Intelligent focusing device and method for laser induced breakdown spectroscopy

Abstract

The present invention discloses an intelligent focusing device and method for laser induced breakdown spectroscopy. The device is arranged in conjunction with a spectrometer, including a laser and its control box, an autofocus module, and an optical path module. The equipment except the laser control box is integrated into a cubic box. At the same time, in order to simplify the sample preprocessing operation and improve the accuracy of laser focusing, an autofocus integration algorithm is written based on the focusing method and control technology of image processing, which is used to predict the focus moving distance, and the autofocus module is used for automatic focusing according to the moving distance. Compared with the traditional manual focusing method, the present invention makes full use of the image learning ability of deep learning, mines the information characteristics of the sample image and the ablation pit left by the LIBS laser hitting the sample, and establishes a multimodal information fusion regression model algorithm, which reduces the human error in manual focusing and effectively improves the collection efficiency of LIBS samples.

Description

Technical Field

[0001] The invention belongs to the technical field of laser induced breakdown spectroscopy, and in particular relates to an intelligent focusing device and method for laser induced breakdown spectroscopy. Background Art

[0002] Laser-Induced Breakdown Spectroscopy (LIBS) technology uses ultrashort pulse lasers to focus on the sample surface to form plasma, and then analyzes the plasma emission spectrum to determine the material composition and content of the sample. It is mainly used for qualitative and quantitative analysis of chemical elements. Its basic principle is to focus high-energy laser pulses on the sample surface to generate high-temperature plasma. Then the laser pulse stops, and the ions and atoms in the high-energy state of the plasma transition to a low-energy state and emit radiation with a certain energy. This radiation has obvious elemental characteristics, and then collects light with a specific wavelength and sends it to the spectrum for analysis. Since the wavelengths of the ion spectrum and the atomic spectrum correspond to specific elements individually and have a certain quantitative relationship, the elemental composition and content in the sample can be determined based on the wavelength and intensity of the emission line.

[0003] LIBS has the advantages of easy sample preparation, low sample loss, high detection efficiency, and online detection. It has been widely used in clinical medicine, archaeology, industry and other fields. However, there are many constraints based on traditional LIBS equipment. The detection samples usually need to be pre-processed such as grinding and tableting, which consumes manpower and reduces the timeliness of detection. When the sample is directly tested, the laser focus needs to be adjusted manually due to the uneven surface and uneven texture of the sample. As a result, the relative position of the focus is not fixed when detecting different samples, which will reduce the stability of the acquired spectral signal, and the detection accuracy is bound to be negatively affected. The above factors seriously restrict the timeliness and accuracy of LIBS detection. Summary of the invention

[0004] In order to solve the shortcomings of the prior art, simplify sample preprocessing, improve the convenience and accuracy of laser focusing, and quickly autofocus on uneven or uneven samples, the present invention adopts the following technical solutions:

[0005] An intelligent focusing method for laser induced breakdown spectroscopy comprises the following steps:

[0006] Step 1: Obtain sample images at different focal planes and ablation pit images of the laser-induced breakdown spectroscopy after focusing as training data;

[0007] The purpose of acquiring sample images on different focal planes is to acquire sample images at different focus distances, so as to provide sufficient learning samples for the model. The purpose of acquiring LIBS ablation pit images is to deeply explore the potential relationship between laser ablation pits and focus. Considering the unevenness of the sample surface, a single sample image cannot fully reflect the accuracy of focus. Since the importance of information around the ablation point is much higher than that of other areas, it is more beneficial for the data set to provide better model training, so that the model can better complete autofocus.

[0008] Step 2: Fusion of features extracted from sample image and ablation crater image;

[0009] The model can fully mine the information contained in the LIBS detection system. This screening and fusion process not only effectively reduces the training time of the model, but also enables the model to understand the surface conditions of the sample more accurately. The optimal model after training can predict the focus adjustment distance accordingly.

