A laser welding element burnout multivariate regression monitoring method based on spectral information

Abstract

The application discloses a laser welding element burnout multivariate regression monitoring method based on spectral information, and belongs to the field of material processing engineering. The method adopts a spectrometer to collect spectral signals in a welding process, adopts a background removing, high-frequency noise filtering and normalization treatment method to automatically reduce noise components in the spectral signals, and constructs an element burnout multivariate regression prediction model database. Then, a laser welding element burnout multivariate regression prediction model based on machine learning is built. A feature screening algorithm is adopted to realize dimension reduction of the spectral signals, and the calculation efficiency and prediction accuracy of the model are improved. The element burnout multivariate regression prediction model is automatically optimized through a grid search, cross-validation and other algorithms, and element burnout prediction in the laser welding process is accurately realized. The application can improve the laser welding quality.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of material processing engineering, and particularly relates to a laser welding element burnout multivariate regression monitoring method based on spectral information. BACKGROUND

[0002] With the rapid development of industrial intelligent manufacturing, intelligent welding systems have become an important part of modern industry. Intelligent welding not only improves production efficiency, but also significantly improves welding quality, which is of great significance to social and economic development. Especially in the manufacturing fields of automobiles, aerospace, etc., the welding quality is directly related to the safety and reliability of the products. In the process of welding automation, the traditional post-welding inspection method has been unable to meet the real-time and accuracy requirements of modern manufacturing. Therefore, online welding quality monitoring technology has emerged as the times require. This technology monitors the physical and chemical changes in the welding process in real time, such as arc spectrum, acoustic spectrum, and welding current, voltage, etc. parameters, providing a reliable basis for the prediction and judgment of welding quality. The application of online monitoring technology not only ensures the high-quality quality of the weld, but also improves the competitiveness and production efficiency of the products. With the continuous progress of technology, online monitoring of welding quality will become an important development direction in the future intelligent manufacturing field, and inject new vitality into the development of industry. However, in the process of laser welding, the partial strengthening elements in the material to be welded have a much higher equilibrium vapor pressure than the base alloy elements, resulting in a lower melting point. In the process of laser welding, they will evaporate first, resulting in a lower content of strengthening elements in the weld than the content of alloy elements in the base material. The loss will result in a lower performance of the welded joint than the base material, resulting in a decrease in the quality of the welded parts.

[0003] Therefore, developing a set of online monitoring system for laser welding process quality, especially involving the quantitative detection system of element burnout, which will provide a scientific basis for the quantitative evaluation of weld performance. Some studies at home and abroad have shown that the loss of elements during welding can be monitored. Chinese patent application CN115615939A discloses a device and method for monitoring manganese evaporation in a high-nitrogen steel GMAW additive process, which collects stable spectral signals during the arc welding process, and selects the manganese spectral line that is not interfered by other characteristic spectra by manual screening, and then qualitatively judges whether there is manganese element burnout phenomenon in the arc welding process according to the obtained manganese spectral line, but this invention is through manual way, only qualitative manganese element burnout analysis. Chinese patent application CN115060708A discloses a method for detecting trace elements in laser plasma, which provides a method for arc excitation laser plasma to ensure that stable spectral signals can be obtained, but does not provide a semi-quantitative or quantitative method for element content. Chinese patent application CN107764798A discloses a metal additive manufacturing quality online detection system, which provides a method for collecting spectral signals during additive process, and proposes to calculate the amount of element burnout by free calibration method, but does not provide a specific method for quantitative calculation of element burnout. In addition, unlike the strict quantitative analysis requirement of laser-induced breakdown spectroscopy, in the laser welding process, due to the influence of factors such as device laser energy density, spectral acquisition method, environmental noise, the collected spectral signal and element burnout is a completely nonlinear relationship, and the free calibration method for calculating element burnout will be greatly disturbed. Chinese patent CN112017183B discloses a welding wire composition online detection method based on arc image and arc spectrum information cooperation, which realizes the prediction of welding wire composition by using deep learning method to fuse image and spectrum information, but this method is not proposed for the composition content of the weld after welding.

