A method and system for identifying defects in high-power laser-arc hybrid welding process

Abstract

The present invention belongs to the field of welding technology, and discloses a method for identifying defects in a high-power laser-arc hybrid welding process. The present invention proposes a high-power laser-arc hybrid welding process defect monitoring system based on the top visual signal of the welding process. This system automatically extracts the characteristics of the high-power laser-arc hybrid welding process based on high-speed camera and deep learning methods, aiming to solve the problem of low quality monitoring accuracy caused by strong noise interference in the high-power hybrid welding process and frequent target overlapping behavior. It is expected to be suitable for the quality monitoring scenario of the high-power laser-arc hybrid welding process of medium and thick-walled components. Improve monitoring accuracy: By combining the top visual signal of high-speed camera and deep learning technology, the present invention effectively solves the strong noise interference and target overlapping behavior in the high-power laser-arc hybrid welding process.

Description

Technical Field

[0001] The present invention belongs to the technical field of welding, and in particular relates to a method for identifying defects in a high-power laser-arc hybrid welding process. Background Art

[0002] High-power laser-arc hybrid welding technology has the advantages of fast welding speed, deep penetration, strong gap adaptability, and good joint quality. It has become one of the key technologies for high-quality and efficient welding of medium-thick wall components. At present, high-power laser-arc hybrid welding of medium-thick wall components has problems such as a narrow process window and the proneness of defects such as incomplete penetration, bottom hump, and collapse under dynamic conditions. Based on this, the study of high-power laser-arc hybrid welding process defect monitoring technology is of great significance to ensure the welding quality of medium-thick wall components. Compared with single laser or single arc welding, the high-power laser-arc hybrid welding process of medium-thick wall components has a greater heat input, more types of heat sources, stronger interference factors such as spatter and metal vapor, and periodic overlapping behavior of targets such as arc, keyhole, and droplet. Noise and periodic dynamic overlapping phenomena make it difficult to extract target features, making it difficult to accurately monitor the quality of the hybrid welding process.

[0003] Currently, the monitoring targets for laser-arc hybrid welding are usually thin-walled components, and multiple high-speed cameras and other sensors are required to simultaneously capture the dynamic behavior of targets such as keyholes and metal vapor during the welding process to reduce the impact of signal interference on welding feature extraction and recognition. However, the joint use of multiple high-speed cameras and other sensors will affect the spatial adaptability of the equipment during the welding process. At the same time, when processing and identifying the signals collected during the hybrid welding of thin-walled components, artificial feature extraction methods based on feature engineering are usually adopted. In the process of high-power laser-arc hybrid welding quality monitoring of medium- and thick-walled components with diverse and complex working conditions, there are many signal interference factors, and manually extracted features are prone to problems such as poor adaptability and low degree of automation.

[0004] Through the above analysis, the problems and defects of the existing technology are as follows:

[0005] (1) The joint use of multiple high-speed cameras and other sensors will affect the spatial adaptability of the equipment during the welding process. At the same time, when processing and identifying the signals collected during the composite welding process of thin-walled components, artificial feature extraction methods based on feature engineering are usually used.

[0006] (2) In the process of high-power laser-arc hybrid welding quality monitoring of medium-thick wall components with diverse and complex working conditions, there are many signal interference factors, and manually extracted features are prone to problems such as poor adaptability and low degree of automation. Summary of the Invention

[0007] In view of the problems existing in the prior art, the present invention provides a method for identifying defects in a high-power laser-arc hybrid welding process.

[0008] The present invention is implemented as follows: a method for identifying defects in a high-power laser-arc hybrid welding process comprises:

[0009] Step 1: Conduct multiple welding experiments using a welding monitoring experimental platform to collect and acquire top visual signals of the high-power laser-arc welding process, including image data of the arc, keyhole, molten droplet, and metal vapor characteristics during the welding process.

[0010] Step 2: Obtain and record the weld morphology characteristics of the upper and lower surfaces of all welds after the welding test, and classify the welded workpieces into three weld types based on the weld morphology characteristics after welding, and obtain the category results of each weld surface forming quality;

[0011] Step 3: Using the pre-processing method of ROI selection, brightness adjustment and gamma change, the metal vapor and spatter interference in the original welding process image are filtered out to obtain the process image with periodic overlap of keyhole, arc and droplet targets;

[0012] Step 4: According to the morphology type of each weld after welding, the pre-processed welding process images of the weld are classified and labeled, and the dataset is divided into training set, validation set and test set;

[0013] The ratio of the training and validation sets is 7:3. The test set uses welding process data of welds that are different from the training and validation sets to verify the generalization ability of the model.

[0014] Step 5: Based on the high noise intensity of the welding process, a deep learning model and channel attention mechanism are used to build a welding process quality monitoring model;

[0015] Step 6: Use the quality monitoring model to train, verify, and test the high-power laser-arc welding process images to obtain three welding state types of the weld, and finally realize online monitoring of welding quality.

[0016] Furthermore, in step 1, a welding experiment of multiple welds is performed, and the parameters of each welding experiment are fixed.

[0017] Furthermore, the three welds in step 2 include: bottom incomplete penetration (I), good weld (II) and bottom hump (III).

[0018] Furthermore, the basis for pre-processing and retaining the arc and keyhole targets in step 3 is that the arc light and keyhole are closely related to the heat input of the arc and laser during the welding process, respectively.

[0019] Furthermore, in step 4, a monitoring model with good defect monitoring performance is obtained through the training set and the validation set, and a generalization performance test is performed on welds of the same three weld types but different process parameters.

[0020] Furthermore, the step 5 constructs and designs a channel attention mechanism module:

[0021] First, the input features are compressed by global maximum and average pooling to extract different feature information;

[0022] Then, the information is sent to the fully connected layer to fit the complex correlation between feature channels and compress the features. The number of neurons is s1 and s2 respectively, and the activation function of ReLU is used.