[0010] Step 3: Train the multimodal information fusion model, and use the training data after feature fusion for model training to predict the displacement distance of image focus;

[0011] Step 4: Obtain the image to be focused and input it into the trained multimodal information fusion model to obtain the predicted distance;

[0012] Step 5: Adjust the distance according to the predicted distance and deviation and move accordingly to complete the focus.

[0013] Furthermore, in step 3, the multimodal information fusion model uses ResNet as the network core. This structure enables the network to learn the residual (i.e., F(x)), thereby making it easier to train very deep networks because it allows gradients to propagate back to earlier layers more easily, avoiding the gradient vanishing problem, and thus better mining the image information of the sample surface and the LIBS ablation pits. The algorithm formula is:

[0014] F(x)=x+F(x)

[0015] Among them, x represents an input feature map received by the residual block. The first convolutional layer applies a convolution operation on the input x, and the second convolutional layer applies the convolution operation again, usually with different parameters from the first convolutional layer. In particular, ResNet uses a skip connection to add the input x to the output of the previous convolutional layer, that is, the output is x+F(x), where F(x) represents the output of the previous convolutional layer.

[0016] Furthermore, in step 3, the multimodal information fusion model adopts the optimization algorithm Adam, and its formula is as follows:

[0017]

[0018] Among them, θ t+1 represents the parameter value of the next iteration, θ t Represents the parameter value of the current iteration, α represents the learning rate, m_hat represents the first-order moment estimate, v_hat represents the second-order moment estimate, ε is a small positive number used for numerical stability (to avoid division by zero errors), and t represents the current iteration number.

[0019] Furthermore, in step 3, the evaluation index determination coefficient R of the optimal multimodal information fusion model is 2 The loss function RMSE is as follows:

[0020]

[0021] Where N represents the number of samples, y i represents the true target value, represents the predicted value of the model, SSE is the sum of squares of the error, which represents the sum of squares of the difference between the model predicted value and the observed data, SST is the total sum of squares, which represents the total variance of the target value (dependent variable). This calculation formula measures the proportion of the difference between the model's predicted value and the observed data in the total variance, R 2 The value range is between 0 and 1. The closer it is to 1, the better the model can explain the variability of the observed data, while the closer it is to 0, the poorer the explanatory power of the model.

[0022] Furthermore, in the step 3, the network structure of the multimodal information fusion model includes a convolution layer (input channel: 3, output channel: 64), a normalization layer (number of features: 64), an activation function (ReLU), a maximum pooling layer, sequence blocks 1 to 4, an adaptive average pooling layer, a flattening layer, and a fully connected layer, wherein each sequence block has multiple basic blocks (the number of channels increases, including downsampling), each basic block contains a convolution layer, a normalization layer, and a ReLU activation function of the basic block, and each sequence block has the same composition, the purpose is to further increase the number of channels, and the feature extraction of the adaptive average pooling layer is converted into a specified size and then input into the flattening layer, the flattening layer flattens the feature into a one-dimensional vector and inputs it into the fully connected layer, and the fully connected regression layer converts and outputs the input multi-dimensional (512-dimensional) feature to obtain a predicted output.

[0023] Furthermore, in step 5, after the stage is moved to the position where the camera can obtain the clearest sampling image, there is a deviation between the clear point of the sample image and the laser ablation point that needs to be adjusted to obtain the best spectral data, that is, the stage needs to be moved up by a distance D (the deviation can be obtained by optical simulation or experimental measurement) to collect spectral data in a negative defocusing manner. The calculation formula for the focal length f of the light beam passing through the plano-convex lens and the deviation adjustment distance D is as follows:

[0024]

[0025] D = f C -f L +d

[0026] Where R is the radius of curvature of the plano-convex lens, n(λ) is the refractive index of the material at different wavelengths of light, and f C represents the focal length of the camera through the plano-convex lens, f L Represents the focal length of the laser after passing through the plano-convex lens. Since the wavelength of the laser is different from that of visible light, there is a deviation in the focal point after passing through the plano-convex lens. C -f L , d is the optimized parameter obtained experimentally, that is, the laser beam is focused at a point d below the sample surface through a plano-convex lens, so that the spectral data is collected in a negative defocusing manner.

