A pressure vessel weld defect detection method and system based on frequency domain orthogonal decomposition

Abstract

This invention discloses a method and system for detecting weld defects in pressure vessels based on frequency domain orthogonal decomposition. The method includes: reading a pressure vessel model and planning a scanning path; calculating exposure parameters and generating a control script; executing elements based on the scanning path and control script to perform multi-degree-of-freedom collaborative scanning and acquire original weld images; performing image preprocessing and model training; based on the trained depth detection model based on multi-resolution frequency domain orthogonal decomposition sampling, separating frequency subbands through two-dimensional orthogonal multi-resolution decomposition, constructing hybrid frequency domain features through tensor reconstruction of frequency domain information, and obtaining enhanced multi-scale features through cross-frequency band feature mapping; and predicting and outputting the defect category, location, and size through hybrid feature encoding and decoding; performing qualification judgment and generating a detection report. This invention fundamentally avoids the information entropy loss caused by traditional downsampling, improving the classification accuracy and positioning precision of pressure vessel weld defect detection.

Description

Technical Field

[0001] This invention relates to the field of unmanned inspection technology, and more specifically to a method and system for detecting weld defects in pressure vessels based on frequency domain orthogonal decomposition. Background Technology

[0002] In the manufacturing and safety assessment of pressure vessels, weld quality is a key factor in determining whether the vessel can withstand high-pressure loads. Welds often contain defects such as tiny pores, cracks, or lack of fusion. These defects are often small in size and have weak edge features.

[0003] Currently, with the development of intelligent manufacturing technology, deep learning methods based on convolutional neural networks (CNN) have been widely used in the automatic detection of weld defects. In order to reduce the amount of computation and extract higher-level semantic features, existing technologies generally use max pooling or large stride convolution operations to downsample the feature maps when building detection models.

[0004] However, the existing processing method has significant drawbacks when detecting minute weld defects in pressure vessels: from a signal processing perspective, max pooling or stride convolution is mathematically equivalent to a low-pass filter. However, defects such as microcracks and pinholes in welds are mainly represented by abrupt high-frequency components in the image signal. The downsampling operation of the existing technology, while reducing the image resolution, inevitably filters out high-frequency information in the image, resulting in a significant attenuation or even complete loss of the energy of key details containing defect edges. This irreversible loss of high-frequency information makes it difficult for subsequent network layers to reconstruct the fine structure of minute defects, thus making it very easy to miss or mislocate minute defects when detecting precision welds in pressure vessels.

[0005] Therefore, how to solve the problem that existing deep learning detection technology filters out high-frequency edge features of tiny defects in pressure vessel welds due to low-pass filtering effect when performing feature dimensionality reduction through pooling layers or stride convolutional layers, resulting in missed detection of microcracks, blurred defect boundary positioning, and inability to adapt to interference from complex weld textures, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of the above problems, the present invention aims to provide a method and system for detecting weld defects in pressure vessels based on frequency domain orthogonal decomposition, which overcomes or at least partially solves the above problems.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for detecting weld defects in pressure vessels based on frequency domain orthogonal decomposition includes the following steps: S1. Read the pressure vessel structural model and plan the scanning path of the actuator. Automatically calculate the exposure parameters based on the weld thickness and generate a control script that includes the robot arm posture and exposure timing. S2. The actuator performs multi-degree-of-freedom collaborative scanning based on the scan path and control script motion to acquire the original weld seam image; S3. The acquired raw weld images are preprocessed and the model is trained. Based on the trained depth detection model based on multi-resolution frequency domain orthogonal decomposition sampling, the frequency sub-bands are separated by two-dimensional orthogonal multi-resolution decomposition. The hybrid frequency domain features are constructed by tensor reconstruction of frequency domain information. The enhanced multi-scale features are obtained by cross-band feature mapping. The category, location and size of the weld defect are predicted by hybrid feature encoding and decoding. S4. Based on the detected defect type and size, determine the pass / fail status by comparing with preset standards, and generate an inspection report containing a defect distribution heatmap.

[0009] Preferably, the specific method for acquiring the original weld image in step S2 is as follows: The actuator executes the control script and moves along the weld seam trajectory. During the movement, based on the real-time feedback of position information, the camera is triggered to expose and collect weld seam images every time a preset step is reached. After marking the current robot arm posture coordinates, the images are stored in the high-speed cache queue.

