Pressure-resistant shell laser-arc hybrid welding quality detection method and system based on machine vision

Abstract

The invention discloses a pressure-resistant shell laser-arc hybrid welding quality detection method and system based on machine vision, and the method comprises the steps: obtaining and processing real-time welding images of pressure-resistant shell laser-arc hybrid welding at different visual angles, and extracting feature images of a molten pool and a groove at different visual angles; carrying out three-dimensional fusion on the feature image, carrying out time sequence conversion, and decomposing the converted time sequence by adopting an ensemble empirical mode to obtain an intrinsic mode component IMF and a residual error; and according to the obtained intrinsic mode function IMF and the residual error, through the trained bidirectional long-short-term memory network model, the welding state and the welding seam quality are recognized, and pressure-resistant shell laser-arc hybrid welding quality detection is achieved. The mode of combining multi-angle visual information collection and efficient image processing is adopted, and the real-time monitoring capacity for the welding quality of the pressure-resistant shell in the welding process is comprehensively improved. Based on an EEMD-BiLSTM deep learning model, precise real-time detection of the laser-arc hybrid welding quality of the pressure-resistant shell is achieved.

Description

Technical Field

[0001] The invention relates to welding quality detection, and in particular to a method and system for detecting the quality of laser-arc composite welding of a pressure shell based on machine vision. Background Art

[0002] The quality of pressure hull welding is directly related to the safety and reliability of the submersible. As a key part of the welding process, the quality of laser-arc hybrid welding plays a decisive role in the formation of the weld and the performance of the overall welded structure. However, defects that may occur during the welding process, such as lack of fusion, pores, cracks, etc., seriously affect the density and structural integrity of the weld.

[0003] Laser-arc hybrid welding technology has been widely used in the welding field due to its advantages such as high efficiency and low heat input. However, its welding process is complex, and real-time monitoring and quality control are difficult, especially for the accurate extraction and analysis of key physical quantities such as molten pool morphology and groove characteristics during welding. Existing monitoring systems often have difficulty in accurately evaluating weld quality in complex environments and cannot meet the increasingly high welding quality requirements. Summary of the invention

[0004] Purpose of the invention: In view of the above problems, the present invention provides a method and system for detecting the quality of pressure-resistant shell laser-arc hybrid welding based on machine vision, which can efficiently and accurately detect the welding quality.

[0005] Technical solution: To solve the above problems, the present invention adopts a pressure hull laser-arc hybrid welding quality detection method based on machine vision, which includes the following steps:

[0006] (1) Obtaining real-time welding images of pressure hull laser-arc hybrid welding at different viewing angles;

[0007] (2) Processing the real-time welding images acquired at different viewing angles to extract characteristic images of the molten pool and groove at different viewing angles;

[0008] (3) performing three-dimensional fusion on the characteristic images of the molten pool and the groove at different viewing angles to obtain a fused characteristic image;

[0009] (4) Performing time series conversion on the fused feature image, and decomposing the converted time series using the ensemble empirical mode to obtain the intrinsic mode component IMF and residual;

[0010] (5) The intrinsic mode component IMF and residual are used as the input of the bidirectional long short-term memory network model, the welding state and weld quality are used as the output of the bidirectional long short-term memory network model, and the trained bidirectional long short-term memory network model is obtained through training samples;

[0011] (6) Based on the obtained intrinsic mode function IMF and residual, the welding state and weld quality are identified through the trained bidirectional long short-term memory network model to realize the quality detection of pressure hull laser-arc hybrid welding.

[0012] Furthermore, the processing of the real-time welding images acquired from different viewing angles in step (2) includes: noise removal, edge enhancement, histogram equalization, grayscale stretching, and morphological processing.

[0013] Furthermore, the noise removal includes Gaussian filtering, edge enhancement uses Canny edge detection, and morphological processing includes dilation and erosion.

[0014] Furthermore, the specific steps of three-dimensionally fusing the characteristic images of the molten pool and the groove at different viewing angles are: converting the coordinate system of the characteristic images of the molten pool and the groove at different viewing angles, and obtaining the fused characteristic images based on feature point matching through a three-dimensional reconstruction algorithm.

