw detection method according to claim 1, characterized in that: The defect classification and label generation includes: The feature fusion representation is input into the classification model to perform a discrimination operation on the categories of cracks, inclusions and pore defects; Conduct credibility assessment and grade classification processing on the discriminant operation results to generate preliminary classification labels; The preliminary classification labels are subjected to position association processing and structured packaging to generate defect identification labels.

7. The flaw detection method according to claim 1, characterized in that: The updating of model parameters and generating detection parameter configuration includes: The defect identification label and the feature fusion representation are obtained, and the regional residual value and the reconstruction error are calculated.

8. The flaw detection method according to claim 7, characterized in that: Also includes: Based on the reconstruction error and historical comparison sample data, the internal parameters of the classification model and feature extraction module are updated.

9. The flaw detection method according to claim 8, characterized in that: Also includes: The updated model parameters are fed back into the path planning parameter structure to generate the detection parameter configuration for subsequent scan optimization.

10. An ultrasonic flaw detection system for carbon steel bars, applied to the flaw detection method according to any one of claims 1 to 9, characterized in that: include: Image and label processing module, used to obtain the shape structure image and cross-sectional geometric labels of the carbon steel bar and construct the initial scanning path configuration of the probe; An acoustic field optimization module, used to generate an acoustic field compensation factor based on the initial scanning path configuration and coupling medium property information; A signal acquisition and processing module is used to acquire and pre-process ultrasonic echo signals based on the sound field compensation factor and the initial scanning path configuration to generate a multimodal signal sequence; Feature extraction module, used to extract temporal and spatial features of multimodal signal sequences and generate feature fusion representation; Defect recognition module, which is used to classify defects and generate labels based on feature fusion representation, and output defect recognition labels; The model update module is used to obtain defect identification labels and feature fusion representations, update model parameters and generate detection parameter configurations.

Images