[0023] The compressed features connected in two ways are operated through the second fully connected layer, and the weight coefficients are obtained through the Sigmoid function to return the number of features to the original number of features;

[0024] Finally, the original features are scaled using the weight coefficients to generate a weighted feature map; specifically:

[0025] The global maximum pooling operation is expressed as

[0026]

[0027] in Represents Y generated by global maximum pooling max The cth value of Y max ∈R c , C is the channel dimension threshold of the input feature map X, X∈R H×W×C ,X=[x1,x2,…,x c ,…,x c ];x c (i, j) represents the value of the cth X at position (i, j); C, H and W are the channel, height and width of the feature map; the relationship between c and C is 1≤c≤C;

[0028] The fully connected layer and ReLU function are used to reduce the dimension of the feature vector and enhance the model representation effect respectively, which can be expressed as

[0029] G max =δ(W1Y max +b1)

[0030] The dimension of W1 is C / s1×C, s1 is the scaling parameter; the fully connected layer can reduce the number of channels, thereby reducing the amount of calculation, δ represents the ReLU function; Gmax The dimensions are 1×1×C;

[0031] The global average pooling can be expressed as

[0032]

[0033] in Represents Y generated by global maximum pooling max The cth value of Y max ∈R c , C is the channel dimension threshold of the input feature map X, X∈R H×W×C ,X=[x1,x2,…,x c ,…,x c ];x c (i, j) represents the value of the cth X at position (i, j); C, H and W are the channel, height and width of the feature map; the relationship between c and C is 1≤c≤C;

[0034] The fully connected layer and ReLU function are used to reduce the dimension of the feature vector and enhance the model representation effect respectively, which can be expressed as

[0035] G avg =δ(W2Y avg +b2)

[0036] The dimension of W1 is C / s2×C, s2 is the scaling parameter; the fully connected layer can reduce the number of channels, thereby reducing the amount of calculation, δ represents the ReLU function; G max The dimensions are 1×1×C;

[0037] In order to obtain compressed feature maps of different pooling operations, the outputs of two parallel feature compression operations are combined using a concatenation operation, which can increase the width of the network and improve the model accuracy; the equation can be expressed as:

[0038] V=f contact (G avg ,G max )

[0039] The dimension of V is 1×1×2C(s1+s2);

[0040] By using fully connected layers and activation functions, the output multiplication operation of the module is expressed as;

[0041]

[0042] Where W3 dimension is C×[(s1+s1) / s1s2], σ is the activation function;

[0043] The high-power laser-arc hybrid welding process defect recognition model constructed in step 4 is specifically:

[0044] The model uses DenseNet as the backbone network; the network consists of five layers of Dense Blocks. Dense Blocks interconnect all layers. The input of each layer in the backbone network comes from the feature maps of all previous layers, and the output of each layer is directly connected to the input of all subsequent layers. Batch normalization improves training stability and standardizes the input of previous layers.

[0045] The convolutional layer then extracts specific feature information from the data. The attention mechanism module is added to the Dense Block in the second and third layers so that each stage of the network can effectively learn important and overall welding process features.

[0046] Identification and classification are performed by performing global average pooling and inputting the ice normalized exponential function into the fully connected layer, and finally outputting the welding state type, the model training and verification process.

[0047] Another object of the present invention is to provide a high-power laser-arc hybrid welding process defect recognition system comprising:

[0048] Welding work platform, arc welding gun, wire feeder, laser head, shielding gas, fiber laser, image acquisition system,

[0049] The image acquisition system includes a high-speed camera, an acquisition card, an auxiliary light source and an image monitoring system;

[0050] The welding sample is fixed on the welding platform, the laser welding gun is fixed above the movable platform with an inclination angle of 5°, the angle between the arc welding gun and the welding workpiece is 40°, and the image acquisition system is fixed above the platform.

[0051] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the high-power laser-arc hybrid welding process defect identification method.

[0052] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the high-power laser-arc hybrid welding process defect identification method.

[0053] Another object of the present invention is to provide an information data processing terminal, which is used to implement the high-power laser-arc hybrid welding process defect identification system.

[0054] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0055] First, this invention proposes a high-power laser-arc hybrid welding process defect monitoring system based on top-view visual signals from the welding process. This system automatically extracts high-power laser-arc hybrid welding process features using high-speed videography and deep learning methods. This system aims to address the challenges of high noise interference and low quality monitoring accuracy caused by frequent target overlap during high-power hybrid welding. It is expected to be applicable to high-power laser-arc hybrid welding process quality monitoring scenarios for medium- and thick-walled components.

[0056] Improve monitoring accuracy: By combining high-speed camera top visual signals and deep learning technology, the present invention effectively solves the strong noise interference and target overlap behavior in the high-power laser-arc hybrid welding process. The use of pre-processing methods such as image gamma filtering not only reduces the noise in the image, but also reduces interference such as metal vapor, spatter and reflection, so that key features in the welding process (such as keyholes and arcs) can be more clearly identified. In addition, the constructed convolutional neural network model framework based on the channel attention mechanism further improves the model's accuracy in identifying welding defects, overcoming the problem of weak generalization ability of traditional monitoring methods in high-noise environments.

[0057] Efficient and rapid monitoring: By employing a highly effective deep learning model to monitor the welding process in real time, this invention can rapidly identify different weld conditions, such as incomplete penetration, moderate penetration, and bottom hump. This efficient and rapid monitoring capability not only reduces the need for manual inspection but also enables timely detection of problems during the welding process, significantly improving production efficiency and welding quality.