[0027] An intelligent focusing device for laser induced breakdown spectroscopy comprises a laser, an automatic focusing module and an optical path module which are arranged in a coordinated manner. The optical path module obtains an image to be focused through a sample to be focused. According to the intelligent focusing method for laser induced breakdown spectroscopy, a predicted distance is obtained through the image to be focused, and then the automatic focusing module is used to perform corresponding movement to complete the focusing of the sample. The laser acts on the focused sample and generates plasma for spectral analysis.

[0028] Furthermore, the autofocus module includes a PLC (programmable logic controller), a driver and a translation stage connected in sequence. A sample to be focused is arranged on the translation stage. The translation stage includes a Z-axis translation stage, a Y-axis translation stage and an X-axis translation stage which are arranged in coordination. The moving path, speed and distance of the translation stage are planned and controlled by control software. The PLC and the computer are connected through a network port. The control program generated by computer programming sends motion instructions to the translation platform driver, thereby controlling the rotation of the motor in the slide to realize the movement of the translation platform. In addition, the translation stage also includes additional hardware such as a power supply and a sample clamping device. The power supply is responsible for supplying power to the PLC and the driver, and the clamping device is used to fix the sample. The X- and Y-axis displacements are used to lock the sample to be focused, and the Z-axis displacement is used to focus on the sample surface. The displacement accuracy meets the requirements for collecting training samples of the multimodal autofocus model and the fine-tuning requirements during autofocusing.

[0029] Furthermore, the optical path module includes a camera and a lens, the lens includes a half-wave plate, a polarization beam splitter, a filter lens, a dichroic mirror, and a plano-convex lens. The camera and the filter lens, the dichroic mirror, and the plano-convex lens are located on the Z optical axis, and the camera obtains an image of the sample to be focused through the filter lens, the dichroic mirror, and the plano-convex lens; the laser and the half-wave plate, the polarization beam splitter, and the dichroic mirror are located on the X optical axis, and the laser passes through the half-wave plate, the polarization beam splitter, the dichroic mirror, and the plano-convex lens in sequence, and then acts on the focused sample surface and generates plasma; the half-wave plate and the polarization beam splitter are arranged in coordination for adjusting the laser intensity, and the polarization beam splitter and the dichroic mirror are used for beam separation and optical path turning. The laser emitted by the laser is coaxial with the optical path of the camera after being reflected by the dichroic mirror and shares the plano-convex lens, and the magnification is increased, the working distance of the camera is adjusted, the imaging quality of the LIBS ablation pit is improved, and the focusing of the laser and the camera is achieved through the plano-convex lens.

[0030] The optical path module includes a camera, a filter lens, a dichroic mirror, and a plano-convex lens connected in sequence in the Z optical axis direction, wherein the filter lens is used to prevent the laser from damaging the camera, the plano-convex lens provided below the dichroic mirror is used to increase the magnification, adjust the working distance of the camera, and improve the imaging quality of the laser-induced breakdown spectroscopy ablation pit, and the laser emitted by the laser is coaxial with the optical path of the camera after being reflected by the dichroic mirror and shares the plano-convex lens.

[0031] Furthermore, the angle between the polarization beam splitter and the X-ray axis is 135°, and the angle between the dichroic mirror and the X-ray axis is 45°.

[0032] The advantages and beneficial effects of the present invention are:

[0033] Aiming at the demand for rapid detection of LIBS, the present invention designs a system structure and selects and integrates components such as lasers, spectrometers and optical path modules, and integrates the equipment except the laser control box into a cubic box, so that it is convenient for acquiring images, autofocusing and sample detection. In order to simplify sample preprocessing operations, fully mine LIBS-specific information and improve the accuracy of laser focusing, the present invention realizes autofocusing of the original sample based on the multimodal information fusion algorithm of ResNet. Compared with the traditional manual focusing method, the present invention makes full use of the image learning ability of deep learning, deeply mines the image information characteristics unique to LIBS in the sample, and establishes a multimodal information fusion regression model with strong generalization ability. It reduces the human error in manual focusing and effectively improves the collection efficiency of LIBS samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flow chart of a method in an embodiment of the present invention.

