Weld joint straightness evaluation method and system based on deep learning

Abstract

The invention provides a method and a system for evaluating the flatness of a welding seam based on deep learning. The method comprises the following steps of: acquiring marking data of a center line of the welding seam through image processing and edge data calibration; fusing the marking data of the center line of the welding seam and the geometric characteristic parameters of the welding seam to construct and obtain structured characteristic data; inputting the structured feature data to a deep learning model to calculate and obtain welding seam straightness deviation data; and comparing the flatness deviation data with an industrial standard, and generating a welding seam flatness evaluation report. According to the method, the center line of the welding seam is accurately positioned through image processing and edge calibration, structural data are constructed by fusing geometric features, the straightness deviation is calculated by adopting a deep learning model, and a quality report is generated, so that the problems of low efficiency and poor precision of a traditional method are solved, non-contact automatic detection is realized, the measurement precision and efficiency are remarkably improved, and the detection efficiency is improved. And the system has strong environment adaptability and system docking capability.

Description

Technical Field

[0001] This application relates to the field of welding technology, specifically to a method and system for evaluating weld straightness based on deep learning. Background Technology

[0002] Currently, the inspection of rail weld straightness in the railway engineering field mainly relies on manual measurement methods and portable equipment, which has prominent problems such as low inspection efficiency, insufficient accuracy, and poor adaptability.

[0003] Traditional methods use a 1m steel ruler and feeler gauge for contact measurement. This method can only estimate the joint height difference by the number of steel strips inserted, and cannot obtain key geometric parameters such as the wavelength and range of the weld. Moreover, the measurement results are heavily dependent on the operator's experience. Although electronic straightedges can measure both the height and wavelength of unevenness simultaneously, their detection range is limited, requiring frequent equipment movement to accommodate long-distance weld inspections, resulting in low work efficiency.

[0004] To improve inspection efficiency, engineering departments often use handcart-type inspection equipment such as CAT and BBM. However, these devices are mainly designed for short-wave irregularities such as rail corrugation. Rail welds are typically distributed at 100m (or 25m) intervals, and their geometric features are easily masked by short-wave noise from the track (such as corrugation and seamed joints), resulting in insufficient signal-to-noise ratio for weld-specific inspections. Furthermore, existing methods generally suffer from three technical bottlenecks: First, it is impossible to achieve continuous automated detection of weld straightness; Secondly, it lacks robustness to complex working conditions (such as corrosion and oil contamination); Third, the measurement data is difficult to directly interface with the digital maintenance system.

[0005] These problems severely restrict the precise control of rail weld quality, and there is an urgent need for an innovative technical solution that is non-contact, high-precision, and adaptable to long-distance inspection. Summary of the Invention

[0006] This application provides a deep learning-based method and system for evaluating weld straightness, which can solve the technical problems of low efficiency, insufficient accuracy, poor adaptability, and inability to achieve automated continuous detection in the existing technology.

[0007] In a first aspect, this application provides a deep learning-based method for evaluating weld straightness, comprising the following steps: Weld centerline annotation data is obtained through image processing and edge data calibration; Structured feature data is constructed by integrating weld centerline annotation data with weld geometric feature parameters; Input structured feature data into a deep learning model to calculate and obtain weld straightness deviation data; By comparing the straightness deviation data with industry standards, a weld straightness assessment report is generated.

[0008] Furthermore, the acquisition of weld centerline annotation data through image processing and edge data calibration specifically includes the following steps: Acquire raw image data by capturing images of the rail weld area; Edge data is obtained by extracting weld contour features from the original image; Calibrate edge data to correct coordinate errors and obtain centerline coordinates; Based on the centerline coordinates, combined with recorded environmental parameters and mileage information, weld centerline annotation data containing multi-dimensional features is obtained.

