Method and system for predicting mechanical properties of welded joints based on cascaded neural networks

Abstract

The application discloses a kind of based on cascading neural network's welding joint mechanical property prediction method and system, collect laser welding process in photo-induced plasma topography image, and pass to the performance test of the weld obtained by welding and obtain a variety of weld performance parameter data;The feature extraction is carried out to the photo-induced plasma topography image collected, the corresponding relationship is established to the image feature extracted with weld performance parameter, and training set, verification set and test set are constructed;The welding joint mechanical property prediction model based on cascading neural network is constructed;The welding joint mechanical property prediction model trained is used to carry out real-time weld mechanical property prediction.The application uses the prediction model based on cascading neural network to fit and predict the mechanical property of titanium-aluminum swing laser powder filling welding joint, can realize welding joint mechanical property real-time prediction, reduce artificial workload, improve welding quality.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of welding performance prediction, in particular to a welding joint mechanical property prediction method and system based on a cascade neural network. BACKGROUND

[0002] In modern manufacturing, titanium-aluminum composite structures have been widely used in aerospace, automobile manufacturing and other fields due to their high strength and low weight, such as the manufacture of aircraft cabin heat sinks, seat rails, and wing honeycomb sandwich components. Titanium-aluminum swing laser powder filling welding technology, as a key means to achieve efficient connection of titanium-aluminum, combines high laser energy density, narrow heat-affected zone, and adjustable weld composition through powder filling, and exhibits significant advantages in precise control of heat input, reduction of joint peak temperature, and suppression of brittle intermetallic compound formation, which is of great significance to improving the quality and performance of titanium-aluminum welded joints.

[0003] However, current research on titanium-aluminum swing laser powder filling welding performance prediction still has many defects. On the one hand, traditional prediction methods based on empirical formulas or simple physical models cannot fully consider the complex thermophysical, metallurgical reactions and multi-parameter coupling in the welding process, resulting in low prediction accuracy and inability to meet actual engineering needs. On the other hand, although existing experimental research can obtain weld performance data under specific process conditions, it is costly, time-consuming, and difficult to systematically evaluate and predict weld performance over a wide range of process parameters. SUMMARY

[0004] The present application provides a welding joint mechanical property prediction method and system based on a cascade neural network to solve the problems of low prediction accuracy and high workload in existing titanium-aluminum swing laser powder filling welding performance prediction methods.

[0005] According to a first aspect, a welding joint mechanical property prediction method based on a cascade neural network is provided in an embodiment, the method comprising:

[0006] Collecting plasma morphology images during laser welding, and obtaining a plurality of weld performance parameter data by testing the welds obtained by welding;

[0007] Extracting features from the collected plasma morphology images, establishing a correspondence between the extracted image features and the weld performance parameters, and constructing a training set, a validation set, and a test set;

[0008] Constructing a welding joint mechanical property prediction model based on a cascade neural network, and training, validating, and testing the welding joint mechanical property prediction model using the established training set, validation set, and test set;

[0009] The trained mechanical property prediction model of the welding joint is used to predict the mechanical property of the welding seam in real time.

[0010] Further, the light-induced plasma topography image in the laser welding process is collected, specifically including:

[0011] A visual test platform is used for multiple welding tests and image collection, the visual test platform including a welding device for welding test and a visual device for image collection;

[0012] The visual device includes a camera and an image processing workstation, the light-induced plasma topography image in the welding process is collected by the camera and fed back to the image processing workstation.

[0013] Further, a plurality of welding performance parameter data is obtained by performance testing of the welding seam, specifically including:

[0014] The welding seam obtained by multiple welding tests is subjected to cross-section morphology measurement, chemical composition scanning and mechanical tensile test, and a plurality of performance parameters including welding seam side penetration, welding seam cross-sectional area, welding seam cross-sectional zirconium content and tensile strength are obtained.

[0015] Further, feature extraction is performed on the collected light-induced plasma topography image, specifically including:

[0016] The plasma centroid, area and deflection angle information are extracted from the light-induced plasma topography image;

[0017] The extracted plasma centroid, area and deflection angle information are filtered and feature signal decomposed;

[0018] The plurality of feature signals obtained by decomposition are subjected to feature screening by correlation analysis, and the screened feature signals are subjected to dimension reduction processing.