[0015] Furthermore, the time series conversion of the fused feature image includes taking the molten pool morphology, groove characteristics and optical changes during the welding process as time series data.

[0016] Furthermore, the specific formula for decomposition using the set empirical mode is:

[0017]

[0018] Among them, IMF i (t) is the final obtained ith intrinsic mode function, x(t) is the original signal, i.e. the time series after conversion; n j (t) is the jth added white noise; K is the number of IMFs obtained by EMD decomposition; EMD(x(t)+n j (t)) is the result of EMD on the noisy signal; (EMD(x(t)+n j (t))) i is the i-th intrinsic mode component IMF; M is the number of white noise samples.

[0019] Furthermore, the bidirectional long short-term memory network model trained in step (5) is optimized and enhanced through regularization and grid search hyperparameter tuning.

[0020] Furthermore, the formula for the bidirectional long short-term memory network model recognition process is:

[0021]

[0022] Among them, h t is the hidden state of the bidirectional long short-term memory network model BiLSTM at time step t, and is the forward hidden state and the reverse hidden state connection; and They are the weight matrices of forward LSTM and reverse LSTM respectively; and is the bias vector of the forward LSTM and the backward LSTM; σ is the sigmoid activation function; Indicates the connection operation, x t For input data.

[0023] Furthermore, the welding state includes molten pool stability and groove integrity; the weld quality includes molten pool stability, groove integrity, weld density and surface finish.

[0024] The present invention also adopts a pressure hull laser-arc hybrid welding quality detection system based on machine vision, comprising:

[0025] Welding image acquisition module, used to obtain real-time welding images of pressure hull laser-arc hybrid welding at different viewing angles;

[0026] The image preprocessing module is used to process the real-time welding images acquired at different viewing angles, extract the characteristic images of the molten pool and the groove at different viewing angles; perform three-dimensional fusion on the characteristic images of the molten pool and the groove at different viewing angles to obtain the fused characteristic images;

[0027] The time series analysis module is used to perform time series conversion on the fused feature image and decompose the converted time series using the set empirical mode to obtain the intrinsic mode component IMF and residual;

[0028] A model building module is used to use the intrinsic mode component IMF and residual as the input of the bidirectional long short-term memory network model, and the welding state and weld quality as the output of the bidirectional long short-term memory network model, and obtain the trained bidirectional long short-term memory network model through training samples;

[0029] The detection module is used to identify the welding state and weld quality according to the obtained intrinsic mode function IMF and residual through the trained bidirectional long short-term memory network model to realize the quality detection of pressure hull laser-arc hybrid welding.

[0030] Beneficial effects: Compared with the prior art, the significant advantage of the present invention is that it adopts a combination of multi-angle visual information acquisition and efficient image processing to comprehensively improve the real-time monitoring capability of the pressure shell welding quality during the welding process. Based on the EEMD-BiLSTM deep learning model, accurate real-time detection of the quality of laser-arc composite welding of the pressure shell is achieved. The time-frequency features are effectively extracted by the ensemble empirical mode decomposition method, and combined with the bidirectional long short-term memory network (BiLSTM), the classification accuracy and robustness of the welding state and weld quality are improved. It reduces the dependence on traditional non-destructive testing technology, realizes rapid detection and feedback of the quality of laser-arc composite welding of the pressure shell, and greatly improves the detection efficiency and reliability in the production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the quality detection method of the present invention;

[0032] Figure 2 It is a structural schematic diagram of the pressure hull laser-arc hybrid welding quality detection system of the present invention;

[0033] Figure 3 This is the processing flow of the image preprocessing module in the present invention;

[0034] Figure 4 It is a schematic diagram of the principle flow of the EEMD-BiLSTM model in the present invention;

[0035] Figure 5 It is the confusion matrix for the quality classification of pressure hull laser-arc hybrid welding in the present invention. DETAILED DESCRIPTION

[0036] Example 1

[0037] Depend on Figure 1 As shown, in this embodiment, a pressure hull laser-arc hybrid welding quality detection method based on machine vision includes the following steps:

[0038] (1) Obtaining real-time welding images of pressure hull laser-arc hybrid welding at different viewing angles; Figure 2 As shown, each high-speed camera is used to collect multi-view real-time images of the pressure shell welding area during the welding process.

