Ultrasonic testing system for pipe weld defects

Abstract

The present application belongs to the technical field of weld detection, and discloses an ultrasonic detection system for pipeline weld defects, which comprises: a data acquisition module for acquiring original echo data generated by multi-angle ultrasonic scanning of pipeline welds through a ring array probe group; a feature extraction module for parallel processing of the original echo data, extracting time domain waveform features, frequency domain spectrum features and spatial domain orientation features, and forming an initial multi-dimensional feature set; a feature space transformation module for inputting the initial multi-dimensional feature set into a dynamic feature space transformation network to generate a deep feature vector; and a collaborative decision module for simultaneously inputting the deep feature vector into a defect type recognition unit and a defect size quantification unit to generate a collaborative decision result, which contains a defect type recognition result and a defect size quantification result. The present application realizes ultrasonic detection of pipeline weld defects through intelligent data acquisition, feature extraction, feature space transformation and collaborative decision functions.

Description

Technical Field

[0001] This invention belongs to the field of weld inspection technology, and particularly relates to an ultrasonic inspection system for pipeline weld defects. Background Technology

[0002] As a critical infrastructure for transporting energy sources such as oil and gas, the quality of pipeline welds directly affects the safety and stable operation of the entire pipeline network. Ultrasonic testing technology, due to its high sensitivity to both volumetric and area defects, has become a core non-destructive testing method for ensuring the integrity of pipeline welds. In particular, phased array ultrasonic testing technology, through electronic control to deflect and focus the sound beam, enables rapid and comprehensive scanning of welds, making it the current mainstream testing method.

[0003] Among related technologies, Chinese invention patent CN120832635B discloses a method, system, equipment, and storage medium for detecting defects in hydrogen transportation pipelines, including: acquiring phased array ultrasonic testing data, digital X-ray imaging data, and hydrogen concentration data of the hydrogen transportation pipeline; extracting defect boundary features based on phased array ultrasonic testing data, extracting defect microstructure features based on digital X-ray imaging data, and extracting spatiotemporal distribution features of hydrogen concentration based on hydrogen concentration data; constructing a modal correlation graph using defect boundary features, defect microstructure features, and spatiotemporal distribution features of hydrogen concentration as nodes, wherein the edge weight between each pair of nodes in the modal correlation graph is determined by the mutual information of the two nodes; generating a multimodal correlation feature vector based on the modal correlation graph; and obtaining the defect detection results of the hydrogen transportation pipeline based on the multimodal correlation feature vector.

[0004] However, the aforementioned existing technical solutions have the following technical defects. Feature extraction relies on manual pre-setting, resulting in insufficient adaptability: The defect boundary features, microstructural features, and spatiotemporal distribution features of hydrogen concentration extracted by this technology are all fixed features designed manually. These features cannot be adaptively adjusted according to the complex morphology and weak signals of weld defects in actual detection. The sensitivity for identifying atypical and hidden defects is low, easily leading to missed detections or misjudgments due to insufficient feature representation. Defect qualitative and quantitative processing are separated, resulting in poor decision consistency: The technical solution does not establish a collaborative constraint mechanism for defect type identification and size quantification; the two tasks are mostly executed independently. This can lead to logical contradictions in the detection results. For example, the identification result may be spherical porosity, but the quantitative result shows that the defect aspect ratio is too large. This cannot be corrected through internal correlation verification, reducing the reliability of the detection conclusion. Lack of online learning capability makes it difficult to adapt to changes in working conditions: The detection model of this technology is a static model, which cannot autonomously identify new clusters in the feature space, i.e., potential unknown defect patterns, during continuous detection. When faced with new defects that emerge under new working conditions, the model cannot be fine-tuned and optimized online using field data. In long-term applications, it lacks robustness and requires manual retraining, resulting in high adaptation costs and low efficiency. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide an ultrasonic testing system for pipeline weld defects, which achieves ultrasonic testing of pipeline weld defects through intelligent data acquisition, feature extraction, feature space transformation, and collaborative decision-making functions.

[0006] The above objectives can be achieved through the following approach: An ultrasonic testing system for pipe weld defects includes the following modules connected in sequence: The data acquisition module is used to acquire the raw echo data generated by multi-angle ultrasonic scanning of pipe welds using a ring array probe group; The feature extraction module is used to process the raw echo data in parallel, extracting time-domain waveform features, frequency-domain spectral features, and spatial-domain steering features to form an initial multi-dimensional feature set. The feature space transformation module is used to input the initial multidimensional feature set into a preset dynamic feature space transformation network. The dynamic feature space transformation network generates adaptive feature space transformation parameters based on the content of the initial multidimensional feature set, and performs spatial reconstruction on the initial multidimensional feature set based on the adaptive feature space transformation parameters to generate a deep feature vector. The collaborative decision-making module is used to simultaneously input deep feature vectors into a preset defect type identification unit and a defect size quantification unit to generate collaborative decision-making results, which include defect type identification results and defect size quantification results.

[0007] Preferably, the data acquisition module includes: The full matrix data acquisition unit is used to control the probes in the ring array probe group to emit ultrasonic waves sequentially and be received synchronously by all probes to obtain full matrix capture data. The signal preprocessing unit, connected to the full matrix data acquisition unit, is used to perform time base correction and channel equalization processing on the full matrix captured data to generate time-domain aligned and amplitude-normalized raw echo data.

[0008] Preferably, the feature extraction module includes: The time-domain waveform feature extraction unit, connected to the signal preprocessing unit, is used to perform time-domain analysis on the raw echo data, extract the echo peak value and pulse width, and form time-domain waveform features. The frequency domain spectral feature extraction unit, connected to the signal preprocessing unit, is used to perform spectral analysis on the raw echo data, extract the main frequency and bandwidth, and form frequency domain spectral features. The spatial domain guidance feature extraction unit, connected to the signal preprocessing unit, is used to extract the defect azimuth and contour information based on the spatial position and echo arrival time information of each probe in the ring array probe group, and to form spatial domain guidance features. The feature fusion unit, connected to the time-domain waveform feature extraction unit, the frequency-domain spectrum feature extraction unit, and the spatial-domain guided feature extraction unit, is used to splice time-domain waveform features, frequency-domain spectrum features, and spatial-domain guided features to form an initial multidimensional feature set.

[0009] Preferably, the feature space transformation module includes: The parameter generation unit, connected to the feature fusion unit, is used to analyze the initial multidimensional feature set through the parameter generation subnetwork in the dynamic feature space transformation network and output adaptive feature space transformation parameters. The projection matrix construction unit, connected to the parameter generation unit, is used to dynamically construct the spatial projection matrix based on the adaptive feature space transformation parameters. The spatial transformation unit, connected to the projection matrix construction unit, is used to perform matrix transformation operations on the initial multidimensional feature set using the spatial projection matrix to obtain the depth feature vector.

