Welding quality detection system

Abstract

The invention relates to a welding quality detection system which comprises a process monitoring layer configured with a multi-mode sensor array and used for collecting original performance data in the welding process in real time; the intelligent analysis layer is provided with a multi-source data fusion module and a defect prediction module, and the multi-source data fusion module adopts an improved D-S evidence theory algorithm to perform fusion processing on the original performance data collected by the process monitoring layer to obtain multi-dimensional performance data; the defect prediction module analyzes and captures abnormal data according to the multi-dimensional performance data, dynamically predicts a defect development trend and outputs defect prediction information; the decision execution layer is provided with a self-adaptive control module and is used for dynamically adjusting welding process parameters according to the defect prediction information output by the intelligent analysis layer; and the data communication bus is used for realizing real-time data interaction and closed-loop feedback control among the process monitoring layer, the intelligent analysis layer and the decision execution layer. The method has the effect of effectively improving the detection precision and the detection efficiency.

Description

Technical Field

[0001] The present application relates to the field of weld detection, and in particular to a detection system for welding quality. Background Art

[0002] Welding, as an indispensable connection technology, plays a key role in many industrial fields. In automobile manufacturing, the assembly of vehicle bodies largely relies on welding technology to ensure the strength and stability of the body structure; in the aerospace field, the connection of key components such as the wings and fuselages of aircraft also requires high-precision welding processes to ensure flight safety; in the power industry, the manufacturing and installation of various power equipment are also inseparable from welding, such as the connection of transformers and transmission lines. Welding quality is directly related to the performance, reliability, and service life of these products, and thus affects the safety and economy of the entire engineering project.

[0003] With the development of industrial products towards high precision and high performance, the requirements for welding quality are getting higher and higher. For example, in fields such as high-end equipment manufacturing and semiconductor manufacturing, even minor defects in weld joints may lead to serious quality problems and safety accidents. Therefore, the detection system is required to accurately detect smaller-sized defects and provide more reliable detection results.

[0004] Chinese Patent No. CN117593298B discloses a laser welding quality detection system based on machine vision. It collects images of devices after several welds in the welding area, matches corresponding preprocessing strategies for each device image, and takes the preprocessed device image as the device target image; according to the image quality coefficient, it screens out the images to be recognized from several target images, performs defect recognition and quality detection on the images to be recognized, constructs the device quality coefficient during the period using the detection data, and thus distinguishes the welding quality of the device; and constructs the batch quality coefficient from the device quality coefficients within the current batch. If the batch quality coefficient is lower than the batch quality threshold, it matches the corresponding process optimization plan for the current welding equipment from the established welding optimization knowledge graph. However, only visually detecting the welding quality by collecting image data has a single test method. For defects such as pores, slag inclusions, and lack of fusion existing inside the weld, it is very difficult to discover them only relying on image data, which may lead to serious quality hazards being ignored. And the defect detection is carried out after the product welding is completed. Offline detection results in a lag in defect discovery and a high cost of manual intervention.

[0005] Regarding the above related technologies, the inventor believes that the existing welding quality detection methods are single and backward, which are likely to cause serious quality hazards to be ignored. Summary of the Invention

[0006] In order to solve the above problems, the present application provides a detection system for welding quality.

[0007] In a first aspect, the present application provides a welding quality detection system, adopting the following technical solutions:

[0008] A welding quality detection system includes:

[0009] A process monitoring layer, configured with a multi-modal sensor array, for real-time acquisition of original performance data during the welding process, where the original performance data includes arc spectrum, molten pool image, acoustic emission signal, and thermodynamic parameters during the welding process;

[0010] An intelligent analysis layer, deployed with a multi-source data fusion module and a defect prediction module. The multi-source data fusion module uses an improved D-S evidence theory algorithm to fuse and process the original performance data collected by the process monitoring layer to obtain multi-dimensional performance data; the defect prediction module constructs a defect evolution model based on an LSTM-GAN hybrid neural network. The defect evolution model analyzes and captures abnormal data according to the multi-dimensional performance data to dynamically predict the defect development trend, and outputs defect prediction information, where the defect prediction information includes defect risk level, defect risk type, defect location information (AR), defect evolution trend heat map, and predicted failure time;

[0011] A decision execution layer, configured with an adaptive control module, for dynamically adjusting welding process parameters according to the defect prediction information output by the intelligent analysis layer;

[0012] A data communication bus, for realizing real-time data interaction and closed-loop feedback control among the process monitoring layer, the intelligent analysis layer, and the decision execution layer.

[0013] Preferably, the multi-modal sensor array includes:

[0014] A high-frequency current probe, for collecting current signals during the welding process at a frequency not lower than 1 MHz;

[0015] A high-speed CMOS camera, for collecting molten pool images during the welding process;

[0016] A broadband acoustic emission sensor, with a frequency band set to 50 kHz - 1.2 MHz, for collecting unique acoustic emission signals generated by various defects;

[0017] A multi-spectral infrared thermal imager, for collecting thermodynamic parameters during the welding process.