[0019] Further, a welding joint mechanical property prediction model based on a cascade neural network is constructed, specifically including:

[0020] The welding joint mechanical property prediction model includes a BPNN network, an XGBoost network and a random forest network arranged in cascade.

[0021] Further, a welding joint mechanical property prediction model based on a cascade neural network is constructed, specifically including:

[0022] A BPNN (Back Propagation neural network) network structure is defined, the BPNN network taking the extracted image features as input and the welding seam side penetration and the welding seam cross-sectional area as output;

[0023] Define the XGBoost (eXtreme Gradient Boosting) network structure, and the XGBoost network takes the weld side penetration and weld cross-sectional area output by the BPNN network as input, and takes the weld cross-sectional zirconium content as output;

[0024] Define the random forest network structure, and the random forest network takes the weld cross-sectional zirconium content output by the XGBoost network as input, and takes the weld mechanical property parameters as output.

[0025] Further, the established training set, validation set and test set are used to train, validate and test the weld joint mechanical property prediction model, specifically including:

[0026] Define the training parameters of the BPNN network and initialize the network;

[0027] Define the training parameters of the XGBoost network and initialize the network;

[0028] Define the training parameters of the random forest network and initialize the network;

[0029] Define the training process, use the training set, validation set and test set to train, validate and test the BPNN, XGBoot and random forest network, and save the trained network parameters after training;

[0030] Cascade the trained BPNN, XGBoot and random forest network to establish a weld joint mechanical property prediction model, and verify and evaluate the model.

[0031] Further, the trained weld joint mechanical property prediction model is used for real-time weld mechanical property prediction, specifically including:

[0032] After image feature extraction, the real-time acquisition of the plasma morphology image in the welding process is input into the trained weld joint mechanical property prediction model, and the weld mechanical property prediction result is output.

[0033] Further, the method further includes:

[0034] According to the weld mechanical property prediction result, the welding parameters are adjusted in real time, including welding laser power, laser swing frequency, laser swing angle and welding speed.

[0035] According to the second aspect, an embodiment provides a welding joint mechanical property prediction system based on a cascade neural network, the system comprising:

[0036] The data acquisition module is used to acquire photo-induced plasma morphology images during the laser welding process and obtain various weld performance parameters by performing performance tests on the weld seam.

[0037] The dataset construction module is used to extract features from the collected photo-induced plasma morphology images, establish a correspondence between the extracted image features and weld performance parameters, and construct training, validation, and test sets.

[0038] The model building and training module is used to build a prediction model of the mechanical properties of welded joints based on cascaded neural networks. The established training set, validation set and test set are used to train, validate and test the prediction model of the mechanical properties of welded joints.

[0039] The prediction module is used to predict the mechanical properties of welds in real time using a trained prediction model of the mechanical properties of welded joints.

[0040] This invention provides a method and system for predicting the mechanical properties of welded joints based on a cascaded neural network. The method involves acquiring photo-induced plasma morphology images during laser welding and obtaining various weld performance parameters through performance testing. Feature extraction is performed on the acquired photo-induced plasma morphology images, establishing a correspondence between the extracted image features and weld performance parameters, and constructing training, validation, and test sets. A welded joint mechanical property prediction model based on a cascaded neural network is then constructed, and the model is trained, validated, and tested using the established training, validation, and test sets. The trained model is then used for real-time prediction of weld mechanical properties. This invention processes and extracts features from photo-induced plasma images obtained using visual methods, and then uses a prediction model based on a cascaded neural network to fit and predict the mechanical properties of titanium-aluminum oscillating laser powder-filled welded joints. This invention enables real-time prediction of welded joint mechanical properties, reduces manual workload, and improves welding quality. Attached Figure Description

[0041] Figure 1 A flowchart illustrating a method for predicting the mechanical properties of welded joints based on a cascaded neural network, as provided in one embodiment of the present invention;

[0042] Figure 2 Photoinduced plasma morphology image in a method for predicting the mechanical properties of welded joints based on cascaded neural networks, provided as an embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram of the logical structure of a welding joint mechanical property prediction system based on a cascaded neural network, provided as an embodiment of the present invention. Detailed Implementation

[0044] The application will be described in further detail below with specific reference to the drawings. Like elements in different embodiments are denoted by like reference numerals. In the following description, numerous specific details are described to provide a thorough understanding of the application. However, one of ordinary skill in the art will recognize that the application can be practiced without one or more of the specific details. In other instances, well-known structures have not been described in order to avoid obscuring the application. The detailed description, therefore, is not to be taken in a limiting sense.

