Intelligent detection method for ultrasonic phased array weld defect

Abstract

The application discloses an ultrasonic phased array weld defect intelligent detection method, which comprises the following steps: acquiring A-scan signals corresponding to a plurality of preset angles at the same detection position, and constructing a multi-angle A-scan signal matrix; performing short-time Fourier transform and logarithmic compression on the A-scan signals of each angle to obtain a two-dimensional time-frequency spectrum; inputting the time-frequency spectrum of all angles into a two-dimensional convolution network for feature extraction, obtaining a characteristic vector of each angle, and arranging the characteristic vector in an angle sequence to form a multi-angle characteristic sequence; inputting the multi-angle characteristic sequence into a Transformer encoder, learning the correlation between angles by using a self-attention mechanism, and performing feature weighted fusion, aggregating the fused characteristic sequence along the angle dimension to obtain a sample embedding vector, and classifying and identifying defects according to the sample embedding vector. Through time-frequency domain transformation of multi-angle A-scan signals, two-dimensional convolution feature extraction and Transformer self-attention fusion, the application realizes automatic and accurate identification of weld defects and determination of defect angle information, and improves the intelligent level and practicability of weld detection.

Description

Technical Field

[0001] This invention relates to an intelligent detection method for ultrasonic phased array weld defects, belonging to the field of artificial intelligence defect recognition technology. Background Technology

[0002] Specialized equipment is a fundamental component of national economic development, widely used in pillar industries such as petroleum, chemical, electric power, and energy. During the manufacturing and service of specialized equipment, welded joints are often weak points where stress concentration and microstructure evolution occur, and are also high-risk areas for various defects such as cracks, lack of fusion, incomplete penetration, slag inclusions, and porosity. Therefore, employing advanced and reliable non-destructive testing technologies for comprehensive and high-precision quality control of welds has become a crucial line of defense for ensuring national industrial security.

[0003] Ultrasonic phased array testing technology is an advanced non-destructive testing method that has developed rapidly in recent years. Compared with traditional single-crystal probe ultrasonic testing, phased array technology utilizes multiple crystals in a single probe assembly. By controlling the excitation delay of each array element, it achieves beam deflection, focusing, and scanning, enabling the inspection of welds from multiple angles and significantly improving the probability of detecting abnormal weld conditions. Electronic focusing can optimize the shape and size of the sound beam at the expected location of defects, thereby further improving the defect detection rate.

[0004] However, ultrasonic phased array weld inspection still has the following shortcomings in practical applications: First, in terms of defect identification, the existing detection model relies entirely on professional technicians to manually interpret the detection patterns on the screen. This method is highly subjective, with significant differences in interpretation among different personnel. Furthermore, staring at the screen for extended periods can easily lead to visual fatigue, resulting in missed detections. This places extremely high demands on the experience and concentration of the inspection personnel.

[0005] Secondly, in the field of intelligent identification research, existing research mainly falls into two categories. One category focuses on intelligent identification of traditional single-angle ultrasonic signals, analyzing a single A-scan signal to determine the presence of defects. However, the core advantage of ultrasonic phased array detection lies in its ability to acquire A-scan signals from multiple angles at the same detection location. Weld defects exhibit differentiated echo response characteristics at different angles, making it difficult to comprehensively characterize defect information and prone to missed detections or misjudgments due to inappropriate angle selection. The other category of research focuses on target detection using images generated from ultrasonic signal imaging. While this method leverages the advantages of imaging visualization, it often loses the precise coordinate information contained in the original signal during the imaging process, failing to determine the spatial location of the defect in the real world and limiting the practical value of the detection results.

[0006] Third, regarding the completeness of defect information, most existing intelligent detection methods can only provide a qualitative judgment on the existence of defects, lacking the ability to determine the specific angle and location of defects. Even if a defect is detected, further manual analysis is still required to obtain the angle information of the defect, failing to achieve truly intelligent full-process detection.

[0007] Therefore, how to fully utilize the complementary information of multi-angle A-scan signals acquired by ultrasonic phased array at the same detection position to achieve automatic and accurate defect identification and further determine the angular position of the defect has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] To address the aforementioned issues, this invention provides an intelligent detection method for ultrasonic phased array weld defects. This method utilizes time-frequency domain transformation of multi-angle A-scan signals, two-dimensional convolutional feature extraction, and Transformer self-attention fusion to solve problems such as incomplete single-angle information, difficulty in effectively fusing multi-angle signals, and missing defect angle information. It achieves automatic and accurate identification of weld defects and can determine the angle information of defects, thereby improving the intelligence level and practicality of ultrasonic phased array weld detection.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides an intelligent detection method for ultrasonic phased array weld defects, comprising the following steps: A-scan signals corresponding to multiple preset angles at the same detection position are acquired, and a multi-angle A-scan signal matrix is ​​constructed. The multi-angle A-scan signal matrix has an angle dimension and a sound path sampling point dimension. A short-time Fourier transform is performed on each A-scan signal in the multi-angle A-scan signal matrix to obtain a power spectrum, and the power spectrum is logarithmically compressed to obtain a two-dimensional time spectrum. The time-spectrum maps of all angles are stacked according to the angle dimension to form a multi-angle time-spectrum map batch. The multi-angle time-spectrum map batch is then input into a two-dimensional convolutional network for parallel feature extraction to obtain a feature vector corresponding to each angle. The feature vectors of all angles are then arranged in a preset angle order to form a multi-angle feature sequence. The multi-angle feature sequence is input into the Transformer encoder, and the correlation between angles is learned by the self-attention mechanism and the features are weighted and fused to obtain the fused feature sequence. The fused feature sequence is aggregated along the angular dimension to obtain a sample embedding vector representing the detection location; Defect classification and identification are performed based on the sample embedding vector, and the identification result of whether a defect exists at the detection location is output.

[0010] In one embodiment of the present invention, the step of acquiring A-scan signals corresponding to multiple preset angles at the same detection position and constructing a multi-angle A-scan signal matrix specifically includes: Parse the original detection file and extract the scanning position and detection angle corresponding to each A-scan record; The scanning position is rounded to a precision of 0.1 mm to obtain the group coordinates, and the A-scan records are grouped according to the group coordinates, with the same group corresponding to the same detection position; For each detection position, the A-scan signal corresponding to each angle is extracted according to the preset angle set. Angles not included in the original data are filled with zero-value waveforms. Each waveform is preprocessed by amplitude truncation, linear interpolation resampling to a uniform fixed length and normalization. The preprocessed waveforms at the same detection position are stacked in angular order to form the multi-angle A-scan signal matrix.