[0035] Figure 2 Schematic diagram of the structure of the device in the embodiment of the present invention.

[0036] In the figure: 1. Laser, 2. Spectrometer, 3. PLC, 4. Driver, 5. Translation stage, 6. Half-wave plate, 7. Polarization beam splitter, 8. Dichroic mirror, 9. Industrial camera, 10. Filter lens, 11. Plano-convex lens, 12. Bi-convex lens. DETAILED DESCRIPTION

[0037] The specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the present invention, and is not used to limit the present invention.

[0038] like Figure 1 As shown, a smart focusing method for laser induced breakdown spectroscopy comprises the following steps:

[0039] Step 1: Obtain the corresponding data set by obtaining sample images on different focal planes and LIBS laser ablation crater images after focusing as training data for the model.

[0040] The purpose of obtaining sample images on different focal planes is to obtain sample images at different focal distances so as to provide sufficient learning samples for the model.

[0041] The acquired LIBS ablation pit images are intended to deeply explore the potential relationship between laser ablation pits and focus. Considering the unevenness of the sample surface, a single sample image cannot fully reflect the accuracy of focus. Since the importance of information around the ablation point is much higher than that of other areas, it is more beneficial for the dataset to provide better model training, so that the model can better complete autofocus.

[0042] The obtained dataset is divided into training set, validation set and test set according to the ratio of 3:1:1, which is used for different model training, adjustment and evaluation.

[0043] Step 2: Preprocess the obtained data, extract features from the image data of each module, and then fuse and splice the feature layer information.

[0044] The data preprocessing process is to extract and fuse the image information characteristics of the sample surface and the LIBS ablation pit, so that the model can fully mine the information contained in the LIBS detection system. This feature extraction and fusion process not only effectively reduces the training time of the model, but also enables the model to understand the sample surface more accurately. The optimal model after training can predict the focus adjustment distance accordingly.

[0045] Step 3: Train the multimodal information fusion model and use the preprocessed training data set as input to complete the model training.

[0046] The multimodal information fusion model uses ResNet as the network core. This structure enables the network to learn the residual (i.e., F(x)), making it easier to train very deep networks because it allows gradients to propagate back to earlier layers more easily, avoiding the gradient vanishing problem, and thus better mining the image information of the sample surface and the LIBS ablation pits. The algorithm formula can be simplified as follows:

[0047] F(x)=x+F(x)

[0048] Here, x represents a residual block that accepts an input feature map. The first convolutional layer applies a convolution operation on the input x, and the second convolutional layer applies the convolution operation again, usually with different parameters from the first convolutional layer. In particular, ResNet uses a skip connection to add the input x to the output of the previous convolutional layer, that is, the output is x+F(x), where F(x) represents the output of the previous convolutional layer.

[0049] The multimodal information fusion model uses the optimization algorithm Adam, and its formula is as follows:

[0050]

[0051] Among them, θ t+1 represents the parameter value of the next iteration, θ t represents the parameter value of the current iteration, α represents the learning rate, m_hat represents the first-order moment estimate, v_hat is the second-order moment estimate, ε is a small positive number used for numerical stability (to avoid division by zero errors), and t represents the current iteration number.

[0052] The evaluation index determination coefficient R of the optimal model 2 And the loss function mean square error loss function RMSE:

[0053]

[0054] Where N represents the number of samples, y i represents the true target value, Represents the predicted value of the model. SSE is the sum of squared errors, which represents the sum of squares of the differences between the model's predicted value and the observed data. SST is the total sum of squares, which represents the total variance of the dependent variable (target value). This calculation formula measures the proportion of the difference between the model's predicted value and the observed data in the total variance. 2 The value range is between 0 and 1. The closer it is to 1, the better the model can explain the variability of the observed data, while the closer it is to 0, the poorer the explanatory power of the model.

