Laser welding spatter real-time identification method based on OCT keyhole depth measurement

Abstract

The invention discloses a laser welding spatter real-time recognition method based on OCT keyhole depth measurement. The method comprises the steps of S1, keyhole depth data collection based on OCT measurement and data set construction and S2, construction training of a deep learning model and spatter defect recognition. According to the method, an SD-OCT measurement technology is combined with a composite deep learning model based on 1DCNN-BiLSTM splash defect recognition, a Dropout technology and adaptive moment estimation are adopted to prevent model overfitting, and a TPE algorithm is adopted to carry out hyper-parameter optimization on the model, so that the trained 1DCNN-BiLSTM splash defect recognition model can accurately predict the splash state in the dynamic laser welding process, and the accuracy of the dynamic laser welding process is improved. Manual intervention is not needed, and the laser welding spatter defect can be rapidly and accurately recognized.

Description

Technical Field

[0001] The invention relates to the technical field of laser welding, and in particular to a real-time recognition method for laser welding spatter based on OCT keyhole depth measurement. Background Art

[0002] Lasers are widely used in welding technology due to their high energy density, precise control and fast processing capabilities. Due to the extremely high energy focusing of laser welding, it can achieve extremely small heat-affected zones and high-precision welds, thereby significantly reducing material deformation and subsequent processing requirements. In addition, the advantages of laser welding on aluminum alloys and the good corrosion resistance of aluminum alloys have made it widely used in the automotive and aerospace manufacturing industries. However, during laser welding of aluminum alloys, the inside of the keyhole is often in an extremely unstable state, and the keyhole is prone to collapse and closure, resulting in metal spatter defects. In power battery welding, the resistance and heat loss of the power battery connector will increase, thereby reducing the overall performance and safety of the battery. These defects may cause internal short circuits in the battery, and then cause thermal runaway or even fire and other safety accidents. Therefore, it is very necessary to identify spatter defects in the aluminum alloy laser welding process. However, previous spatter identification methods such as visual sensing methods, acoustic emission sensing methods, photoelectric sensing methods and temperature sensing methods have problems such as weak anti-interference ability and inaccurate recognition, and cannot directly obtain direct information inside the keyhole, which brings certain limitations to accurately reflect the internal dynamics of the keyhole and accurately judge the spatter state.

[0003] A high-resolution optical coherence tomography (OCT) technology that has emerged in recent years can allow the measuring beam to smoothly enter the bottom of the keyhole during the welding process, realizing online and accurate measurement of the keyhole depth. However, the keyhole depth fluctuation measured by OCT technology currently only reflects the change in weld penetration, which cannot be directly equated with the actual welding quality. The identification of spatter is still a subjective judgment by humans, and there are certain errors. Summary of the invention

[0004] In response to the technical problems existing in the background technology, the present invention provides a real-time identification method for laser welding spatter based on OCT keyhole depth measurement, which can accurately predict the spatter state of the dynamic laser welding process without human intervention, and realize fast and accurate identification of laser welding spatter defects.

[0005] According to one aspect of the present application, a method for real-time identification of laser welding spatter based on OCT keyhole depth measurement is provided, which comprises the following steps:

[0006] S1. Keyhole depth data acquisition and dataset construction based on OCT measurement;

[0007] S1.1. Obtaining OCT measurement data that can reflect the change in keyhole depth in a series of laser welding process tests under different laser powers, wherein the OCT measurement data is obtained by coaxially measuring the OCT measurement beam and the laser beam using spectral domain OCT technology;

[0008] S1.2, constructing a keyhole depth-splash state data set based on the OCT measurement data and taking whether there is a splash defect as an evaluation index;

[0009] S2. Construction and training of deep learning models and identification of splash defects;

[0010] S2.1, randomly dividing the constructed keyhole depth-splash state data set into training samples and test samples according to a set ratio, and performing data normalization preprocessing;

[0011] S2.2, establish a 1DCNN-BiLSTM composite deep learning model, train the 1DCNN-BiLSTM composite deep learning model through the training samples in the keyhole depth-splash state data set, use the keyhole depth feature as the input feature vector of the composite deep learning model, and use the presence of splash defect as the output of the composite deep learning model;