[0010] Preferably, step S3, the method for obtaining the enhanced multi-scale features, specifically includes: Using a set of conjugate orthogonal mirror filter banks, convolution and intermittent sampling operations are performed on each channel of the input feature tensor in the row and column directions, respectively, to output a frequency sub-band feature map, including four frequency sub-bands: low-frequency approximate component, horizontal detail component, vertical detail component and diagonal detail component; The separated frequency information is fused into the convolutional network, and a channel-dimensional tensor concatenation operation is performed to construct hybrid frequency domain features. The convolution kernel is used to linearly combine the mixed frequency domain features, and then batch normalization and nonlinear activation function are used to map them back to the target channel space to output enhanced multi-scale features.

[0011] Preferably, the specific content of hybrid feature encoding and decoding prediction includes: After being enhanced, the multi-scale features are fed into the hybrid encoder, where the attention mechanism is used to aggregate semantics within the same scale and perform feature fusion between different scales to generate high-level semantic features containing rich details of defects. The decoder interacts with image features through a set of learnable query vectors and finally outputs two branches through a fully connected layer: the classification branch outputs the category probability of the defect, and the regression branch outputs the center coordinates and width and height offsets of the defect bounding box.

[0012] Preferably, the low-frequency approximation component for:

[0013] Horizontal detail component for:

[0014] Vertical detail component for:

[0015] Diagonal detail component for:

[0016] in, This refers to the coordinate indices of the output feature map. It is a low-pass filter. For high-pass filters, For the input feature tensor, For channels, r represents rows, and c represents columns. This indicates the Kronecker product or filter cascade.

[0017] Preferably, the hybrid frequency domain characteristics are:

[0018] The enhanced multi-scale features output are as follows:

[0019] in, It is a non-linear activation function. Represents convolution operation. The statistical mean of the current training micro-batch data in the feature channel dimension, The statistical variance of the current training micro-batch data in the feature channel dimension. The smallest positive constant introduced to ensure the stability of numerical calculations. , For learnable affine parameters, For convolution kernel, It is a hybrid frequency domain feature.

[0020] Preferably, preprocessing the acquired raw weld images includes: The original weld images were cleaned and formatted. Defects in the images were labeled with rectangular boxes using a labeling tool to generate a standard dataset containing category labels and coordinate information. Mosaic enhancement, hybrid enhancement, and random photometric distortion were introduced for online data augmentation. The input image size was uniformly adjusted to the standard resolution required by the network input through adaptive image normalization. The pixel intensity was standardized to make the distribution meet the zero mean and unit variance.

[0021] Preferably, the deep detection model based on multi-resolution frequency domain orthogonal decomposition sampling includes: a feature extraction backbone network, a hybrid feature encoder, and a defect decoding prediction head; the feature extraction backbone network is embedded with a multi-resolution frequency domain orthogonal decomposition sampling module.

[0022] A pressure vessel weld defect detection system based on frequency domain orthogonal decomposition, based on the aforementioned pressure vessel weld defect detection method based on frequency domain orthogonal decomposition, includes: a sensing layer, an execution layer, and a calculation layer; The sensing layer is configured as a ray imaging sensing unit to acquire raw weld images; The execution layer is configured as an automated collaborative scanning unit, which is used for the execution elements to perform multi-degree-of-freedom collaboration based on the scan path and control script movement; The computing layer is configured as a central control and image processing unit, used to read the pressure vessel structural model and plan the scanning path of the actuators. It automatically calculates exposure parameters based on the weld thickness and generates a control script that includes the robotic arm posture and exposure timing. The acquired raw weld images are preprocessed and the model is trained. Based on the trained depth detection model based on multi-resolution frequency domain orthogonal decomposition sampling, frequency subbands are separated through two-dimensional orthogonal multi-resolution decomposition. Hybrid frequency domain features are constructed through tensor reconstruction of frequency domain information, and enhanced multi-scale features are obtained through cross-frequency band feature mapping. The category, location, and size of weld defects are predicted through hybrid feature encoding and decoding. Based on the detected defect type and size, the qualified condition is judged against preset standards, and an inspection report including a defect distribution heatmap is generated.