[0055] In an embodiment of the present invention, the network structure of the multimodal information fusion model consists of a convolution layer (input channel: 3, output channel: 64), a normalization layer (number of features: 64), an activation function (ReLU), a maximum pooling layer, sequence blocks 1 to 4, an adaptive average pooling layer, a flattening layer, and a fully connected layer. Each sequence block has multiple basic blocks (the number of channels increases, including downsampling), and each basic block contains a convolution layer, a normalization layer, and a ReLU activation function. The composition of each sequence block is the same, and the purpose is to further increase the number of channels. Finally, the features are extracted by the adaptive average pooling layer and converted into a specified size and input into the flattening layer. The flattening layer flattens the features into a one-dimensional vector and inputs them into the fully connected layer. Finally, the fully connected regression layer converts and outputs the input 512-dimensional features, which is the predicted output.

[0056] In one embodiment of the present invention, white peony slices are used as experimental objects, the focal plane label is given 0mm, and an image acquisition plane is selected every 1mm above and below the focal plane. Among them, 7 image acquisition planes are determined in the range of -3mm to 3mm (the label corresponding to each plane is the distance relative to the focal plane), and 50 images are collected for each plane, for a total of 350 images, which are used to establish PLSR and ResNet quantitative models. In this embodiment, the training of the multimodal information fusion model requires additional ablation images, and 50 images of LIBS laser-impacted ablation pits after focusing are collected. The training of the multimodal information fusion model requires feature extraction of each module data first, and then feature layer information fusion and splicing. In this embodiment, a method is proposed that only an image of a sample to be focused and an image of the ablation pit are obtained as the input of the model, and the model will output the corresponding predicted value. The subsequent displacement stage 5 adjusts the distance 2mm (obtained according to experimental measurement) according to the predicted value and the deviation to move the specified distance to complete the automatic focus. In order to test the prediction performance of the model, classical machine learning, deep learning methods and the multimodal information fusion model proposed in this invention were used. Table 1 shows the prediction results of each model.

[0057] Table 1. Prediction results of representative models

[0058]

[0059] Note: R 2 represents the coefficient of determination, and RMSE represents the root mean square error.

[0060] In another embodiment of the present invention, rice leaves are used as experimental objects, the focal plane label is given 0mm, and an image acquisition plane is selected every 0.5mm above and below the focal plane. Among them, 7 image acquisition planes are determined in the range of -1.5mm to 1.5mm (the label corresponding to each plane is the distance relative to the focal plane), and 20 images are collected for each plane, for a total of 140 images, which are used to establish PLSR and ResNet quantitative models. In this embodiment, the training of the multimodal information fusion model requires additional ablation images, and 20 images of LIBS laser-impacted ablation pits after focusing are collected. The training of the multimodal information fusion model requires feature extraction of each module data first, and then feature layer information fusion and splicing. In this embodiment, a method is proposed that only an image of a sample to be focused and an image of the ablation pit are obtained as the input of the model, and the model will output the corresponding predicted value. The subsequent displacement stage 5 adjusts the distance 2mm (obtained according to experimental measurement) according to the predicted value and the deviation to move the specified distance to complete automatic focusing. In order to test the prediction performance of the model, classical machine learning, deep learning methods and the multimodal information fusion model proposed in this invention were used. Table 2 shows the prediction results of each model.

[0061] Table 2. Prediction results of representative models

[0062]

[0063] Note: R 2 represents the coefficient of determination, and RMSE represents the root mean square error.

[0064] Step 4: Predict the displacement distance based on the given image, obtain the image information to be focused and input it into the trained multimodal information fusion model, and the model automatically gives the predicted distance;

[0065] Step 5: The translation stage 5 moves accordingly based on the model predicted distance and the deviation adjustment distance of 2 mm (obtained from experimental measurements), and the automatic focusing is completed.