[0012] S2.3, using Dropout technology and adaptive moment estimation to prevent the composite deep learning model from overfitting, optimizing the model's hyperparameters through the TPE algorithm, selecting the largest x for the new hyperparameters by maximizing EI (Expected Improvement) in the sampling phase, and continuously updating the best hyperparameter combination by selecting the largest x, and obtaining four indicators of the composite deep learning model: accuracy, precision, recall, and F1 value;

[0013] S2.4. When the four indicators meet the requirements or the maximum number of TPE algorithm parameter searches is reached, the hyperparameter optimization process is terminated, otherwise, return to step S2.3;

[0014] S2.5. According to the optimal number of convolution kernels, number of LSTM layers, number of hidden layers, Dropout value and learning rate obtained by optimization, a 1DCNN-BiLSTM spatter defect recognition model is established;

[0015] S2.6. Use the trained 1DCNN-BiLSTM spatter defect recognition model to predict the test sample data and output the classification results.

[0016] Furthermore, the OCT measurement data reflecting the change in the keyhole depth is obtained by building an online measurement system for the laser welding keyhole depth based on spectral domain OCT;

[0017] The system comprises a low-coherence light source, a spectrometer linear array camera, a fiber coupler, a sample arm, a reference arm, a laser processing head, a laser and a computer, wherein the laser processing head is connected to the laser to obtain a laser beam for welding, the low-coherence light source, the spectrometer linear array camera, the sample arm and the reference arm are respectively connected to the fiber coupler through optical fibers, the sample arm is connected to the laser processing head through optical fibers so that the OCT measurement beam is coaxial with the laser beam, and the spectrometer linear array camera is connected to the computer;

[0018] The low coherence light source is divided into two light beams through a fiber coupler. One beam passes through a reference arm to hit the workpiece and returns along the original path. The other beam passes through a sample arm and hits the bottom of the keyhole coaxially with the laser beam and returns along the original path. The two returning light beams interfere with each other through a fiber coupler. A spectrometer linear array camera is used to obtain a spectral interference fringe image and send it to a computer. The power spectral density of the spectral interference fringe image is inversely Fourier transformed by the computer to obtain keyhole depth information and form OCT measurement data.

[0019] Furthermore, the low coherence power supply adopts a superluminescent diode wide spectrum light source.

[0020] Furthermore, the keyhole depth-splash state dataset is randomly divided into training samples and test samples according to a ratio of 7:3.

[0021] Furthermore, the series of laser welding process tests at different laser powers are a series of aluminum alloy laser welding process tests at different laser powers.

[0022] Furthermore, the computer includes a memory and a processor, wherein the memory is used to store one or more computer instructions, and wherein the one or more computer instructions are executed by the processor to implement the above method steps.

[0023] According to another aspect of the present application, a readable storage medium is provided, on which computer instructions are stored, wherein the computer instructions implement the above method when executed by a processor.

[0024] Compared with the prior art, the real-time identification method of laser welding spatter based on OCT keyhole depth measurement provided by the present invention has the following beneficial effects:

[0025] (1) The present invention builds an online measurement system for the laser welding keyhole depth based on spectral domain OCT, and uses spectral domain OCT (SD-OCT) technology to obtain keyhole depth data in real time.

[0026] (2) By combining the keyhole depth feature and the deep learning model, the present invention proposes a 1DCNN-BiLSTM spatter quantitative evaluation model for the laser welding process to realize the online prediction of welding spatter defects. Compared with the traditional method, it has the advantages of good real-time compatibility and high recognition rate;

[0027] (3) The composite deep learning model of 1DCNN-BiLSTM for spatter defect recognition constructed by the present invention is effectively combined with SD-OCT measurement and TPE algorithm, which greatly improves the recognition accuracy of spatter defects without the need for human subjective intervention and has high universality and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a complete flow chart of the method described in the present invention;

[0029] Figure 2 is a schematic diagram of an online measurement system for laser welding keyhole depth based on spectral domain OCT in the present invention;

[0030] Figure 3 It is a schematic diagram of the relationship between the fluctuation characteristics of the keyhole depth data and the spatter and the weld surface in different spatter states in the present invention;

[0031] Figure 4 It is a schematic diagram of the architecture of the 1DCNN-BiLSTM spatter defect recognition model in the present invention;

[0032] Figure 5 It is a schematic diagram of the laser welding spatter defect recognition method based on the TPE algorithm optimization model; DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Unless otherwise defined, the technical terms or scientific terms used herein should be the usual meanings understood by people with general skills in the field to which the present disclosure belongs. What is not described in detail in the present invention is the known technology of technicians in this technical field.