[0058] Strong adaptability and generalization: Because this method utilizes a method based on visual signals and deep learning technology, it is not only applicable to monitoring high-power laser-arc hybrid welding processes for medium and thick plates, but can also be extended to defect monitoring in other welding processes (such as laser welding and arc welding). This method's strong adaptability and generalization capabilities mean it can maintain high monitoring accuracy under varying welding environments and process parameters, thus promising broad application prospects.

[0059] Significant economic benefits: By reducing the incidence of weld defects, this invention helps reduce rework and scrap, thereby directly reducing production costs. Furthermore, rapid and accurate defect identification reduces downtime, improves production line efficiency, and further enhances the company's economic benefits.

[0060] By combining high-speed camera top visual signals with deep learning methods, the present invention not only significantly improves the accuracy and efficiency of high-power laser-arc hybrid welding process monitoring, but also has good adaptability and generalization capabilities, providing the welding manufacturing industry with a low-cost, high-performance, and easy-to-use high-quality welding process monitoring technology solution.

[0061] Second, this invention addresses the problem of defect identification during high-power laser-arc hybrid welding by proposing a deep learning-based approach. This approach significantly improves the accuracy and efficiency of weld defect identification through specific model construction and data processing procedures. The following is a detailed explanation of the technical problems solved by this invention and the technological advancements achieved:

[0062] During high-power laser-arc hybrid welding, various weld defects, such as incomplete bottom penetration and bottom humps, are prone to occur due to a variety of factors, including welding parameters, material properties, and equipment status. Traditional defect detection methods often rely on manual visual inspection or simple sensor detection. These methods are not only inefficient but also susceptible to human factors, resulting in inaccurate detection results. Therefore, how to accurately and quickly identify defects during high-power laser-arc hybrid welding has become a pressing technical challenge.

[0063] 1. Improved defect identification accuracy: This paper constructs a deep learning-based defect identification model that automatically learns and extracts characteristic information from the welding process, accurately identifying different types of welding defects. Compared to traditional methods, this model has greater generalization and robustness, adapting to defect identification tasks under varying process parameters and welding conditions.

[0064] 2. Improved defect identification efficiency: The method of this invention achieves real-time monitoring and defect identification of the welding process through automated data processing and analysis. Compared with manual visual inspection, the method of this invention can significantly shorten inspection time and improve production efficiency.

[0065] 3. Introducing a channel attention mechanism module: This invention introduces a channel attention mechanism module in model construction. By allocating weights to different feature channels, the model's attention to key features is enhanced, further improving the accuracy of defect recognition.

[0066] 4. Using DenseNet as the backbone network: DenseNet has powerful feature extraction capabilities and can effectively process complex data in the welding process. By combining the inter-layer connection method of Dense Block, the model can fully utilize the feature information of the previous layer and improve the recognition accuracy.

[0067] By incorporating deep learning technology and a specific model-building approach, this paper successfully addresses the technical challenges of defect identification during high-power laser-arc hybrid welding, achieving significant technological advancements. This not only improves welding quality and reduces production costs, but also provides strong technical support for the intelligent development of the welding industry.

[0068] Third, as auxiliary evidence for the inventiveness of the claims of the present invention, it is also reflected in the following important aspects:

[0069] (1) The expected benefits and commercial value of the technical solution of the present invention after transformation are:

[0070] In monitoring applications of this technical solution, the images of the welding process often contain repetitive monitoring target behavior and strong noise such as metal vapor. This method can efficiently extract key features from these complex scenes, thereby accurately determining the welding quality. In addition, this method is not only suitable for quality monitoring of laser-arc hybrid welding processes, but can also be widely applied to other welding scenarios, such as single laser welding or arc welding, where the thermal response is relatively mild, showing good versatility and adaptability.

[0071] (2) The technical solution of the present invention fills the technical gap in the industry at home and abroad:

[0072] This technical solution can monitor the laser-arc hybrid welding process by using a single process image. Existing monitoring technologies generally rely on feature engineering techniques based on manually defined features for process control. This method is highly subjective and lacks robustness and timeliness. In contrast, this technical solution uses a deep learning algorithm based on an attention mechanism to automatically identify and extract signal image features during the welding process. By assigning higher weights to key features through the attention mechanism, it significantly improves robustness and timeliness compared to traditional feature engineering methods.

[0073] (3) The technical solution of the present invention solves the technical problems that people have been eager to solve but have never been able to solve successfully:

[0074] Traditional welding process monitoring methods are limited by interference such as high-intensity noise, metal vapor, spatter, and reflections, making it difficult to achieve high-precision and real-time weld quality assessment. In addition, the particularities of high-power laser-arc welding, such as the periodic overlap of the keyhole and the arc, further increase the complexity of monitoring, making it difficult for traditional methods to accurately identify and classify welding defects. The present invention effectively overcomes the above-mentioned difficulties by introducing high-speed camera top visual signal acquisition and deep learning technology, combined with gamma filter preprocessing and a convolutional neural network model based on a channel attention mechanism, significantly improving the accuracy and real-time performance of defect monitoring during the welding process, and solving a technical challenge that the industry has long faced but has not been able to effectively solve.

[0075] (4) The technical solution of the present invention overcomes technical prejudice:

[0076] Traditional methods usually rely on simple image processing techniques and shallow learning models. These methods have limited generalization capabilities and accuracy when dealing with complex welding scenarios, especially when faced with unique interference in high-power laser-arc hybrid welding processes (such as strong noise, spatter, and reflections). Their performance is greatly reduced. The present invention effectively identifies and classifies specific defects in complex welding scenarios, such as incomplete penetration, moderate penetration, and bottom hump welding states, by adopting deep learning and channel attention mechanisms, demonstrating an effective overcoming of biases in traditional technologies. In addition, by using high-speed cameras combined with image preprocessing technologies such as gamma filtering, the present invention has made breakthroughs in reducing interference and improving image quality, which further illustrates the effectiveness of this technical solution in overcoming technical biases.