[0018] Preferably, the multi-source data fusion module fusing and processing the original performance data to obtain multi-dimensional performance data specifically includes the following steps:

[0019] Performing wavelet packet transform noise reduction processing on the arc spectrum, acoustic emission signal, and thermodynamic parameters in the original performance data to remove noise interference and obtain effective data reflecting the characteristics of the welding process;

[0020] Based on the periodic characteristic points of the welding process, spatio-temporal alignment and normalization processing are performed on the effective data after noise reduction through a spatio-temporal alignment algorithm based on dynamic time warping to obtain standardized aligned data;

[0021] An improved D-S evidence theory algorithm is used to calculate the confidence weight of each sensor in the multi-modal sensor array, and the standardized aligned data after spatio-temporal alignment is weighted and fused through an attention mechanism to obtain a signal-level fusion feature vector;

[0022] The molten pool image in the original performance data is processed by a pre-set improved YOLOv5s model to obtain an image-level feature vector;

[0023] The signal-level fusion feature vector and the image-level feature vector are feature spliced to generate multi-dimensional performance data.

[0024] Preferably, the specific improvements of the improved YOLOv5s model include:

[0025] Adding a deformable convolutional layer in the Backbone part;

[0026] Replacing the object detection loss function CIoU Loss with EIOU Loss;

[0027] The output layer fuses the molten pool area volatility and morphology symmetry index.

[0028] Preferably, the improvement of the D-S evidence theory algorithm lies in introducing an adaptive basic probability assignment function:

[0029]

[0030] where m i (A) is the basic probability assignment of the i-th sensor to the proposition A, representing the degree of trust in A; w i is the dynamic weight coefficient of the i-th sensor, which is updated in real time through online learning; Confidence i is the confidence of the i-th sensor; λ is the conflict factor adjustment term.

[0031] Preferably, the defect prediction module captures abnormal data according to the multi-dimensional performance data analysis and dynamically predicts the defect development trend. The output defect prediction information specifically includes the following steps:

[0032] Slice the multi-dimensional performance data according to the time series and perform normalization processing to construct a time series data set;

[0033] The signal-level fusion feature vector and the image-level feature vector are spliced and weighted and fused through an attention mechanism to enhance the pre-annotated key features;

[0034] The defect evolution model is used to capture the time-dependent relationship of defect evolution for predicting future defect characteristics, and defect prediction information is output.

[0035] Preferably, the adaptive control module dynamically adjusts the welding process parameters according to the defect prediction information output by the intelligent analysis layer, and specifically includes the following steps:

[0036] Receiving and parsing the defect prediction information in real time, and matching the corresponding welding process parameter adjustment strategy from the preset process knowledge base;

[0037] Through the improved NSGA-Ⅲ algorithm, with the goal of minimizing the defect risk probability, energy consumption, and maximizing production efficiency, multi-objective optimization is carried out under the constraints of the process knowledge base and equipment physical limitations;

[0038] Initializing the parameter population through Latin hypercube sampling, and generating 3-5 Pareto optimal parameter combinations through non-dominated sorting and crowding degree comparison;

[0039] Selecting the parameter combination according to the defect risk level, adjusting the welding process parameters in a small amplitude in stages, and sending control instructions to the welding equipment through the data communication bus.

[0040] Preferably, the defect position information includes the transverse position inside the weld, the longitudinal position inside the weld, the depth position inside the weld, the three-dimensional space positioning based on augmented reality, and the position based on the fusion of augmented reality and the actual scene.

[0041] In summary, the present application includes at least one of the following beneficial technical effects:

[0042] 1. Adopting a three-level distributed structure to construct a "perception - analysis - decision - execution" dynamic closed loop, realizing real-time detection of the welding process and self-optimization of process parameters. Through multi-dimensional data fusion, covering 4 types of physical field data, the defect detection rate is greatly improved. Then, through the defect evolution model, the defect development trend is accurately predicted, and the upcoming defects are eliminated by timely interfering with and optimizing the parameters, avoiding the one-sidedness of traditional detection methods and the irreversible defects easily caused by off-line lag detection, avoiding equipment damage, and achieving the effect of effectively improving the detection accuracy and detection efficiency;

[0043] 2. Denoising through wavelet packet transform can remove noise interference and improve data quality. Spatiotemporal alignment and normalization solve the problem of inconsistency in multi-source data. The non-image data and the molten pool image are processed independently. The combination of the improved D-S evidence theory algorithm and the attention mechanism can reasonably allocate the weights of the data of each signal sensor, make full use of the information of multi-source data, and improve the fusion effect of signal data. The improved YOLOv5s model is used to perform image processing and feature enhancement on the molten pool image to obtain image-level feature vectors. Finally, the signal-level and image-level features are spliced together, integrating the advantages of different types of data to generate more comprehensive multi-dimensional representation data, achieving the effect of accurately and comprehensively reflecting the welding process.