[0009] Furthermore, the process of fusing the weld centerline annotation data with the weld geometric feature parameters to construct structured feature data specifically includes the following steps: The coordinates of the weld centerline are normalized, and the normalized coordinate data is obtained. By fusing normalized coordinates and mileage information, coordinate features with mileage information are obtained; weld width and inclination angle parameters are calculated based on weld edge contour features to obtain weld geometric feature parameters. The weld locations are coded to obtain coding features with spatial relationships; Structured feature data is constructed by integrating coordinate features with mileage information, weld geometric feature parameters, and coded features with spatial relationships.

[0010] Furthermore, the input of structured feature data to the deep learning model to calculate and obtain weld straightness deviation data specifically includes the following steps: Input structured feature data into the U-Net network architecture, and perform multi-scale feature extraction through the encoder-decoder structure to obtain multi-level feature representations; Based on multi-level feature representation, a channel attention mechanism is embedded in the residual block to dynamically weight the key geometric feature responses and obtain optimized features; A smooth loss function is used to train and optimize the features to predict the straightness deviation of the weld.

[0011] Furthermore, the multi-level feature representation includes low-level features of weld contour details, high-level features of global weld semantics, and multi-scale features.

[0012] Furthermore, the step of training and optimizing the features using a smoothing loss function to predict the straightness deviation of the weld seam specifically includes the following steps: Compare the prediction deviation with the actual value, calculate the preliminary error between the prediction deviation and the actual value, and obtain the original loss value; Based on the original loss value, the error is segmented according to a preset threshold to obtain the adjusted loss value. Based on the adjusted loss value, the squared loss is calculated for the small error region to obtain the optimized loss value for the accurate region; Based on the adjusted loss value, linear loss calculation is used for the large error region to obtain the optimized loss value for the robust region. The optimized loss values ​​of the precise region and the robust region are weighted and summed to obtain the final optimized flatness deviation data.

[0013] Furthermore, the comparison of straightness deviation data with industry standards to generate a weld straightness assessment report specifically includes the following steps: Compare the flatness deviation data with the industry standard threshold to obtain the weld quality grade judgment result; Generate a structured report containing deviation values, quality grade, and location information based on the weld quality grade determination results; Analyze historical data to determine the trend of weld quality changes and obtain trend prediction data; Integrate the judgment results, structured reports, and trend prediction data to generate a weld straightness assessment report.

[0014] Secondly, this application provides a deep learning-based weld straightness evaluation system, including: The annotation data acquisition module is used to acquire weld centerline annotation data through image processing and edge data calibration. The structured feature data acquisition module is communicatively connected to the annotation data acquisition module and is used to fuse the weld centerline annotation data and weld geometric feature parameters to construct structured feature data. The straightness deviation data acquisition module is communicatively connected to the structured feature data acquisition module and is used to input structured feature data into the deep learning model to calculate and obtain weld straightness deviation data. The evaluation report generation module is connected in communication with the straightness deviation data acquisition module to compare the straightness deviation data with industry standards and generate a weld straightness evaluation report.

[0015] Furthermore, the flatness deviation data acquisition module includes: The multi-level feature representation acquisition unit is used to input structured feature data into the U-Net network architecture, and to extract multi-scale features through the encoder-decoder structure to obtain multi-level feature representations. The optimized feature acquisition unit is communicatively connected to the multi-level feature representation acquisition unit. It is used to embed a channel attention mechanism in the residual block based on the multi-level feature representation, dynamically weight the key geometric feature response, and acquire the optimized features. The straightness deviation data acquisition unit is communicatively connected to the optimized feature acquisition unit and is used to train and optimize the optimized features using a smoothing loss function to predict the straightness deviation of the weld.

[0016] Furthermore, the evaluation report generation module includes: The grade determination result acquisition unit is used to compare the straightness deviation data with the industry standard threshold to obtain the weld quality grade determination result; The structured report acquisition unit is used to generate a structured report containing deviation values, quality grades, and location information based on the weld quality grade determination results. The trend prediction data acquisition unit is used to analyze the trend of weld quality changes based on historical data and acquire trend prediction data. The report generation unit is communicatively connected to the grade determination result acquisition unit, the trend prediction data acquisition unit, and the trend prediction data acquisition unit, and is used to integrate the determination result, the structured report, and the trend prediction data to generate a weld straightness evaluation report.