[0034] As shown in the accompanying drawings, the present invention provides a real-time identification method for laser welding spatter based on OCT keyhole depth measurement, comprising the following steps:

[0035] Step S1, collecting keyhole depth data based on OCT measurement and constructing a data set;

[0036] Obtain OCT measurement data that can reflect the change in keyhole depth in a series of laser welding process tests at different laser powers. The OCT measurement data is obtained by measuring the OCT measurement beam coaxially with the laser beam using spectral domain OCT (SD-OCT) technology;

[0037] like Figure 2 As shown, the present application realizes the acquisition of OCT measurement data by building an online measurement system for the depth of a laser welding keyhole based on spectral domain OCT. The system is mainly composed of a low-coherence light source, a spectrometer linear array camera, a fiber coupler, a sample arm, a reference arm, a laser processing head, a laser, and a computer. The low-coherence power source adopts a superradiant light-emitting diode (Superradiant Light-Emitting Diode) wide-spectrum light source with a central wavelength of 850nm and a spectrum half-maximum full width of 45nm. The laser processing head is connected to the laser to obtain a laser beam for welding. The low-coherence light source, the spectrometer linear array camera, the sample arm, and the reference arm are respectively connected to the fiber coupler through optical fibers to form an OCT hardware system. The sample arm is connected to the laser processing head through optical fibers so that the OCT measurement beam is coaxial with the laser beam. The spectrometer linear array camera is connected to the computer. The computer stores relevant programs for processing spectral images, which constitute an OCT software system.

[0038] The low-coherence light source SLD emits a wide spectrum light source and is divided into two beams through a fiber coupler. Beam 1 (reference beam) passes through a collimator to form collimated light and enters the reference arm, hits the workpiece plane and returns along the original path; beam 2 (measurement beam) passes through a collimated light and enters the sample arm, and enters the laser processing head at the same time. It is combined with the laser beam through a beam splitter and focused on the workpiece through a focusing lens. After a stable keyhole is formed, beam 2 hits the bottom of the keyhole to form scattered light and then returns along the original path. Beam 2 hitting the bottom of the keyhole and beam 1 hitting the surface of the workpiece interfere with each other through a fiber coupler, and the original spectral interference fringe image is obtained by taking a photo with a spectrometer linear array camera. Finally, the image is collected by an image acquisition card and transmitted to a computer for data processing. Due to the different transmission distances of the two beams, the frequency wavelength of the measurement light wave and the frequency of the reference wave have a phase difference. SD-OCT performs inverse Fourier transform on the power spectrum density of the interference signal to obtain the phase difference between the two beams, and then the keyhole depth information can be obtained to form OCT measurement data.

[0039] In view of the fact that laser power directly determines the keyhole depth variation and actual spatter defects, the present invention takes aluminum alloy welding as an example to illustrate, and conducts a series of adjustable annular laser welding process tests with different inner ring / outer ring powers to obtain OCT measurement data under various conditions. The specific process parameters are shown in Table 1.

[0040] Table 1 Adjustable annular laser welding process parameters

[0041]

[0042] Figure 3 The relationship between the keyhole depth data fluctuation characteristics and spatter and the schematic diagram of the weld surface with different spatter states are shown. According to the fluctuation degree of the keyhole depth and the actual metallographic analysis results, the non-splash and spatter ranges are determined. In order to better analyze the intrinsic relationship between the keyhole dynamic change and the spatter state, the classic percentile filtering method is used to smooth the keyhole depth data for the original keyhole depth information. Figure 3 It can be seen that the length of spatter defects is generally 2-6mm. According to the welding speed of 150mm / s, the spatter occurrence cycle is generally 23-40 keyhole depth data. The experimental weld length is 60mm, and the amount of data collected for each weld is 400. The OCT measurement data is used to construct a keyhole depth-spatter state data set with the presence of spatter defects as the evaluation index, and is randomly divided into training sample data and test sample data in a ratio of 7:3, and data normalization preprocessing is performed.