[0023] Preferably, the X-ray imaging sensing unit includes an amorphous silicon digital flat panel detector, an X-ray generator, and a synchronization controller; the X-ray generator is placed inside or on one side of the pressure vessel to emit X-rays, and the amorphous silicon digital flat panel detector is installed at the end of an industrial robot. The detector located at the corresponding position on the outside of the container receives the X-ray energy after penetrating the weld and converts it into a digital grayscale image. The automated collaborative scanning unit includes a six-axis industrial robot and a motion controller. The industrial robot is the motion carrier of the amorphous silicon digital flat panel detector. It receives scanning path instructions from the central control and image processing unit, drives the detector to perform a fitting motion along the weld seam trajectory, and maintains a preset focal length and alignment relationship with the X-ray generator. The central control and image processing unit interfaces with the factory's MES or PACS system to upload test reports.

[0024] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for detecting weld defects in pressure vessels based on frequency domain orthogonal decomposition, which has the following beneficial effects: By adopting a detection algorithm based on frequency domain orthogonal decomposition sampling, the industry problem of detecting microcracks and pinholes in pressure vessel welds is solved. In practical applications, it helps to avoid high-pressure vessel leakage or explosion accidents caused by missed detection of micro defects, and greatly improves the safety of pressure vessels before leaving the factory and during service. This invention can effectively distinguish between normal fish scale texture of welds and abnormal defect signals, solving the problem of high false alarm rate of traditional visual inspection on complex weld surfaces. In automated production lines, it helps to reduce downtime for inspection or manual secondary verification caused by misjudgment, and improves the continuity of production cycle. This invention replaces manual evaluation based on subjective experience, establishes an objective and unified defect judgment standard, and achieves full lifecycle digital traceability of weld quality data through seamless integration with the MES system, providing reliable data support for process optimization and quality control of pressure vessels. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0026] Figure 1 This is a flowchart of a pressure vessel weld defect detection method based on frequency domain orthogonal decomposition provided in an embodiment of the present invention; Figure 2 This is a flowchart of the depth detection model prediction process based on multi-resolution frequency domain orthogonal decomposition sampling provided in an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] This invention discloses a method for detecting weld defects in pressure vessels based on frequency domain orthogonal decomposition, such as... Figure 1 As shown, it includes the following steps: S1. Read the pressure vessel structural model and plan the scanning path of the actuator. Automatically calculate the exposure parameters based on the weld thickness and generate a control script that includes the robot arm posture and exposure timing. S2. The actuator performs multi-degree-of-freedom collaborative scanning based on the scan path and control script motion to acquire the original weld seam image; S3. The acquired raw weld images are preprocessed and the model is trained. Based on the trained depth detection model based on multi-resolution frequency domain orthogonal decomposition sampling, the frequency sub-bands are separated by two-dimensional orthogonal multi-resolution decomposition. The hybrid frequency domain features are constructed by tensor reconstruction of frequency domain information. The enhanced multi-scale features are obtained by cross-band feature mapping. The category, location and size of the weld defect are predicted by hybrid feature encoding and decoding. S4. Based on the detected defect type and size, determine the pass / fail status by comparing with preset standards, and generate an inspection report containing a defect distribution heatmap.

[0029] To further implement the above technical solution, the specific method for acquiring the original weld image in step S2 is as follows: The actuator executes the control script and moves along the weld seam trajectory. During the movement, based on the real-time feedback of position information, the camera is triggered to expose and collect weld seam images every time a preset step is reached. After marking the current robot arm posture coordinates, the images are stored in the high-speed cache queue.

[0030] To further implement the above technical solutions, such as Figure 2 As shown, step S3, the method for obtaining the enhanced multi-scale features, specifically includes: Using a set of conjugate orthogonal mirror filter banks, convolution and intermittent sampling operations are performed on each channel of the input feature tensor in the row and column directions, respectively, to output a frequency sub-band feature map, including four frequency sub-bands: low-frequency approximate component, horizontal detail component, vertical detail component and diagonal detail component; The separated frequency information is fused into the convolutional network, and a channel-dimensional tensor concatenation operation is performed to construct hybrid frequency domain features. The convolution kernel is used to linearly combine the mixed frequency domain features, and then batch normalization and nonlinear activation function are used to map them back to the target channel space to output enhanced multi-scale features.