[0011] In one embodiment of the present invention, the amplitude truncation range is 0 to 255; the resampling refers to resampling the waveform to a uniform fixed length of 512; the normalization refers to dividing the amplitude of each point of the resampled waveform by 255.

[0012] In one embodiment of the present invention, the number of Fourier transform points of the short-time Fourier transform is 64, and the frame shift is 32; the logarithmic compression is performed according to the formula... In progress, among which The power spectrum is shown. It is a constant.

[0013] In one embodiment of the present invention, the two-dimensional convolutional network is composed of multiple convolutional modules and a global average pooling layer connected in series; each convolutional module includes a convolutional layer, a batch normalization layer and a ReLU activation function, and the other convolutional modules except the last one also include a max pooling layer; the global average pooling layer is used to compress its input feature map into a feature vector of fixed length.

[0014] In one embodiment of the present invention, the structure of the two-dimensional convolutional network is specifically as follows: The first convolutional module has 1 input channel, 16 output channels, a kernel size of 3×3, a stride of 1, zero padding of 1, and includes a ReLU activation function and a max pooling layer. The second convolutional module has 16 input channels, 32 output channels, a kernel size of 3×3, a stride of 1, zero padding of 1, and includes a ReLU activation function and a max pooling layer. The third convolutional module has 32 input channels, 64 output channels, a kernel size of 3×3, a stride of 1, zero padding of 1, and includes the ReLU activation function. The global average pooling layer is used to compress the spatial dimension of the feature map output by the third convolutional module to 1×1, and after flattening, it yields a 64-dimensional feature vector for each angle.

[0015] In one embodiment of the present invention, the Transformer encoder generates a query matrix, a key matrix, and a value matrix from the multi-angle feature sequences through linear mapping, and calculates the correlation weights between angles through a scaling dot product attention mechanism to perform the feature weighted fusion; the aggregation along the angle dimension is average pooling along the angle dimension.

[0016] In one embodiment of the present invention, the step of performing defect classification and identification based on the sample embedding vector and outputting the identification result of whether a defect exists at the detection location specifically includes: The sample embedding vector is processed through a fully connected layer, a batch normalization layer, a ReLU activation function, and a Dropout layer, and then mapped to corresponding two-dimensional classification outputs for defect-free and defective samples. The two-dimensional classification output is converted into probability values ​​using the Softmax function, and the category corresponding to the one with the higher probability is output as the recognition result.

[0017] In one embodiment of the present invention, the method further includes: When the identification result indicates the presence of a defect, the following defect angle determination steps are performed: Extract the echo amplitude at each detection angle of the detection location, and determine the region with the largest amplitude that shows a decreasing trend towards both increasing and decreasing angles as the candidate angle region; Structural echo regions belonging to weld corner reflection waves and mountain-shaped waves are eliminated from the candidate angle regions. Specifically, candidate angle regions located at the intersection of two echo wave sub-regions and horizontally positioned at the weld boundary or fusion zone boundary are identified as weld corner reflection wave regions and eliminated. Candidate angle region groups containing weld corner reflection echoes, longitudinal reflection echoes, and transverse reflection echoes appearing sequentially, where the longitudinal reflection echo is located approximately halfway between the transverse reflection echoes and the path difference between the longitudinal and transverse reflection echoes does not exceed one plate thickness, are identified as mountain-shaped wave regions and eliminated. Candidate areas whose echo amplitude points are located outside the preset weld zone and fusion zone spatial range are excluded; In the remaining candidate regions, the angle with the foremost echo wave number and the largest corresponding amplitude is selected as the angle information of the defect.

[0018] Secondly, the present invention provides an intelligent detection system for ultrasonic phased array weld defects, used to implement the aforementioned intelligent detection method for ultrasonic phased array weld defects, the system comprising: The signal matrix construction module is used to acquire A-scan signals corresponding to multiple preset angles at the same detection position and construct a multi-angle A-scan signal matrix, wherein the multi-angle A-scan signal matrix has an angle dimension and a sound path sampling point dimension. The time-frequency transformation module is used to perform short-time Fourier transform on each A-scan signal in the multi-angle A-scan signal matrix to obtain a power spectrum, and to perform logarithmic compression on the power spectrum to obtain a two-dimensional time-frequency spectrum. The convolutional feature extraction module is used to stack the time spectrum maps of all angles according to the angle dimension to form a multi-angle time spectrum map batch, and input the multi-angle time spectrum map batch into a two-dimensional convolutional network for parallel feature extraction to obtain a feature vector corresponding to each angle. The feature vectors of all angles are arranged in a preset angle order to form a multi-angle feature sequence. An angle feature fusion module is used to input the multi-angle feature sequence into the Transformer encoder, learn the correlation between angles using a self-attention mechanism and perform feature weighted fusion to obtain a fused feature sequence. The feature aggregation module is used to aggregate the fused feature sequence along the angular dimension to obtain a sample embedding vector representing the detection position; The defect classification and identification module is used to classify and identify defects based on the sample embedding vector and output the identification result of whether a defect exists at the detection location.

[0019] The beneficial effects achieved by this invention are as follows: (1) This invention proposes a multi-angle A-scan signal joint modeling and recognition framework for ultrasonic phased array weld inspection. It unifies the A-scan signals acquired by the phased array at multiple preset angles at the same detection position into a multi-angle A-scan signal matrix, fully utilizing the complementary information from multi-angle phased array detection. Compared with existing technologies that rely solely on a single-angle A-scan signal for defect judgment, this invention effectively solves the problem of missed detections and misjudgments caused by insignificant or missing defect echo characteristics at a single angle through the joint representation of multi-angle information, significantly improving the robustness and accuracy of defect identification.

[0020] (2) This invention proposes a time-spectrum representation and logarithmic compression strategy based on short-time Fourier transform, which transforms the one-dimensional A-scan signal at each angle into a two-dimensional time-spectrum. This transformation not only preserves the time-domain localization capability of the original signal, but also introduces a frequency-domain feature dimension, enabling the time-frequency texture structure of the defect echo to be explicitly presented. Compared with the method of directly extracting features from the original one-dimensional waveform, the two-dimensional time-spectrum is more conducive to the extraction of discriminative local texture and energy distribution features by two-dimensional convolutional networks, enhancing the ability to distinguish different types of defect echoes.