[0066] After the stage 5 is moved to the position where the industrial camera 9 can obtain the clearest sampling image, there is a deviation between the clear point of the sample image and the laser ablation point that needs to be adjusted to obtain the best spectral data, that is, the stage 5 needs to be moved up by a distance D (the deviation can be obtained by optical simulation or experimental measurement) to collect spectral data in a negative defocusing manner. The calculation formula of the focal length f of the light beam passing through the plano-convex lens 11 and the deviation adjustment distance D is as follows:

[0067]

[0068] D = f C -f L +d

[0069] Wherein, R represents the curvature radius of the plano-convex lens 11, n(λ) represents the refractive index of the material at different light wavelengths, and f C represents the focal length of the industrial camera 9 through the plano-convex lens 11, f L represents the focal length of the laser after passing through the plano-convex lens 11. Since the wavelength of the laser is different from that of visible light, there is a deviation in the focal point after passing through the plano-convex lens 11. C -f L , d is the optimized parameter obtained experimentally, that is, the laser beam is focused at a point d below the sample surface through the plano-convex lens 11, so that the spectral data is collected in a negative defocusing manner.

[0070] In the embodiment of the present invention, the LIBS autofocus device is carried out under normal pressure conditions, and the parameters need to be adjusted before use, wherein the laser wavelength is 1064nm, the pulse energy is 100mJ, the pulse frequency is 1Hz, the spectrometer acquisition delay time is 2μs, and the integration time is 10μs. This acquisition method can further collect more sample surface information and reduce the influence of other factors.

[0071] The surface of the research object sample is uneven, and when focusing, only the ablation point is focused, the laser focus, i.e. the ablation point, is selected as the center, and the 215×215 pixel matrix is ​​the focus area. The research objects in the two embodiments are different, and the multimodal information fusion model proposed in the present invention shows good prediction performance, reflecting its excellent generalization.

[0072] like Figure 2 As shown, an intelligent focusing device for laser induced breakdown spectroscopy integrates the above method into a LIBS autofocus device with a sample image acquisition function, and integrates an industrial camera 9 into the coaxial optical path of the LIBS, thereby deeply mining the image information unique to LIBS, simplifying the sample preprocessing operation and improving the laser focusing accuracy, so as to meet the requirements of the LIBS device for detection timeliness and accuracy, forming the LIBS autofocus hardware and supporting control software, and adopting the autofocus method of related control technology and multimodal information fusion to realize the regional spectrum and image data acquisition of the original sample, simplify the sample preprocessing operation and improve the laser focusing accuracy. The device includes: a cubic box and a pulse laser 1 and its control box, a spectrometer 2, a switching power supply 3, an autofocus module and an optical path module arranged in the cubic box.

[0073] like Figure 2As shown, the autofocus module includes PLC3, electric displacement stage 5, and driver 4. The displacement stage 5 is connected to the driver 4, and the driver 4 is connected to PLC3. The displacement stage 5 includes a Z-axis electric displacement stage, a Y-axis electric displacement stage, and an X-axis electric displacement stage. The moving path, speed, and distance of the displacement stage 5 are planned and controlled by the control software. PLC3 and the computer are connected through the network port, and the control program generated by computer programming sends motion instructions to the displacement platform driver 4, thereby controlling the rotation of the motor in the slide to realize the movement of the displacement platform. In addition, the displacement stage 5 also includes additional hardware such as a power supply and a sample clamping device, wherein the power supply is responsible for supplying power to PLC3 and the driver 4, and the clamping device is used to fix the sample. The electric displacement slide can move in three-dimensional space, with a travel of 100mm in the X and Y directions, a travel of 12mm in the Z direction, a repeat positioning accuracy of ±1μm, a load of 20kg, and X and Y axis displacements are used to lock the sample, and the Z axis displacement is used to focus on the sample surface. The displacement accuracy meets the requirements for collecting samples for multimodal autofocus model training and the requirements for fine-tuning during autofocusing.