[0004] Current studies have shown that the method of using spectral signals can realize the monitoring of element burnout during welding process, but the related invention methods are concentrated on arc, or qualitative analysis of element burnout phenomenon. SUMMARY

[0005] In order to overcome the defects of the prior art, the present application proposes a laser welding element burnout multivariate regression monitoring method based on spectral information. In the arc welding process, the elements evaporated from the welding material are restrained in the arc, and the obtained spectral signal is relatively stable. The present application proposes a method for the plasma generated in the laser welding process, which has a process of eruption, expansion and dissipation. This process is extremely unstable, and the obtained signal contains noise components caused by factors such as equipment, human, environment and material. In order to solve the more nonlinear factors between the element burnout quantitative analysis and the spectral signal caused by these factors as much as possible, the present application proposes a burnout element multivariate regression monitoring method based on machine learning. Compared with the calibration method and the free calibration method, the multivariate regression monitoring method based on multivariate can fully extract and learn the spectral signal characteristics in the welding process, and accurately select the characteristic spectral lines related to the burnout elements according to the importance of the characteristics, reduce the monitoring error, and realize the accurate monitoring of the burnout elements of the laser welding joint.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0007] A laser welding element burnout multivariate regression monitoring method based on spectral information, comprising the following steps:

[0008] S1: In order to obtain spectral signals of different intensities and corresponding different degrees of element burnout, laser welding experiments with different welding process parameters are designed, and the process parameters mainly include laser power and welding speed. Then, laser welding experiments are carried out, spectral signal acquisition parameters are set, and spectral signals under different welding experiments are collected. At the same time, the corresponding different element burnout states of the weld are obtained, and the element content detection method is used to detect the post-weld weld burnout element content to obtain the post-weld weld burnout element content;

[0009] S2: In order to minimize the noise influence caused by equipment, environment, human and base value effect, the spectral signals obtained in S1 are sequentially denoised and pretreated and normalized. The spectral lines of the burnout elements in the normalized spectral signals are selected and linearly fitted with the burnout element content in the post-weld weld, a univariate fitting model between the burnout element spectral lines and the post-weld weld burnout element under different normalization methods is established, and then the determination coefficient R 2 of the univariate fitting model is selected to select the optimal normalization method;

[0010] S3: using a feature screening algorithm to screen the normalized spectral signal: first, calculate the correlation coefficient or importance score between the normalized spectral signal at different wavelengths and the post-weld weld burn-off element content; then set different correlation coefficients or importance scores as feature screening thresholds, discard the normalized spectral signal at wavelengths less than the feature screening threshold, and retain the normalized spectral signal at the remaining wavelengths, and construct a multivariate regression monitoring model dataset, and divide the dataset into a training set, a test set, and a validation set; then build a multivariate regression prediction model of the screened spectral signal and the measured post-weld weld burn-off element, then automatically optimize the multivariate regression prediction model of different feature screening thresholds, and calculate the accuracy of the optimized multivariate regression prediction model of different feature screening thresholds; select the feature screening threshold corresponding to the multivariate regression prediction model with the maximum accuracy as the optimal feature screening threshold, and simultaneously select the multivariate regression prediction model with the maximum accuracy as the laser welded joint element burn-off multivariate regression monitoring model.

[0011] Preferably, in S1, the spectral signal of the laser plasma focused by the collimating mirror at a position of more than 1 mm above the welding area, because the continuous spectrum of the plasma decays faster than the characteristic spectrum, so at more than 1 mm, the interference of the continuous spectrum on the characteristic spectrum can be reduced.

[0012] Preferably, in S1, the collimating mirror is used to change the divergent spectral signal emitted by the welding plasma into a parallel light beam, and then the spectral signal is transmitted to the spectrometer through the optical fiber coupled with the collimating mirror, and then the acquisition parameters of the spectrometer are set to record the welding spectral signal at different welding times online.