[0021] (3) This invention introduces a Transformer self-attention mechanism for multi-angle feature association modeling and adaptive fusion. By arranging the feature vectors extracted from each angle by a two-dimensional convolutional network in angular order to form a multi-angle feature sequence, and utilizing the self-attention mechanism of the Transformer encoder to learn the correlation between different detection angles, higher weights are assigned to key angles containing defect-sensitive information, while redundant or low-contribution angles are suppressed, thus achieving effective fusion of complementary information from multiple angles. Compared with the method of simply concatenating or averaging the features of each angle, this adaptive weighted fusion mechanism can highlight the angle features that contribute the most to defect identification, further improving the identification accuracy.

[0022] (4) Based on the identification of the presence of defects, this invention further proposes a method for determining the defect angle. This method extracts candidate angle regions by utilizing the distribution law of multi-angle echo amplitudes, and performs structural echo elimination by combining the typical characteristic rules of weld corner reflection waves and mountain-shaped waves. Then, it filters effective regions by constraining the spatial range of the weld area and fusion area, and finally determines the specific angle information of the defect based on the principles of wave priority and amplitude priority. This method makes up for the deficiency of not being able to provide the defect angle position in simple defect presence identification, provides key information for subsequent quantitative defect assessment, and improves the practicality and completeness of the detection results. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0024] Figure 1 This is an overall flowchart of the method of the present invention.

[0025] Figure 2 This is a schematic diagram illustrating the construction of a multi-angle A-scan signal matrix at the same detection position according to the present invention.

[0026] Figure 3 This is a network structure diagram of the multi-angle A-scan signal defect identification based on time-frequency diagram and Transformer in this invention.

[0027] Figure 4 This is a schematic diagram of the defect angle determination method of the present invention.

[0028] Figure 5 This is a schematic diagram of the results annotated by human experts on a D-scan.

[0029] Figure 6 This is a schematic diagram showing the results of defect detection performed only in the horizontal axis direction.

[0030] Figure 7 A schematic diagram showing the results of further angle screening to determine if there are defects in the horizontal axis direction. Detailed Implementation

[0031] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0032] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terms “comprising” and “having”, and any variations thereof, in the specification, claims and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0033] In this invention, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least some embodiments of the invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this invention can be combined with other embodiments.

[0034] The relevant content of this application is first introduced as follows: 1. Principle of ultrasonic phased array weld inspection and multi-angle A-scan acquisition.

[0035] 1.1 Principle of Ultrasonic Phased Array Weld Inspection: Ultrasonic phased array testing is a non-destructive testing method that utilizes the delayed excitation and delayed reception of multiple array elements in an array transducer to achieve beam deflection, focusing, and scanning. During weld inspection, by applying different excitation delays to each array element, ultrasonic waves can be incident on the weld area at different refraction angles, thereby obtaining echo information under different propagation paths within the weld.

[0036] For the same detection location, the phased array system is set according to a preset angle.

[0037] By sequentially transmitting the sound beam and receiving the echo, A-scan signal sequences corresponding to different angles can be obtained.

[0038] in, Indicates the first From one angle The one-dimensional A-scan signal was collected.

[0039] Defects in welds reflect, scatter, and diffract ultrasonic waves, resulting in varying echo amplitudes, arrival times, and energy distribution characteristics at different angles. Joint analysis using multi-angle A-scan signals provides a more comprehensive characterization of internal weld defects, improving the accuracy and robustness of the identification results.

[0040] 1.2 Multi-angle A-scan input modeling principle: The A-scan signal is essentially a one-dimensional discrete sequence of ultrasonic echo amplitude varying with propagation time. Let's assume that at the same detection location, the... The length of the A-scan signal corresponding to each angle is: Then it can be represented as

[0041] in Indicates the first Echo values ​​at each sampling point.

[0042] Under the same detection location The A-scan signals at each angle are stacked in angular order to obtain the network input tensor.

[0043] in For batch size, For degrees, This represents the number of sampling points for each A-scan.

[0044] 2. A-sweep time-frequency transformation principle: While A-scan signals can directly reflect the time-domain variation characteristics of echoes, weld defect echoes are often affected by material structure noise, interface echoes, and coupling state fluctuations. Simply relying on time-domain amplitude is insufficient to fully characterize the differences between different defects. To simultaneously extract the temporal local features and frequency component features of the signal, this invention employs Short-Time Fourier Transform (STFT) to convert the one-dimensional A-scan signal into a two-dimensional time-spectrum diagram.

[0045] For the A-scan signal at various angles Its short-time Fourier transform is defined as

[0046] in For window functions, For frame shift, The number of points in the Fourier transform. For time indexing, This refers to the index frequency.

[0047] The power spectrum can be obtained from the short-time Fourier transform:

[0048] To reduce the dynamic range of the power spectrum and improve the stability of subsequent neural network training, logarithmic compression is performed on the power spectrum to obtain a logarithmic spectrum.

[0049] in, To prevent logarithmic overflow using extremely small constants, after the above transformation, each original A-scan signal is mapped to a two-dimensional time-spectrum graph. This spectrum can simultaneously reflect the distribution of ultrasonic echo energy on both the time and frequency axes, which helps to improve the separability of defect features.

[0050] 3. Principle of 2D Convolution Feature Extraction: Convolutional neural networks (CNNs) exhibit excellent performance in feature extraction from two-dimensional images and spectra. Therefore, this invention employs a two-dimensional convolutional network to extract features from the time-spectral maps corresponding to A-scan signals at various angles. Let the input spectra be...

[0051] in, and These represent the frequency dimension and time dimension of the spectrum, respectively.

[0052] The basic operations of a two-dimensional convolutional layer can be represented as follows:

[0053] in and They represent the first Layer and first Layer feature map, For convolution kernel parameters, For bias terms, It is a non-linear activation function.

[0054] This invention extracts local texture features, energy distribution features, and high-level semantic features from the spectral image through multi-layer two-dimensional convolution, batch normalization, nonlinear activation, and pooling operations. To ensure that features from different angles have a unified dimension, global average pooling is finally used to compress the convolutional features from each angle into a one-dimensional vector.