[0044] 3. Through the improved YOLOv5s model, first, adding a deformable convolutional layer in the Backbone part can adaptively adjust the sampling position of the convolutional kernel, better adapt to the irregular shapes and pose changes of objects in the molten pool image, and improve the accuracy of feature extraction. Second, it is improved based on CIOU Loss to enhance the sensitivity to small targets. Compared with the traditional EIOU, the penalty for the width and height differences of small targets is weak, which is likely to cause training difficulties. After the improvement, the supervision signal of the width and height differences is enhanced through a dynamic denominator, and the positioning accuracy of tiny molten pools during the welding process is improved by targeted optimization of width and height matching, thereby improving the detection accuracy of welding quality. Finally, when outputting the features of the molten pool image, the area volatility and morphology symmetry of the molten pool are additionally output, which helps to more comprehensively and deeply evaluate the actual state of the molten pool and effectively improve the detection ability and accuracy of welding defects. Brief Description of the Drawings

[0045] Figure 1 is the architecture block diagram of a welding quality detection system in an embodiment of the present application;

[0046] Figure 2 is the system block diagram of a multi-modal sensor array in an embodiment of the present application;

[0047] Figure 3 is the method flowchart for fusing and processing the original performance data in an embodiment of the present application;

[0048] Figure 4 is the method flowchart for dynamically predicting the development trend of defects in an embodiment of the present application;

[0049] Figure 5 is the method flowchart for dynamically adjusting welding process parameters in an embodiment of the present application.

[0050] Description of the drawing reference numerals: 1. Process monitoring layer; 11. Multimodal sensor array; 111. High-frequency current probe; 112. High-speed CMOS camera; 113. Wideband acoustic emission sensor; 114. Multispectral infrared thermal imager; 2. Intelligent analysis layer; 21. Multisource data fusion module; 22. Defect prediction module; 3. Decision execution layer; 31. Adaptive control module; 4. Data communication bus. Detailed implementation manners

[0051] The following further describes the present application with reference to the Figures 1-5 accompanying drawings.

[0052] An embodiment of the present application discloses a welding quality detection system. Referring to Figure 1 , a welding quality detection system includes a process monitoring layer 1, an intelligent analysis layer 2, a decision execution layer 3, and a data communication bus 4. Real-time data interaction and closed-loop feedback control are realized between the process monitoring layer 1, the intelligent analysis layer 2, and the decision execution layer 3 through the data communication bus 4. The process monitoring layer 1 is configured with a multimodal sensor array 11 for real-time acquisition of the original performance data during the welding process, and the original performance data includes the arc spectrum, molten pool image, acoustic emission signal, and thermodynamic parameters during the welding process. The intelligent analysis layer 2 is deployed with a multisource data fusion module 21 and a defect prediction module 22. The multisource data fusion module 21 uses an improved D-S evidence theory algorithm to fuse and process the original performance data collected by the process monitoring layer 1 to obtain multi-dimensional performance data. The defect prediction module 22 constructs a defect evolution model based on an LSTM-GAN hybrid neural network. The defect evolution model analyzes and captures abnormal data according to the multi-dimensional performance data to dynamically predict the defect development trend and outputs defect prediction information. The defect prediction information includes the defect risk level (low, medium, high), defect risk type, defect location information, defect evolution trend heat map, and estimated failure time. The decision execution layer 3 is configured with an adaptive control module 31 for dynamically adjusting the welding process parameters according to the defect prediction information output by the intelligent analysis layer 2. A three-level distributed structure is adopted to construct a "perception - analysis - decision - execution" dynamic closed loop, realizing real-time detection of the welding process and self-optimization of process parameters. Through multi-dimensional data fusion, covering 4 types of physical field data, the defect detection rate is greatly improved. Then, through the defect evolution model, the defect development trend is accurately predicted, and the upcoming defects are eliminated by timely interfering and optimizing the parameters, avoiding the one-sidedness of traditional detection methods and the irreversible defects easily caused by offline lag detection, avoiding equipment damage, and achieving the effect of effectively improving the detection accuracy and detection efficiency.

[0053] Referring to Figure 2 , the multimodal sensor array 11 includes:

[0054] A high-frequency current probe 111 is used to collect current signals during the welding process at a frequency not lower than 1 MHz. It can collect current signals during the welding process at a high frequency. The high sampling rate ensures that subtle changes in the current signal can be captured, thus accurately reflecting the stability of the arc and the input of welding energy during the welding process, providing high-precision current data for subsequent analysis.