[0013] Further, in the actual spectral signal acquisition process, in order to prevent the spectral signal intensity from being oversaturated, an attenuation sheet with a corresponding transmittance is needed to set the integration time and integration times of signal acquisition, etc.

[0014] Preferably, in S1, the post-weld element acquisition first uses a wire cutting or numerical control machining device to prepare a weld metallographic piece perpendicular to the weld, and then uses sandpaper for grinding and other operations, and then uses a composition detection technology to measure the content of the post-weld weld burn-off element, but not limited to: XRF, EDS, LIBS, etc. Composition detection technology.

[0015] Preferably, in S2, due to the existence of composite radiation and inverse bremsstrahlung radiation during welding, the collected spectral signals have continuous spectral lines with certain intensity, which are more background noise and will seriously affect the spectral signal intensity of the wavelength related to the element to be predicted. Background correction is needed to remove it, including but not limited to least square method, polynomial fitting algorithm, etc. At the same time, due to factors such as environment, equipment, signal acquisition method, etc., the obtained signal has a lot of high-frequency noise, which also needs to be removed by using the corresponding high-frequency filtering denoising algorithm, including but not limited to Fourier transform, wavelet transform, wavelet packet transform, etc. Subsequently, the normalization method can further remove the interference caused by equipment, human factors and base value effect, including but not limited to maximum and minimum value normalization, background normalization, internal standard normalization, area normalization, etc. Then the spectral lines of the burnout element in the normalized spectral signal are selected and the content of the burnout element in the measured post-weld weld is monovariant linearly fitted, and the normalized method with the largest correlation determination coefficient R 2 is selected as the optimal normalization method of the spectral signal.

[0016] Further, there are a large number of signals with low correlation with the burnout element in the actual obtained spectral signal, which will seriously interfere with the calculation speed and accuracy of the multivariate regression monitoring model, not only affecting the monitoring accuracy of the element burnout in the actual welding process, but also causing a certain delay in the monitoring. Therefore, it is necessary to remove the spectral signals unrelated to the burnout element through feature selection. The feature correlation coefficient or importance score between the spectral signal and the post-weld weld burnout element in the training set, validation set and test set is calculated to represent the close degree between the burnout element and the spectral signal, wherein the calculation method of the feature correlation coefficient or importance score is not limited to variance threshold, correlation coefficient method, principal component analysis, model-based feature selection, etc. Then the feature correlation coefficient or importance score is set as the feature selection threshold, and the wavelength spectral signal less than the threshold is discarded, and the wavelength spectral signal greater than the threshold is retained. By setting different feature selection thresholds, the spectral signal is screened, and data sets with different feature selection thresholds are constructed. Then each spectral signal data set is divided into training set, validation set and test set.

[0017] Further, a machine learning-based element burn loss multiple regression prediction model is constructed, and the spectrum signals of the training set and the validation set screened by different thresholds are taken as inputs, and the measured post-weld weld burn loss element content is taken as the output of the model. Then, the model hyperparameter optimization under different feature screening thresholds is automatically performed by a multiple regression prediction model optimization algorithm, wherein the multiple regression prediction model optimization algorithm is not limited to: Bayesian optimization, grid search method; the precisions of the multiple regression monitoring models under different feature screening thresholds are calculated, and the threshold corresponding to the maximum precision is selected as the optimal feature screening threshold for spectrum signal screening, and the model with the maximum prediction precision is selected as the optimal laser welded joint element burn loss multiple regression monitoring model.

[0018] Further, the model is trained by cross-validation, and then the performance of the model is verified by using the test set, the prediction precision of the constructed multiple regression monitoring model is calculated, and the prediction precision and generalization ability of the model are verified.