[0039] (2) Figure 3 As shown, the real-time welding images acquired at different viewing angles are processed to extract the characteristic images of the molten pool and the groove at different viewing angles; the collected images are preprocessed including noise removal, edge enhancement, histogram equalization, grayscale stretching, and morphological processing.

[0040] (3) The characteristic images of the molten pool and the groove under different viewing angles are fused in three dimensions to obtain the fused characteristic image; the two-dimensional images taken by each camera are mapped into a unified three-dimensional coordinate system, and the internal and external parameters of the camera are used for correction and matching; based on stereo vision and depth estimation algorithms, a three-dimensional point cloud of the weld area is generated; by fusing multi-angle data, the three-dimensional shape of the weld groove is reconstructed, and the characteristics such as weld width, depth, and groove integrity are identified.

[0041] (4) Figure 4 As shown in the figure, the fused feature image is converted into a time series, and the converted time series is decomposed using the set empirical mode to obtain the intrinsic mode component IMF and residual; key features related to welding quality, such as molten pool shape, groove edge, weld width, etc., are extracted from each frame of the fused image. These features are input into the subsequent analysis model as time series data. These feature data can directly reflect the quality changes in the welding process. Through multi-dimensional analysis, the system can capture the key indicators of welding quality and provide accurate basic data for subsequent time series analysis and classification.

[0042] Combined with the real-time collected welding data, the status of different time periods in the welding process is tracked and monitored. The deep learning time series model can effectively capture the long-term dependencies in the time series data and predict the morphological changes that may occur during the welding process. This module can not only detect the stability of the welding process, but also capture abnormal changes in real time, such as uneven welds or abnormal temperature fluctuations during welding. By processing time series data, the system can predict possible welding problems in advance and provide timely feedback to the control system to ensure that the welding quality is effectively guaranteed.

[0043] (5) The intrinsic mode component IMF and residual are used as the input of the bidirectional long short-term memory network model, and the welding state and weld quality are used as the output of the bidirectional long short-term memory network model. The trained bidirectional long short-term memory network model is obtained through training samples. The welding quality is divided into different levels, and the evaluation criteria include the integrity of the welding groove, the stability of the molten pool, the density of the weld and the surface finish. The quality status of the welding process is detected and classified in real time to determine whether the welding result meets the set process standards.

[0044] (6) Based on the obtained intrinsic mode function IMF and residual, the welding state and weld quality are identified through the trained bidirectional long short-term memory network model to realize the quality detection of pressure hull laser-arc hybrid welding. The system supports multiple classification levels, such as groove integrity can be divided into complete, partially missing and severely missing, and weld density can be divided into high, medium and low density. The detection classification results will be directly fed back to the welding operators to help them make adjustments according to the actual situation and further optimize the welding quality.

[0045] like Figure 5 As shown in the figure, after the detection and classification are completed, the weld quality assessment module conducts a comprehensive analysis of the welding results, and combines the various feature data extracted in the early stage to evaluate the overall effect of welding. Combined with the three-dimensional morphological data and timing analysis results after multi-angle image fusion, the system can realize real-time monitoring and feedback adjustment of the welding process. The system not only pays attention to the surface state of the weld, but also makes a comprehensive judgment on the internal welding defects, weld uniformity and heat-affected zone. The evaluation report generated by this module can provide an important decision-making basis for subsequent welding quality control and process optimization. The system feeds back the evaluation results to the welding control system through an automated feedback mechanism, so as to automatically adjust the parameters in the next welding process to further improve welding efficiency and quality.

[0046] Through the above steps, all-round detection and evaluation of the quality of laser-arc hybrid welding of the pressure hull is realized, which effectively improves the automation degree and detection accuracy of welding quality detection, reduces the necessity of manual intervention, has a higher degree of automation and real-time performance, and helps to improve the quality control level of the welding process.

[0047] Example 2

[0048] In this embodiment, a pressure hull laser-arc hybrid welding quality detection system based on machine vision includes:

[0049] The welding image acquisition module obtains real-time image data during the laser-arc hybrid welding process. Figure 2 As shown in the figure, the module is equipped with a high-speed camera, which can accurately capture the details of the welding area. In order to ensure the stability of the acquired image, the system uses autofocus and anti-vibration technology to reduce the interference of the image caused by environmental factors. In this embodiment, the resolution of the high-speed camera used by the system reaches 1920×1080 pixels, ensuring the high accuracy and clarity of the image. The image acquisition rate is set to 30 frames per second to ensure that the dynamic changes of the welding process can be accurately captured.