[0055] A high-speed CMOS camera 112 is used to collect molten pool images during the welding process. In this embodiment, a high-speed CMOS camera 112 with a frame rate not lower than 2000 fps is adopted, which can quickly and continuously capture molten pool images, clearly record the dynamic change process of the molten pool. For some rapidly changing molten pool phenomena, such as droplet transfer and molten pool fluctuation, it can provide clear image information, which helps to timely detect potential welding defects.

[0056] A broadband acoustic emission sensor 113 with a frequency band set to 50 kHz - 1.2 MHz is used to collect unique acoustic emission signals generated by various defects. It can capture acoustic emission signals of various frequencies generated during the welding process, and can detect acoustic emission signals generated by different types and sizes of defects, providing richer information for the identification and location of defects.

[0057] A multi-spectral infrared thermal imager 114 is used to collect thermodynamic parameters during the welding process. In this embodiment, the multi-spectral infrared thermal imager 114 has 8 spectral channels and can simultaneously obtain infrared thermal images of multiple spectral bands. The multi-spectral design enables the thermal imager to more comprehensively monitor the temperature distribution in the welding area. Data from different spectral channels can complement each other, improving the monitoring accuracy of thermodynamic parameters during the welding process. Through the setting of the multi-modal sensor array 11, multi-dimensional data of electricity, light, sound, and heat during the welding process are synchronously collected, which can comprehensively and accurately reflect the actual situation of the welding process, avoid the limitations of single-sensor data, and improve the detection accuracy of welding quality.

[0058] In addition, in this embodiment, the defect position information of the defect prediction information includes the transverse position inside the weld, the longitudinal position inside the weld, the depth position inside the weld, the three-dimensional space positioning based on augmented reality, and the position based on the fusion of augmented reality and the actual scene. Among them, the transverse position inside the weld: the position of the defect in the weld width direction, such as the distance from the center line to the left or right. The longitudinal position inside the weld: the position of the defect in the weld length direction, such as the length from the starting point or a specific marking point. The depth position inside the weld: the position of the defect in the weld thickness direction, such as a certain depth below the surface. The three-dimensional space positioning based on augmented reality (AR): visually displays the specific position of the defect in space in a three-dimensional form. The position based on the fusion of augmented reality (AR) and the actual scene: fuses and displays the defect position with the actual geometric shape, surface features, etc. of the welded workpiece. The detailed transverse, longitudinal, and depth position information locates the defect from multiple dimensions, enabling accurate knowledge of the specific position of the defect in the weld. The three-dimensional space positioning based on augmented reality and the position based on the fusion with the actual scene visually display the defect position, improving the understanding and judgment speed of the defect position.

[0059] Referring to Figure 3 , the multi-source data fusion module 21 uses an improved D-S evidence theory algorithm to fuse and process the original performance data collected by the process monitoring layer to obtain multi-dimensional performance data. The specific steps are as follows:

[0060] A1. Wavelet packet transform noise reduction: Perform wavelet packet transform noise reduction processing on the arc spectrum, acoustic emission signal, and thermodynamic parameters in the original performance data to remove noise interference and obtain effective data reflecting the characteristics of the welding process;

[0061] A2. Data spatio-temporal alignment: Based on the periodic feature points of the welding process, perform spatio-temporal alignment and normalization processing on the effective data after noise reduction processing through a spatio-temporal alignment algorithm based on dynamic time warping to obtain standardized aligned data;

[0062] Different types of sensors may have differences in data acquisition time and transmission process, which can lead to inconsistencies in data in terms of time and space. For example, a certain sensor may have its data out of sync with that of other sensors in terms of time due to its own response time or transmission delay. Such inconsistencies will bring difficulties to data fusion because out-of-sync data cannot be directly compared and analyzed. The spatio-temporal alignment algorithm based on dynamic time warping (DTW) uses the method of dynamic programming to find the optimal matching path between the data of different sensors, eliminate the influence brought by the transmission delay of multi-source signals, and achieve precise alignment of the data of each sensor in time and space. In this way, it provides a unified and standard time and space benchmark for subsequent fusion processing, ensuring that the data of different sensors can be fused under the same spatio-temporal framework. In addition, the periodic feature points on which the spatio-temporal alignment algorithm based on dynamic time warping (DTW) is based are some key events or features with obvious rules and repetitiveness, such as droplet transfer related feature points, arc combustion related feature points, welding equipment operation related feature points, mechanical movement related feature points (applicable to automated welding), etc. The periodic feature point selected in the embodiment of this application is the droplet transfer frequency (droplet transfer related feature point).