[0055]

[0056] in, For feature dimensions.

[0057] All The feature vectors of each angle are arranged in angular order to form an angle feature sequence.

[0058] 4. Transformer Angle-Related Modeling Principle: In ultrasonic phased array weld inspection, A-scan signals at different angles exhibit a clear complementary relationship. Certain angles more readily capture specular reflections from defects, while others better highlight diffraction or scattering characteristics at defect edges. Relying solely on a single angle can easily lead to missed detections or misjudgments due to inappropriate angle selection. Therefore, this invention introduces a Transformer encoder to perform correlation modeling of multi-angle feature sequences.

[0059] Let the angular feature sequence output by the two-dimensional convolutional network be...

[0060] First, the query matrix, key matrix, and value matrix are obtained through linear mapping:

[0061] in, , , This is the learnable parameter matrix.

[0062] The formula for calculating self-attention is:

[0063] in, This is the feature scaling factor. Through this mechanism, the model can learn the correlation between features from different perspectives, assigning higher weights to key perspectives and lower weights to redundant or low-contribution perspectives.

[0064] After passing through the Transformer encoder, an output feature sequence fusing multi-angle information is obtained.

[0065] To obtain a sample-level overall feature representation, the output sequence A is average-pooled along the angular dimension to obtain the embedding vector.

[0066] in, This embedding vector comprehensively represents defect information from multiple detection angles at the same detection location, and can be used for subsequent classification and recognition.

[0067] 5. Determining the angle of the defect: In ultrasonic phased array weld inspection, after determining the location of the defect along the weld direction, it is necessary to further determine the angle information of the defect. If the model identifies a defect at a location along the weld direction, it can be assumed that there is one and only one main defect response source at that location. Whether it is a real defect, a pseudo-defect, or a structural echo, its signal behavior in multi-angle detection is usually the same: the amplitude reaches its maximum at a certain angle, and the echo amplitude decreases as the detection angle increases or decreases. Therefore, this pattern can be used to first determine the candidate angle region of the defect.

[0068] After obtaining the candidate angle regions, it is necessary to further remove the candidate regions corresponding to typical structural echoes. The typical structural echoes include weld corner reflections and ridge-shaped waves. Wherein: Weld corner reflected waves refer to the echo signals that occur when ultrasonic waves are incident on a weld corner, and the normal direction at the weld corner is the same as or similar to the direction of the incident ultrasonic beam. Part of the ultrasonic waves return along the original propagation path and are received by the probe, thus forming an echo signal on the flaw detector screen. Its characteristics are: in the depth direction, it is usually located at the junction of two waves; in the horizontal position, it is usually at the weld boundary or the boundary of the fusion zone.

[0069] A mountain-shaped wave refers to a wave pattern where, when the main ultrasonic beam is incident on the weld bead contour and its transverse wave incident angle is less than the third critical angle, in addition to the reflected transverse wave, a deformed longitudinal wave is generated in the workpiece due to waveform conversion. When the deformed longitudinal wave or the reflected transverse wave is reflected again on the upper surface of the weld and returns along the original path, the probe receives multiple related echoes, thus forming a mountain-shaped wave signal on the screen. Its characteristics are: the probe typically receives the weld bead reflected echo signal, the reflected longitudinal wave echo signal, and the reflected transverse wave echo signal sequentially. The first signal has the same characteristics as the weld bead reflected wave, the second echo is located approximately halfway down the middle of the third echo, and the difference in sound path between them does not exceed one plate thickness.

[0070] After removing the candidate regions corresponding to weld corner reflection waves and ridge waves, it is necessary to further filter the candidate regions by considering the spatial distribution range of the weld zone and the fusion zone. If the amplitude point of a candidate region appears outside the weld zone and the fusion zone, the candidate region is removed; for the remaining candidate regions, the angle with the earlier wave order and larger amplitude is preferentially selected as the amplitude point of the defect, thereby determining the angle information of the defect.

[0071] like Figures 1 to 4 As shown, this invention provides an intelligent detection method for ultrasonic phased array weld defects based on multi-angle A-scan signal merging, the method comprising: Step 1: Read the original angle and load the label task table; Read the raw CSV file of the ultrasonic phased array inspection to be processed, and read the label task table. The label task table includes at least the filename and the starting position of the defect region. End position of defect area And the label field. For each original detection file, the corresponding label task information is matched according to the filename. In this embodiment, only areas with a label value of 1 are considered defect areas. When the scan coordinate x of a certain detection position falls into any defect area... If the value is 1, then the sample label at that location is determined to be 1; otherwise, it is determined to be 0.

[0072] Step 2: Construct multi-angle A-scan samples at the same X-scan location; The original CSV file is parsed to extract the X-scan position corresponding to each A-scan record. and detection angle .in, Obtained by parsing the first column of the original table. The coordinates are obtained from the third column of the original table. To reduce the impact of minor measurement errors on grouping, the scanning positions are rounded to a precision of 0.1 mm, resulting in the grouped coordinates:

[0073] Subsequently, according to The original A-scan records are grouped. For each group, it is assumed that it corresponds to the same spatial location in the weld inspection, and the A-scan signals corresponding to different angles at that location are extracted from it.

[0074] In this embodiment, the preset angle range is 35° to 75°, and the angle step size is 0.5°. Therefore, the angle set is represented as follows:

[0075] The total number of corresponding angles is 81.

[0076] For each target angle at the same location If the original data contains an A-scan record for the corresponding angle, then the waveform data for that angle is extracted; otherwise, it is padded with a waveform of all zeros. Let the first... The original waveforms corresponding to each angle are Then the multi-angle A-scan sequence corresponding to this position is represented as:

[0077] Step 3: Preprocessing and unifying the length of the original waveform; Preprocess each raw A-scan waveform extracted in step 2. First, read the raw waveform data from the amplitude column of each record and replace non-numerical items with 0. Then, truncate the waveform's amplitude to limit its range to the full screen height of the system.