[0050] Image preprocessing module, the image preprocessing module optimizes the original image to improve the analysis effect. Use Gaussian filtering to remove noise. Edge enhancement is achieved through Canny edge detection, and the Canny algorithm accurately detects complex edges. Contrast is improved through histogram equalization technology, and grayscale stretching enhances image details. Morphological processing includes corrosion and expansion to remove noise, connect objects, and adjust morphological features. Precise timestamps or hardware synchronization signals are used to ensure that the data collected by each camera is synchronized in time to avoid image mismatch due to time differences.

[0051] For the image data collected by cameras at different angles, coordinate system conversion is performed to uniformly map the images at each angle into a standard three-dimensional coordinate system. The adopted method includes multi-camera calibration based on the calibration plate, and coordinate mapping is achieved through the camera internal and external parameter matrices. Through the three-dimensional reconstruction algorithm, the image data at different angles are fused to generate a three-dimensional model of the molten pool and groove. The method based on feature point matching is adopted to ensure that the images at different angles can be accurately aligned and fused, and finally obtain the complete three-dimensional morphological changes of the welding process.

[0052] The time series analysis module uses the ensemble empirical mode decomposition (EEMD) algorithm to extract time-frequency domain features from the preprocessed image data. Key features related to welding quality, such as molten pool shape, groove edge, weld width, etc., are extracted from each frame of the fused image. The multi-dimensional data transmitted by the feature extraction module is further analyzed. This module monitors the dynamic changes of the welding state by tracking the time series data during the welding process, captures the abnormal fluctuations generated during the welding process, and provides timely feedback for subsequent quality assessment.

[0053] The model building module deeply analyzes the dynamic changes of the molten pool and groove morphology during welding, and adopts a deep learning time series model (EEMD-BiLSTM). The image feature sequence after multi-angle fusion is used as the input of the time series model, and the morphological prediction results at each time step (such as the position of the molten pool boundary, the trend of groove morphological changes, etc.) are output. These models can effectively capture the long-term dependencies in the time series data and predict the morphological changes that may occur during the welding process.

[0054] The optimization and adjustment module automatically adjusts the model through Dropout and L2 regularization and grid search (GS) hyperparameter tuning based on the image analysis and time series monitoring results to ensure continuous optimization of welding quality. The time series model can perform morphological prediction on multi-angle image features input in real time. Through the prediction results of continuous time steps, the dynamic changes of the molten pool and groove morphology can be captured and compared with the actual observation data in real time. If abnormal morphological changes are found (such as groove deformation or uneven molten pool), the feedback mechanism is triggered to dynamically adjust the welding parameters. This module can effectively identify unstable phenomena in the welding process and ensure continuous monitoring of weld quality.

[0055] The detection module detects and classifies the real-time input welding image data based on the trained EEMD-BiLSTM model to determine the quality status of the welding. The system feeds back the classification results to the operator or control system in real time to guide the adjustment and optimization of the welding process. In this embodiment, the molten pool stability is divided into two levels: stable and unstable; the welding groove integrity is divided into three levels: complete, partially missing, and severely missing; the weld density is divided into three levels: high density, medium density, and low density; the surface finish is divided into three levels: smooth, medium smooth, and rough. This classification is based on a comprehensive evaluation of multiple quality features, such as the uniformity of the weld, the distribution of welding defects, and the width of the heat-affected zone.