[0017] The beneficial effects of the technical solutions provided in this application include at least the following: Precise positioning of the weld centerline is achieved through image processing and edge data calibration. Structured feature data is constructed by combining weld geometric features. Straightness deviation is calculated using a deep learning model. Finally, a quality assessment report is generated by comparing with industry standards. This achieves non-contact automated inspection, significantly improving measurement accuracy and work efficiency. It also has strong adaptability to complex working conditions and the ability to interface with digital systems. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the deep learning-based weld straightness evaluation method of this application. Figure 2 This is a functional block diagram of the deep learning-based weld straightness evaluation system of this application. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0020] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0021] In the description of the embodiments in this application, terms such as "exemplary," "for example," or "for instance" are used as examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a specific manner.

[0022] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0023] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0025] Firstly, such as Figure 1 As shown, this application provides a deep learning-based method for evaluating weld straightness, including the following steps: Step S1: Obtain weld centerline annotation data through image processing and edge data calibration; Step S2: Integrate the weld centerline annotation data with the weld geometric feature parameters to construct structured feature data; Step S3: Input structured feature data into the deep learning model to calculate and obtain weld straightness deviation data; Step S4: Compare the straightness deviation data with industry standards to generate a weld straightness assessment report.

[0026] This application uses image processing and edge calibration to accurately locate the weld centerline, integrates geometric features to construct structured data, and uses a deep learning model to calculate the straightness deviation and generate a quality report. It solves the problems of low efficiency and poor accuracy of traditional methods, realizes non-contact automated inspection, significantly improves measurement accuracy and efficiency, and has strong environmental adaptability and system integration capabilities.

[0027] In one embodiment, step S1: obtaining weld centerline annotation data through image processing and edge data calibration specifically includes the following steps: Step S11: Acquire images of the rail weld area to obtain raw image data; specifically, use an industrial camera to acquire rail images; Step S12: Extract weld contour features from the original image to obtain edge data; Step S13: Calibrate edge data, correct coordinate errors, and obtain centerline coordinates; Step S14: Based on the centerline coordinates, and combined with the recorded environmental parameters and mileage information, obtain weld centerline annotation data containing multi-dimensional features.

[0028] This embodiment acquires rail images and combines edge detection with manual calibration to accurately locate the starting and ending coordinates of the weld centerline, generating weld centerline annotation data to provide reliable benchmark input for subsequent deep learning models.

[0029] In one embodiment, step S1 is specifically implemented as follows: First, a high-resolution industrial camera was used to photograph the rail weld area under stable lighting conditions to ensure that the image was clear and free of motion blur. Subsequently, edge detection algorithms such as Canny and Sobel were used to initially extract the weld contour, and Hough transform or sub-pixel level fitting techniques were combined to optimize the edge positioning accuracy and reduce noise interference. Because the surface of the rail may be corroded, worn, or subject to environmental interference (such as oil stains or reflections), pure algorithm detection may have errors. Therefore, a manual calibration process needs to be introduced, whereby professionals fine-tune the automatic detection results to ensure that the coordinate error of the start and end points of the weld centerline is controlled within ±0.1mm, meeting the requirements of industrial-grade measurement. The labeled data not only includes coordinate information, but also needs to record metadata such as the rail mileage position at the time of collection (e.g., offset relative to sleepers), imaging angle, and environmental parameters (e.g., temperature, humidity) for subsequent feature fusion. The final labeled dataset must undergo rigorous quality verification (e.g., repeatability testing, cross-validation) to ensure its reliability and consistency as a benchmark for model training.