[0019] The present application can realize online real-time monitoring of laser post-weld weld burn loss elements, and has the following beneficial effects compared with traditional element measurement methods:

[0020] 1. The present application applies LIBS related technical theory to the welding process, avoiding the use of complete LIBS non-destructive testing equipment in the welding process, so the method and device are a simpler and more integrated device and non-destructive testing method.

[0021] 2. The present application is a real-time online non-destructive testing method, which can realize in-process monitoring and eliminate post-weld detection methods, thereby significantly reducing production manufacturing costs and related detection processes.

[0022] 3. The present application is a machine learning-based element burn loss multiple regression monitoring method. Since machine learning has strong non-linear fitting capability, the interference caused by factors such as equipment, human, and matrix effect can be largely ignored. In addition, there is no need for excessive non-destructive testing field and professional knowledge and experience related to spectrum. Denoising, grid search, and other optimization algorithms are used to autonomously and fully capture the spectrum signals strongly correlated with the predicted burn loss elements, without manually querying and searching for burn loss element characteristic spectral lines, and automatically optimizing the element burn loss multiple regression prediction model to realize quantitative monitoring of element burn loss in the welding process. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 A flowchart of a laser welding element burn loss multiple regression monitoring method based on spectrum information according to the present application;

[0024] Figure 2 A schematic diagram of a welding plasma spectrum information acquisition device according to the present application;

[0025] Figure 3 The schematic diagram of the spectrum signal and the post-weld weld magnesium element measurement obtained by the application; wherein a is the post-weld collected spectrum signal, b is the average value of the spectrum signal in the 0.5mm window of a, c is the post-weld weld magnesium element collection schematic diagram, and d is the post-weld weld magnesium element content EDS point scanning schematic diagram; Figure 3 The average value of the spectrum signal in the 0.5mm window of a, c is the post-weld weld magnesium element collection schematic diagram, and d is the post-weld weld magnesium element content EDS point scanning schematic diagram;

[0026] Figure 4 The spectrum signal denoising process and result schematic diagram of the application;

[0027] Figure 5 The magnesium element univariate loss monitoring fitting result diagram of the application using different normalization methods;

[0028] Figure 6 The importance score vertical line diagram of the spectrum signal of the application at different wavelengths;

[0029] Figure 7 The model fitting result corresponding to the feature screening using different threshold values of the application.

[0030] In the figure, the reference signs are: 1, industrial computer; 2, spectrometer; 3, display; 4, optical fiber; 5, collimating mirror; 6, neutral attenuation sheet; 7, protective lens; 8, plasma; 9, welding workbench; 10, laser head; 11, laser. DETAILED DESCRIPTION

[0031] In order to make the features, technical solutions and advantages of the application more easily understood, the application is further described in detail in conjunction with the drawings and specific implementation solutions. In addition, the specific implementation solutions described herein are only used to explain the gist of the application and do not limit the application. The application can be properly improved according to the specific actual use situation.

[0032] As shown in the figure, the laser welding element burnout multivariate regression monitoring method based on spectrum information of the application includes the following steps: Figure 1

[0033] S1: In order to obtain spectrum signals of different intensities and element burnout laser welding tests of different degrees, laser welding experiments with different welding process parameters are designed, and the different welding process parameters include welding power and welding speed. Then laser welding is carried out, spectrum signal acquisition parameters are set, and spectrum signals under different laser welding tests are collected; at the same time, the welds under different element burnout states are obtained, and element content detection methods are used to detect the post-weld weld element content, and the post-weld weld burnout element content is obtained.

[0034] ​S2: In order to minimize the noise influence brought by equipment, environment, human and base value effect, the spectral signal obtained in S1 is sequentially denoised and preprocessed, and normalized, and the spectral line of the burnout element after normalization is linearly fitted with the burnout element content in the weld after welding, a univariate fitting model between the spectral line of the burnout element and the burnout element in the weld after welding under different normalization methods is established, and then the optimal normalization method is selected according to the determination coefficient R 2 of the univariate fitting model.