[0045] In addition, the features, operations or characteristics described in the specification can be combined in any appropriate manner to form various embodiments. Meanwhile, the steps or actions in the method description can also be sequentially adjusted or changed in a manner that is obvious to those skilled in the art. Therefore, the various sequences in the specification and drawings are only for the purpose of clearly describing a certain embodiment, and do not mean that the sequence is necessary, unless otherwise stated that a certain sequence must be followed.

[0046] The first embodiment of the application provides a welding joint mechanical property prediction method based on a cascade neural network. The following will be described in detail in combination with Figure 1 the drawings.

[0047] As shown in Figure 1 step S100, the plasma morphology image in the laser welding process is collected, and a plurality of weld performance parameter data is obtained by testing the performance of the weld.

[0048] The above steps specifically include:

[0049] S110, collecting the plasma morphology image

[0050] In this embodiment, a visual test platform is built to perform multiple welding tests and image collection. The visual test platform includes a welding device for welding test and a visual device for image collection. The visual device includes a camera and an image processing workstation. The plasma morphology image in the welding process is collected by the camera and fed back to the image processing workstation.

[0051] In this embodiment, the welding device includes a 10kw disc scanning galvanometer laser, a laser light source, a powder feeder, and a specially treated welding part. Specifically, the titanium-aluminum swing laser powder filling welding uses a swing laser, and the welding form is lap welding. The specific assembly method is that the upper titanium plate is below the aluminum plate, and a 1mm wide small groove is opened above the aluminum plate in the laser welding path, and Zr powder is filled in the groove to control the weld performance.

[0052] The plasma generated by laser welding is essentially a photo-induced plasma generated by the vaporization of metal materials in the laser welding process. Its form has strong periodicity and changes with the change of the molten pool form. The test results show that during the welding process, the increase of the molten pool volume will cause the increase of the volume of the plasma group. At the same time, the change of the molten depth of the molten pool side wall will directly affect the angle of the plasma generation, and then affect the force of the plasma to make it deflect. Therefore, the plasma area is positively correlated with the volume of the molten pool, and the plasma deflection angle and the centroid can reflect the size of the molten depth of the molten pool side wall. The plasma morphology is shown in Figure 2

[0053] In this embodiment, a high-speed camera is used to collect the plasma morphology image of the titanium-aluminum laser powder filling welding process and feed back to the image processing workstation. The image processing workstation includes a model training module and a model deployment module. The model training module is used to train the initial model by inputting the training set obtained by division. The model deployment module is used to deploy the trained model, input the image data stream in real time through the high-speed camera, and predict the mechanical properties of the weld in real time to feed back the welding state.

[0054] S120, weld test

[0055] In this embodiment, the weld obtained by welding is geometrically measured, chemically scanned and mechanically tested. Specifically, the cross-sectional morphology of different parts of the weld is measured, the composition is scanned, and the tensile test is performed to obtain various performance parameters such as weld side penetration h, weld cross-sectional area A, weld cross-sectional Zr element content, tensile strength, etc.

[0056] As shown in Figure 1 In step S200, the collected photo-induced plasma morphology image is feature extracted, the extracted image features are correlated with the weld performance parameters, and the training set, the validation set and the test set are constructed.

[0057] The above steps specifically include:

[0058] S210, image feature extraction, the specific steps are as follows:

[0059] S211, the plasma centroid, area and deflection angle information are extracted from the photo-induced plasma morphology image by image feature extraction algorithm;

[0060] ​S212, the extracted plasma centroid, area and deflection angle information are filtered and feature signal decomposed;

[0061] S213, the multiple feature signals obtained from the decomposition are screened by correlation analysis, and then the screened feature signals are subjected to dimensionality reduction processing.