[0030] The technical challenge of this step lies in balancing automation efficiency with human precision, while also adapting to robustness requirements under complex working conditions. The quality of its output directly determines the upper limit of prediction for subsequent models.

[0031] In one embodiment, step S2: fusing weld centerline annotation data with weld geometric feature parameters to construct structured feature data, specifically includes the following steps: Step S21: Normalize the coordinates of the weld centerline to eliminate scale differences and obtain normalized coordinate data; Step S22: Merge normalized coordinates and mileage information to perform feature fusion and obtain coordinate features with mileage information; Step S23: Calculate the weld width and inclination angle parameters based on the weld edge contour features to obtain the weld geometric feature parameters; Step S24: Encode the weld location to strengthen the spatial relationship and obtain the coding features with spatial relationship; Step S25: Integrate coordinate features with mileage information, weld geometric feature parameters (including weld width and weld inclination angle, etc.) and coded features with spatial relationships to construct structured feature data.

[0032] This embodiment transforms the labeled weld coordinates and weld geometric feature parameters into structured feature vectors, which are then input into a deep learning model for straightness prediction. The specific process includes key operations such as coordinate normalization, position encoding, and auxiliary feature fusion. Mathematical transformations enhance the model's understanding of spatial relationships, while physical parameters such as width and tilt angle are embedded to improve prediction accuracy.

[0033] In one embodiment, step S21: normalize the coordinates of the weld centerline to obtain normalized coordinate data, specifically implemented as follows: First, the starting point of the weld centerline. and the end point Normalization is performed to map it to the [0, 1] interval, thus eliminating the impact of differences in rail dimensions.

[0034] In the formula: , These are the original pixel coordinates of the start and end points of the weld, respectively; These are the minimum and maximum boundary coordinates of the effective detection region of the image, respectively. These are the normalized coordinates, used to eliminate scale differences.

[0035] In one embodiment, step S24: Position encoding is performed on the weld location to obtain encoding features with spatial relationships, specifically implemented as follows: Since deep learning models (such as Transformer or CNN) are not sensitive to absolute position information, positional encoding needs to be introduced to enhance the spatial distribution features of the weld. Specifically, a sine-cosine encoding method is used to embed the normalized coordinates in a high dimension, as shown in the following equation:

[0036] In the formula: This refers to the position of the weld point in the sequence; For dimensional indexing, ; The total dimension of the encoded vector; The output position encoding vector is used to enhance the model's perception of the weld geometry order.

[0037] In one embodiment, step S25: fusing coordinate features with mileage information, weld geometric feature parameters, and coded features with spatial relationships to construct structured feature data, specifically implemented as follows: To improve model robustness, the following weld geometric feature parameters are embedded with spatially related coding features, taking into account the actual working conditions of the rails: weld width Calculated and normalized using pixel distance; inclination The angle between the weld centerline and the horizontal direction is calculated using the following formula:

[0038] Mileage information 's' represents the absolute position of the rail weld relative to the sleeper (unit: meters), and needs to be standardized.

[0039] In the formula: and These are the mean and standard deviation of the training set mileage, respectively.

[0040] Finally, the above features are concatenated into the model input vector F:

[0041] in, These are the normalized x and y coordinates of the first coordinate point; These are the normalized x and y coordinates of the second coordinate point; Encode the initial position of the weld; The code represents the weld termination position; w is the weld width parameter; θ is the weld inclination angle parameter. This is a normalized weld length parameter.

[0042] In one embodiment, step S3, which involves inputting structured feature data into a deep learning model to calculate and obtain weld straightness deviation data, specifically includes the following steps: Step S31: Input structured feature data into the U-Net network architecture, and perform multi-scale feature extraction through the encoder-decoder structure to obtain multi-level feature representation; Step S32: Based on multi-level feature representation, embed a channel attention mechanism in the residual block, dynamically weight the key geometric feature response, and obtain the optimized features; Step S33: Use a smooth loss function to train and optimize the optimization features to predict the straightness deviation of the weld.