[0035] S3: The normalized spectral signal is screened by using a feature screening algorithm: firstly, the correlation coefficient between the normalized spectral signal at different wavelengths and the burnout element content in the weld after welding or the importance score of different wavelengths is calculated; then different correlation coefficients or importance scores are set as feature screening thresholds, the spectral signal at the wavelength less than the feature screening threshold is discarded, the normalized spectral signal at the remaining wavelength is retained, a multivariate regression monitoring model data set is constructed, and the data set is divided into a training set, a test set and a validation set; then a multivariate regression prediction model of the screened spectral signal and the measured burnout element in the weld after welding is constructed, and the automatic optimization of the multivariate regression prediction model with different feature screening thresholds is performed, and the accuracy of the optimized multivariate regression prediction model with different feature screening thresholds is calculated; the feature screening threshold corresponding to the multivariate regression prediction model with the maximum accuracy is selected as the optimal feature screening threshold, and the multivariate regression prediction model with the maximum accuracy is selected as the burnout element burnout multivariate regression monitoring model of the laser welded joint.

[0036] Preferably, in S1, the spectral signal of the laser plasma focused on the welding area about 1mm above by the collimating mirror 5, because the continuous spectrum of the plasma 8 attenuates faster than the characteristic spectrum, so that the spectral signal collected at more than 1mm can reduce the interference of the continuous spectrum on the characteristic spectrum.

[0037] Preferably, in S2, the collimating mirror 5 changes the divergent spectral signal emitted by the welding plasma into a parallel light beam, and then the spectral signal is transmitted to the spectrometer 2 through the optical fiber 4 coupled with the collimating mirror, and then the acquisition parameters of the spectrometer 2 are set to record the welding spectral signal at different welding moments online.

[0038] Further, in the actual spectral signal acquisition process, in order to prevent the spectral signal intensity from being oversaturated, a neutral attenuation sheet 6 with a corresponding transmittance needs to be used, and the integral time and integral times of the spectrometer 2 in the signal acquisition process are set.

[0039] Preferably, in S2, the post-weld element is first obtained by using a wire cutting or numerical control machining equipment to prepare a weld metallographic sample perpendicular to the weld, and then sequentially using sandpaper to grind and other operations, and then measuring the post-weld weld burn-off element content by methods such as X-ray fluorescence spectroscopy (XRF), energy dispersive X-ray spectroscopy (EDS), laser-induced breakdown spectroscopy (LIBS), etc.

[0040] Preferably, in S3, due to the existence of composite radiation and inverse bremsstrahlung radiation during welding, the collected spectral signals have continuous spectra with certain intensity, which are more background noise and will seriously affect the spectral signal intensity of the wavelength related to the element to be predicted. Background correction is needed to remove it, including but not limited to least squares method, polynomial fitting algorithm, etc. At the same time, due to factors such as environment, equipment, signal acquisition method, etc., the obtained signal has a lot of high-frequency noise, which also needs to be removed by using the corresponding high-frequency filtering denoising algorithm, including but not limited to Fourier transform, wavelet transform, wavelet packet transform, etc. Then, using normalization method can further remove the interference caused by equipment, human factors and base value effect, including but not limited to maximum and minimum value normalization, background normalization, internal standard normalization, area normalization, etc. Then, the burn-off element spectral line in the normalized spectral signal is selected and single-variable linear fitting is performed with the post-weld weld burn-off element content obtained by measurement. The normalized method with the largest correlation determination coefficient R 2 is selected as the optimal normalization method of the spectral signal.