[0031] To further implement the above technical solution, the specific content of hybrid feature encoding and decoding prediction includes: After being enhanced, the multi-scale features are fed into the hybrid encoder, where the attention mechanism is used to aggregate semantics within the same scale and perform feature fusion between different scales to generate high-level semantic features containing rich details of defects. The decoder interacts with image features through a set of learnable query vectors and finally outputs two branches through a fully connected layer: the classification branch outputs the category probability of the defect, and the regression branch outputs the center coordinates and width and height offsets of the defect bounding box.

[0032] To further implement the above technical solution, the low-frequency approximation component... for:

[0033] Horizontal detail component for:

[0034] Vertical detail component for:

[0035] Diagonal detail component for:

[0036] in, This refers to the coordinate indices of the output feature map. It is a low-pass filter. For high-pass filters, For the input feature tensor, These represent height, width, and number of channels, respectively. For channels, r represents rows, and c represents columns. This indicates the Kronecker product or filter cascade.

[0037] This embodiment addresses the problem of high-frequency information loss in traditional downsampling for microcracks in pressure vessels. It designs a frequency domain downsampling operator based on two-dimensional discrete wavelet transform (2D-DWT), and achieves higher resolution through two-dimensional orthogonal multiresolution decomposition. arrive It achieves physical dimensionality reduction without losing frequency domain information.

[0038] To further implement the above technical solution, the hybrid frequency domain characteristics are as follows:

[0039] The enhanced multi-scale features output are as follows:

[0040] in, It is a non-linear activation function. Represents convolution operation. The statistical mean of the current training micro-batch data in the feature channel dimension, This represents the statistical variance of the current training micro-batch data along the feature channel dimension. The smallest positive constant introduced to ensure the stability of numerical calculations. , For learnable affine parameters, For convolution kernel, For mixed frequency domain features, the dimension of the mixed frequency domain feature tensor is... .

[0041] To further implement the above technical solution, the preprocessing of the acquired raw weld images includes: The acquired raw weld images are cleaned and formatted. Through quality screening, motion-blurred or overexposed samples are removed. Defects in the images are labeled with rectangular boxes using a labeling tool to generate a standard dataset containing category labels and coordinate information. Mosaic enhancement, hybrid enhancement, and random photometric distortion are introduced for online data augmentation. Adaptive image normalization is used to uniformly adjust the input image size to the standard resolution required by the network input. Pixel intensity is standardized to ensure that the distribution meets the zero mean and unit variance.

[0042] In this embodiment, the method for standardizing pixel intensity is as follows:

[0043] in, and These represent the channel mean and standard deviation of the dataset, respectively.

[0044] To further implement the above technical solution, the deep detection model based on multi-resolution frequency domain orthogonal decomposition sampling includes: a feature extraction backbone network, a hybrid feature encoder, and a defect decoding prediction head; the feature extraction backbone network is embedded with a multi-resolution frequency domain orthogonal decomposition sampling module to replace the general max pooling layer or stride convolutional layer.

[0045] In this embodiment, a multi-task joint loss function is used during the training phase of the depth detection model based on multi-resolution frequency domain orthogonal decomposition sampling. Perform end-to-end optimization:

[0046] in, For classification loss, weighted binary cross-entropy or focal loss is used to address the imbalance between positive and negative samples. For regression loss, combine L1 distance loss. Compared with the generalized intersection and comparison loss This ensures that the predicted bounding box highly overlaps with the actual defect bounding box in both location and scale.

[0047] A pressure vessel weld defect detection system based on frequency domain orthogonal decomposition, which is based on a pressure vessel weld defect detection method based on frequency domain orthogonal decomposition, includes: a sensing layer, an execution layer, and a calculation layer; The sensing layer is configured as a ray imaging sensing unit to acquire raw weld images; The execution layer is configured as an automated collaborative scanning unit, which is used for the execution elements to perform multi-degree-of-freedom collaboration based on the scan path and control script movement; The computing layer is configured as a central control and image processing unit, used to read the pressure vessel structural model and plan the scanning path of the actuators. It automatically calculates exposure parameters based on the weld thickness and generates a control script that includes the robotic arm posture and exposure timing. The acquired raw weld images are preprocessed and the model is trained. Based on the trained depth detection model based on multi-resolution frequency domain orthogonal decomposition sampling, frequency subbands are separated through two-dimensional orthogonal multi-resolution decomposition. Hybrid frequency domain features are constructed through tensor reconstruction of frequency domain information, and enhanced multi-scale features are obtained through cross-frequency band feature mapping. The category, location, and size of weld defects are predicted through hybrid feature encoding and decoding. Based on the detected defect type and size, the qualified condition is judged against preset standards, and an inspection report including a defect distribution heatmap is generated.