[0078] Since the number of sampling points in different original A-scan records may vary, this embodiment uses a pure numerical resampling method, directly linearly interpolating each original waveform to a fixed length of 512. Let the length of the original waveform be L, and the resampled waveform is represented as follows:

[0079] Specifically, the original sampling point coordinates and the target sampling point coordinates are established in the normalized coordinate interval [0,1]. A waveform with uniform length is obtained through linear interpolation. Finally, the resampled waveform is normalized according to the system's full-screen height to obtain:

[0080] After the above processing, each angle corresponds to a normalized A-scan signal with a length of 512.

[0081] Step 4: Constructing input tensors using multi-angle A-scan; The normalized A-scan signals corresponding to all 81 angles at the same X-scan position are stacked in angular order to obtain the input matrix of a single sample:

[0082] In this matrix, the first dimension corresponds to the angle dimension, and the second dimension corresponds to the sampling point dimension of the A-scan waveform sound path. Furthermore, samples from multiple detection positions are combined into a batch input tensor:

[0083] Where B represents the batch size. Simultaneously, based on the label determination method in step 1, a category label is assigned to each sample. , where 0 represents no defects and 1 represents defects.

[0084] Step 5: Time-frequency diagram transformation of A-scan signal; To simultaneously extract the time-domain and frequency-domain features of the A-scan signal, this embodiment performs a short-time Fourier transform on the A-scan signal at each angle in the input tensor X. First, the input tensor is expanded as follows:

[0085] Then, a short-time Fourier transform is performed on each A-scan signal, assuming the number of Fourier transform points is . Frame shift is The two-dimensional power spectrum is obtained:

[0086] The frequency dimension is 33. Logarithmic compression of the power spectrum yields the logarithmic spectrum:

[0087] in, .

[0088] After adding channel dimension to all spectral graphs, the input tensor of the two-dimensional convolutional network is constructed as follows:

[0089] Step 6: Two-dimensional convolution feature extraction; This embodiment employs a two-dimensional convolutional neural network to extract features from the time-spectral maps obtained in step 5 at various angles. The two-dimensional convolutional neural network consists of three convolutional layers and an adaptive global average pooling layer, used to progressively extract local texture features, energy distribution features, and defect-related pattern features from the spectra.

[0090] Step 61, Extraction of the first convolutional layer: The spectral map obtained in step 5 is input into the first-layer 2D convolutional module, which includes convolutional layers, batch normalization layers, ReLU activation functions, and max pooling layers. The convolutional kernel size is set to 3×3, the number of input channels is 1, the number of output channels is 16, the stride is 1, and the padding is 1. After the first convolutional layer, the feature map is obtained:

[0091] After the max pooling operation, we get:

[0092] in and These represent the spatial dimensions after the first pooling layer.

[0093] Step 62, Extraction of the second convolutional layer: Will The second 2D convolutional module is input, which includes convolutional layers, batch normalization layers, ReLU activation functions, and max pooling layers. The kernel size is set to 3×3, the number of input channels is 16, the number of output channels is 32, the stride is 1, and the padding is 1. The feature map is obtained after the second convolutional layer.

[0094] After the max pooling operation, we get:

[0095] in and These represent the spatial dimensions after the second pooling layer.

[0096] Step 63, Extraction of the third convolutional layer: Will The input is a third-layer 2D convolutional module, which includes convolutional layers, batch normalization layers, and a ReLU activation function. The kernel size is set to 3×3, the number of input channels is 32, the number of output channels is 64, the stride is 1, and the padding is 1. The resulting feature map is:

[0097] Step 64, Global Average Pooling: right Adaptive global average pooling is performed to compress the spatial dimension to 1×1, resulting in:

[0098] Then flatten it and reorganize it according to the angular dimension to obtain a multi-angle feature sequence:

[0099] in, This represents the high-level feature representation sequence corresponding to 81 angles at the same detection location.

[0100] Step 7: Multi-angle feature association and fusion; Since the same weld defect will exhibit different echo response characteristics at different detection angles, and there is a clear complementary relationship between the features at each angle, this embodiment uses the multi-angle feature sequence obtained in step 6. Perform correlation fusion to obtain a more complete defect characterization.

[0101] Specifically, multi-angle feature sequences The input feature association modeling module calculates the correlation between features from different perspectives through a self-attention mechanism, and adaptively weights the features from each perspective based on the correlation to obtain the fused multi-angle feature sequence:

[0102] In this process, higher weights are assigned to angular features with more obvious defect responses and stronger discriminative capabilities, while lower weights are assigned to angular features with redundant information or weak responses, thereby achieving effective fusion of multi-angle features. To obtain a sample-level overall feature representation, the fused feature sequence is... Aggregating along the angular dimension yields an embedding vector representing the overall state of the detection location:

[0103] in,

[0104] This embedded vector comprehensively reflects the defect response information at different detection angles under the same detection position, and can be used as a global representation for subsequent defect determination.

[0105] Step 8: Defect identification output; The fusion feature sequence obtained in step 7 The input is to the classification and determination module, which outputs the defect identification result corresponding to the current detection location. Specifically, the fused feature sequence is first... The input features are expanded along the angular and feature dimensions to form sample-level classification input features; then, through fully connected mapping, batch normalization, ReLU activation, and Dropout processing, the final classification output is obtained.

[0106] The output consists of two components corresponding to the categories "no defects" and "defective." Further, the classification output is converted into probability values ​​for each category using the Softmax function, and the category with the higher probability is used as the recognition result for the current detection location.

[0107] Step 9: Determine the defect angle.

[0108] When the classification result of step 8 indicates the presence of a defect at a certain X-scan location, the multi-angle A-scan response at that location is further analyzed to determine the angular information of the defect. Specifically, the amplitude distribution curve of that location across all detection angles is extracted. Since true defects, false defects, and structural echoes typically exhibit the characteristic of having the largest amplitude at a certain angle and gradually decreasing amplitude as the angle increases or decreases during multi-angle detection, the candidate angular region of the defect is first determined based on this pattern.

[0109] After determining the candidate angle regions, candidate regions corresponding to typical structural echoes are further removed. These typical structural echoes include weld corner reflections and mountain-shaped waves. Weld corner reflections are typically located at the boundary between two waves, and horizontally at the weld boundary or fusion zone boundary. Mountain-shaped waves are typically characterized by the probe sequentially receiving weld corner reflection echo signals, reflected longitudinal wave echo signals, and reflected transverse wave echo signals, with the second echo located approximately halfway down the middle of the third echo, and the path difference between them not exceeding one plate thickness. After removing these typical structural echoes, candidate regions whose amplitude points appear outside the weld zone and fusion zone are then removed. For the final retained candidate regions, angles with earlier waves and larger amplitudes are preferentially selected as the amplitude points of the defects, thereby determining the angle information of the defects.