[0041] Further, there are a lot of signals in the normalized spectral signal that are not closely related to element burn-off. These signals in these bands will seriously interfere with the calculation speed and accuracy of the multivariate regression monitoring model, not only affecting the monitoring accuracy of element burn-off in the actual welding process, but also causing a certain delay in monitoring. Therefore, it is necessary to remove the spectral signals not related to the burn-off element by feature selection. The feature correlation coefficient or importance score between the spectral signal and the post-weld weld burn-off element is calculated to represent the close degree between the burn-off element and the spectral signal. The calculation method of the feature correlation coefficient or importance score is not limited to variance threshold, correlation coefficient method, principal component analysis, model-based feature selection, etc. Then, by setting the feature correlation coefficient or importance score as the feature selection threshold, the wavelength spectral signal less than the threshold is discarded, and the wavelength spectral signal greater than the threshold is retained. By setting different feature selection thresholds, the spectral signal is screened, and data sets with different feature selection thresholds are constructed. Then, each spectral signal data set is divided into training set, validation set and test set in the ratio of 7:2:1.

[0042] Further, in the S3, a machine learning-based multi-element burn-off regression prediction model is constructed, the spectral signals of the training set and the validation set are taken as the input of the multi-element burn-off regression prediction model, and the measured post-weld weld burn-off element content is taken as the output of the multi-element burn-off regression prediction model.

[0043] Subsequently, model hyperparameter optimization under different feature screening thresholds is automatically performed through a model optimization algorithm, wherein the multi-element burn-off regression prediction model optimization algorithm is not limited to Bayesian optimization and grid search method; meanwhile, the precision of the multi-element burn-off regression prediction model optimized under different feature screening thresholds is calculated, and the threshold corresponding to the maximum precision is selected as the optimal feature screening threshold for spectral signal screening; at the same time, the model with the maximum prediction precision is selected as the optimal laser welded joint burn-off element burn-off multi-element burn-off regression prediction model.

[0044] Further, the model is trained by cross-validation, and then the performance of the model is verified by using the test set, the prediction precision of the constructed multi-element burn-off regression prediction model is calculated, and the prediction precision and generalization ability of the model are verified.

[0045] Embodiment:

[0046] The laser welded element burn-off multi-element burn-off regression prediction method based on spectral information of the embodiment includes:

[0047] First, an Al-Mg alloy with a brand of 5251 is selected for laser welding experiment. During the welding process, a collimating lens is used to change the divergent light beam into a parallel light beam, which is then transmitted to the 314nm-418nm channel of the 8-channel spectrometer through an optical fiber to collect and obtain the spectral signal; a protective lens 7 is installed in front of the collimating lens to prevent splashing damage to the lens, and a 20% neutral attenuation sheet 6 is used to prevent the spectral signal from being oversaturated, the integration time is set to 5ms, and the integration number is set to 2, as shown in Figure 2The laser welding work is completed by the industrial computer 1, the display 3, the welding workbench 9, the laser head 10 and the laser 11. The industrial computer 1 is connected with the laser 11 through a signal line to control the process parameters of the laser. The laser 11 is connected with the laser head 10 through an optical fiber to realize the transmission of the laser beam from the laser 11 to the laser head 10. Meanwhile, the industrial computer 1 is connected with the welding workbench 9 through a signal line to control the welding workbench. The collection and analysis of the spectral signal during the welding process are completed by the industrial computer 1, the spectrometer 2, the display 3, the optical fiber 4, the collimating mirror 5, the neutral attenuation sheet 6, the protective lens 7 and the welding plasma 8. The protective lens 7 and the neutral attenuation sheet 6 are fastened on the collimating mirror 5 through threads, and then the collimating mirror and the spectrometer 2 are coupled together through the optical fiber 4. The spectral signal emitted by the plasma 8 is transmitted to the spectrometer 2 and recorded in the industrial computer 1, and displayed on the display 3. Subsequently, the analysis work of the laser welding element burnout monitoring task is performed on the industrial computer 1.