Claims

1. A method for detecting the quality of laser-arc hybrid welding of a pressure hull based on machine vision, characterized in that: The following steps are involved: (1) Obtaining real-time welding images of pressure hull laser-arc hybrid welding at different viewing angles; (2) Processing the real-time welding images acquired at different viewing angles to extract characteristic images of the molten pool and groove at different viewing angles; (3) performing three-dimensional fusion on the characteristic images of the molten pool and the groove at different viewing angles to obtain a fused characteristic image; (4) Performing time series conversion on the fused feature image, and decomposing the converted time series using the ensemble empirical mode to obtain the intrinsic mode component IMF and residual; (5) The intrinsic mode component IMF and residual are used as the input of the bidirectional long short-term memory network model, the welding state and weld quality are used as the output of the bidirectional long short-term memory network model, and the trained bidirectional long short-term memory network model is obtained through training samples; (6) Based on the obtained intrinsic mode function IMF and residual, the welding state and weld quality are identified through the trained bidirectional long short-term memory network model to realize the quality detection of pressure hull laser-arc hybrid welding.

2. The pressure hull laser-arc hybrid welding quality detection method according to claim 1 is characterized in that: The processing of the real-time welding images acquired at different viewing angles in step (2) includes: noise removal, edge enhancement, histogram equalization, grayscale stretching, and morphological processing.

3. The pressure hull laser-arc hybrid welding quality detection method according to claim 2 is characterized in that: The noise removal includes Gaussian filtering, and the edge enhancement adopts Canny edge detection; the morphological processing includes expansion and corrosion.

4. The pressure hull laser-arc hybrid welding quality detection method according to claim 3 is characterized in that: The specific steps of three-dimensionally fusing the characteristic images of the molten pool and the groove at different viewing angles are: converting the characteristic images of the molten pool and the groove at different viewing angles into coordinate systems, and obtaining the fused characteristic images based on feature point matching through a three-dimensional reconstruction algorithm.

5. The pressure hull laser-arc hybrid welding quality detection method according to claim 4 is characterized in that: The time series conversion of the fused feature image includes taking the molten pool morphology, groove characteristics and optical changes in the welding process as time series data.

6. The pressure hull laser-arc hybrid welding quality detection method according to claim 5 is characterized in that: The specific formula for decomposition using the set empirical mode is: Among them, IMF i (t) is the final obtained ith intrinsic mode function, x(t) is the original signal, i.e. the time series after conversion; n j (t) is the jth added white noise; K is the number of IMFs obtained by EMD decomposition; EMD(x(t)+n j (t)) is the result of EMD on the noisy signal; (EMD(x(t)+n j (t))) i is the i-th intrinsic mode component IMF; M is the number of white noise samples.

7. The pressure hull laser-arc hybrid welding quality detection method according to claim 6 is characterized in that: The bidirectional long short-term memory network model trained in step (5) is optimized and enhanced through regularization and grid search hyperparameter tuning.

8. The pressure hull laser-arc hybrid welding quality detection method according to claim 7 is characterized in that: The formula for the bidirectional long short-term memory network model recognition process is: Among them, h t is the hidden state of the bidirectional long short-term memory network model BiLSTM at time step t, and is the forward hidden state and the reverse hidden state connection; and They are the weight matrices of forward LSTM and reverse LSTM respectively; and is the bias vector of the forward LSTM and the backward LSTM; σ is the sigmoid activation function; Indicates the connection operation, x t For input data.

9. The pressure hull laser-arc hybrid welding quality detection method according to claim 8 is characterized in that: The welding state includes molten pool stability and groove integrity; the weld quality includes molten pool stability, groove integrity, weld density and surface finish.

10. A pressure hull laser-arc hybrid welding quality inspection system based on machine vision, characterized in that: include: Welding image acquisition module, used to obtain real-time welding images of pressure hull laser-arc hybrid welding at different viewing angles; The image preprocessing module is used to process the real-time welding images acquired at different viewing angles, extract the characteristic images of the molten pool and the groove at different viewing angles; perform three-dimensional fusion on the characteristic images of the molten pool and the groove at different viewing angles to obtain the fused characteristic images; The time series analysis module is used to perform time series conversion on the fused feature image and decompose the converted time series using the set empirical mode to obtain the intrinsic mode component IMF and residual; A model building module is used to use the intrinsic mode component IMF and residual as the input of the bidirectional long short-term memory network model, and the welding state and weld quality as the output of the bidirectional long short-term memory network model, and obtain the trained bidirectional long short-term memory network model through training samples; The detection module is used to identify the welding state and weld quality according to the obtained intrinsic mode function IMF and residual through the trained bidirectional long short-term memory network model to realize the quality detection of pressure hull laser-arc hybrid welding.

Images