[0010] Preferably, the collaborative decision-making module includes: The defect type identification unit is connected to the spatial transformation unit and the defect size quantification unit respectively. It is used to generate a first type of attention weight map to characterize the contribution of classification decision based on the activation value of its internal features. The first type of attention weight map is passed to the defect size quantification unit to guide it to focus on the features related to the defect type in the deep feature vector for size calculation. The defect size quantification unit is connected to the spatial transformation unit and the defect type identification unit respectively. It is used to generate a second type of attention weight map to characterize size sensitivity based on the gradient information of its regression loss. The second type of attention weight map is fed back to the defect type identification unit to enhance its ability to identify size-sensitive features in the depth feature vector. The attention interaction unit is connected to the defect type identification unit and the defect size quantification unit respectively. It is used to manage the bidirectional transmission and weighted fusion of the first type of attention weight map and the second type of attention weight map between the defect type identification unit and the defect size quantification unit to generate collaborative attention features. The decision fusion unit, connected to the attention interaction unit, receives the defect type identification result from the defect type identification unit and the defect size quantification result from the defect size quantification unit. Based on the collaborative attention feature, it performs consistency verification and weighted fusion on the two types of results to generate the final collaborative decision result.

[0011] Preferably, the dynamic feature space transformation network is trained in the following way: An unlabeled dataset was constructed by acquiring ultrasonic data containing only normal weld samples. Data augmentation is performed on samples in the unlabeled dataset to generate positive and negative sample pairs for contrastive learning. The dynamic feature space transformation network is trained by self-supervised contrastive learning. The training aims to bring positive sample pairs closer together and push negative sample pairs further apart in the feature space, thus mapping all normal weld features into compact feature clusters.

[0012] Preferably, after feeding the second type of attention weight map back to the defect type identification unit, the defect type identification unit uses the second type of attention weight map to recalibrate and generate calibrated classification features; The defect size quantification unit uses the first type of attention weight map for weighted calculation to generate a weighted regression loss; The classification loss is calculated based on the calibrated classification features and combined with the weighted regression loss to form a joint loss function. The synergistic effect of the defect type identification unit and the defect size quantification unit is optimized and solidified by optimizing the joint loss function.

[0013] Preferably, in the collaborative decision-making module, after generating the collaborative decision-making result, a consistency judgment is made between the defect type identification result and the defect size quantitative result in the collaborative decision-making result; When it is determined that there is a logical contradiction between the defect type identification result and the physical form inferred from the quantitative result of defect size, a decision consistency verification signal is generated. Based on the decision consistency verification signal, the transformation strategy in the dynamic feature space transformation network is adjusted, and the adjusted transformation strategy is used to re-perform spatial reconstruction of the initial multidimensional feature set.

[0014] Preferably, the system also includes: During continuous detection, the distribution of deep feature vectors in the feature space is monitored, and new clusters that are distinct from known feature clusters are identified. When the preset collaborative decision-making module has uncertainty in the output results of samples within a new cluster, the new cluster is marked as a mode to be confirmed. Upload the raw echo data associated with the mode to be confirmed to obtain feedback tags, and use the feedback tags to fine-tune the dynamic feature space transformation network or collaborative decision-making module online.

[0015] The present invention has the following advantages: This invention utilizes a dynamic feature space transformation network to adaptively reconstruct the feature space based on the content of the input raw echo data. This approach surpasses traditional methods that rely on fixed, manually designed features, automatically learning and generating deep feature representations optimized for defect classification and quantification tasks, thereby improving the ability to identify and detect complex, weak, or atypical defect signals.

[0016] The collaborative decision-making module designed in this invention processes the two tasks of defect type identification and defect size quantification in parallel, and uses an attention interaction mechanism to guide and constrain the two tasks. This ensures that the final output decision results have a high degree of internal logical consistency in terms of type and size, overcomes the problem of contradictory qualitative and quantitative conclusions that may occur in traditional separate processing flows, and enhances the reliability of the detection results.

[0017] This invention integrates online learning and adaptive model evolution capabilities, enabling it to autonomously discover new clusters in the feature space during continuous detection, i.e., potential unknown defect patterns. After obtaining feedback labels through human-computer interaction, the system can fine-tune the model online, giving it the ability to continuously learn and improve itself. This allows it to adapt to constantly changing working conditions and defect morphologies, ensuring the system's advanced nature and robustness in long-term applications. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the modules of the system of the present invention; Figure 2 This is a detailed architectural diagram of the system of the present invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0020] like Figure 1 As shown, the ultrasonic testing system for pipe weld defects includes a data acquisition module, a feature extraction module, a feature space transformation module, and a collaborative decision-making module connected in sequence, wherein: The data acquisition module is used to acquire the raw echo data generated by multi-angle ultrasonic scanning of pipe welds using a ring array probe group; The data acquisition module includes a full matrix data acquisition unit and a signal preprocessing unit, wherein: The full matrix data acquisition unit is used to control the probes in the ring array probe group to emit ultrasonic waves sequentially and be received synchronously by all probes to obtain full matrix capture data. The signal preprocessing unit, connected to the full matrix data acquisition unit, is used to perform time base correction and channel equalization processing on the full matrix captured data to generate time-domain aligned and amplitude-normalized raw echo data.

[0021] Data acquisition is performed by a full-matrix data acquisition unit, utilizing a ring array probe group to achieve comprehensive, multi-angle acoustic illumination and information reception of the weld area. The workflow of the full-matrix data acquisition unit strictly follows preset timing control logic. Within one detection cycle, the unit first designates the first probe element in the ring array probe group as the transmission source. This probe element is excited and emits a narrowband ultrasonic pulse with a center frequency between 2.5 MHz and 10 MHz. At the instant of this pulse emission, all N probe elements in the array, including the transmitting element, synchronously switch to receiving mode and continuously record the echo signal at a sampling rate of 50 MHz to 100 MHz within a preset time window, such as 20 microseconds. The length of this time window is precisely calculated based on the pipe diameter, wall thickness, and material sound velocity to ensure complete reception of reflected signals from the weld interior and the far-end boundary. After one reception cycle, the N A-scan signals are stored as a single data packet.

[0022] Subsequently, the full matrix data acquisition unit automatically switches the transmission task to the second probe element and repeats the above transmission and full reception process. This process is repeated until all N probe elements have completed one excitation as transmission sources, thereby obtaining a complete dataset consisting of N×N A-scan signals, i.e., full matrix acquisition data. This data can be represented as a three-dimensional matrix with dimensions N×N×T, where N is the number of probe elements and T is the number of time sampling points in a single A-scan.

[0023] The full-matrix acquisition data is immediately fed into the signal preprocessing unit to eliminate systematic errors introduced by physical hardware differences and propagation paths, generating raw echo data with precise time-domain alignment and consistent amplitude response. This unit's execution involves two key processing algorithms. The first is time-base correction, designed to calibrate time delay inconsistencies between channels, typically caused by minor differences in probe wedges, cable length variations, or transit time differences in electronics. The correction algorithm is first calibrated using a standard test block with known geometric characteristics, measuring each transmit-receive pair... The actual arrival time of the surface wave or bottom echo is used to calculate the time delay correction for each channel by comparing it with the theoretical or reference time. This correction is applied to actual detection data, and its mathematical expression is: ; In the formula, For channel The original A-scan signal, This is the time correction amount for this channel. This is the signal after time base correction. It is obtained by subtracting the arrival time from the reference time measured during the calibration phase.