[0063] A3. Obtain the signal-level fusion feature vector through weighted fusion: Use the improved D-S evidence theory algorithm to calculate the confidence weights of each sensor in the multi-modal sensor array, and obtain the signal-level fusion feature vector by weighted fusion of the standardized aligned data after spatio-temporal alignment through the attention mechanism. In this embodiment, the signal-level features include 128-dimensional features such as arc stability index, acoustic emission band energy, and thermodynamic gradient. The attention mechanism is used to evaluate the importance of real-time data, and the importance weight is a i , such as when the molten pool fluctuates, the image data of the high-speed CMOS camera is more important for defect prediction, a 相机 will increase significantly, and its training process is the prior art and will not be elaborated here; weighted fusion is the confidence weight m i multiplied by the importance weight a i , m i reflects the historical reliability of the sensor, and a i is adjusted according to the importance of real-time data. Through the above mechanism, the intelligent fusion of signal-level data is realized, significantly improving the accuracy and adaptability of the welding quality detection system.

[0064] A4. Obtain the image-level feature vector: Process the molten pool image in the original performance data with the pre-set improved YOLOv5s model to obtain the image-level feature vector. In this embodiment, the image-level features include 32-dimensional features such as molten pool area volatility, morphology symmetry, and defect position coordinates.

[0065] A5. Generate multi-dimensional representation data by splicing: Splice the signal-level fusion feature vector and the image-level feature vector to generate multi-dimensional representation data. Denoising by wavelet packet transform can remove noise interference and improve data quality. Spatial-temporal alignment and normalization solve the problem of inconsistency of multi-source data. Independently process non-image data and molten pool images respectively, and adopt the combination of the improved D-S evidence theory algorithm and the attention mechanism, which can reasonably allocate the weights of each signal sensor data, make full use of the information of multi-source data, and improve the fusion effect of signal data. Obtain the image-level feature vector through image processing and feature enhancement processing of the molten pool image by the improved YOLOv5s model. Finally, splice the signal-level and image-level features, synthesize the advantages of different types of data, generate more comprehensive multi-dimensional representation data, and achieve the effect of accurately and comprehensively reflecting the welding process.

[0066] The improvement of the above D-S evidence theory algorithm lies in introducing an adaptive basic probability assignment function:

[0067]

[0068] where m i (A) is the basic probability assignment of the i-th sensor to proposition A, representing the degree of trust in A; w i is the dynamic weight coefficient of the i-th sensor, which is updated in real time through online learning. For example, if sensor i frequently detects real defects in the current welding stage, its w i increases. If the data of sensor i fluctuates greatly, then w i decreases. Confidence i is the confidence of the i-th sensor, which is preset by the management personnel; λ is the conflict factor adjustment term, and its value range in this embodiment is within [0, 0.25]. When the conflict between sensors (such as the classification results of two sensors for the same defect are contradictory) is greater, the larger λ will reduce the basic probability assignment of all sensors, avoiding wrong decisions caused by conflicts. When the conflict is smaller, the smaller λ will retain more original information. In this embodiment, λ is determined by calculating the conflict (Jousselme) distance between each pair of sensors, and then taking the average of the conflict distances of all sensor pairs. How to calculate the conflict (Jousselme) distance between two sensors is prior art and will not be elaborated here. Based on the existing D-S evidence theory algorithm, an adaptive basic probability assignment function is introduced. Through adaptive weight allocation and conflict adjustment, the trust degree assignment of each sensor to proposition A is dynamically adjusted, and the problem that the reliability changes due to the change of sensor performance in the dynamic welding scenario of the traditional D-S evidence theory is solved. w iUpdated in real time through online learning to adapt to changes in sensor performance during the welding process (such as probe aging and environmental interference), improving the reliability of fusion; through λ adjustment, misjudgment in high-conflict scenarios is avoided, and the improved algorithm combined with multi-sensor fusion detection can greatly improve the defect recognition accuracy.

[0069] For example: Assume that in aluminum alloy MIG welding, the high-frequency current probe detects current fluctuations (confidence level 0.8), while the acoustic emission sensor does not detect defects (confidence level 0.4):

[0070] Traditional method: If the weight of the current probe is fixed at 0.6 and that of the acoustic emission is 0.4, then m 缺陷 = 0.6 * 0.8 = 0.48, which may ignore the actual impact of current fluctuations.

[0071] In this application, if online learning discovers that current fluctuations are strongly correlated with defects, w 电流 = 0.7 and the conflict factor λ = 0.2, then:

[0072] At this time, the current probe has a higher degree of trust, which is more in line with the actual situation.