[0077] Fourth, the present invention provides a method for identifying defects in a high-power laser-arc hybrid welding process, which solves several technical problems in the prior art and achieves significant technological progress.

[0078] Technical problems solved by existing technologies:

[0079] 1. Process monitoring is difficult: Traditional welding quality monitoring usually relies on post-weld inspection, such as X-ray or ultrasonic inspection. These methods cannot achieve real-time monitoring and immediate feedback of the welding process.

[0080] 2. Image processing complexity: During high-power laser-arc welding, due to interference from factors such as the arc, keyholes, molten droplets, and metal vapor, the acquired image data is of low quality and difficult to use directly for quality assessment.

[0081] 3. Quality judgment standards are highly subjective: In traditional methods, the evaluation of welding quality often relies on the operator's experience and judgment, and lacks objective and unified evaluation standards, which limits the accuracy and consistency of the evaluation.

[0082] Significant technological advancements achieved:

[0083] 1. Real-time monitoring and preprocessing technology: By collecting the top visual signals of the high-power laser-arc welding process in real time and using preprocessing methods such as ROI selection, brightness adjustment, and gamma change, the interference of metal vapor and spatter is effectively filtered out, the usability of the image is improved, and high-quality input data is provided for subsequent image analysis.

[0084] 2. Deep Learning and Channel Attention Mechanism: A welding process quality monitoring model was constructed using a deep learning model and a channel attention mechanism. This approach not only handles images with high noise intensity during welding but also improves recognition accuracy by learning subtle features in the image, thereby achieving accurate identification of weld defects.

[0085] 3. Online Monitoring of Welding Quality: By integrating the aforementioned technologies, this invention enables online monitoring and real-time feedback of welding quality, significantly improving welding process control efficiency and weld product qualification rates. This is crucial for improving production efficiency, reducing costs, and ensuring product quality.

[0086] 4. Model generalization and wide applicability: By validating the model's generalization across different weld types, the technology's applicability and reliability across a variety of industrial applications are ensured, enabling its widespread application in industries such as aviation, automotive manufacturing, and shipbuilding.

[0087] In summary, this invention effectively addresses the monitoring challenges associated with high-power laser-arc hybrid welding by incorporating advanced image processing techniques and sophisticated machine learning algorithms, providing an efficient, reliable, and objective method for evaluating weld quality. These technological advancements not only enhance the automation and intelligence of welding processes but also significantly improve production efficiency and product quality in related industries. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 This is a flow chart of a defect identification method for a high-power laser-arc hybrid welding process provided by an embodiment of the present invention.

[0089] Figure 2 This is a framework diagram of the defect monitoring method based on the deep learning network of image denoising and channel attention mechanism provided by the implementation of the present invention.

[0090] Figure 3 3. It is a schematic diagram comparing the noise reduction effect of welding process pictures provided by the embodiment of the present invention.

[0091] Figure 4 It is a schematic diagram of the deep learning model structure based on the channel attention mechanism provided by the implementation of the present invention.

[0092] Figure 5 It is a schematic diagram of an online monitoring system device for a high-power laser-arc hybrid welding process defect identification method provided by the present invention.

[0093] Figure 6 This is an identification and classification effect diagram of the defect identification method in the high-power laser-arc hybrid welding process provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0094] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0095] Specific embodiments of the high-power laser-arc hybrid welding process defect identification method can be described based on the above characteristics and working principles. The following are two embodiments:

[0096] Example 1:

[0097] In the welding workshop of a certain automobile manufacturer, high-power laser-arc hybrid welding technology is used to connect body parts. To ensure weld quality, the workshop introduced the above-mentioned defect identification method.

[0098] First, a welding monitoring experimental platform was installed on the welding equipment to collect real-time image data of the welding process. These images contain key features of the welding process, such as the arc, keyhole, and droplet.

[0099] Subsequently, the technicians in the workshop used a preprocessing algorithm to process the collected image data, removing interference factors such as metal vapor and spatter, and obtaining a clear image of the welding process.

[0100] Next, technicians categorize and label the pre-processed welding process images based on the weld seam morphology. They then train a welding process quality monitoring model built using a deep learning model and a channel attention mechanism.

[0101] After training, the model was integrated into the welding equipment's control system, enabling real-time online monitoring of the welding process. During each welding session, the model analyzes images of the welding process, determines the weld's condition, and provides real-time feedback to the operator. If a defect is identified, the system immediately issues an alarm, prompting the operator to make adjustments or repairs.

[0102] Example 2:

[0103] In the welding workshop of a shipbuilding company, in order to improve the welding quality and efficiency of ship structures, high-power laser-arc hybrid welding technology was introduced and the above-mentioned defect identification method was applied.

[0104] First, a welding monitoring experimental platform is placed next to the welding station to capture real-time image data of the welding process. These image data are transmitted to a computer for preprocessing and feature extraction.

[0105] Using deep learning algorithms and a channel attention mechanism, technicians trained a welding process quality monitoring model tailored to the characteristics of ship welding. The model can accurately identify defects such as porosity and lack of fusion during the welding process.

[0106] During the actual welding process, the model analyzes the welding process image in real time. Once a defect is detected, an alarm is immediately issued, indicating the location and type of the defect. Based on this information, the operator can adjust the welding parameters or make repairs in a timely manner, thus avoiding defects and improving welding quality.