[0074] An industrial camera 9 is installed on the optical path module so that the focus of the camera coincides with the focus of the laser on the sample surface. The optical path module is composed of an industrial camera 9 and a lens, wherein the lens contains a half-wave plate 6, a polarization beam splitter 7, a filter lens 10, a dichroic mirror 8, and a plano-convex lens 11. The laser 1 and the half-wave plate 6, polarization beam splitter 7, and dichroic mirror 8 of the lens are located on the X-ray axis. The angle between the polarization beam splitter 7 and the X-ray axis is 135°, and the angle between the dichroic mirror 8 and the X-ray axis is 45°. The laser intensity is adjusted by the coordinated setting of the half-wave plate 6 and the polarization beam splitter 7. The beam splitter 7 and the dichroic mirror 8 realize beam separation and light path turning. The laser passes through the half-wave plate 6, the polarization beam splitter 7, the dichroic mirror 8 and the plano-convex lens 11 in sequence, and then acts on the focused sample surface to generate plasma; the industrial camera 9, the dichroic mirror 8 and the plano-convex lens 11 are located on the Z optical axis, and the laser and the industrial camera 9 are focused by the plano-convex lens; the industrial camera 9 obtains the sample image by the filter lens 10, the dichroic mirror 8 and the plano-convex lens 11, and the laser and the industrial camera 9 share the plano-convex lens, and the beam separation and light path turning are realized by adding the polarization beam splitter and the dichroic mirror.

[0075] The functions of the optical path module include adjusting and reflecting the laser so that it illuminates the object to be measured and receiving the light. The module adopts a cage structure and optical support rods to ensure the coaxiality of each optical element.

[0076] The supporting control software has the functions of acquiring camera images, controlling laser 1 and translation stage 5, acquiring data from spectrometer 2, etc. It can acquire image data captured by industrial camera 9 and store it in the path specified by the user. The resolution of the acquired image is 5472×3648; it can acquire spectral data captured by spectrometer 2 and store it in the path specified by the user. The spectral data will be saved in the default storage format of spectrometer 2.

[0077] In addition, the double convex lens 12 for collecting signals of the spectrometer 2 is a fiber focusing lens, which is adjusted so that its focus coincides with the laser focus to collect the plasma emission light.

[0078] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent focusing device for laser induced breakdown spectroscopy, comprising a laser (1), a spectrometer (2), an autofocus module and an optical path module that are arranged in a coordinated manner, characterized in that: The optical path module obtains an image to be focused through a sample to be focused. According to an intelligent focusing method for laser induced breakdown spectroscopy, a predicted distance is obtained through the image to be focused, and then the automatic focusing module is used to perform corresponding movement to complete the focusing of the sample. The laser (1) acts on the focused sample and generates plasma for spectral analysis. The intelligent focusing method for laser induced breakdown spectroscopy comprises the following steps: Step 1: Obtain sample images at different focal planes and ablation pit images of the laser-induced breakdown spectroscopy after focusing as training data; Step 2: Fusion of features extracted from sample image and ablation crater image; Step 3: Train the multimodal information fusion model, and use the training data after feature fusion for model training to predict the displacement distance of image focus; Step 4: Obtain the image to be focused and input it into the trained multimodal information fusion model to obtain the predicted distance; Step 5: Adjust the distance according to the predicted distance and deviation and move accordingly to complete the focus; After the translation stage (5) is moved to the position where the camera can obtain the clearest sampling image, the deviation between the clear point of the sample image and the laser ablation point is adjusted. The calculation formula of the focal length f of the light beam passing through the plano-convex lens (11) and the deviation adjustment distance D is as follows: D=f C -f L +d Wherein, R represents the radius of curvature of the plano-convex lens (11), n(λ) represents the refractive index of the material at different wavelengths of light, and f C represents the focal length of the camera through the plano-convex lens (11), f L represents the focal length of the laser after passing through the plano-convex lens (11). Since the wavelength of the laser is different from that of visible light, there is a deviation in the focal point after passing through the plano-convex lens (11). C -f L , d is the optimized parameter obtained experimentally, that is, the laser beam is focused at a point d below the sample surface through the plano-convex lens (11).

2. The intelligent focusing device for laser induced breakdown spectroscopy according to claim 1, characterized in that: In step 3, the multimodal information fusion model uses ResNet as the network core, and its algorithm formula is: F(x)=x+F(x) Among them, x represents an input feature map received by the residual block, the first convolutional layer applies a convolution operation on the input x, and the second convolutional layer applies the convolution operation again. ResNet uses a skip connection to add the input x to the output of the previous convolutional layer, that is, the output is x+F(x), where F(x) represents the output of the previous convolutional layer.