[0073] The specific improvements of the above improved YOLOv5s model include:

[0074] Add a deformable convolutional layer in the Backbone part; specifically, insert a deformable convolutional layer (DeformableConvolutionalLayer) after the C3 module of CSPDarknet; each deformable convolutional layer contains a 3×3 convolutional kernel, learning 2 offset channels (Δx, Δy), and the output feature map size is the same as that of a conventional convolution; when a traditional convolutional layer processes an image, the sampling position of the convolutional kernel is fixed and unchanged. For a molten pool image with complex shapes and dynamic changes, it is often difficult to fully adapt. The deformable convolutional layer is like an "adaptive sampler". By learning the offsets of each sampling position, it can flexibly adjust the sampling position adaptively according to the specific content of the image, thereby more accurately capturing the subtle feature information in the molten pool image and greatly improving the model's sensitivity to changes in the molten pool morphology.

[0075] Replace the target detection loss function CIoU Loss with EIOU Loss; among them, CIoU Loss is a commonly used target detection loss function for measuring the difference between the predicted bounding box and the ground truth bounding box, while EIOU Loss is deeply optimized and improved on the basis of CIoU Loss. It not only comprehensively considers the overlapping area, center point distance, and aspect ratio between the predicted bounding box and the ground truth bounding box, but also further carefully considers the precise distance between the predicted bounding box and the ground truth bounding box in each direction, and can more accurately and delicately reflect the positional relationship between the predicted bounding box and the ground truth bounding box, thus significantly improving the localization accuracy and detection accuracy of the model; the EIOU Loss formula is:

[0076]

[0077] where IoU is the intersection over union of the predicted bounding box and the ground truth bounding box, b, b gt are the center point coordinates of the predicted bounding box and the ground truth bounding box, and ρ 2 (b, b gt ) represents the square of the Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box, w, w gt are the widths of the predicted bounding box and the ground truth bounding box, h, h gt are the heights of the predicted bounding box and the ground truth bounding box, and c is the diagonal length of the minimum enclosing box of the predicted bounding box and the ground truth bounding box; C w , C h are the width and height of the minimum enclosing box respectively;

[0078] The output layer fuses the indicators of the molten pool area volatility and the morphology symmetry. In the output layer of the model, in addition to outputting the traditional object detection results, the indicators of the molten pool area volatility and the morphology symmetry are innovatively fused; the molten pool area volatility intuitively reflects the dynamic change of the molten pool area over time and is an important indicator to measure the stability of the welding process; the morphology symmetry indicator accurately measures whether the shape of the molten pool is symmetric. Generally, a symmetric molten pool means better welding quality. By fusing these key indicators, the model can more comprehensively and deeply evaluate the actual state of the molten pool, effectively improving the detection ability and accuracy of welding defects. First, by adding a deformable convolutional layer in the Backbone part, the sampling position of the convolutional kernel can be adaptively adjusted to better adapt to the irregular shapes and pose changes of objects in the molten pool image, improving the accuracy of feature extraction; second, an improvement is made based on the CIOU Loss to enhance the sensitivity to small targets. Compared with the traditional EIOU, the penalty for the width and height differences of small targets is weak, which is likely to lead to difficult training. After the improvement, the supervision signal of the width and height differences is enhanced through a dynamic denominator, and the positioning accuracy of tiny molten pools in the welding process is improved by targeted optimization of the width and height matching, thereby improving the detection accuracy of welding quality. Finally, when outputting the features of the molten pool image, the molten pool area volatility and the morphology symmetry are additionally output, which helps to more comprehensively and deeply evaluate the actual state of the molten pool, effectively improving the detection ability and accuracy of welding defects.

[0079] It should be emphasized that the above-mentioned traditional algorithm models are all existing technologies known to those skilled in the art, and their specific application steps and training steps will not be elaborated here. In the embodiments of the present application, only the improvements of these algorithm models in the actual application of the present application are explained.

[0080] Refer to Figure 4 , the defect prediction module 22 constructs a defect evolution model based on the LSTM-GAN hybrid neural network. The defect evolution model captures abnormal data according to the analysis of multi-dimensional performance data and dynamically predicts the defect development trend. The output defect prediction information specifically includes the following steps:

[0081] B1. Construct a time series data set: slice the multi-dimensional performance data according to the time series and perform normalization processing to construct a time series data set;

[0082] B2. Dynamically splice signal and image features: splice the signal-level fusion feature vector and the image-level feature vector and perform weighted fusion through the attention mechanism to enhance the pre-annotated key features; in the embodiments of the present application, the pre-annotated key feature is the molten pool area feature. The method of sliding window difference is used to calculate the difference in the molten pool area within adjacent time windows to more clearly reflect the dynamic change of the molten pool area;

[0083] B3. Perform defect feature prediction: Use the defect evolution model to capture the time-dependent relationship of defect evolution for future defect feature prediction and output defect prediction information. By constructing a time series dataset and performing time series analysis, the dynamic evolution law of defects during the welding process can be captured, improving the accuracy and timeliness of defect prediction. Secondly, through the attention mechanism, weighted fusion of signal-level fusion features and image-level features highlights important features, and combined with the enhancement processing of pre-annotated key features, efficient and accurate dynamic feature enhancement is achieved, further improving the model's ability to identify and predict defects. Finally, use the defect evolution model to output comprehensive and accurate defect prediction information, providing a detailed basis for the detection and decision-making of the welding process, and helping to take timely measures to prevent and handle defects.