[0043] Step S2, construction and training of deep learning model and splash defect recognition;

[0044] A 1DCNN-BiLSTM composite deep learning model was established. The 1DCNN-BiLSTM composite deep learning model was trained using training samples from the keyhole depth-splash state dataset. The keyhole depth feature was used as the input feature vector of the composite deep learning model, and the presence of splash defects was used as the output of the composite deep learning model.

[0045] Dropout technology and adaptive moment estimation are used to prevent the composite deep learning model from overfitting. The hyperparameters of the model are optimized through the TPE algorithm. In the sampling stage, the new hyperparameters select the largest x by maximizing EI (Expected Improvement), which means that x is the probability of a hyperparameter with a good effect. The best hyperparameter combination is continuously updated by selecting the largest x, and the four indicators of the composite deep learning model are obtained: accuracy, precision, recall, and F1 value.

[0046] When the four indicators meet the requirements or the maximum number of TPE algorithm parameter searches is reached, the hyperparameter optimization process is terminated, otherwise return to the above steps;

[0047] According to the optimal number of convolution kernels, number of LSTM layers, number of hidden layers, Dropout value and learning rate obtained by optimization, a 1DCNN-BiLSTM spatter defect recognition model is established;

[0048] Figure 4The 1DCNN-BiLSTM spatter defect recognition model architecture is shown. The model first extracts local features from the keyhole depth data through a one-dimensional convolution layer. The convolution kernel slides on the time dimension of the keyhole depth data, and reduces the time complexity of the keyhole depth data by extracting local patterns. Then it is downsampled through the maximum pooling layer to reduce the length of the keyhole depth feature. After two such operations, it passes through a bidirectional long short-term memory network. At the same time, the Adam optimizer is introduced during the network training process, and then a part of the neurons are randomly discarded during the training process through the Dropout layer. Finally, the splash fluctuation feature representation of the keyhole depth is converted into a probability distribution of splash or no splash through the fully connected layer, and the final classification result of splash and no splash is output.

[0049] The BiLSTM unit consists of two layers of opposite long short-term memory (LSTM) units. t Determines how much of the current input should be stored in the cell state; the output gate o t Determines what information should be output from the current unit state; forget gate f t Determines how much of the previous cell state to retain; cell state C t It is updated based on the results of the forget gate and the input gate, and it is the key part of maintaining the long-term state; the unit output h t It is the filtered output of the unit state, which is used to pass to the next layer of the network or used as the output of the current layer. Let W(·) and b(·) be the weight coefficient and bias coefficient of each gate respectively, and h t is the output of the hidden layer, σ is the kernel function, then the signal update method of each gate is as follows:

[0050]

[0051] Figure 5 The laser welding spatter defect recognition method based on the TPE algorithm optimization model is shown. The TPE algorithm replaces the Gaussian process in Bayesian optimization with a Parzen window estimation method based on a tree structure. The kernel function is usually a function that is symmetrical around the data point and has an integral of 1. The TPE algorithm uses a Gaussian kernel, and its mathematical expression is:

[0052]

[0053] Where u is the offset of the kernel center. In the Parzen window estimation of the TPE algorithm, each sample point x i The center is x, and the width of the kernel is controlled by adjusting the bandwidth, that is:

[0054]

[0055] Then, given a data sample X = {x 1 ,x 2 ,…,x n Its probability density function is It can be expressed as:

[0056]

[0057] Where K is the kernel function and n is the number of samples. The bandwidth h controls the degree of smoothing.

[0058] The TPE algorithm builds a probability model through existing evaluation results in the modeling phase by smoothing existing data points, and establishes the probability density function when x belongs to the hyperparameter of the better result. And the probability density function when x belongs to the hyperparameter with poor results During the sampling phase, new hyperparameters are selected by maximizing EI (Expected Improvement), where EI is calculated based on the ratio of two models:

[0059]

[0060] The TPE algorithm will select the x that maximizes EI(x), which means that x is likely to be a hyperparameter with good effect. When the four indicators meet the requirements or the maximum number of TPE algorithm parameter searches is reached, the hyperparameter optimization process is terminated. A 1DCNN-BiLSTM spatter defect recognition model is established based on the optimal hyperparameter group, and the trained 1DCNN-BiLSTM spatter defect recognition model is used to predict the test sample data and output the classification results.