[0024] The second is channel equalization, designed to compensate for differences in sensitivity and amplification among various probe elements and receiving channels. This algorithm also requires calibration data, such as recording the peak echo amplitudes of all channels to the same stable reflection source. These amplitude values ​​are used to calculate the gain equalization factor for each channel, which is used to bring the responses of all channels to a uniform reference, ultimately achieving amplitude normalization. This typically involves scaling the amplitude value of each time-base corrected A-scan signal to the interval [-1, 1], mathematically expressed as: ; In the formula, This is the signal after time-base correction. It is the maximum absolute amplitude of the signal within the entire time window. This is the signal that is finally normalized.

[0025] One application example is the inspection of a 12mm thick pipe weld. The full-matrix data acquisition unit of the data acquisition module first activates the cyclic excitation mode of the ring array probe group. When the fifth probe element in the array emits an ultrasonic pulse, the entire array probe synchronously receives the echo signal and records it as full-matrix capture data. The signal preprocessing unit then performs processing on specific transmit channel i (corresponding to the i-th transmit probe) and receive channel j (corresponding to the j-th receive probe). This is based on the time delay correction amount calibrated on the standard test block. The time is 0.25 microseconds, and the unit applies the formula. Perform time base correction. If the sampling time t in the current analysis is 12.00 microseconds, then the corrected signal value... Value taken from the raw data at 12.25 microseconds. .

[0026] After time alignment is completed, the signal preprocessing unit identifies the maximum absolute amplitude of the channel within the entire time window. The value is 1.5. This is the signal value after time-base correction. It is 1.2, calculated The normalized value obtained is 0.8. This process eliminates systematic errors caused by hardware differences and finally generates raw echo data that is time-aligned and has an amplitude in the range of -1 to 1, providing standardized input for the feature extraction module.

[0027] The feature extraction module is used to process the raw echo data in parallel, extracting time-domain waveform features, frequency-domain spectral features, and spatial-domain steering features to form an initial multi-dimensional feature set. The feature extraction module includes a time-domain waveform feature extraction unit, a frequency-domain spectrum feature extraction unit, a spatial-domain guided feature extraction unit, and a feature fusion unit, wherein: The time-domain waveform feature extraction unit, connected to the signal preprocessing unit, is used to perform time-domain analysis on the raw echo data, extract the echo peak value and pulse width, and form time-domain waveform features. The frequency domain spectral feature extraction unit, connected to the signal preprocessing unit, is used to perform spectral analysis on the raw echo data, extract the main frequency and bandwidth, and form frequency domain spectral features. The spatial domain guidance feature extraction unit, connected to the signal preprocessing unit, is used to extract the defect azimuth and contour information based on the spatial position and echo arrival time information of each probe in the ring array probe group, and to form spatial domain guidance features. The feature fusion unit, connected to the time-domain waveform feature extraction unit, the frequency-domain spectrum feature extraction unit, and the spatial-domain guided feature extraction unit, is used to splice time-domain waveform features, frequency-domain spectrum features, and spatial-domain guided features to form an initial multidimensional feature set.

[0028] The time-domain waveform feature extraction unit receives the raw echo data and quantifies the energy and morphological characteristics of the defect echo signal on the time axis. For each filtered valid echo signal, i.e., an A-scan signal whose amplitude exceeds a specific noise threshold within a preset time window (e.g., 10% of the signal peak value), the unit performs the following calculations: First, it extracts the echo peak value, i.e., the maximum absolute amplitude value found within the time window of the A-scan signal. Then, it extracts the pulse width, a parameter characterizing the duration of the echo signal, often defined as the time difference between two time points corresponding to the signal envelope dropping to -6 dB from the peak value. These two parameters, the echo peak value and the pulse width, together constitute the time-domain waveform features, directly reflecting preliminary information about the defect's reflectivity and axial dimensions.

[0029] The frequency domain spectral feature extraction unit also processes the raw echo data to reveal the differences in the response of defects to ultrasonic waves of different frequencies. This is usually closely related to the geometry and properties of the defects, such as smooth cracks or rough pores. This unit first applies a Fast Fourier Transform to the selected A-scan signal segment, transforming it from the time domain to the frequency domain to obtain a spectrum. Based on this spectrum, two core features are extracted: the dominant frequency, which corresponds to the frequency at which the spectral energy reaches its maximum value; and the bandwidth, defined as the frequency range covered by a specific decibel decrease in spectral energy from its peak, typically -6 dB. The shift in dominant frequency and the change in bandwidth are key indicators for identifying the scattering characteristics of different types of defects, together constituting the frequency domain spectral features.

[0030] The spatial domain guidance feature extraction unit processes the data, utilizing the spatial layout information of the array probes to achieve preliminary defect localization and contour mapping. This unit executes a synthetic aperture focusing algorithm based on the precise physical coordinates of each probe element in the ring array probe group and the echo arrival time information of each channel extracted from the raw echo data. Its core step is to calculate the theoretical sound path for each spatial point within the detection area and coherently superimpose the full matrix capture data along this sound path. Its mathematical expression is: ; In the formula, It is the image intensity of spatial point p. It represents the signal amplitude of the i-th transmitting channel and the j-th receiving channel. This is the actual arrival time of the echo from that channel. This is the theoretical flight time of the ultrasonic wave from the transmitting probe i to the spatial point p and then to the receiving probe j. f is a weighting function used to enhance coherence. The basic form of the weighting function is as follows: the weight is 1 when the time difference is within a preset tolerance range, and 0 otherwise: ; in The time tolerance is determined by the pulse width of the ultrasonic signal and the required detection accuracy. A two-dimensional ultrasonic C-scan image is generated by pixelating the entire weld cross-section and performing the above calculations. From this image, image processing techniques, such as edge detection and centroid calculation, can be used to extract the defect azimuth and contour information. The defect azimuth refers to the angle of the defect centroid relative to the pipe coordinate system, while the contour information can be parameterized as descriptors such as the defect's length, height, or shape factor. These together constitute the spatial domain guidance features.

[0031] The feature fusion unit performs the final integration step, combining the heterogeneous features extracted by the three parallel units into a single feature vector. This unit concatenates the extracted time-domain waveform features (echo peak value and pulse width), frequency-domain spectral features (dominant frequency and bandwidth), and spatial-domain steering features (defect azimuth angle and contour information parameters) in a predetermined order. This operation forms a fixed-length numerical vector, the initial multidimensional feature set. Each dimension of this vector corresponds to a specific physical measurement, providing a comprehensive, multi-faceted description of the defect for subsequent modules.

[0032] In one application example, the feature extraction module receives the preprocessed raw echo data and initiates a parallel processing task. The time-domain waveform feature extraction unit first quantizes the A-scan signal, setting the maximum absolute amplitude of the echo signal to 0.85. The two time points where the signal envelope drops to -6 dB from the peak are 12.2 microseconds and 12.8 microseconds, respectively. Calculations show the echo peak value to be 0.85 and the pulse width to be 0.6 microseconds; these two parameters constitute the time-domain waveform feature. The frequency-domain spectrum feature extraction unit simultaneously applies a Fast Fourier Transform to the signal segment to obtain the spectrum. The frequency corresponding to the maximum spectral energy is 5.0 MHz, and the frequency range covered by the spectral energy dropping -6 dB from the peak is from 3.5 MHz to 6.5 MHz. Based on this, the main frequency of 5.0 MHz and the bandwidth of 3.0 MHz are extracted; these two constitute the frequency-domain spectrum feature.