[0048] To further implement the above technical solution, the X-ray imaging sensing unit includes an amorphous silicon digital flat panel detector, an X-ray generator, and a synchronization controller; the X-ray generator is placed inside or on one side of the pressure vessel to emit X-rays, and the amorphous silicon digital flat panel detector is installed at the end of an industrial robot. The detector located at the corresponding position on the outside of the container receives the X-ray energy after penetrating the weld and converts it into a digital grayscale image. The automated collaborative scanning unit includes a six-axis industrial robot and a motion controller. The industrial robot is the motion carrier of the amorphous silicon digital flat panel detector. It receives scanning path instructions from the central control and image processing unit, drives the detector to perform a fitting motion along the weld seam trajectory, and maintains a preset focal length and alignment relationship with the X-ray generator. The central control and image processing unit interfaces with the factory's MES or PACS system to upload test reports.

[0049] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for detecting weld defects in pressure vessels based on frequency domain orthogonal decomposition.

[0050] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements a method for detecting weld defects in pressure vessels based on frequency domain orthogonal decomposition.

[0051] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0052] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting defects in a weld of a pressure vessel based on frequency domain orthogonal decomposition, characterized by, The method comprises the following steps: S1. Read the pressure vessel structure model and plan the execution element scanning path, automatically calculate the exposure parameters according to the weld thickness, and generate a control script containing the mechanical arm posture and exposure timing; S2. The execution element moves based on the scanning path and control script, performs multi-degree-of-freedom collaborative scanning, and collects original weld images; S3. The collected original weld images are preprocessed and trained, based on the trained deep detection model based on multi-resolution frequency domain orthogonal decomposition sampling, the two-dimensional orthogonal multi-resolution decomposition is used to separate the frequency subbands, the tensor reorganization of the frequency domain information is used to construct the mixed frequency domain features, and the cross-frequency band feature mapping is used to obtain the enhanced multi-scale features, and the mixed feature encoding and decoding prediction is used to output the category, position and size of the weld defects; S4. According to the detected defect type and size, the pre-set standard is used for qualified judgment, and a detection report containing a defect distribution heat map is generated.

2. The method for detecting a weld defect of a pressure vessel based on frequency domain orthogonal decomposition according to claim 1, wherein, The specific method for collecting original weld images in step S2 is: The execution element executes the control script and moves along the weld trajectory. During the movement, based on the real-time feedback position information, every time a preset step is reached, the camera is triggered to collect weld images, and the current mechanical arm posture coordinates are marked and stored in the cache queue.

3. The method for detecting a weld defect of a pressure vessel based on frequency domain orthogonal decomposition according to claim 1, wherein, The method for obtaining the enhanced multi-scale features in step S3 specifically includes: A set of conjugate orthogonal mirror filter banks are used to perform convolution and point sampling operations on each channel of the input feature tensor in the row and column directions respectively, outputting frequency subband feature maps including low-frequency approximation components, horizontal detail components, vertical detail components and diagonal detail components; The separated frequency information is fused into the convolution network to perform channel dimension tensor splicing operations to construct mixed frequency domain features; The mixed frequency domain features are linearly combined using a convolution kernel, and are mapped back to the target channel space through batch normalization and a nonlinear activation function to output enhanced multi-scale features.

4. The method for detecting a weld defect of a pressure vessel based on frequency domain orthogonal decomposition according to claim 3, wherein, The specific content of the mixed feature encoding and decoding prediction includes: The enhanced multi-scale features enter the mixed encoder, use the attention mechanism to aggregate semantics within the same scale, and fuse features between different scales to generate high-level semantic features containing rich defect details; The decoder interacts with the image features through a set of learnable query vectors, and finally outputs two branch results through a fully connected layer, the classification branch outputs the category probability of the defect, and the regression branch outputs the center coordinates and height and width offset of the defect bounding box.