[0110] Through the above steps, the joint analysis of A-scan signals from multiple angles at the same X-scan position is realized, which not only enables automatic identification of weld defects, but also further determines the angular position of the defects.

[0111] Furthermore, the present invention also provides an intelligent detection system for ultrasonic phased array weld defects, used to implement the aforementioned intelligent detection method for ultrasonic phased array weld defects, the system comprising: The signal matrix construction module is used to acquire A-scan signals corresponding to multiple preset angles at the same detection position and construct a multi-angle A-scan signal matrix, wherein the multi-angle A-scan signal matrix has an angle dimension and a sound path sampling point dimension. The time-frequency transformation module is used to perform short-time Fourier transform on each A-scan signal in the multi-angle A-scan signal matrix to obtain a power spectrum, and to perform logarithmic compression on the power spectrum to obtain a two-dimensional time-frequency spectrum. The convolutional feature extraction module is used to stack the time spectrum maps of all angles according to the angle dimension to form a multi-angle time spectrum map batch, and input the multi-angle time spectrum map batch into a two-dimensional convolutional network for parallel feature extraction to obtain a feature vector corresponding to each angle. The feature vectors of all angles are arranged in a preset angle order to form a multi-angle feature sequence. An angle feature fusion module is used to input the multi-angle feature sequence into the Transformer encoder, learn the correlation between angles using a self-attention mechanism and perform feature weighted fusion to obtain a fused feature sequence. The feature aggregation module is used to aggregate the fused feature sequence along the angular dimension to obtain a sample embedding vector representing the detection position; The defect classification and identification module is used to classify and identify defects based on the sample embedding vector and output the identification result of whether a defect exists at the detection location.

[0112] To verify the effectiveness of the intelligent detection method for ultrasonic phased array weld defects proposed in this invention, the following experiment was conducted: 1. Experimental conditions and parameters: The dataset used in this invention consists of ultrasonic phased array weld inspection signal data acquired by an HSPA30-E device. The raw data includes six different categories of ultrasonic signals: normal, slag inclusion, incomplete penetration, lack of fusion, crack, and porosity. The normal signal represents the echo response when there are no obvious defects in the weld; the slag inclusion, incomplete penetration, lack of fusion, crack, and porosity signals correspond to five typical weld defect types. The normal category is considered defect-free, while the other five categories are collectively classified as defects. All signals are derived from A-scan signal data obtained during actual inspection, effectively reflecting the differences in ultrasonic echo characteristics under different weld conditions, and providing a data foundation for the subsequent training and validation of the defect identification model.

[0113] The CPU used in this experiment was a 24 vCPU AMD EPYC 9754 128-Core Processor, with 72GB of memory, a vGPU-32GB GPU, PyTorch version 2.3.0, and CUDA version 12.1.

[0114] 2. Experiment and Results Analysis: To verify the performance of the method of this invention, detection was performed on a dataset with the goal of detecting whether there is a defect at position X. The evaluation metrics included accuracy (Acc), precision (Precision), recall (Recall), and F1-score. To verify the superiority of the proposed model, the method of this invention was compared with a method that directly classifies the original waveform without using time-frequency graph transformation, as shown in Table 1.

[0115] Table 1. Performance comparison between this method and conversion without time-frequency diagrams.

[0116] As shown in Table 1, the method of this invention outperforms the method that directly classifies the original A-scan waveform without time-frequency conversion in all four metrics: accuracy, precision, recall, and F1 score. Specifically, the accuracy of this invention reaches 92.59%, an improvement of 2.97 percentage points compared to the method without time-frequency conversion; the precision reaches 75.79%, an improvement of 6.86 percentage points; the recall reaches 78.25%, an improvement of 15.59 percentage points; and the F1 score reaches 77.00%, an improvement of 11.35 percentage points.

[0117] The above results demonstrate that converting the one-dimensional A-scan signal into a two-dimensional time-frequency spectrum map can more fully characterize the energy distribution features of the defect echo in both the time and frequency dimensions, enabling the two-dimensional convolutional network to extract more discriminative local texture features and energy aggregation features. In particular, the significant improvement in recall indicates that the method of this invention can effectively reduce the risk of missed detections caused by the indistinct appearance of defect echoes in the original time-domain waveform; the improved F1 score indicates that this invention achieves a better balance between defect detection capability and false detection control. Therefore, Table 1 verifies that the technique of combining time-frequency spectrum conversion with two-dimensional convolutional feature extraction in this invention can effectively improve the performance of ultrasonic phased array weld defect identification.

[0118] To verify the effectiveness of the method of the present invention on actual weld inspection data, typical inspection data were selected for verification. Figure 5 A schematic diagram showing the results of annotations made by human experts on a D-scan. Figure 6 This is a schematic diagram showing the results of defect detection performed only in the horizontal axis direction; Figure 7 This diagram illustrates the results of angle screening to determine if there are defects along the horizontal axis; the horizontal axis represents the scanning position, the vertical axis represents the projection depth, and the color represents the echo amplitude. Red marks indicate the defect locations identified by the method of this invention.

[0119] Depend on Figure 5 It can be seen that the defects marked by human experts on the D scan are mainly distributed in several consecutive scan position intervals, and the corresponding areas usually show a more obvious high-amplitude echo response. Figure 6This is the result of defect detection performed first in the horizontal scanning direction by the method of the present invention. The red marked area is... Figure 5 The expert-annotated areas in the scan are basically consistent, indicating that the present invention can automatically identify the X-scan location of suspected defects without relying on manual frame-by-frame interpretation, and has good defect detection capability.

[0120] Figure 6 The red vertical markers in the image mainly indicate that the corresponding X-scan position is determined to have a defect. In the depth direction, they are still coarse-grained markers and have not yet distinguished between real defect echoes, structural echoes, and other non-defect high-amplitude responses. Figure 7 After confirming the presence of defects along the horizontal axis, angle screening is further performed based on the multi-angle echo amplitude distribution pattern. This is combined with structural echo elimination rules such as weld corner reflection waves and mountain-shaped waves, as well as the spatial constraints of the weld zone and fusion zone to obtain candidate defect points. It can be seen that the screened red defect points are concentrated in the expert-annotated area and near the high-amplitude defect response, reducing the mislabeling of irrelevant depth areas and structural echo areas.