[0048] Further, the collected spectral signal is as shown in FIG. 8a. Subsequently, the wire cutting machine bed is used to obtain the weld longitudinal section metallographic piece along the center of the weld, and then the metallographic piece is ground to obtain the weld longitudinal section metallographic piece as shown in FIG. 8c. Subsequently, the energy dispersive X-ray spectroscopy (EDS) point scanning method is used as shown in FIG. 8d. Subsequently, the magnesium element content of multiple points in a 0.5mm window is measured every 5mm along the weld, and the average value is taken. From the initial time T0 to the final time T1, the magnesium element content of multiple window regions is measured and obtained as shown in FIG. 8c, which is the output of the subsequent constructed multiple regression prediction model. Correspondingly, the average spectral signal in the same interval and the same width window is obtained as shown in FIG. 8b, which is the input of the subsequent constructed multiple regression prediction model. Figure 3 Figure 3 Figure 3 Figure 3 Figure 3

[0049] Further, the collected spectral signal is denoised, and the process and result are as shown in FIG. 9. Polynomial fitting algorithm and wavelet packet algorithm are used in sequence to obtain the background and high-frequency noise components. Then, the original signal is subtracted from the noise components to obtain the denoised spectral signal. Figure 4

[0050] ​​​​​​Further, the denoised spectral signals are subjected to different normalization processing. In this embodiment, no normalization processing, area normalization processing, and background normalization processing are compared. The area normalization processing uses the integral definition algorithm to calculate the total area of the spectral signals in the wavelength range of 314 nm-418 nm, and then compares the spectral intensity at different wavelengths with the integral total area. The background normalization processing compares the spectral signal intensity at each wavelength after denoising with the intensity of the background noise at the corresponding wavelength obtained during denoising. Then, the normalized magnesium spectral line at 383.829 nm is linearly fitted with the element content obtained by EDS measurement to construct a univariate fitting model, and the results are shown in Figure 5 . As can be seen from Figure 5 , the area normalization method can effectively improve the determination coefficient R 2 of the univariate model. Therefore, the obtained spectral signals are subjected to area normalization processing to construct a data set for the magnesium element burn loss multivariate regression prediction model of the weld.

[0051] Further, a magnesium element burn loss monitoring model based on LightGBM (Light Gradient Boosting Machine Learning) machine learning is constructed. Subsequently, based on the embedded feature screening principle, the LightGBM model is used to automatically calculate the relationship between the spectral signals at different wavelengths and the element content, and the importance of different wavelengths. The results are shown in Figure 6 .

[0052] Further, by setting different importance thresholds, feature screening of different wavelengths is performed, and wavelengths lower than the set threshold are discarded, and wavelengths greater than or equal to the set threshold are retained. The screened spectral signals are used as the input of the LightGBM model, and the magnesium element content in the post-weld weld pool obtained by measurement is used as the output of the model. Then, a grid search algorithm is used to search for the optimal parameters of the model, including boosting_type (model boosting algorithm), colsample_bytree (sampling rate), learning_rate (learning rate), max_depth (tree maximum depth), n_estimators (estimator number), num_leaves (leaf node number), reg_alpha (L1 regularization coefficient), and reg_lambda (L2 regularization coefficient). Finally, the optimal model under different importance thresholds is obtained, and the accuracy of the optimal model under different thresholds is calculated. The results are shown in Figure 7 . As can be seen from Figure 7 , under this embodiment, the model importance threshold is set to 0.05, and the optimal weld element burn loss multivariate regression prediction model is constructed.