[0107] like Figure 1 As shown, a method for identifying defects in a high-power laser-arc hybrid welding process provided by an embodiment of the present invention includes the following steps:

[0108] S101: Conduct multiple welding experiments using a welding monitoring experimental platform to collect and acquire top visual signals of the high-power laser-arc welding process, and collect image data of the welding process including arc, keyhole, molten droplet, and metal vapor characteristics;

[0109] S102: Obtain and record the weld morphology characteristics of the upper and lower surfaces of all welds after the welding test, and classify the welded workpieces into three weld types based on the weld morphology characteristics after welding, and obtain the category result of each weld surface forming quality;

[0110] S103: Preprocessing methods such as selecting a region of interest (ROI), adjusting brightness, and changing gamma are used, such as Figure 3 As shown in the figure, the metal vapor and spatter interference in the original welding process image are filtered out, and a process image with periodic overlap of keyhole, arc and droplet is obtained;

[0111] S104: Classify and label the pre-processed welding process images of each weld according to the morphology type of each weld after welding, and divide the data set into a training set, a validation set, and a test set;

[0112] The ratio of the training and validation sets is 7:3. The test set uses welding process data of welds that are different from the training and validation sets to verify the generalization ability of the model.

[0113] S105: Based on the high noise intensity of the welding process, a welding process quality monitoring model is built using a deep learning model and a channel attention mechanism;

[0114] S106: The quality monitoring model is used to train, verify, and test high-power laser-arc welding process images to obtain three welding state types of the weld, ultimately achieving online monitoring of welding quality.

[0115] In S101 provided in the embodiment of the present invention, a welding experiment of multiple welds is performed, and the parameters of each welding experiment are fixed.

[0116] The three weld types in S102 provided in the embodiment of the present invention include: bottom incomplete penetration (I), good weld (II) and bottom hump (III).

[0117] The basis for pre-processing and retaining the arc and keyhole targets in S103 provided by the embodiment of the present invention is that the arc light and the keyhole are closely related to the heat input of the arc and the laser during the welding process, respectively.

[0118] In S104 provided by the embodiment of the present invention, a monitoring model with good defect monitoring performance is obtained through the training set and the validation set, and a generalization performance test is performed on welds of the same three weld types but with different process parameters.

[0119] S105 provided in an embodiment of the present invention constructs and designs a channel attention mechanism module:

[0120] First, the input features are compressed by global maximum and average pooling to extract different feature information;

[0121] Then, the information is sent to the fully connected layer to fit the complex correlation between feature channels and compress the features. The number of neurons is s1 and s2 respectively, and the activation function of ReLU is used.

[0122] The compressed features connected in two ways are operated through the second fully connected layer, and the weight coefficients are obtained through the Sigmoid function to return the number of features to the original number of features;

[0123] Finally, the original features are scaled using the weight coefficients to generate a weighted feature map; specifically:

[0124] The global maximum pooling operation is expressed as

[0125]

[0126] in Represents Y generated by global maximum pooling max The cth value of Y max ∈Rc , C is the channel dimension threshold of the input feature map X, X∈R H×W×C ,X=[x1,x2,…,x c ,…,x c ];x c (i, j) represents the value of the cth X at position (i, j); C, H and W are the channel, height and width of the feature map; the relationship between c and C is 1≤c≤C;

[0127] The fully connected layer and ReLU function are used to reduce the dimension of the feature vector and enhance the model representation effect respectively, which can be expressed as

[0128] G max =δ(W1Y max +b1)

[0129] The dimension of W1 is C / s1×C, s1 is the scaling parameter; the fully connected layer can reduce the number of channels, thereby reducing the amount of calculation, δ represents the ReLU function; G max The dimensions are 1×1×C;

[0130] The global average pooling can be expressed as

[0131]

[0132] in Represents Y generated by global maximum pooling max The cth value of Y max ∈R c , C is the channel dimension threshold of the input feature map X, X∈R H×W×C ,X=[x1,x2,…,x c ,…,x c ];x c (i, j) represents the value of the cth X at position (i, j); C, H and W are the channel, height and width of the feature map; the relationship between c and C is 1≤c≤C;

[0133] The fully connected layer and ReLU function are used to reduce the dimension of the feature vector and enhance the model representation effect respectively, which can be expressed as

[0134] G avg =δ(W2Y avg +b2)

[0135] The dimension of W1 is C / s2×C, s2 is the scaling parameter; the fully connected layer can reduce the number of channels, thereby reducing the amount of calculation, δ represents the ReLU function; G max The dimensions are 1×1×C;

[0136] In order to obtain compressed feature maps of different pooling operations, the outputs of two parallel feature compression operations are combined using a concatenation operation, which can increase the width of the network and improve the model accuracy; the equation can be expressed as:

[0137] V=f contact (G avg ,G max )

[0138] The dimension of V is 1×1×2C(s1+s2);

[0139] By using fully connected layers and activation functions, the output multiplication operation of the module is expressed as;

[0140]

[0141] Where W3 dimension is C×[(s1+s1) / s1s2], σ is the activation function;

[0142] The high-power laser-arc hybrid welding process defect recognition model constructed by S104 is specifically:

[0143] The model uses DenseNet as the backbone network; the network consists of five layers of Dense Blocks. Dense Blocks interconnect all layers. The input of each layer in the backbone network comes from the feature maps of all previous layers, and the output of each layer is directly connected to the input of all subsequent layers. Batch normalization improves training stability and standardizes the input of previous layers.

[0144] The convolutional layer then extracts specific feature information from the data. The attention mechanism module is added to the Dense Block in the second and third layers so that each stage of the network can effectively learn important and overall welding process features.

[0145] Identification and classification are performed by performing global average pooling and inputting the ice normalized exponential function into the fully connected layer, and finally outputting the welding state type, the model training and verification process.

[0146] like Figure 5 As shown, an embodiment of the present invention provides a high-power laser-arc hybrid welding process defect recognition system comprising:

[0147] Welding work platform, arc welding gun, wire feeder, laser head, shielding gas, fiber laser, image acquisition system,

[0148] The image acquisition system includes a high-speed camera, an acquisition card, an auxiliary light source and an image monitoring system;

[0149] The welding sample is fixed on the welding platform, the laser welding gun is fixed above the movable platform with an inclination angle of 5°, the angle between the arc welding gun and the welding workpiece is 40°, and the image acquisition system is fixed above the platform.