[0033] The spatial domain guidance feature extraction unit performs imaging based on the probe's spatial position and echo arrival time information. For the weld spatial point with coordinate p, the signal amplitude composed of the first transmitting channel and the second receiving channel is set. The actual arrival time of the echo is 0.7. The calculated theoretical flight time is 15.0 microseconds. With a time interval of 15.0 microseconds and a weighting function f of 1, the image intensity at that point is calculated using the formula. The value is 0.7. Subsequent image processing extracts a defect azimuth angle of 45 degrees and a contour length of 2.5 mm, thus forming a spatial domain guiding feature. Finally, the feature fusion unit performs an integration step, concatenating the extracted echo peak value (0.85), pulse width (0.6), main frequency (5.0), bandwidth (3.0), defect azimuth angle (45 degrees), and contour information (2.5 mm) in a predetermined order to generate a fixed-length numerical vector, i.e., the initial multidimensional feature set.

[0034] The feature space transformation module is used to input the initial multidimensional feature set into a preset dynamic feature space transformation network. The dynamic feature space transformation network generates adaptive feature space transformation parameters based on the content of the initial multidimensional feature set, and performs spatial reconstruction on the initial multidimensional feature set based on the adaptive feature space transformation parameters to generate a deep feature vector. The feature space transformation module includes a parameter generation unit, a projection matrix construction unit, and a spatial transformation unit, wherein: The parameter generation unit, connected to the feature fusion unit, is used to analyze the initial multidimensional feature set through the parameter generation subnetwork in the dynamic feature space transformation network and output adaptive feature space transformation parameters. The projection matrix construction unit, connected to the parameter generation unit, is used to dynamically construct the spatial projection matrix based on the adaptive feature space transformation parameters. The spatial transformation unit, connected to the projection matrix construction unit, is used to perform matrix transformation operations on the initial multidimensional feature set using the spatial projection matrix to obtain the depth feature vector.

[0035] The parameter generation unit receives an initial multidimensional feature set from the feature fusion unit and dynamically generates a set of transformation rules best suited to each input defect sample. At the core of this unit is a parameter generation subnetwork within a pre-trained dynamic feature space transformation network. This subnetwork is typically a shallow, fully connected neural network, containing, for example, 2 to 4 hidden layers with 32 to 128 neurons per layer. When an initial multidimensional feature set is fed into this subnetwork, it undergoes forward propagation through a series of non-linear activation functions, outputting a fixed-length vector, which represents the adaptive feature space transformation parameters. These parameters are not directly used for the transformation but rather encode all the information needed to construct the transformation matrix, such as rotation angles, scaling factors, and shearing coefficients.

[0036] The projection matrix construction unit receives adaptive feature space transformation parameters output by the parameter generation unit, and concretizes these abstract transformation parameters into a mathematical entity capable of performing matrix operations, namely, the spatial projection matrix. This unit parses the input adaptive feature space transformation parameters according to preset construction rules and uses these parameters to fill the elements of a matrix. For example, some parameters may be used to construct a rotation matrix, while others may be used to construct a scaling matrix; the final spatial projection matrix is ​​then composed of these fundamental matrices. This dynamic construction method ensures that the transformation for each input sample is customized, thereby achieving content-adaptive feature space reconstruction.

[0037] The spatial transformation unit performs the final mapping operation, applying a dynamically constructed spatial projection matrix to linearly transform the initial multidimensional feature set from its original space to a new depth feature space. This unit receives the initial multidimensional feature set and the constructed spatial projection matrix, and performs a matrix multiplication operation, mathematically expressed as: ; In the formula, The initial multidimensional feature set vector represents the input, and its dimension is determined by the number of features in the feature extraction stage. The spatial projection matrix generated by the projection matrix construction unit based on adaptive parameters determines the dimension of the output feature vector. It is the depth feature vector obtained after the operation. Obtained through feature fusion unit It is generated collaboratively by the parameter generation unit and the projection matrix construction unit. The essence of this operation is to project the original feature vectors onto a new coordinate system, which is optimized to better separate defects of different categories.

[0038] Through the above steps, the final output of the feature space transformation module is a deep feature vector. Compared with the initial multidimensional feature set, the dimension of the deep feature vector may be the same, higher, or lower, but its internal structure has been fundamentally reshaped. Through adaptive spatial reconstruction, the complex nonlinear relationships between the original features are decoupled, key features that help distinguish different defect types are enhanced, while redundant information or noise unrelated to the essence of the defect is effectively suppressed. In this new feature space, the distance between feature clusters of different defect categories is increased, while features within the same category become more compact.

[0039] One application example is that the feature space transformation module obtains the initial multidimensional feature set vector generated by the feature fusion unit. Then, spatial transformation is performed. Assuming this is for a specific weld porosity sample, the initial multidimensional feature set extracted... Given a column vector with 6 elements, its transpose is represented as: The parameter generation unit inputs the vector into the parameter generation subnetwork within the preset dynamic feature space transformation network. Through nonlinear mapping of multiple fully connected neurons, it outputs a set of adaptive feature space transformation parameters. The projection matrix construction unit then parses and dynamically fills this set of parameters to generate a 2x6 spatial projection matrix. In this verification example, let the constructed spatial projection matrix be... The first row of elements is The second row of elements is The spatial transformation unit performs operations using matrix transformation formulas.

[0040] Deep feature vector The first dimension is obtained by the dot product of the first row of the matrix and the eigenvector, i.e. Deep feature vectors The second dimension is obtained by the dot product of the second row of the matrix and the eigenvector, i.e. The final deep feature vector obtained After transpose, it becomes This process enables the reconstruction from the physical feature space to the deep feature space. By using an adaptive projection matrix, the original features are decoupled and projected onto a more discriminative coordinate system, providing high-purity feature input for subsequent collaborative decision-making.

[0041] The dynamic feature space transformation network is trained in the following way: An unlabeled dataset was constructed by acquiring ultrasonic data containing only normal weld samples. Data augmentation is performed on samples in the unlabeled dataset to generate positive and negative sample pairs for contrastive learning. The dynamic feature space transformation network is trained by self-supervised contrastive learning. The training aims to bring positive sample pairs closer together and push negative sample pairs further apart in the feature space, thus mapping all normal weld features into compact feature clusters.

[0042] To obtain unlabeled datasets and provide the foundation for self-supervised learning, a large amount of ultrasonic data needs to be collected from certified, defect-free pipe weld segments. This data is acquired through a data acquisition module, forming a massive raw echo database containing only normal weld samples without any manual annotation.

[0043] Data augmentation generates contrastive learning sample pairs, creating multiple perspectives for a single data sample, allowing the model to learn which changes are insignificant and which are essential. For each raw echo data sample randomly drawn from the unlabeled dataset, a series of pre-defined data augmentation operators are applied to generate one or more transformed samples. These operators simulate subtle perturbations that may occur in real-world detection; for example, adding Gaussian white noise with a signal-to-noise ratio in the range of 20 to 40 dB to simulate electronic noise, performing a small time shift on the signal to simulate slight probe position deviations, or slightly scaling its amplitude to simulate changes in coupling agent thickness. The original sample and its augmented sample form a positive sample pair because these samples essentially originate from the same normal weld area. The original sample and any other unrelated sample in the dataset form a negative sample pair.