[0084] Refer to Figure 5 , the adaptive control module 31 dynamically adjusts the welding process parameters according to the defect prediction information output by the intelligent analysis layer, specifically including the following steps:

[0085] C1. Match adjustment strategy: Receive and parse the defect prediction information in real time, and match the corresponding welding process parameter adjustment strategy from the preset process knowledge base; the process knowledge base contains the optimal adjustment strategies, equipment physical limit data, and historical successful cases corresponding to various types of typical defects;

[0086] C2. Multi-objective optimization: Through the improved NSGA-Ⅲ algorithm, with the goals of minimizing the defect risk probability, energy consumption, and maximizing production efficiency, multi-objective optimization is carried out under the constraints of the process knowledge base and equipment physical limits;

[0087] C3. Generate parameter combinations: Initialize the parameter population through Latin hypercube sampling, and generate 3-5 Pareto optimal parameter combinations through non-dominated sorting and crowding degree comparison;

[0088] C4. Welding process parameter adjustment: Select parameter combinations according to the defect risk level, adjust the welding process parameters in a small amplitude in stages, and send control instructions to the welding equipment through the data communication bus. In this embodiment, the single adjustment amplitude is controlled within ±10%. When a high risk level is detected, those combinations that can most effectively reduce the defect risk are preferentially selected from the generated parameter combinations. In the medium risk level, a trade-off will be made among reducing the defect risk, improving production efficiency, and controlling energy consumption, and a relatively balanced parameter combination will be selected. In the low risk state, the parameter combinations that can optimize production efficiency and energy consumption will be emphasized.

[0089] Through the above steps, the adjustment strategy is matched from the process knowledge base according to the defect prediction information. The improved NSGA-Ⅲ algorithm is used to seek a balance with the goals of minimizing the defect risk probability, energy consumption, and maximizing production efficiency. While ensuring the welding quality, energy consumption is reduced and production efficiency is improved. Multiple Pareto optimal parameter combinations are generated, and the weights of each target selection are optimized according to the actual risk level. The best parameter combination is intelligently selected, and the welding process parameters are adjusted in stages in small amounts to avoid adverse effects on welding quality due to parameter mutations. Intelligent and refined control of welding process parameters is achieved, and defect prediction and intervention are achieved in advance, which avoids welding defects as much as possible and effectively improves the welding quality.

[0090] For example: When the defect prediction module outputs high risk (crack growth rate 0.2 mm / s), the adaptive control module triggers the NSGA-III algorithm and generates the following parameter combination:

[0091] Solution 1: Current 240A, voltage 28V, welding speed 30cm / min (defect risk reduced by 40%, energy consumption increased by 5%)

[0092] Solution 2: Current 220A, voltage 25V, welding speed 25cm / min (defect risk reduced by 35%, energy consumption increased by 2%)

[0093] Select solution 1 and monitor in real time whether the melt pool image recovers its stable shape.

[0094] The above-mentioned multi-objective optimization algorithm is an improved NSGA-Ⅲ algorithm. The algorithm has been improved in many aspects on the original basis to improve the optimization effect of welding process parameters. The specific improvements include:

[0095] Introducing welding process knowledge base to constrain solution space: The welding process knowledge base contains a large amount of welding process experience and expert knowledge. By introducing this knowledge base, the solution space of the optimization algorithm is constrained to ensure that the searched process parameter combination meets the actual welding process requirements and avoid unreasonable parameter settings;

[0096] Use Latin hypercube sampling to initialize the population: Latin hypercube sampling is an efficient sampling method that can uniformly sample in the solution space to generate a representative initial population. Compared with traditional random sampling methods, Latin hypercube sampling can converge to the optimal solution faster and improve the search efficiency of the algorithm.

[0097] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting the protection scope of the invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art can still, without conflict and without creative efforts, combine, add, delete or make other adjustments to the features in the various embodiments of the present invention according to the circumstances, so as to obtain different technical solutions that essentially do not deviate from the concept of the present invention, and these technical solutions also fall within the scope of protection of the present invention.