[0150] An embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the high-power laser-arc hybrid welding process defect identification method.

[0151] An embodiment of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the high-power laser-arc hybrid welding process defect identification method.

[0152] An embodiment of the present invention provides an information data processing terminal, which is used to implement the high-power laser-arc hybrid welding process defect recognition system.

[0153] The present invention is specifically implemented:

[0154] This embodiment is based on a self-built dataset of a high-power laser-arc hybrid welding process and adopts a defect monitoring method of a deep learning network based on image denoising and channel attention mechanism in the welding process according to the present invention;

[0155] Defect detection methods based on deep learning networks with image denoising and channel attention mechanisms, such as Figure 2 As shown, the following steps are included:

[0156] A method for identifying defects in a high-power laser-arc hybrid welding process mainly comprises the following steps:

[0157] Step 1: Conduct multiple welding experiments using the welding monitoring experimental platform to collect and acquire top visual signals of the high-power laser-arc welding process, and obtain welding process image data including targets such as arc, keyhole, molten droplet, and metal vapor during the welding process.

[0158] Step 2: Obtain and record the weld morphology characteristics of the upper and lower surfaces of all welds after the welding test, and classify the welded workpieces into three weld types based on the weld morphology characteristics after welding, and obtain the category results of each weld surface forming quality;

[0159] Step 3: Use pre-processing methods such as region of interest (ROI) selection, brightness adjustment and gamma change, such as Figure 3 As shown in the figure, the interferences such as metal vapor and splash in the original process image are filtered out, and a process image with overlapping periodic targets such as small holes, arcs and droplets is obtained.

[0160] Step 4: According to the morphology type of each weld after welding, the welding process images of the preprocessed weld are classified and labeled, and the dataset is divided into training set, validation set and test set. The training and validation set division ratio is 7:3. At the same time, the test set uses welding process data of welds different from the training and validation sets to verify the generalization ability of the model.

[0161] Step 5: Based on the high noise intensity of the high-power laser-arc hybrid welding process, a deep learning model and channel attention mechanism are used to build a welding process quality monitoring model.

[0162] Step 6: Use the quality monitoring model to train, verify, and test the high-power laser-arc welding process images to obtain three welding state types of the weld, and finally realize online monitoring of welding quality.

[0163] In step 1, a welding experiment of multiple welds was carried out. The parameters of each welding experiment were fixed, so the heat input during the welding process was certain. At the same time, there was a certain gap in the settings of each heat input, so the weld type of each weld after welding was certain and unique.

[0164] The three weld types in step 2 include: bottom incomplete penetration (I), good weld (II), and bottom hump (III), and multiple welding experiments with different process parameters were conducted for each weld type;

[0165] The basis for pre-processing and retaining targets such as arc and keyhole in step 3 is that the arc light and keyhole are closely related to the heat input of the arc and laser during the welding process, respectively, and the heat input is an important factor in determining the type of weld.

[0166] In step 4, a monitoring model with good defect monitoring performance is obtained through the training set and the validation set, and the generalization performance is tested on welds of the same three weld types but with different process parameters, and then applied to subsequent actual production monitoring.

[0167] Step 5: Build and design a channel attention mechanism module: Figure 4 As shown, the specific method is:

[0168] First, the input features are compressed by global maximum and average pooling to extract different feature information; then, the information is sent to the fully connected layer to better fit the complex correlation between feature channels and compress the features. The number of neurons is s1 and s2 respectively, and the Relu function is used for activation; next, the compressed features connected in two ways are operated through the second fully connected layer, and the weight coefficient is obtained through the Sigmoid function to regress the number of features to the original number of features. Finally, the original features are scaled using the weight coefficient to generate a weighted feature map. This module can automatically weight the features of different channels to adapt to various welding conditions and situations, improve the robustness and generalization ability of the network, reduce the sensitivity to noise, and improve the performance of processing high-power laser-arc hybrid welding processes. Specifically:

[0169] The global maximum pooling operation is expressed as

[0170]

[0171] in Represents Y generated by global maximum pooling max The cth value of Y max ∈R c , C is the channel dimension threshold of the input feature map X, X∈R H×W×C ,X=[x1,x2,…,x c ,…,x c ].x c (i, j) represents the value of the cth X at position (i, j). C, H, and W are the channel, height, and width of the feature map. The relationship between c and C is 1≤c≤C.

[0172] The fully connected layer and ReLU function are used to reduce the dimension of the feature vector and enhance the model representation effect respectively, which can be expressed as

[0173] G max =δ(W1Y max +b1)

[0174] The dimension of W1 is C / s1×C, and s1 is the scaling parameter. The fully connected layer can reduce the number of channels, thereby reducing the amount of calculation. δ represents the ReLU function. max The dimensions are 1×1×C.

[0175] The global average pooling can be expressed as

[0176]

[0177] in Represents Y generated by global maximum poolingmax The cth value of Y max ∈R c , C is the channel dimension threshold of the input feature map X, X∈R H×W×C ,X=[x1,x2,…,x c ,…,x c ].x c (i, j) represents the value of the cth X at position (i, j). C, H, and W are the channel, height, and width of the feature map. The relationship between c and C is 1≤c≤C.

[0178] The fully connected layer and ReLU function are used to reduce the dimension of the feature vector and enhance the model representation effect respectively, which can be expressed as

[0179] G avg =δ(W2Y avg +b2)

[0180] The dimension of W1 is C / s2×C, and s2 is the scaling parameter. The fully connected layer can reduce the number of channels, thereby reducing the amount of calculation. δ represents the ReLU function. max The dimensions are 1×1×C.