[0121] therefore, Figures 5 to 7 The invention not only enables automatic defect identification along the weld scanning direction, but also further determines the angle / amplitude response point related to the defect after the defect location is identified. This improves the detection result from simply indicating whether a defect exists to a more complete result that includes scanning position and angle information, thereby enhancing the interpretability and engineering practical value of the detection result.

[0122] In summary, this invention addresses the problems in ultrasonic phased array weld inspection, such as incomplete A-scan information from a single angle, differences in defect echoes due to variations in refraction angle / beam direction, and difficulties in effectively fusing signals from different angles. It proposes an intelligent ultrasonic phased array weld defect detection method based on multi-angle A-scan signal merging and attention sequence modeling. This method takes A-scan signal sequences obtained from the phased array at multiple angles at the same detection location as input. It maps the one-dimensional signal to a two-dimensional spectrum through time-frequency transformation, then uses a two-dimensional convolutional network to extract local time-frequency texture features at each angle, and introduces a Transformer encoder to perform correlation modeling of multi-angle features, ultimately achieving automatic identification of weld defects. After identifying a defect at a location along the weld direction, it further combines multi-angle amplitude response patterns and structural echo rejection rules to determine the defect angle information.

[0123] In the defect identification process of this invention, the set of multi-angle A-scan signals obtained from the phased array weld inspection at the same detection position is recorded as follows: First, according to the preset angle set, The organization is of length Angle sequence} ,in To merge the number of angles, each Indicates the corresponding number A-scan signals acquired at various angles, and Let L be a one-dimensional sampling sequence. After inputting this multi-angle sequence into the recognition network of this invention, the network input tensor can be represented as: ,in For batch size, For degrees, This represents the number of sampling points for each A-scan.

[0124] To enhance the characterizability and consistency of defect features, this invention first performs time-frequency transformation on the A-scan signal at each angle, mapping the one-dimensional signal to a two-dimensional time-frequency spectrum. The specific steps are as follows: For each angle signal... A short-time Fourier transform is performed, and the frequency domain resolution parameter and the frame shift parameter of the angle signal are set to obtain the power spectrum. To suppress feature instability caused by excessive dynamic range, logarithmic compression is performed on the power spectrum to obtain a logarithmic spectrum. Finally, each angle signal is converted into a two-dimensional spectrum and used as the input of a two-dimensional convolutional feature extractor.

[0125] Two-dimensional convolutional networks can effectively extract features such as local texture, energy clusters, and edge structures from spectral images. Therefore, this invention uses a two-dimensional convolutional neural network to extract features from the spectral image at each angle. Specifically, Each angular spectral map is input into a two-dimensional convolutional feature extractor consisting of multiple layers of convolution, normalization, nonlinear activation, and pooling. Multi-scale time-frequency texture and energy distribution features are extracted layer by layer. Subsequently, global average pooling is used to compress the two-dimensional feature map of each angle into a fixed-length feature vector, yielding the angular feature representation. , where C is the feature dimension. An angle feature sequence is obtained by stacking the angle feature vectors in angle order. Its tensor representation is .

[0126] In weld defect identification, the echo information of the same defect at different angles is complementary: some angles are more likely to produce obvious reflected echoes, while others may better reflect diffraction, scattering, or boundary effects. Therefore, adaptive fusion of multi-angle features is needed to highlight the angle information most sensitive to defects. To this end, this invention introduces a Transformer encoder based on a self-attention mechanism to model the angle feature sequence, learn the correlation between angles, and perform weighted fusion. The specific steps are as follows: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] The input is a Transformer encoder, which calculates the correlation weights between angles through multi-head self-attention and performs feature reweighting to obtain sequence features after fusing complementary angle information. .

[0127] To obtain a sample-level merged representation, this invention aggregates the angular sequence features output by the Transformer to obtain an embedding vector for defect detection. Specifically, for Mean pooling is performed along the angular dimension to obtain the embedding vector. This embedding vector represents multiple angles. The compact features resulting from merging scan signals can be used for defect identification.

[0128] After the model identifies a defect at a location along the weld direction, this invention further determines the angle information of the defect. Based on the principle that true defects, false defects, and structural echoes all exhibit the pattern of "maximum amplitude at a certain angle, gradually decreasing with increasing or decreasing angle" in multi-angle responses, the multi-angle echo amplitudes at this location are first analyzed to extract candidate angle regions for defects. Then, candidate regions corresponding to typical structural echoes, including weld corner reflections and ridge waves, are removed. Subsequently, candidate regions whose amplitude regions appear outside the weld zone and fusion zone are eliminated. For the remaining candidate regions, angles with earlier wave orders and larger amplitudes are preferentially selected as the amplitude points of the defects, thereby determining the angular location and corresponding depth location of the defects.

[0129] Although the invention has been described with reference to preferred embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, the technical features mentioned in the various embodiments can be combined in any manner as long as there is no structural conflict. The invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for intelligent detection of defects in ultrasonic phased array welds, characterized in that, Includes the following steps: A-scan signals corresponding to multiple preset angles at the same detection position are acquired, and a multi-angle A-scan signal matrix is ​​constructed. The multi-angle A-scan signal matrix has an angle dimension and a sound path sampling point dimension. A short-time Fourier transform is performed on each A-scan signal in the multi-angle A-scan signal matrix to obtain a power spectrum, and the power spectrum is logarithmically compressed to obtain a two-dimensional time spectrum. The time-spectrum maps of all angles are stacked according to the angle dimension to form a multi-angle time-spectrum map batch. The multi-angle time-spectrum map batch is then input into a two-dimensional convolutional network for parallel feature extraction to obtain a feature vector corresponding to each angle. The feature vectors of all angles are then arranged in a preset angle order to form a multi-angle feature sequence. The multi-angle feature sequence is input into the Transformer encoder, and the correlation between angles is learned by the self-attention mechanism and the features are weighted and fused to obtain the fused feature sequence. The fused feature sequence is aggregated along the angular dimension to obtain a sample embedding vector representing the detection location; Defect classification and identification are performed based on the sample embedding vector, and the identification result of whether a defect exists at the detection location is output.