[0043] In one embodiment, the multi-level feature representation in step S31 includes low-level features of weld contour details, high-level features of global weld semantics, and multi-scale features.

[0044] In one embodiment, step S33: training and optimizing the optimization features using a smoothing loss function to predict the straightness deviation of the weld seam, specifically includes the following steps: Compare the prediction deviation with the actual value, calculate the preliminary error between the prediction deviation and the actual value, and obtain the original loss value; Based on the original loss value, the error is segmented according to a preset threshold to obtain the adjusted loss value. Based on the adjusted loss value, the squared loss is calculated for the small error region to obtain the optimized loss value for the accurate region; Based on the adjusted loss value, linear loss calculation is used for the large error region to obtain the optimized loss value for the robust region. The optimized loss values ​​of the precise region and the robust region are weighted and summed to obtain the final optimized flatness deviation data.

[0045] In one embodiment, step S3 is specifically implemented as follows: The core task of this step is to build a deep learning model that maps the structured feature vector F generated in step S2 to weld straightness deviation Δ. High-precision prediction is achieved by optimizing the network structure and training strategy. The specific implementation includes three key steps: model architecture design, input feature adaptation, and loss function optimization. The final output is straightness deviation data that can be used for engineering judgment.

[0046] 1. Model Input and Feature Adaptation The structured feature vector F output from step S2 contains normalized coordinates, position encoding, and weld geometric feature parameters, and needs to be adapted to the input requirements of deep learning models. For improved architectures based on U-Net or ResNet, the input needs to be transformed into a form suitable for processing by convolutional or fully connected layers. Spatial feature extraction (applicable to U-Net): Reshape the coordinates and positions into a 2D feature map M:

[0047] Where H and W are the feature map height and width, respectively; C is the number of channels, which is determined by the dimension of F.

[0048] The conversion formula is shown below:

[0049] In the formula: i, j are the pixel indices of the feature map; k is the channel index; F is the feature vector output in step S2.

[0050] Fully connected input (for ResNet): Directly flatten F into a vector Dimensionality reduction through fully connected layers:

[0051] In the formula: The feature vector output from step two (dimension n = len(F)); d is the hidden layer dimension; The feature vector after dimensionality reduction is nonlinearly introduced by the ReLU activation function (ReLU(z)=max(0,z)).

[0052] 2. Improve the model architecture The following U-Net and ResNet improvements are introduced on top of the standard U-Net or ResNet: 2.1 Multi-scale feature fusion (U-Net improvement): Add skip connections to the encoder-decoder architecture to fuse low-level details with high-level semantic features:

[0053] In the formula: The features of the encoder's l-th layer; This is the upsampled output of the (l+1)th layer of the decoder; This is an element-wise addition; This is the output of the decoder's layer l.

[0054] 2.2 Residual Attention Module (ResNet Improvement): Embedding a channel attention mechanism in the residual block enhances the response to key features:

[0055] In the formula: The input features are GAP, which stands for Global Average Pooling. , Here are the trainable weights; σ is the Sigmoid activation function. This is channel multiplication.

[0056] 3. Output Layer and Loss Function The model ultimately outputs the flatness deviation. During training, Huber loss is used to balance outlier effects:

[0057] In the formula: Flatness deviation predicted by the deep learning model (unit: mm); The threshold value is used.

[0058] In one embodiment, step S4: comparing the straightness deviation data with industry standards to generate a weld straightness assessment report, specifically includes the following steps: Compare the flatness deviation data with the industry standard threshold to obtain the weld quality grade judgment result; Generate a structured report containing deviation values, quality grade, and location information based on the weld quality grade determination results; Analyze historical data to determine the trend of weld quality changes and obtain trend prediction data; Integrate the judgment results, structured reports, and trend prediction data to generate a weld straightness assessment report.