[0062] Specifically, in this embodiment, the plasma centroid, area, and deflection angle feature information are extracted from the photoinduced plasma morphology image. Then, these three feature information are filtered by EEMD (Ensemble Empirical Mode Decomposition) and the feature signals are decomposed, resulting in 48 feature signals. Then, MIC (Maximal information coefficient) correlation analysis is used to extract 23 feature signals that are highly correlated with the weld side penetration depth h and weld cross-sectional area A. Finally, PCA (Principal Components Analysis) is used to reduce the dimensionality to 10 composite abstract feature signals.

[0063] S220, Dataset Partitioning

[0064] After multiple experiments, plasma images of titanium-aluminum oscillating laser powder-filled welding were obtained. Cross-sectional morphology measurements, composition scanning, and tensile tests were performed on different parts of the weld to obtain performance parameters such as weld side penetration depth h, weld cross-sectional area A, Zr element content in the weld cross-section, and tensile strength. These parameters were then matched one-to-one with the plasma images generated when welding to that part, establishing a plasma morphology image and weld performance database. Training, validation, and test sets were then packaged and saved in a ratio of 0.7:0.15:0.15.

[0065] like Figure 1 As shown, in step S300, a prediction model for the mechanical properties of welded joints based on a cascaded neural network is constructed, and the established training set, validation set, and test set are used to train, validate, and test the prediction model for the mechanical properties of welded joints.

[0066] The above steps specifically include:

[0067] S310, Predictive Model Construction

[0068] In this embodiment, the mechanical property prediction model for welded joints includes a BPNN network, an XGBoost network, and a random forest network cascaded in sequence.

[0069] Constructing a prediction model for the mechanical properties of welded joints based on cascaded neural networks, specifically including:

[0070] S311 defines the BPNN network structure. The BPNN network takes 10 extracted composite abstract feature signals as input and the weld side penetration depth and weld cross-sectional area as output.

[0071] Specifically, in this embodiment, the number of hidden layers in the BPNN network is set to 2, the number of neurons in each layer is set to 10-15-10-2, the ReLU (Linear rectification function) function is used as the activation function between the input layer and the hidden layer, and between the hidden layers, and the linear activation function is set as the activation function between the hidden layer and the output layer.

[0072] S312 defines the XGBoost network structure. The XGBoost network takes the weld side penetration depth and weld cross-sectional area output by the BPNN network as input and the zirconium content of the weld cross-section as output.

[0073] S313 defines the structure of a random forest network. The random forest network takes the zirconium content of the weld section output by the XGBoost network as input and the mechanical property parameters of the weld as output.

[0074] S320, model training, specifically includes:

[0075] S321, Define the training parameters of the BPNN network and initialize the network.

[0076] In this embodiment, the BPNN network is trained with an initial learning rate of 0.001. In order to train the neural network more efficiently and robustly and improve the final performance of the model, the Adam optimizer and cosine annealing learning rate scheduling strategy are adopted.

[0077] S322, Define the training parameters of the XGBoost network and initialize the network.

[0078] In this embodiment, the evaluation metric is set as MAE (Mean Absolute Error); the network training parameters include the number of trees (300, 500), the maximum tree depth (10, 15), the learning rate (0.05, 0.1), the sample ratio (0.8, 1.0), and the feature ratio (0.8, 1.0); and grid search is used to select hyperparameters, and 3-fold cross-validation is used to reduce the random dependence of the evaluation results on a single data split, providing a more robust estimate of the model's generalization ability.

[0079] S323 defines the training parameters of the random forest network and initializes the network.

[0080] In this embodiment, the evaluation metric is set as MAE; the network training parameters include the number of trees (500, 1000), the maximum depth of the trees (15, 20, 15), the minimum number of samples required for node splits (2, 5, 10), the minimum number of samples required for leaf nodes (1, 2), and the maximum number of features during tree splits (1, 2); and grid search is used to select the optimal parameters, and 3-fold cross-validation is used to reduce the model's reliance on randomness in the data.

[0081] S324 defines the training process, using training, validation, and test sets to train, validate, and test BPNN, XGBoot, and random forest networks, and saving the trained network parameters after training is completed.