Claims

1. A detection system for welding quality, characterized in that Including: A process monitoring layer (1) equipped with a multi-modal sensor array (11) for real-time collection of raw performance data during the welding process. The raw performance data includes arc spectrum, molten pool image, acoustic emission signal, and thermodynamic parameters during the welding process. An intelligent analysis layer (2) deployed with a multi-source data fusion module (21) and a defect prediction module (22). The multi-source data fusion module (21) uses an improved D-S evidence theory algorithm to fuse and process the raw performance data collected by the process monitoring layer (1) to obtain multi-dimensional performance data. The defect prediction module (22) constructs a defect evolution model based on an LSTM-GAN hybrid neural network. The defect evolution model analyzes and captures abnormal data according to the multi-dimensional performance data to dynamically predict the development trend of defects and outputs defect prediction information, including defect risk level, defect risk type, defect location information, defect evolution trend heat map, and estimated failure time. A decision execution layer (3) configured with an adaptive control module (31) for dynamically adjusting welding process parameters according to the defect prediction information output by the intelligent analysis layer (2). A data communication bus (4) for realizing real-time data interaction and closed-loop feedback control among the process monitoring layer (1), the intelligent analysis layer (2), and the decision execution layer (3).

2. The detection system for welding quality according to claim 1, characterized in that, The multi-modal sensor array (11) includes: A high-frequency current probe (111) for collecting current signals during the welding process at a frequency not lower than 1 MHz. A high-speed CMOS camera (112) for collecting molten pool images during the welding process. A broadband acoustic emission sensor (113) with a frequency band set to 50 kHz - 1.2 MHz for collecting unique acoustic emission signals generated by various defects. A multi-spectral infrared thermal imager (114) for collecting thermodynamic parameters during the welding process.

3. The detection system for welding quality according to claim 1, characterized in that, The specific steps for the multi-source data fusion module to fuse and process the raw performance data to obtain multi-dimensional performance data are as follows: Perform wavelet packet transform denoising on the arc spectrum, acoustic emission signal, and thermodynamic parameters in the raw performance data to remove noise interference and obtain effective data reflecting the characteristics of the welding process. Based on the periodic feature points of the welding process, perform spatio-temporal alignment and normalization processing on the denoised effective data through a spatio-temporal alignment algorithm based on dynamic time warping to obtain standardized aligned data. Use an improved D-S evidence theory algorithm to calculate the confidence weight of each sensor in the multi-modal sensor array, and weight-fuse the standardized aligned data after spatio-temporal alignment through an attention mechanism to obtain a signal-level fusion feature vector. Process the molten pool image in the raw performance data using a pre-set improved YOLOv5s model to obtain an image-level feature vector. Perform feature splicing on the signal-level fusion feature vector and the image-level feature vector to generate multi-dimensional performance data.

4. The detection system for welding quality according to claim 3, wherein: The specific improvements of the improved YOLOv5s model include: Adding a deformable convolutional layer in the Backbone part. Replacing the object detection loss function CIoU Loss with EIOU Loss. The output layer fuses the indicators of the molten pool area volatility and the morphology symmetry degree.

5. The detection system for welding quality according to claim 3, characterized in that, The improvement of the D-S evidence theory algorithm lies in introducing an adaptive basic probability assignment function: where m i (A) is the basic probability assignment of the i-th sensor to proposition A, representing the degree of belief in A; w i is the dynamic weight coefficient of the i-th sensor, which is updated in real time through online learning; Confidence i is the confidence of the i-th sensor; λ is the conflict factor adjustment term.

6. The detection system for welding quality according to claim 1, wherein, The defect prediction module captures abnormal data based on the analysis of multi-dimensional performance data and dynamically predicts the development trend of defects. The specific steps for outputting defect prediction information are as follows: Slice the multi-dimensional performance data according to the time series and perform normalization processing to construct a time series data set; Concatenate the signal-level fusion feature vector and the image-level feature vector and perform weighted fusion through the attention mechanism to enhance the pre-labeled key features; Use the defect evolution model to capture the time-dependent relationship of defect evolution for future defect feature prediction and output defect prediction information.

7. The detection system for welding quality according to claim 1, wherein: The adaptive control module dynamically adjusts the welding process parameters according to the defect prediction information output by the intelligent analysis layer. The specific steps are as follows: Receive and parse the defect prediction information in real time, and match the corresponding welding process parameter adjustment strategy from the preset process knowledge base; Through the improved NSGA-Ⅲ algorithm, with the goal of minimizing the defect risk probability, energy consumption, and maximizing production efficiency, perform multi-objective optimization under the constraints of the process knowledge base and equipment physical limitations; Initialize the parameter population through Latin hypercube sampling, and generate 3-5 Pareto optimal parameter combinations through non-dominated sorting and crowding degree comparison; Select the parameter combination according to the defect risk level, adjust the welding process parameters in a small amplitude in stages, and send control instructions to the welding equipment through the data communication bus.

8. The detection system for welding quality according to claim 1, wherein, The defect location information includes the transverse position inside the weld, the longitudinal position inside the weld, the depth position inside the weld, the three-dimensional space positioning based on augmented reality, and the position based on the fusion of augmented reality and the actual scene.

Images