[0061] The computer in the above method includes a memory and a processor, wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the above method steps.

[0062] Based on the same inventive concept, a readable storage medium stores computer instructions thereon, wherein the computer instructions implement the above method when executed by a processor.

Claims

1. A real-time identification method for laser welding spatter based on OCT keyhole depth measurement, characterized in that: The steps include: S1. Keyhole depth data acquisition and dataset construction based on OCT measurement; S1.

1. Obtaining OCT measurement data that can reflect the change in keyhole depth in a series of laser welding process tests under different laser powers, wherein the OCT measurement data is obtained by coaxially measuring the OCT measurement beam and the laser beam using spectral domain OCT technology; S1.2, constructing a keyhole depth-splash state data set based on the OCT measurement data and taking whether there is a splash defect as an evaluation index; S2. Construction and training of deep learning models and identification of splash defects; S2.1, randomly dividing the constructed keyhole depth-splash state data set into training samples and test samples according to a set ratio, and performing data normalization preprocessing; S2.2, establish a 1DCNN-BiLSTM composite deep learning model, train the 1DCNN-BiLSTM composite deep learning model through the training samples in the keyhole depth-splash state data set, use the keyhole depth feature as the input feature vector of the composite deep learning model, and use the presence of splash defect as the output of the composite deep learning model; S2.3, using Dropout technology and adaptive moment estimation to prevent the composite deep learning model from overfitting, optimizing the model's hyperparameters through the TPE algorithm, selecting the largest x for the new hyperparameters by maximizing EI (Expected Improvement) in the sampling phase, and continuously updating the best hyperparameter combination by selecting the largest x, and obtaining four indicators of the composite deep learning model: accuracy, precision, recall, and F1 value; S2.

4. When the four indicators meet the requirements or the maximum number of TPE algorithm parameter searches is reached, the hyperparameter optimization process is terminated, otherwise, return to step S2.3; S2.

5. According to the optimal number of convolution kernels, number of LSTM layers, number of hidden layers, Dropout value and learning rate obtained by optimization, a 1DCNN-BiLSTM spatter defect recognition model is established; S2.

6. Use the trained 1DCNN-BiLSTM spatter defect recognition model to predict the test sample data and output the classification results.

2. The method according to claim 1, characterized in that The OCT measurement data reflecting the change in the keyhole depth is obtained by building an online measurement system for the laser welding keyhole depth based on spectral domain OCT; The system comprises a low-coherence light source, a spectrometer linear array camera, a fiber coupler, a sample arm, a reference arm, a laser processing head, a laser and a computer, wherein the laser processing head is connected to the laser to obtain a laser beam for welding, the low-coherence light source, the spectrometer linear array camera, the sample arm and the reference arm are respectively connected to the fiber coupler through optical fibers, the sample arm is connected to the laser processing head through optical fibers so that the OCT measurement beam is coaxial with the laser beam, and the spectrometer linear array camera is connected to the computer; The low coherence light source is divided into two light beams through a fiber coupler. One beam passes through a reference arm to hit the workpiece and returns along the original path. The other beam passes through a sample arm and hits the bottom of the keyhole coaxially with the laser beam and returns along the original path. The two returning light beams interfere with each other through a fiber coupler. A spectrometer linear array camera is used to obtain a spectral interference fringe image and send it to a computer. The power spectral density of the spectral interference fringe image is inversely Fourier transformed by the computer to obtain keyhole depth information and form OCT measurement data.

3. The method according to claim 2, characterized in that The low coherence power source adopts a superluminescent diode wide spectrum light source.

4. The method according to claim 1, characterized in that: The keyhole depth-splash state dataset is randomly divided into training samples and test samples according to a ratio of 7:

3.

5. The method according to claim 1, characterized in that The series of laser welding process tests under different laser powers are a series of aluminum alloy laser welding process tests under different laser powers.

6. The method according to claim 1, wherein the computer comprises a memory and a processor, wherein: The memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the method steps described in any one of claims 1 to 5.

7. A readable storage medium having computer instructions stored thereon, wherein: When the computer instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images