[0082] In this embodiment, MAE, MSE (Mean Square Error), and coefficient of determination (R²) are used for each layer of the network module. 2 This is used to evaluate the network's prediction accuracy, error magnitude, and ability to interpret data variability (sensitivity).

[0083] S325: The trained BPNN, XGBoot and random forest networks are cascaded to build a prediction model for the mechanical properties of welded joints, and the model is validated and evaluated.

[0084] In this embodiment, the trained BPNN, XGBoot, and random forest network modules are cascaded and validated using a brand-new dataset. The model accuracy is evaluated using the coefficient of determination and mean absolute error.

[0085] like Figure 1 As shown, in step S400, the mechanical properties of the weld are predicted in real time using the trained mechanical property prediction model of the weld joint.

[0086] The above steps specifically include:

[0087] S410 extracts image features from photo-induced plasma morphology images acquired in real time during the welding process, inputs them into the trained welding joint mechanical property prediction model, and outputs the predicted results of weld mechanical properties.

[0088] The S420 adjusts welding parameters in real time based on the predicted mechanical properties of the weld. The adjusted welding parameters include welding laser power, laser oscillation frequency, laser oscillation angle, and welding speed.

[0089] This invention processes and extracts features from photoinduced plasma images obtained using visual methods. Then, it employs a cascaded neural network composed of BPNN, XGBoost, and random forest networks to fit and predict the mechanical properties of titanium-aluminum oscillating laser-filled powder welded joints. The MAE of the training set is calculated to be 0.0714 kN, and the model R0 is [not specified]. 2 With a value of 0.91332, this invention can achieve real-time prediction of welding status, reduce manual workload, and improve welding quality.

[0090] Corresponding to the aforementioned method for predicting the mechanical properties of welded joints based on cascaded neural networks, this invention also discloses a system for predicting the mechanical properties of welded joints based on cascaded neural networks, such as... Figure 3 As shown, it specifically includes:

[0091] The data acquisition module is used to acquire photo-induced plasma morphology images during the laser welding process and obtain various weld performance parameters by performing performance tests on the weld seam.

[0092] The dataset construction module is used to extract features from the collected photo-induced plasma morphology images, establish a correspondence between the extracted image features and weld performance parameters, and construct training, validation, and test sets.

[0093] The model building and training module is used to build a prediction model of the mechanical properties of welded joints based on cascaded neural networks. The established training set, validation set and test set are used to train, validate and test the prediction model of the mechanical properties of welded joints.

[0094] The prediction module is used to predict the mechanical properties of welds in real time using a trained prediction model of the mechanical properties of welded joints.

[0095] It should be noted that for a detailed description of the welding joint mechanical property prediction system based on cascaded neural networks provided in the embodiments of the present invention, please refer to the relevant description of the welding joint mechanical property prediction method based on cascaded neural networks provided in the embodiments of the present invention, which will not be repeated here.

[0096] The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.

Claims

1. A method for predicting the mechanical properties of welded joints based on cascaded neural networks, characterized in that, The method includes: Photo-induced plasma morphology images were acquired during the laser welding process, and various weld performance parameters were obtained by performing performance tests on the welds obtained. Feature extraction was performed on the acquired photo-induced plasma morphology images. The extracted image features were then correlated with weld performance parameters, and training, validation, and test sets were constructed. A mechanical property prediction model for welded joints based on cascaded neural networks was constructed, and the model was trained, validated, and tested using the established training, validation, and test sets. Real-time prediction of weld mechanical properties is performed using a trained prediction model for the mechanical properties of welded joints. The mechanical property prediction model for the welded joint includes a BPNN network, an XGBoost network, and a random forest network cascaded in sequence. Constructing a prediction model for the mechanical properties of welded joints based on cascaded neural networks, specifically including: Define the BPNN network structure. The BPNN network takes the extracted image features as input and the weld side penetration depth and weld cross-sectional area as output. Define the XGBoost network structure. The XGBoost network takes the weld side penetration depth and weld cross-sectional area output by the BPNN network as input and the zirconium content of the weld cross-section as output. Define a random forest network structure. The random forest network takes the zirconium content of the weld section output by the XGBoost network as input and the mechanical property parameters of the weld as output.