2. The intelligent detection method for ultrasonic phased array weld defects according to claim 1, characterized in that, The step of acquiring A-scan signals corresponding to multiple preset angles at the same detection position and constructing a multi-angle A-scan signal matrix specifically includes: Parse the original detection file and extract the scanning position and detection angle corresponding to each A-scan record; The scanning position is rounded to a precision of 0.1 mm to obtain the group coordinates, and the A-scan records are grouped according to the group coordinates, with the same group corresponding to the same detection position; For each detection position, the A-scan signal corresponding to each angle is extracted according to the preset angle set. Angles not included in the original data are filled with zero-value waveforms. Each waveform is preprocessed by amplitude truncation, linear interpolation resampling to a uniform fixed length and normalization. The preprocessed waveforms at the same detection position are stacked in angular order to form the multi-angle A-scan signal matrix.

3. The intelligent detection method for ultrasonic phased array weld defects according to claim 2, characterized in that, The amplitude truncation range is 0 to 255; the resampling refers to resampling the waveform to a uniform fixed length of 512; the normalization refers to dividing the amplitude of each point of the resampled waveform by 255.

4. The intelligent detection method for ultrasonic phased array weld defects according to claim 1, characterized in that, The short-time Fourier transform has 64 Fourier transform points and a frame shift of 32; the logarithmic compression is performed according to the formula... In progress, among which The power spectrum is shown. It is a constant.

5. The intelligent detection method for ultrasonic phased array weld defects according to claim 4, characterized in that, The two-dimensional convolutional network consists of multiple convolutional modules and a global average pooling layer connected in series. Each convolutional module includes a convolutional layer, a batch normalization layer, and a ReLU activation function. Except for the last convolutional module, the other convolutional modules also include a max pooling layer. The global average pooling layer is used to compress the input feature map into a feature vector of fixed length.

6. The intelligent detection method for ultrasonic phased array weld defects according to claim 5, characterized in that, The specific structure of the two-dimensional convolutional network is as follows: The first convolutional module has 1 input channel, 16 output channels, a kernel size of 3×3, a stride of 1, zero padding of 1, and includes a ReLU activation function and a max pooling layer. The second convolutional module has 16 input channels, 32 output channels, a kernel size of 3×3, a stride of 1, zero padding of 1, and includes a ReLU activation function and a max pooling layer. The third convolutional module has 32 input channels, 64 output channels, a kernel size of 3×3, a stride of 1, zero padding of 1, and includes the ReLU activation function. The global average pooling layer is used to compress the spatial dimension of the feature map output by the third convolutional module to 1×1, and after flattening, it yields a 64-dimensional feature vector for each angle.

7. The intelligent detection method for ultrasonic phased array weld defects according to claim 1, characterized in that, The Transformer encoder generates a query matrix, a key matrix, and a value matrix from the multi-angle feature sequences through linear mapping, and calculates the correlation weights between angles through a scaling dot product attention mechanism to perform weighted feature fusion; the aggregation along the angle dimension is average pooling along the angle dimension.

8. The intelligent detection method for ultrasonic phased array weld defects according to claim 1, characterized in that, The step of classifying and identifying defects based on the sample embedding vector and outputting the identification result of whether a defect exists at the detection location specifically includes: The sample embedding vector is processed through a fully connected layer, a batch normalization layer, a ReLU activation function, and a Dropout layer, and then mapped to corresponding two-dimensional classification outputs for defect-free and defective samples. The two-dimensional classification output is converted into probability values ​​using the Softmax function, and the category corresponding to the one with the higher probability is output as the recognition result.

9. The intelligent detection method for ultrasonic phased array weld defects according to claim 1, characterized in that, The method further includes: When the identification result indicates the presence of a defect, the following defect angle determination steps are performed: Extract the echo amplitude at each detection angle of the detection location, and determine the region with the largest amplitude that shows a decreasing trend towards both increasing and decreasing angles as the candidate angle region; Structural echo regions belonging to weld corner reflection waves and mountain-shaped waves are eliminated from the candidate angle regions. Specifically, candidate angle regions located at the intersection of two echo wave sub-regions and horizontally positioned at the weld boundary or fusion zone boundary are identified as weld corner reflection wave regions and eliminated. Candidate angle region groups containing weld corner reflection echoes, longitudinal reflection echoes, and transverse reflection echoes appearing sequentially, where the longitudinal reflection echo is located approximately halfway between the transverse reflection echoes and the path difference between the longitudinal and transverse reflection echoes does not exceed one plate thickness, are identified as mountain-shaped wave regions and eliminated. Candidate areas whose echo amplitude points are located outside the preset weld zone and fusion zone spatial range are excluded; In the remaining candidate regions, the angle with the foremost echo wave number and the largest corresponding amplitude is selected as the angle information of the defect.

10. An intelligent detection system for ultrasonic phased array weld defects, characterized in that, The system for implementing the intelligent detection method for ultrasonic phased array weld defects according to any one of claims 1-9 includes: The signal matrix construction module is used to acquire A-scan signals corresponding to multiple preset angles at the same detection position and construct a multi-angle A-scan signal matrix, wherein the multi-angle A-scan signal matrix has an angle dimension and a sound path sampling point dimension. The time-frequency transformation module is used to perform short-time Fourier transform on each A-scan signal in the multi-angle A-scan signal matrix to obtain a power spectrum, and to perform logarithmic compression on the power spectrum to obtain a two-dimensional time-frequency spectrum. The convolutional feature extraction module is used to stack the time spectrum maps of all angles according to the angle dimension to form a multi-angle time spectrum map batch, and input the multi-angle time spectrum map batch into a two-dimensional convolutional network for parallel feature extraction to obtain a feature vector corresponding to each angle. The feature vectors of all angles are arranged in a preset angle order to form a multi-angle feature sequence. An angle feature fusion module is used to input the multi-angle feature sequence into the Transformer encoder, learn the correlation between angles using a self-attention mechanism and perform feature weighted fusion to obtain a fused feature sequence. The feature aggregation module is used to aggregate the fused feature sequence along the angular dimension to obtain a sample embedding vector representing the detection position; The defect classification and identification module is used to classify and identify defects based on the sample embedding vector and output the identification result of whether a defect exists at the detection location.

Images