5. The method for detecting a weld defect of a pressure vessel based on frequency domain orthogonal decomposition according to claim 3, wherein, Low frequency approximation component is: horizontal detail component is: Vertical detail component Is: diagonal detail component is: wherein, is a coordinate index of the output feature map, is a low-pass filter, is a high-pass filter, is an input feature tensor, is a channel, r is a row, and c is a column, denotes a Kronecker product or filter concatenation.

6. A method of detecting a weld defect of a pressure vessel based on frequency domain orthogonal decomposition according to claim 4, wherein, The mixed frequency domain features are: The output enhanced multi-scale features are: wherein, is a non-linear activation function, represents a convolution operation, is a statistical mean of the current training mini-batch data over the feature channel dimension, is a statistical variance of the current training mini-batch data over the feature channel dimension is a small positive number introduced for numerical stability, , is a learnable affine parameter, is a convolution kernel, is a mixed frequency domain feature.

7. The method of claim 1, wherein the method is characterized by: The preprocessing of the collected original weld images includes: Data cleaning and formatting are performed on the collected original weld images, a labeling tool is used to label the defects in the images with a rectangular frame, a standard data set containing category labels and coordinate information is generated, mosaic enhancement, hybrid enhancement and random photometric distortion are introduced for online data enhancement, adaptive image normalization is used to uniformly adjust the input image size to the standard resolution required by the network input, and the pixel intensity is standardized to make the distribution meet the zero mean and unit variance.

8. The method of claim 1, wherein the method is characterized by: The deep detection model based on multi-resolution frequency domain orthogonal decomposition sampling includes: a feature extraction backbone network, a hybrid feature encoder, and a defect decoding prediction head; the feature extraction backbone network is embedded with the multi-resolution frequency domain orthogonal decomposition sampling module.

9. A frequency domain orthogonal decomposition based pressure vessel weld defect detection system, characterized by, A method for detecting weld defects in pressure vessels based on frequency domain orthogonal decomposition according to any one of claims 1-8, comprising: a sensing layer, an execution layer, and a calculation layer; The sensing layer is configured as a ray imaging sensing unit to acquire raw weld images; The execution layer is configured as an automated collaborative scanning unit, which is used for the execution elements to perform multi-degree-of-freedom collaboration based on the scan path and control script movement; The computing layer is configured as a central control and image processing unit, used to read the pressure vessel structural model and plan the scanning path of the actuators. It automatically calculates exposure parameters based on the weld thickness and generates a control script that includes the robotic arm posture and exposure timing. The acquired raw weld images are preprocessed and the model is trained. Based on the trained depth detection model based on multi-resolution frequency domain orthogonal decomposition sampling, frequency subbands are separated through two-dimensional orthogonal multi-resolution decomposition. Hybrid frequency domain features are constructed through tensor reconstruction of frequency domain information, and enhanced multi-scale features are obtained through cross-frequency band feature mapping. The category, location, and size of weld defects are predicted through hybrid feature encoding and decoding. Based on the detected defect type and size, the qualified condition is judged against preset standards, and an inspection report including a defect distribution heatmap is generated.

10. A frequency domain orthogonal decomposition based pressure vessel weld defect detection system as claimed in claim 9, wherein, The X-ray imaging sensing unit includes an amorphous silicon digital flat panel detector, an X-ray generator, and a synchronization controller. The X-ray generator is placed inside or on one side of the pressure vessel to emit X-rays. The amorphous silicon digital flat panel detector is installed at the end of an industrial robot. The detector located at the corresponding position on the outside of the container receives the X-ray energy after penetrating the weld and converts it into a digital grayscale image. The automated collaborative scanning unit includes a six-axis industrial robot and a motion controller. The industrial robot is the motion carrier of the amorphous silicon digital flat panel detector. It receives scanning path instructions from the central control and image processing unit, drives the detector to perform a fitting motion along the weld seam trajectory, and maintains a preset focal length and alignment relationship with the X-ray generator. The central control and image processing unit interfaces with the factory's MES or PACS system to upload test reports.

Images