[0059] In one embodiment, step S4 is specifically implemented as follows: Finally, the straightness deviation of the weld output by the model is decoded into actual coordinates, the centerline curve is fitted and the offset is calculated, the straightness is judged by comparing with industry standards, and a visual report is generated to support process optimization and quality traceability.

[0060] The core task of this step is to convert the normalized straightness deviation Δ predicted by the model in step S3 into an actual physical quantity, and to determine whether the weld quality is qualified through engineering standards, and finally generate a visualization report to guide process optimization.

[0061] First, based on the calibration parameters from step S1 (such as pixel scale and system error compensation value), the normalized deviation Δ of the model output is converted into the actual millimeter-level deviation. To ensure that the data is aligned with the physical world; Subsequently, using the weld start and end coordinates marked in step S1, the actual spatial curve of the weld centerline is fitted, and its maximum offset from the ideal straight line is calculated. The weld straightness is then compared with industry standards (such as ±0.3mm for high-speed rail rails) to determine whether it is up to standard. Finally, the system automatically generates a comprehensive report containing deviation values, judgment results, fitting curve comparison charts, and process suggestions, supporting quality traceability and production optimization. The entire process strictly relies on the prediction output of step S3 and the weld centerline annotation data of step S1 to ensure logical closed loop, while taking into account complex factors such as thermal deformation and inspection errors in actual engineering, so that the prediction results of the deep learning model have practical industrial application value.

[0062] This application provides a multimodal feature fusion method for weld characterization, and innovatively proposes a method for constructing multimodal weld features by integrating geometric coordinates, spatial position encoding, and process parameters. By normalizing the coordinates of weld images acquired by an industrial camera, and combining this with sine-cosine position encoding to enhance spatial relationship perception, auxiliary process parameters such as width and tilt angle are simultaneously embedded to construct a physically meaningful structured feature vector.

[0063] This application provides an attention-enhanced straightness prediction model, pioneering a dynamic assessment and closed-loop optimization mechanism for the health status of high-speed rail infrastructure. This application innovatively introduces a residual attention mechanism into the U-Net / ResNet architecture, proposing a dedicated deep learning model for weld straightness prediction. By embedding channel attention modules in the residual blocks, the importance of multi-source features is dynamically weighted, enabling the model to focus on key geometric features (such as the curvature variation of the weld centerline).

[0064] Secondly, such as Figure 2 As shown, this application provides a deep learning-based weld straightness evaluation system, including a labeled data acquisition module 100, a structured feature data acquisition module 200, a straightness deviation data acquisition module 300, and an evaluation report generation module 400. The labeled data acquisition module 100 is used to acquire weld centerline labeled data through image processing and edge data calibration. The structured feature data acquisition module 200 is communicatively connected to the labeled data acquisition module 100 and is used to fuse the weld centerline labeled data with weld geometric feature parameters to construct structured feature data. The straightness deviation data acquisition module 300 is communicatively connected to the structured feature data acquisition module 200 and is used to input the structured feature data into a deep learning model to calculate and obtain weld straightness deviation data. The evaluation report generation module 400 is communicatively connected to the straightness deviation data acquisition module 300 and is used to compare the straightness deviation data with industry standards to generate a weld straightness evaluation report.

[0065] In one embodiment, the flatness deviation data acquisition module includes: The multi-level feature representation acquisition unit is used to input structured feature data into the U-Net network architecture, and to extract multi-scale features through the encoder-decoder structure to obtain multi-level feature representations. The optimized feature acquisition unit is communicatively connected to the multi-level feature representation acquisition unit. It is used to embed a channel attention mechanism in the residual block based on the multi-level feature representation, dynamically weight the key geometric feature response, and acquire the optimized features. The straightness deviation data acquisition unit is communicatively connected to the optimized feature acquisition unit and is used to train and optimize the optimized features using a smoothing loss function to predict the straightness deviation of the weld.