3. The intelligent focusing device for laser induced breakdown spectroscopy according to claim 1, characterized in that: In step 3, the multimodal information fusion model uses the optimization algorithm Adam, and its formula is as follows: Among them, θ t+1 represents the parameter value of the next iteration, θ t Represents the parameter value of the current iteration, α represents the learning rate, m_hat represents the first-order moment estimate, v_hat represents the second-order moment estimate, ε is a small positive number for numerical stability, and t represents the current iteration number.

4. The intelligent focusing device for laser induced breakdown spectroscopy according to claim 1, characterized in that: In step 3, the evaluation index determination coefficient R of the optimal multimodal information fusion model 2 The loss function RMSE is as follows: Where N represents the number of samples, y i represents the true target value, It represents the predicted value of the model, SSE is the sum of squared errors, which represents the sum of squares of the differences between the model predicted value and the observed data, and SST is the total sum of squares, which represents the total variance of the target value (dependent variable).

5. The intelligent focusing device for laser induced breakdown spectroscopy according to claim 1, characterized in that: In the step 3, the network structure of the multimodal information fusion model includes a convolution layer, a normalization layer, an activation function, a maximum pooling layer, a sequence block, an adaptive average pooling layer, a flattening layer, and a fully connected layer, wherein each sequence block has multiple basic blocks, each basic block contains a convolution layer, a normalization layer, and an activation function of the basic block, and the feature extraction of the adaptive average pooling layer is converted into a specified size and then input into the flattening layer, the flattening layer flattens the feature into a one-dimensional vector and inputs it into the fully connected layer, and the fully connected regression layer converts and outputs the input multi-dimensional features to obtain a predicted output.

6. The intelligent focusing device for laser induced breakdown spectroscopy according to claim 1, characterized in that: The automatic focusing module comprises a PLC (3), a driver (4) and a displacement stage (5) which are connected in sequence, a sample to be focused is arranged on the displacement stage (5), and the displacement stage (5) is an XYZ three-axis displacement stage, X-axis and Y-axis displacements are used to lock the sample to be focused, and Z-axis displacement is used to focus on the sample surface.

7. The intelligent focusing device for laser induced breakdown spectroscopy according to claim 1, characterized in that: The optical path module comprises a camera and a lens, wherein the lens comprises a half-wave plate (6), a polarization beam splitter (7), a filter lens (10), a dichroic mirror (8), and a plano-convex lens (11); the camera, the dichroic mirror (8), and the plano-convex lens (11) are located on the Z optical axis, and the camera obtains an image of the sample to be focused through the filter lens (10), the dichroic mirror (8), and the plano-convex lens (11); the laser (1), the half-wave plate (6), the polarization beam splitter (7), and the dichroic mirror (8) are located on the X optical axis, and the laser passes through the half-wave plate (6), the polarization beam splitter (7), the dichroic mirror (8), and the plano-convex lens (11) in sequence, and acts on the surface of the focused sample to generate plasma; the half-wave plate (6) and the polarization beam splitter (7) are arranged in coordination to adjust the laser intensity, and the polarization beam splitter (7) and the dichroic mirror (8) are used for light beam separation and light path turning; The optical path module is provided with a filter lens (10) below the camera in the Z optical axis direction. A plano-convex lens (11) is provided below the dichroic mirror (8) for increasing the magnification, adjusting the working distance of the camera and improving the imaging quality of laser-induced breakdown spectroscopy ablation pits. The laser light emitted by the laser (1) is coaxial with the optical path of the camera after being reflected by the dichroic mirror (8) and shares the plano-convex lens (11).

8. The intelligent focusing device for laser induced breakdown spectroscopy according to claim 7, characterized in that: The angle between the polarization beam splitter (7) and the X-ray axis is 135°, and the angle between the dichroic mirror (8) and the X-ray axis is 45°.

Images