[0181] In order to obtain compressed feature maps of different pooling operations, the outputs of two parallel feature compression operations are combined using a concatenation operation, which can increase the width of the network and improve the model and accuracy. The equation can be expressed as:

[0182] V=f contact (G avg ,G max )

[0183] The dimension of V is 1×1×2C(s1+s2).

[0184] By using fully connected layers and activation functions, the output multiplication operation of the module is expressed as.

[0185]

[0186] Where W3 dimension is C×[(s1+s1) / s1s2], and σ is the activation function.

[0187] The high-power laser-arc hybrid welding process defect recognition model constructed in the further described step 4 is specifically as follows: the model uses Densenet as the backbone network. The network contains 4 layers of Dense Block. It can be seen that Dense Block is interconnected with all layers. The input of each layer in the backbone network comes from the feature maps of all the previous layers. The output of each layer is directly connected to the input of all the layers behind it. Batch normalization is used to improve the stability of training and standardize the input of the previous layers. Then, the specific feature information of the data is extracted through the convolution layer. The attention mechanism module is added to the Dense Block of the 2nd and 3rd layers so that each stage of the network can effectively learn important and overall welding process features. Identification and classification are performed by performing global average pooling and inputting the normalized exponential function to the fully connected layer, and finally the welding state type is output. The training and verification process of the model is as follows Figure 6 The model converged and achieved a test set accuracy of 95.71%.

[0188] A monitoring system for defects in a high-power laser-arc hybrid welding process, the system being used in a method for identifying defects in a high-power laser-arc hybrid welding process, is characterized in that the monitoring system is as follows: Figure 5 As shown, the system includes a welding work platform, arc welding gun, wire feeder, laser head, shielding gas, fiber laser, and an image acquisition system consisting of a high-speed camera, acquisition card, auxiliary light source, and image monitoring system. The welding specimen is fixed to the welding platform, the laser welding gun is fixed above the movable platform at a 5° tilt angle, the angle between the arc welding gun and the welding workpiece is 40°, and the image acquisition system is fixed above the platform. In subsequent production practice, the acquisition card is connected to a computer equipped with a high-power laser-arc hybrid welding process defect monitoring method to output the final welding monitoring quality results.

[0189] In summary: The present invention provides a method for online monitoring of the quality of high-power laser-arc welding process based on the top visual signal of a high-speed camera and deep learning technology, which provides an effective solution for high-power laser-arc composite welding defect monitoring. Its scope of application covers root hump, incomplete penetration and defect monitoring, that is, during the composite welding process, the welding quality after the weld is formed is monitored and identified. The method for identifying defects in a high-power laser-arc composite welding process proposed by the present invention is universal and can identify the welding status of different materials (such as stainless steel, high-strength steel, aluminum alloy, etc.) and materials with different material thicknesses and welding methods (laser welding, arc welding). It can effectively reduce the interference of strong interference factors on the welding quality identification results, effectively identify the welding quality in the high-power laser-arc welding process, and timely discover and effectively reduce the generation of defective welds in the welding process.

[0190] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0191] In an embodiment of the present invention, a high-power laser-arc hybrid welding process defect identification system is introduced, which aims to improve welding quality and efficiency through advanced monitoring technology. The core components of the system include a welding work platform, an arc welding gun, a wire feeder, a laser head, a shielding gas, a fiber laser, and an image acquisition system. The image acquisition system specifically includes a high-speed camera, an acquisition card, an auxiliary light source, and an image monitoring system, which are designed to capture key image data during the welding process in real time. In addition, the present invention also provides a supporting computer device, which includes a memory and a processor. The memory has a built-in computer program for executing the defect identification method. When the processor executes the program, it can automatically perform the defect identification steps based on the collected image data, thereby achieving efficient and accurate welding quality monitoring.

[0192] The present invention also includes a computer-readable storage medium storing the same computer program. When executed by a processor, the program can also prompt the processor to perform relevant steps of defect identification, further enhancing the practicality and flexibility of the system.

[0193] By integrating high-speed camera technology with intelligent computing, this invention provides a comprehensive quality monitoring solution for high-power laser-arc hybrid welding processes, which is expected to significantly improve the quality and efficiency of welding operations.

[0194] Existing laser arc hybrid welding process quality monitoring technologies often use multiple sensors and machine learning methods based on manual feature extraction. The monitoring methods are usually based on thin plates. This has the following problems and limitations:

[0195] The use of multiple high-speed cameras and other sensors will limit the spatial layout flexibility of the equipment during the welding process. The present invention can realize the monitoring process of laser arc hybrid welding only through the top image, and the monitoring method has strong spatial adaptability.

[0196] Current hybrid welding monitoring technologies primarily target thin-plate welding applications. Monitoring methods often rely on manual feature extraction based on feature engineering when processing welding signals from thin-walled components. However, high-power laser arc hybrid welding of medium and thick plates creates more intense thermal reactions and correspondingly more signal interference. To address this issue, this paper proposes a method utilizing convolutional neural networks and an attention mechanism to effectively extract and focus on key features in process images. This is particularly important for managing variable welding conditions and enhancing the adaptability of the monitoring environment.

[0197] Traditional manual feature extraction methods are often based on experience within specific welding conditions, resulting in limited adaptability to monitoring environments and production automation. In contrast, the proposed method overcomes this limitation by automatically extracting key features from welding images, improving process adaptability and automation, and providing a more effective solution for monitoring high-power laser arc hybrid welding of medium and thick plates.