[0066] In one embodiment, the evaluation report generation module includes: The grade determination result acquisition unit is used to compare the straightness deviation data with the industry standard threshold to obtain the weld quality grade determination result; The structured report acquisition unit is used to generate a structured report containing deviation values, quality grades, and location information based on the weld quality grade determination results. The trend prediction data acquisition unit is used to analyze the trend of weld quality changes based on historical data and acquire trend prediction data. The report generation unit is communicatively connected to the grade determination result acquisition unit, the trend prediction data acquisition unit, and the trend prediction data acquisition unit, and is used to integrate the determination result, the structured report, and the trend prediction data to generate a weld straightness evaluation report.

[0067] The functions of each module in the above-mentioned deep learning-based weld straightness evaluation system correspond to the steps in the above-mentioned deep learning-based weld straightness evaluation method embodiment, and their functions and implementation processes will not be described in detail here.

[0068] Thirdly, embodiments of this application provide a weld straightness measurement device based on deep learning. The weld straightness measurement device based on deep learning can be a personal computer (PC), laptop computer, server, or other device with data processing capabilities.

[0069] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces for interconnecting internal components of the deep learning-based weld straightness measurement equipment, as well as interfaces for interconnecting the deep learning-based weld straightness measurement equipment with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0070] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0071] The processor can be a general-purpose processor, which can call a deep learning-based weld straightness measurement program stored in memory and execute the deep learning-based weld straightness evaluation method provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the deep learning-based weld straightness measurement program is called can be referred to in the various embodiments of the deep learning-based weld straightness evaluation method of this application, and will not be repeated here.

[0072] Fourthly, embodiments of this application also provide a readable storage medium.

[0073] This application has a readable storage medium storing a deep learning-based weld straightness measurement program, wherein when the deep learning-based weld straightness measurement program is executed by a processor, it implements the steps of the deep learning-based weld straightness evaluation method described above.

[0074] The method implemented when the deep learning-based weld straightness measurement program is executed can be referred to in the various embodiments of the deep learning-based weld straightness evaluation method of this application, and will not be repeated here.

[0075] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0076] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0077] The above are merely preferred embodiments of this application and do not limit the scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the scope of protection of this application.

Claims

1. A deep learning-based method for evaluating weld straightness, characterized in that, Includes the following steps: Weld centerline annotation data is obtained through image processing and edge data calibration; Structured feature data is constructed by integrating weld centerline annotation data with weld geometric feature parameters; Input structured feature data into a deep learning model to calculate and obtain weld straightness deviation data; By comparing the straightness deviation data with industry standards, a weld straightness assessment report is generated.

2. The deep learning-based weld straightness evaluation method as described in claim 1, characterized in that, The process of obtaining weld centerline annotation data through image processing and edge data calibration specifically includes the following steps: Acquire raw image data by capturing images of the rail weld area; Edge data is obtained by extracting weld contour features from the original image; Calibrate edge data to correct coordinate errors and obtain centerline coordinates; Based on the centerline coordinates, combined with recorded environmental parameters and mileage information, weld centerline annotation data containing multi-dimensional features is obtained.

3. The deep learning-based weld straightness evaluation method as described in claim 1, characterized in that, The fusion of weld centerline annotation data and weld geometric feature parameters to construct structured feature data specifically includes the following steps: The coordinates of the weld centerline are normalized, and the normalized coordinate data is obtained. By fusing normalized coordinates and mileage information, coordinate features with mileage information are obtained; weld width and inclination angle parameters are calculated based on weld edge contour features to obtain weld geometric feature parameters. The weld locations are coded to obtain coding features with spatial relationships; Structured feature data is constructed by integrating coordinate features with mileage information, weld geometric feature parameters, and coded features with spatial relationships.