[0044] A dynamic feature space transformation network (VLS) is trained using self-supervised contrastive learning to optimize its internal parameters. This enables the VLS to map similar inputs to nearby feature space locations and dissimilar inputs to more distant locations. During training, positive and negative sample pairs are fed into the VLS in batches. The VLS outputs a feature vector for each sample. The optimization objective is to maximize the similarity of positive sample pairs in the feature space while minimizing the similarity of negative sample pairs. This is typically achieved by minimizing a contrastive loss function, such as the InfoNCE loss, mathematically expressed as: ; In the formula, L is the loss value of a single sample. It is the cosine similarity between two feature vectors in a positive sample pair, calculated by the dynamic feature space transformation network. This is the cosine similarity between the feature vector of the sample and the feature vector of the k-th negative sample. T1 is a temperature-related hyperparameter, typically ranging from 0.05 to 0.2, used to adjust the sensitivity of the loss function to similarity differences; the smaller the value, the more the model is driven to learn more discriminative features. The network iterates repeatedly using optimization algorithms such as gradient descent to continuously minimize the average loss of all samples until the dynamic feature space transformation network converges.

[0045] One application example is the use of self-supervised contrastive learning training on a dynamic feature space transformation network. This involves constructing an unlabeled dataset using normal weld samples and generating contrastive sample pairs for validation. The system first extracts a segment of certified defect-free ultrasonic data from the original echo database to form an unlabeled dataset. For one original sample in this dataset, a data augmentation operator is applied: 30 dB of Gaussian white noise is added to generate augmented sample one, and a small time shift is performed to generate augmented sample two. These two samples, originating from the same region, constitute a positive sample pair. Simultaneously, another unrelated normal sample from the dataset is randomly selected as a negative sample.

[0046] The cosine similarity of positive sample pairs is obtained through forward propagation calculation using a dynamic feature space transformation network. The cosine similarity between this sample and the negative sample is 0.9928. The value is 0.736. The temperature hyperparameter T1 is set to 0.1, and the optimization calculation is performed using the defined contrast loss function formula. The numerator term is calculated. Calculate the negative sample exponent term in the denominator. The sum of the denominators, when added together, yields 22067.96. Substituting this into the formula... The score within the logarithm is approximately 0.9288. The final loss value L for this sample is approximately 0.0739. By iteratively minimizing this loss function, the dynamic feature space transformation network achieves the goal of bringing positive sample pairs closer together and pushing negative sample pairs further apart, mapping all normal weld features into a compact feature cluster.

[0047] The collaborative decision-making module is used to simultaneously input deep feature vectors into a preset defect type identification unit and a defect size quantification unit to generate collaborative decision-making results, which include defect type identification results and defect size quantification results.

[0048] The collaborative decision-making module includes a defect type identification unit, a defect size quantification unit, an attention interaction unit, and a decision fusion unit, wherein: The defect type identification unit is connected to the spatial transformation unit and the defect size quantification unit respectively. It is used to generate a first type of attention weight map to characterize the contribution of classification decision based on the activation value of its internal features. The first type of attention weight map is passed to the defect size quantification unit to guide it to focus on the features related to the defect type in the deep feature vector for size calculation. The defect size quantification unit is connected to the spatial transformation unit and the defect type identification unit respectively. It is used to generate a second type of attention weight map to characterize size sensitivity based on the gradient information of its regression loss. The second type of attention weight map is fed back to the defect type identification unit to enhance its ability to identify size-sensitive features in the depth feature vector. The attention interaction unit is connected to the defect type identification unit and the defect size quantification unit respectively. It is used to manage the bidirectional transmission and weighted fusion of the first type of attention weight map and the second type of attention weight map between the defect type identification unit and the defect size quantification unit to generate collaborative attention features. The decision fusion unit, connected to the attention interaction unit, receives the defect type identification result from the defect type identification unit and the defect size quantification result from the defect size quantification unit. Based on the collaborative attention feature, it performs consistency verification and weighted fusion on the two types of results to generate the final collaborative decision result.

[0049] The defect type identification unit and the defect size quantification unit start in parallel, both receiving the same deep feature vector from the feature space transformation module as input. The defect type identification unit, typically a classification sub-network consisting of several fully connected layers and a Softmax activation layer, outputs the probability distribution of the defect's category and generates a first-class attention weight map. After one forward propagation, this unit calculates the importance of each dimension of the deep feature vector to the current classification result based on the activation values ​​of the internal features that contribute most to the final classification decision, using a technique similar to gradient-weighted class activation mapping. This importance distribution map is the first-class attention weight map, indicating which features are key to distinguishing the current defect type. This weight map is immediately passed to the attention interaction unit.

[0050] The defect size quantification unit, typically a regression subnetwork composed of fully connected layers, predicts the specific geometric dimensions of the defect, such as length and height, and generates a second-type attention weight map. After outputting initial size predictions, this unit calculates the regression loss between the predicted and true values, such as mean squared error. Subsequently, it calculates the gradient of this loss function with respect to each element in the deep feature vector through backpropagation. The absolute magnitude of these gradient values ​​reflects the sensitivity of each feature to size prediction; a larger gradient indicates a greater influence of that feature on size calculation. This gradient map is the second-type attention weight map, which labels features in the deep feature vector that are related to size and height. This weight map is also sent to the attention interaction unit.

[0051] The attention interaction unit performs bidirectional information exchange and fusion. It manages the information flow between the defect type identification unit and the defect size quantification unit, and generates a collaborative attention feature that integrates both type and size considerations. First, this unit passes the first type of attention weight map received from the defect type identification unit to the defect size quantification unit. The defect size quantification unit uses this weight map to weight its internal features, focusing on features meaningful to the current defect type for size calculation, avoiding interference from irrelevant features. Simultaneously, this unit feeds back the second type of attention weight map received from the defect size quantification unit to the defect type identification unit, making it pay more attention to size-sensitive features during classification, improving the ability to distinguish between defects of different sizes but similar types. Based on this, the unit weights and fuses the two attention weight maps to generate the collaborative attention feature, mathematically expressed as: ; In the formula, It is a collaborative attention feature vector. It is the first type of attention weight map generated by the defect type identification unit. This is a second-class attention weight map generated by the defect size quantification unit. a and b are learnable fusion weight coefficients that are adaptively adjusted during training to balance the importance of type and size information.

[0052] The decision fusion unit generates the final collaborative decision result. It performs consistency verification and final determination on the independent outputs of the defect type identification unit and the defect size quantification unit. This unit receives preliminary type identification results from the defect type identification unit and preliminary size quantification results from the defect size quantification unit. First, it performs a logical consistency verification. For example, if the type identification result is "pore," its physical morphology is usually near-spherical, but the size quantification result shows a slender shape with a length much greater than its height, then it is judged as a logical contradiction. Subsequently, this unit uses the collaborative attention features generated by the attention interaction unit to perform weighted fusion and correction on the two preliminary results. In cases of logical contradictions, it can selectively believe the result of one unit based on the confidence level indicated by the collaborative attention features, or mark the sample as having low confidence, and output the final calibrated defect type identification result and defect size quantification result, which together constitute the collaborative decision result.