[0053] The part of the present application not described in detail belongs to the known technology of the person skilled in the art. The above-described embodiments are only used to describe the preferred embodiments of the present application, and the preferred embodiments do not describe all the details and limit the application to the specific embodiments described. Without departing from the design spirit of the present application, various modifications and improvements of the technical solutions of the present application made by the person skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A laser welding element burnout multiple regression monitoring method based on spectral information, characterized in that, The method comprises the following steps: S1: design laser welding experiments with different welding process parameters, including laser power and welding speed; then perform laser welding, set spectral signal acquisition parameters, and collect spectral signals under different laser welding experiments; simultaneously obtain welds with different degrees of element burning loss, and detect the content of the burned elements in the welds after welding by using an element content detection method to obtain the content of the burned elements in the welds after welding; S2: the spectral signal obtained in S1 is denoised and preprocessed, and normalized, and the spectral line of the burnout element in the normalized spectral signal is linearly fitted with the content of the burnout element in the weld after welding to establish a univariate fitting model between the spectral line of the burnout element and the burnout element in the weld after welding under different normalization methods, and then the determination coefficient R 2 of the univariate fitting model is determined to select the optimal normalization method; S3: use a feature screening algorithm to screen the normalized spectral signals: first, calculate the correlation coefficient or importance score between the normalized spectral signals at different wavelengths and the content of the burned elements in the welds after welding; then set different correlation coefficients or importance scores as feature screening thresholds, discard the spectral signals at wavelengths less than the feature screening thresholds, retain the normalized spectral signals at the remaining wavelengths, construct a multivariate regression prediction model dataset, and divide the dataset into a training set, a test set, and a validation set; then build a multivariate regression prediction model of the screened spectral signals and the measured burned elements in the welds after welding, and then automatically optimize the multivariate regression prediction models with different feature screening thresholds, calculate the accuracy of the optimized models with different feature screening thresholds, select the feature screening threshold corresponding to the multivariate regression prediction model with the maximum accuracy as the optimal feature screening threshold, and simultaneously select the multivariate regression prediction model with the maximum accuracy as the optimal element burning multivariate regression prediction model.

2. The method of claim 1, wherein the method is a spectral information based laser welding element burnout multivariate regression monitoring method. In S1, the spectral signals of the laser plasma above 1 mm above the welding area during the welding process are collected to reduce the interference of the continuous spectrum on the characteristic spectrum; a collimating mirror is used to change the divergent spectral signals emitted by the welding plasma into parallel light beams, and then the spectral signals are transmitted to a spectrometer through an optical fiber coupled with the collimating mirror, and then the acquisition parameters of the spectrometer are set to record the welding spectral signals at different welding times online.

3. The method of claim 2, wherein the method is a multivariate regression monitoring method based on spectral information of laser welding elements burnout. During the spectral signal acquisition process, a neutral attenuation sheet with a corresponding transmittance is used to prevent the spectral signal intensity from being oversaturated, and the integration time and integration times of the spectrometer acquisition are set.

4. The method of claim 1, wherein the method is a multivariate regression monitoring method based on spectral information of laser welding elements burnout. In S2, the spectral signal denoising preprocessing includes the removal of background and high-frequency noise, and then the decision coefficients of the established single-variable linear fitting models of the burned elements are compared to select the normalization method with the maximum decision coefficient as the optimal normalization method.

5. The method of claim 1, wherein the method is a spectral information based laser welding element burnout multivariate regression monitoring method. In S3, a post-weld weld element burning multivariate regression prediction model based on machine learning is constructed; the spectral signals screened by different feature screening thresholds are used as the input of the multivariate regression prediction model, and the content of the burned elements in the post-weld welds is used as the output of the multivariate regression prediction model; then the model optimization algorithm is used to automatically optimize the model hyperparameters under different feature screening thresholds, calculate the accuracy of the optimized models with different feature screening thresholds, select the threshold corresponding to the maximum accuracy as the optimal threshold for spectral signal screening, and simultaneously select the model with the maximum prediction accuracy as the optimal element burning multivariate regression prediction model.

6. The method of claim 1, wherein the method is a spectral information based laser welding element burnout multivariate regression monitoring method. In the S3, the feature screening algorithm includes a variance threshold method, a correlation coefficient method, a principal component analysis method, a recursive feature elimination method, and a model-based feature selection method.

7. The method of claim 5, wherein the method further comprises: The model optimization algorithm includes a Bayesian optimization and a grid search method.

Images