2. The method for predicting the mechanical properties of welded joints based on cascaded neural networks as described in claim 1, characterized in that, Acquiring images of the photoinduced plasma morphology during laser welding, specifically including: Multiple welding tests and image acquisitions were conducted using a constructed visual testing platform, which includes a welding device for welding tests and a visual device for image acquisition. The vision device includes a camera and an image processing workstation. The camera captures images of the photoinduced plasma morphology during the welding process and feeds them back to the image processing workstation.

3. The method for predicting the mechanical properties of welded joints based on cascaded neural networks as described in claim 1, characterized in that, Various weld performance parameters were obtained through performance testing, including: The welds obtained from multiple welding tests were subjected to cross-sectional morphology measurement, chemical composition scanning, and mechanical tensile tests to obtain various performance parameters, including weld side penetration depth, weld cross-sectional area, zirconium content in weld cross-section, and tensile strength.

4. The method for predicting the mechanical properties of welded joints based on cascaded neural networks as described in claim 3, characterized in that, Feature extraction is performed on the acquired photoinduced plasma morphology images, specifically including: Extract information on plasma centroid, area, and deflection angle from photoinduced plasma morphology images; The extracted plasma centroid, area, and deflection angle information are filtered and feature signal decomposed. The multiple feature signals obtained from the decomposition are filtered through correlation analysis, and then the filtered feature signals are subjected to dimensionality reduction processing.

5. The method for predicting the mechanical properties of welded joints based on cascaded neural networks as described in claim 1, characterized in that, The established training, validation, and test sets are used to train, validate, and test the prediction model for the mechanical properties of welded joints. Specifically, this includes: Define the training parameters of the BPNN network and initialize the network; Define the training parameters of the XGBoost network and initialize the network; Define the training parameters of the random forest network and initialize the network; Define the training process, use training set, validation set and test set to train, validate and test BPNN, XGBoot and random forest networks, and save the trained network parameters after training is completed; The trained BPNN, XGBoot and random forest networks are cascaded to build a prediction model for the mechanical properties of welded joints, and the model is validated and evaluated.

6. The method for predicting the mechanical properties of welded joints based on cascaded neural networks as described in claim 1, characterized in that, Real-time prediction of weld mechanical properties is performed using a trained weld joint mechanical property prediction model, specifically including: The photo-induced plasma morphology images acquired in real time during the welding process are used to extract image features and then input into the trained welding joint mechanical property prediction model, and the predicted results of the weld mechanical properties are output.

7. The method for predicting the mechanical properties of welded joints based on cascaded neural networks as described in claim 1, characterized in that, The method further includes: Welding parameters are adjusted in real time based on the predicted mechanical properties of the weld. These parameters include welding laser power, laser oscillation frequency, laser oscillation angle, and welding speed.

8. A system for predicting the mechanical properties of welded joints based on cascaded neural networks, characterized in that, The system includes: The data acquisition module is used to acquire photo-induced plasma morphology images during the laser welding process and obtain various weld performance parameters by performing performance tests on the weld seam. The dataset construction module is used to extract features from the collected photo-induced plasma morphology images, establish a correspondence between the extracted image features and weld performance parameters, and construct training, validation, and test sets. The model building and training module is used to build a prediction model of the mechanical properties of welded joints based on cascaded neural networks. The established training set, validation set and test set are used to train, validate and test the prediction model of the mechanical properties of welded joints. The prediction module is used to predict the mechanical properties of welds in real time using the trained prediction model of the mechanical properties of welded joints. The mechanical property prediction model for the welded joint includes a BPNN network, an XGBoost network, and a random forest network cascaded in sequence. Constructing a prediction model for the mechanical properties of welded joints based on cascaded neural networks, specifically including: Define the BPNN network structure. The BPNN network takes the extracted image features as input and the weld side penetration depth and weld cross-sectional area as output. Define the XGBoost network structure. The XGBoost network takes the weld side penetration depth and weld cross-sectional area output by the BPNN network as input and the zirconium content of the weld cross-section as output. Define a random forest network structure. The random forest network takes the zirconium content of the weld section output by the XGBoost network as input and the mechanical property parameters of the weld as output.

Images