[0053] One application example is the use of spatially transformed depth feature vectors. for During analysis, the collaborative decision-making module simultaneously activates the defect type identification unit and the defect size quantification unit for processing. The defect type identification unit performs forward propagation on the vector, calculating the first-class attention weight map to characterize the contribution of the classification decision based on the activation values ​​of its internal features. for Simultaneously, the defect size quantification unit outputs preliminary size predictions and calculates the regression loss. Through backpropagation, it obtains the absolute values ​​of the gradients of this loss with respect to each feature dimension, generating a second type of attention weight map to characterize size sensitivity. for .

[0054] Subsequently, the attention interaction unit receives these two weight maps and performs bidirectional propagation, setting the fusion weight coefficients a to 0.6 and b to 0.4, and applying the formula for weighted fusion. Collaborative attention feature vector. The first dimension value equals The second dimension value equals The final generated collaborative attention features for Finally, the decision fusion unit receives the defect type identification result as pores with a probability of 0.95, and the defect size quantitative result as a length of 2.2 mm and a height of 2.0 mm. This unit performs consistency verification based on a built-in physical rule logic library, calculating the aspect ratio... The defect is determined to conform to the physical morphology definition of a near-spherical pore-type defect. Therefore, the two types of results are calibrated by combining the collaborative attention feature and the final collaborative decision result is output.

[0055] After feeding the second type of attention weight map back to the defect type identification unit, the defect type identification unit uses the second type of attention weight map to recalibrate and generate calibrated classification features. The defect size quantification unit uses the first type of attention weight map for weighted calculation to generate a weighted regression loss; The classification loss is calculated based on the calibrated classification features and combined with the weighted regression loss to form a joint loss function. The synergistic effect of the defect type identification unit and the defect size quantification unit is optimized and solidified by optimizing the joint loss function.

[0056] Feature recalibration and loss weighting calculation directly reflect the guiding role of the attention interaction mechanism in the calculation of the loss function, making the optimization process explicitly oriented towards the collaborative goal. When the second type of attention weight map, i.e., the information representing size sensitivity, is fed back from the defect size quantification unit to the defect type identification unit, the defect type identification unit does not directly use the original features to calculate the classification probability. Instead, it first uses the weight map to perform an element-wise multiplication operation on its internal feature map. This is equivalent to a feature channel recalibration, which enhances the attention to size-sensitive features, suppresses the contribution of irrelevant features, and generates calibrated classification features.

[0057] When the first type of attention weight map, which represents the contribution of classification decisions, is passed to the defect size quantification unit, the original regression loss calculated by the unit, such as the mean squared error, is weighted by this attention weight map. Features with higher weights receive a larger penalty in their loss terms, thus generating a weighted regression loss. This operation forces the defect size quantification unit to prioritize features crucial for correct classification during optimization.

[0058] A joint loss function is constructed, fusing the optimization objectives of the defect type identification unit and the defect size quantification unit into a single scalar value that the optimizer can process. The classification loss calculated based on the calibrated classification features, typically the cross-entropy loss, is linearly combined with the weighted regression loss generated by the defect size quantification unit through a balancing coefficient to form the joint loss function. Its mathematical expression is: ; In the formula, It is the total value of the joint loss function. The classification loss is calculated based on the calibrated classification features. It is the weighted regression loss. This is a hyperparameter between 0 and 1, used to balance the importance of classification and regression tasks in the overall optimization process. Its value is usually set empirically, such as 0.5, or obtained through experimental grid search. Joint optimization and collaborative solidification, by minimizing the joint loss function, synchronously update all learnable parameters of the collaborative decision-making module and previous modules, thereby solidifying the collaborative working pattern into the weights of the collaborative decision-making module.

[0059] In each iteration of training, the following is calculated: The gradients of all parameters of the collaborative decision-making module are calculated, and the parameters are updated along the direction of gradient descent using an optimizer such as Adam or SGD. Because It simultaneously incorporates requirements for classification accuracy and size quantification precision, and both of these losses are modulated with attention information from the other. Therefore, the optimization process is essentially searching for an optimal solution that simultaneously satisfies both collaborative tasks. This process continuously strengthens the positive coupling between the defect type identification unit and the defect size quantification unit.

[0060] One application example is that, during the joint training phase of the system, the feature recalibration and loss weighting calculation unit performs parameter optimization for specific defect samples. The defect type identification unit first receives a second type of attention weight map fed back by the defect size quantification unit. The weights are 0.3 and 0.7, respectively. This unit uses this weight map to perform element-wise multiplication recalibration on the original internal classification features 1.72 and 2.65, calculating the first dimension value of the calibrated classification features as equal to... The second dimension value equals Meanwhile, the defect size quantification unit receives the first type of attention weight map from the defect type identification unit. Furthermore, their weight values ​​are 0.8 and 0.2, respectively. Let the original regression loss for length prediction be 0.1 and the original regression loss for height prediction be 0.05. The weighted regression loss is then calculated by weighting the values ​​using attention weights. The system then calculates the classification loss based on the calibrated classification features. Set the hyperparameter to 0.25. This is used to balance the two tasks. The value is 0.5. Substitute this value into the defined joint loss function formula to solve the problem. By synchronously updating the learnable parameters of the collaborative decision-making module through gradient descent, the guiding effect between the defect type identification unit and the defect size quantification unit is solidified into the network weights to achieve a joint improvement in detection consistency.

[0061] In the collaborative decision-making module, after generating the collaborative decision-making results, a consistency judgment is made between the defect type identification results and the defect size quantitative results in the collaborative decision-making results; When it is determined that there is a logical contradiction between the defect type identification result and the physical form inferred from the quantitative result of defect size, a decision consistency verification signal is generated. Based on the decision consistency verification signal, the transformation strategy in the dynamic feature space transformation network is adjusted, and the adjusted transformation strategy is used to re-perform spatial reconstruction of the initial multidimensional feature set.

[0062] A consistency check is performed by cross-validating the two results output by the collaborative decision-making module using existing knowledge of defect physical morphology from human experts. After the system outputs a collaborative decision result containing both defect type identification and defect size quantification, the decision-making module is immediately activated. It has a built-in logic library based on physical rules. For example, the rule library might define that if the defect type identification result is "spherical pore," then its ideal physical morphology should be a geometry with an aspect ratio close to 1. Based on the defect size quantification result, i.e., the predicted length and height, a morphological descriptor, such as the aspect ratio, is calculated. Then, this calculated aspect ratio is compared with the preset reasonable aspect ratio range for the "spherical pore" type in the logic library, for example, 0.8 to 1.2.

[0063] A decision consistency verification signal is generated, issuing a clear and actionable alarm to the system when a logical contradiction is detected. If, during consistency judgment, the calculated morphological descriptor exceeds a preset reasonable range—for example, the system identifies it as a "spherical pore" but calculates an aspect ratio of 3.5—this constitutes a clear logical contradiction. In this case, the system immediately determines that the current collaborative decision result has internal inconsistencies and generates a decision consistency verification signal. This signal is essentially a Boolean flag or a numerical value representing the severity of the contradiction, which will serve as a trigger to initiate subsequent adjustment procedures.