4. The deep learning-based weld straightness evaluation method as described in claim 1, characterized in that, The process of inputting structured feature data into a deep learning model to calculate weld straightness deviation data specifically includes the following steps: Input structured feature data into the U-Net network architecture, and perform multi-scale feature extraction through the encoder-decoder structure to obtain multi-level feature representations; Based on multi-level feature representation, a channel attention mechanism is embedded in the residual block to dynamically weight the key geometric feature responses and obtain optimized features; A smooth loss function is used to train and optimize the features to predict the straightness deviation of the weld.

5. The deep learning-based weld straightness evaluation method as described in claim 4, characterized in that, The multi-level feature representation includes low-level features of weld contour details, high-level features of global weld semantics, and multi-scale features.

6. The deep learning-based weld straightness evaluation method as described in claim 4, characterized in that, The method of training and optimizing the features using a smooth loss function to predict the straightness deviation of the weld seam specifically includes the following steps: Compare the prediction deviation with the actual value, calculate the preliminary error between the prediction deviation and the actual value, and obtain the original loss value; Based on the original loss value, the error is segmented according to a preset threshold to obtain the adjusted loss value. Based on the adjusted loss value, the squared loss is calculated for the small error region to obtain the optimized loss value for the accurate region; Based on the adjusted loss value, linear loss calculation is used for the large error region to obtain the optimized loss value for the robust region. The optimized loss values ​​of the precise region and the robust region are weighted and summed to obtain the final optimized flatness deviation data.

7. The deep learning-based weld straightness evaluation method as described in claim 1, characterized in that, The comparison of straightness deviation data with industry standards generates a weld straightness assessment report, which specifically includes the following steps: Compare the flatness deviation data with the industry standard threshold to obtain the weld quality grade judgment result; Generate a structured report containing deviation values, quality grade, and location information based on the weld quality grade determination results; Analyze historical data to determine the trend of weld quality changes and obtain trend prediction data; Integrate the judgment results, structured reports, and trend prediction data to generate a weld straightness assessment report.

8. A deep learning-based weld straightness evaluation system, characterized in that, include: The annotation data acquisition module is used to acquire weld centerline annotation data through image processing and edge data calibration. The structured feature data acquisition module is communicatively connected to the annotation data acquisition module and is used to fuse the weld centerline annotation data and weld geometric feature parameters to construct structured feature data. The straightness deviation data acquisition module is communicatively connected to the structured feature data acquisition module and is used to input structured feature data into the deep learning model to calculate and obtain weld straightness deviation data. The evaluation report generation module is connected in communication with the straightness deviation data acquisition module to compare the straightness deviation data with industry standards and generate a weld straightness evaluation report.

9. The deep learning-based weld straightness measurement system as described in claim 8, characterized in that, The flatness deviation data acquisition module includes: The multi-level feature representation acquisition unit is used to input structured feature data into the U-Net network architecture, and to extract multi-scale features through the encoder-decoder structure to obtain multi-level feature representations. The optimized feature acquisition unit is communicatively connected to the multi-level feature representation acquisition unit. It is used to embed a channel attention mechanism in the residual block based on the multi-level feature representation, dynamically weight the key geometric feature response, and acquire the optimized features. The straightness deviation data acquisition unit is communicatively connected to the optimized feature acquisition unit and is used to train and optimize the optimized features using a smoothing loss function to predict the straightness deviation of the weld.

10. The deep learning-based weld straightness evaluation system as described in claim 8, characterized in that, The assessment report generation module includes: The grade determination result acquisition unit is used to compare the straightness deviation data with the industry standard threshold to obtain the weld quality grade determination result; The structured report acquisition unit is used to generate a structured report containing deviation values, quality grades, and location information based on the weld quality grade determination results. The trend prediction data acquisition unit is used to analyze the trend of weld quality changes based on historical data and acquire trend prediction data. The report generation unit is communicatively connected to the grade determination result acquisition unit, the trend prediction data acquisition unit, and the trend prediction data acquisition unit, and is used to integrate the determination result, the structured report, and the trend prediction data to generate a weld straightness evaluation report.

Images