[0198] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for identifying defects in a high-power laser-arc hybrid welding process, characterized in that: The following steps are involved: Step 1: Conduct multiple welding experiments through the welding monitoring experimental platform to collect and obtain the top visual signals of the high-power laser-arc welding process, and collect image data of the welding process including arc, keyhole, molten droplet and metal vapor characteristics; Step 2: Obtain and record the weld morphology characteristics of the upper and lower surfaces of all welds after the welding test, and classify the welded workpieces into three weld types based on the weld morphology characteristics after welding, and obtain the category results to which the surface forming quality of each weld belongs; Step 3: Use the preprocessing methods of ROI selection, brightness adjustment and gamma change to filter out the metal vapor and spatter interference in the original welding process image, and obtain the process image with periodic overlap of keyhole, arc and droplet targets; Step 4: According to the morphology type of each weld after welding, the welding process image of the preprocessed weld is classified and marked, and the data set is divided into a training set, a validation set, and a test set; The ratio of the training and validation sets is 7:

3. The test set uses welding process data of welds different from those in the training and validation sets to verify the generalization ability of the model. Step 5: Based on the characteristics of high noise intensity in the welding process, a deep learning model and channel attention mechanism are used to build a welding process quality monitoring model; Step 6: Use the quality monitoring model to train, verify and test the high-power laser-arc welding process images to obtain three welding state types of the weld, and finally realize the online monitoring of welding quality; Step 5 described above constructs and designs a channel attention mechanism module: First, the input features are compressed by global maximum pooling and global average pooling to extract different feature information; Then, the information is sent to the fully connected layer to fit the complex correlation between feature channels and compress the features. The number of neurons is s1 and s2 respectively, and the activation of the Relu function is used; The compressed features connected in two ways are operated through the second fully connected layer, and the weight coefficients are obtained through the Sigmoid function to return the number of features to the original number of features; Finally, the weight coefficients are used to scale the original features to generate a weighted feature map; specifically: The global maximum pooling operation is expressed as in Represents Y generated by global maximum pooling max The cth value of Y max ∈R c , C is the channel dimension threshold of the input feature map X, X∈R H×W×C ,X=[x1,x2,...,x c ,...,x C ]; x c (i,j) represents the value of the cth X at position (i,j); C, H and W are the channel, height and width of the feature map; the relationship between c and C is 1≤c≤C; The fully connected layer and ReLU function are used to reduce the dimension of the feature vector and enhance the model representation effect respectively, which can be expressed as follows: G max =δ(W1Y max +b1) The dimension of W1 is C / s1×C, s1 is the scaling parameter; the fully connected layer reduces the number of channels, thereby reducing the amount of calculation, and δ represents the ReLU function; G max The dimensions are 1×1×C; The global average pooling is expressed as The fully connected layer and ReLU function are used to reduce the dimension of the feature vector and enhance the model representation effect respectively, which can be expressed as follows: G avg =δ(W2Y avg +b2) The dimension of W1 is C / s2×C, s2 is the scaling parameter; the fully connected layer reduces the number of channels, thereby reducing the amount of calculation, and δ represents the ReLU function; G max The dimensions are 1×1×C; In order to obtain compressed feature maps of different pooling operations, the outputs of two parallel feature compression operations are combined using a concatenation operation to increase the width of the network and improve the model and accuracy; the equation can be expressed as: V=f contact (G avg ,G max ) The dimension of V is 1×1×2C(s1+s2); By using fully connected layers and activation functions, the output multiplication operation of the module is expressed as; Where W3 dimension is C×[(s1+s1) / s1s2], σ is the activation function; The welding process quality monitoring model is specifically as follows: The model uses Densenet as the backbone network; the network contains 5 layers of Dense Block; Dense Block connects all layers to each other. The input of each layer in the backbone network comes from the feature maps of all previous layers, and the output of each layer is directly connected to the input of all subsequent layers. Batch normalization is used to improve the stability of training and standardize the input of the previous layers. Then, the convolutional layer is used to extract the specific feature information of the data; the attention mechanism module is added to the Dense Block in the 2nd and 3rd layers so that each stage of the network can effectively learn the important and overall welding process features; The model is trained and validated by performing global average pooling, inputting to the fully connected layer, and normalizing the exponential function for identification and classification, and finally outputting the welding state type.

2. The high power laser-arc hybrid welding process defect identification method according to claim 1 is characterized in that: In the step 1, a welding experiment of multiple welds is performed, and the parameters of each welding experiment are fixed.

3. The high power laser-arc hybrid welding process defect identification method according to claim 1, characterized in that: The three types of welds in step 2 include: bottom incomplete penetration (I), good weld (II) and bottom hump (III).

4. The high power laser-arc hybrid welding process defect identification method according to claim 1, characterized in that: The basis for pre-processing and retaining the arc and keyhole targets in step 3 is that the arc light and keyhole are closely related to the heat input of the arc and laser during the welding process, respectively.

5. The high power laser-arc hybrid welding process defect identification method according to claim 1, characterized in that: In step 4, a monitoring model with good defect monitoring performance is obtained through a training set and a validation set, and a generalization performance test is performed on welds of the same three weld types but with different process parameters.

6. A high power laser-arc hybrid welding process defect recognition system implementing the high power laser-arc hybrid welding process defect recognition method according to any one of claims 1 to 5, characterized in that: The high-power laser-arc hybrid welding process defect recognition system comprises: Welding work platform, arc welding gun, wire feeder, laser head, shielding gas, fiber laser, image acquisition system, The image acquisition system includes a high-speed camera, an acquisition card, an auxiliary light source and an image monitoring system; The welding sample is fixed on the welding platform, the laser welding gun is fixed on the top of the movable platform with an inclination angle of 5°, the angle between the arc welding gun and the welding workpiece is 40°, and the image acquisition system is fixed above the platform.

7. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the high-power laser-arc hybrid welding process defect identification method as described in any one of claims 1-5.

8. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the high-power laser-arc hybrid welding process defect identification method as described in any one of claims 1 to 5.

9. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the high-power laser-arc hybrid welding process defect identification system as described in claim 6.

Images