[0064] The transformation strategy is adjusted and spatial reconstruction is re-executed to immediately correct decision results that lead to logical contradictions, attempting to re-examine the original features from another perspective. When the feature space transformation module receives the decision consistency verification signal, it applies a temporary, minor perturbation to its internal transformation strategy. Specifically, this is done by the parameter generation subnetwork in the dynamic feature space transformation network, which adds a small, preset, or randomly generated perturbation vector to the original adaptive feature space transformation parameters. This perturbation aims to guide a slight change in the spatial projection matrix, thereby projecting the initial multidimensional feature set to a slightly different depth feature space region. Its mathematical expression is: ; In the formula, It is the original initial multidimensional feature set. It is the initially generated spatial projection matrix. It is the tiny adjustment matrix generated by this disturbance. It is the regenerated deep feature vector after adjustment. The generation of this deep feature vector is controlled by the decision consistency verification signal. Subsequently, this adjusted deep feature vector... The sample is then sent back to the collaborative decision-making module for a new round of defect identification and size quantification. A consistency check is performed again. If the new collaborative decision is logically consistent, it is adopted; otherwise, the process can be repeated a limited number of times, or the sample can be marked as "high uncertainty" for manual review.

[0065] One application example is that the system immediately initiates a consistency verification process after generating collaborative decision results. It performs cross-validation between the defect type identification result and the defect size quantification result using a built-in physical rule logic library. Suppose that in the collaborative decision results output by the system for a certain sample, the defect type identification unit determines the defect to be a spherical pore, while the defect size quantification unit outputs geometric dimensions of 3.5 mm in length and 1.0 mm in height. The system then calls the morphological descriptor calculation rules in the logic library to calculate the aspect ratio of the defect to be 3.5. Since the logic library presets a reasonable aspect ratio range for spherical pores to be between 0.8 and 1.2, the system determines that the currently calculated value of 3.5 significantly exceeds the preset range, thus identifying a clear logical contradiction between the defect type and physical morphology and generating a decision consistency verification signal.

[0066] After receiving the decision consistency verification signal, the feature space transformation module adds a preset perturbation vector to the original adaptive feature space transformation parameters through its internal parameter generation subnetwork, generating a small adjustment matrix. The spatial transformation unit then applies the adjusted transformation strategy, using the formula... Perform spatial reconstruction again on the initial multidimensional feature set. Let the original spatial projection matrix be... With adjustment matrix The sum of features has a mapping coefficient of 0.25 in a specific dimension, which is the initial multidimensional feature set. If the input value for this dimension is 0.8, then the adjusted deep feature vector is calculated. The corresponding dimension value is The system uses this perturbation mechanism to project features into slightly different deep feature space regions, and then feeds the regenerated deep feature vectors back into the collaborative decision-making module for secondary inference until a logically consistent final detection conclusion is output.

[0067] The system also includes: During continuous detection, the distribution of deep feature vectors in the feature space is monitored, and new clusters that are distinct from known feature clusters are identified. When the preset collaborative decision-making module has uncertainty in the output results of samples within a new cluster, the new cluster is marked as a mode to be confirmed. Upload the raw echo data associated with the mode to be confirmed to obtain feedback tags, and use the feedback tags to fine-tune the dynamic feature space transformation network or collaborative decision-making module online.

[0068] The system monitors feature distribution and identifies new clusters, discovering potential novel patterns in real-time and unsupervised. During continuous system detection, the generated deep feature vectors for each sample identified as an anomaly are cached. Online clustering algorithms, such as density-based DBSCAN or its variants, are applied to continuously monitor the distribution of these deep feature vectors in the feature space. The algorithm dynamically maintains known feature clusters corresponding to existing defect types in the training set. When a new batch of deep feature vectors forms a new cluster in the feature space that is far from any known feature cluster and whose density reaches a preset threshold, it is identified. For example, if the feature vectors of more than 10 consecutively collected samples have Euclidean distances between them that are less than a certain radius uncertainty threshold, but the distances of each feature vector to the center of any known feature cluster are greater than 5 times the uncertainty threshold, then this new cluster will be marked as a potential new cluster.

[0069] Uncertainty assessment is performed on new clusters to quantify the system's understanding of these newly discovered clusters and determine whether human intervention is necessary. Once a new cluster is identified, the deep feature vectors of all samples within that cluster are fed into a pre-defined collaborative decision-making module for inference. For each sample, the system evaluates the uncertainty of its output. For the defect type identification unit, uncertainty can be measured by the entropy of the probability distribution output by Softmax; a higher entropy value indicates greater uncertainty. For the defect size quantification unit, it can be evaluated by enabling Dropout during inference and performing multiple forward propagations to calculate the variance of the predicted size. When the entropy value of the type identification result for most samples in a cluster, such as more than 80% of the samples, is higher than a certain threshold, or the variance of the size prediction is greater than a certain range, the collaborative decision-making module is deemed to have a high degree of uncertainty regarding that new cluster.

[0070] The process of marking unconfirmed patterns and requesting feedback transforms the machine's uncertainty into a clear human annotation request. Once the system determines that there is high uncertainty regarding a new cluster, it marks the entire new cluster as an unconfirmed pattern. Simultaneously, it automatically packages all raw echo data associated with this cluster, intermediate features from the processing, and preliminary uncertainty decision results, and uploads them via network interface to a remote expert system or the human-machine interface of the on-site operator, along with a clear labeling request.

[0071] Online fine-tuning utilizes feedback labels to integrate newly acquired human-generated knowledge into the existing model, enabling incremental learning. Upon receiving expert-confirmed and labeled feedback labels—for example, confirming the pattern to be identified as a novel "hydrogen-induced crack" and providing its typical size range—the dynamic feature space transformation network or collaborative decision-making module is immediately fine-tuned online using this newly labeled data. Fine-tuning typically employs a small learning rate, such as 10 to 100 times smaller than the initial training rate, training the model only a few epochs. If the features of the new pattern differ significantly from old patterns, the dynamic feature space transformation network may be fine-tuned first to learn a new spatial mapping, opening up an independent region for the new pattern in the feature space. If the features of the new pattern are similar to some old patterns but the decision boundary needs adjustment, the top-level classifier and regressor of the collaborative decision-making module may be fine-tuned more significantly.

[0072] One application example is the continuous online inspection of welds in long-distance oil and gas pipelines. The system generates depth feature vectors in real time using a dynamic feature space transformation network and monitors the feature space using an online clustering algorithm. When the system collects depth feature vectors from 15 consecutive samples, and the Euclidean distance between these vectors is less than a preset radius uncertainty threshold of 0.5, but the distance from these vectors to the center of known defect feature clusters is greater than 5 times the uncertainty threshold (i.e., the distance exceeds 2.5), the system identifies this set of samples as a new cluster distinct from the known feature clusters.

[0073] Subsequently, the system initiates an uncertainty assessment, sending samples from the new cluster one by one into the collaborative decision-making module for inference. If the maximum probability value in the predicted probability distribution output by the defect type identification unit is only 0.4, failing to reach the preset confidence threshold of 0.8, the system determines that the output result of the new cluster has high uncertainty and marks it as a mode to be confirmed. At this time, the system automatically uploads the original echo data associated with the mode to be confirmed to the cloud expert system, obtaining the true feedback label corresponding to the new cluster as an unseen tungsten inclusion defect. Finally, the system uses this feedback label to fine-tune the collaborative decision-making module online, updating the network weights by optimizing the loss function that includes the new defect category, enabling the system to identify tungsten inclusion defects and achieving adaptive evolution of the model.

[0074] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections of lines. Any connection that achieves the purpose of this invention is applicable to the embodiments of this invention. The above descriptions are merely exemplary embodiments of this invention and should not be construed as limiting the scope of the invention.

Claims

1. An ultrasonic testing system for defects in pipe welds, characterized in that, Includes the following modules connected in sequence: The data acquisition module is used to acquire the raw echo data generated by multi-angle ultrasonic scanning of pipe welds using a ring array probe group; The feature extraction module is used to process the raw echo data in parallel, extracting time-domain waveform features, frequency-domain spectrum features, and spatial-domain steering features. These features are then combined by the feature fusion unit to form an initial multidimensional feature set. The spatial-domain steering features consist of defect azimuth angle and contour information. The feature space transformation module is used to input the initial multidimensional feature set into a preset dynamic feature space transformation network. The dynamic feature space transformation network generates adaptive feature space transformation parameters based on the content of the initial multidimensional feature set, and performs spatial reconstruction on the initial multidimensional feature set based on the adaptive feature space transformation parameters to generate a deep feature vector. The feature space transformation module includes: The parameter generation unit, connected to the feature fusion unit, is used to analyze the initial multidimensional feature set through the parameter generation subnetwork in the dynamic feature space transformation network and output adaptive feature space transformation parameters. The projection matrix construction unit, connected to the parameter generation unit, is used to dynamically construct the spatial projection matrix based on the adaptive feature space transformation parameters. The spatial transformation unit, connected to the projection matrix construction unit, is used to perform matrix transformation operations on the initial multidimensional feature set using the spatial projection matrix to obtain the depth feature vector. The dynamic feature space transformation network is trained in the following way: An unlabeled dataset was constructed by acquiring ultrasonic data containing only normal weld samples. Data augmentation is performed on samples in the unlabeled dataset to generate positive and negative sample pairs for contrastive learning. The dynamic feature space transformation network is trained by self-supervised contrastive learning. The training aims to bring positive sample pairs closer together and push negative sample pairs further apart in the feature space, thus mapping all normal weld features into compact feature clusters. The collaborative decision-making module is used to simultaneously input deep feature vectors into the preset defect type identification unit and defect size quantification unit to generate collaborative decision-making results, which include defect type identification results and defect size quantification results. The collaborative decision-making module includes: The defect type identification unit is connected to the spatial transformation unit and the defect size quantification unit respectively. It is used to generate a first type of attention weight map to characterize the contribution of classification decision based on the activation value of its internal features. The first type of attention weight map is passed to the defect size quantification unit to guide it to focus on the features related to the defect type in the deep feature vector for size calculation. The defect size quantification unit is connected to the spatial transformation unit and the defect type identification unit respectively. It is used to generate a second type of attention weight map to characterize size sensitivity based on the gradient information of its regression loss. The second type of attention weight map is fed back to the defect type identification unit to enhance its ability to identify size-sensitive features in the depth feature vector. The attention interaction unit is connected to the defect type identification unit and the defect size quantification unit respectively. It is used to manage the bidirectional transmission and weighted fusion of the first type of attention weight map and the second type of attention weight map between the defect type identification unit and the defect size quantification unit to generate collaborative attention features. The decision fusion unit, connected to the attention interaction unit, receives the defect type identification result from the defect type identification unit and the defect size quantification result from the defect size quantification unit. Based on the collaborative attention feature, it performs consistency verification and weighted fusion on the two types of results to generate the final collaborative decision result.

2. The ultrasonic testing system for pipe weld defects according to claim 1, characterized in that, The data acquisition module includes: The full matrix data acquisition unit is used to control the probes in the ring array probe group to emit ultrasonic waves sequentially and be received synchronously by all probes to obtain full matrix capture data. The signal preprocessing unit, connected to the full matrix data acquisition unit, is used to perform time base correction and channel equalization processing on the full matrix captured data to generate time-domain aligned and amplitude-normalized raw echo data.

3. The ultrasonic testing system for pipe weld defects according to claim 2, characterized in that, The feature extraction module includes: The time-domain waveform feature extraction unit, connected to the signal preprocessing unit, is used to perform time-domain analysis on the raw echo data, extract the echo peak value and pulse width, and form time-domain waveform features. The frequency domain spectral feature extraction unit, connected to the signal preprocessing unit, is used to perform spectral analysis on the raw echo data, extract the main frequency and bandwidth, and form frequency domain spectral features. The spatial domain guidance feature extraction unit, connected to the signal preprocessing unit, is used to extract the defect azimuth and contour information based on the spatial position and echo arrival time information of each probe in the ring array probe group, and to form spatial domain guidance features. The feature fusion unit, connected to the time-domain waveform feature extraction unit, the frequency-domain spectrum feature extraction unit, and the spatial-domain guided feature extraction unit, is used to splice time-domain waveform features, frequency-domain spectrum features, and spatial-domain guided features to form an initial multidimensional feature set.

4. The ultrasonic testing system for pipe weld defects according to claim 1, characterized in that, After feeding the second type of attention weight map back to the defect type identification unit, the defect type identification unit uses the second type of attention weight map to recalibrate and generate calibrated classification features. The defect size quantification unit uses the first type of attention weight map for weighted calculation to generate a weighted regression loss; The classification loss is calculated based on the calibrated classification features and combined with the weighted regression loss to form a joint loss function. The synergistic effect of the defect type identification unit and the defect size quantification unit is optimized and solidified by optimizing the joint loss function.

5. The ultrasonic testing system for pipe weld defects according to claim 1, characterized in that, In the collaborative decision-making module, after generating the collaborative decision-making results, a consistency judgment is made between the defect type identification results and the defect size quantitative results in the collaborative decision-making results; When it is determined that there is a logical contradiction between the defect type identification result and the physical form inferred from the quantitative result of defect size, a decision consistency verification signal is generated. Based on the decision consistency verification signal, the transformation strategy in the dynamic feature space transformation network is adjusted, and the adjusted transformation strategy is used to re-perform spatial reconstruction of the initial multidimensional feature set.

6. The ultrasonic testing system for pipe weld defects according to claim 1, characterized in that, The system also includes: During continuous detection, the distribution of deep feature vectors in the feature space is monitored, and new clusters that are distinct from known feature clusters are identified. When the preset collaborative decision-making module has uncertainty in the output results of samples within a new cluster, the new cluster is marked as a mode to be confirmed. Upload the raw echo data associated with the mode to be confirmed to obtain feedback tags, and use the feedback tags to fine-tune the dynamic feature space transformation network or collaborative decision-making module online.

Images