A method for monitoring deposition impurities on the first mirror of a tokamak based on a convolutional neural network

Abstract

The present invention relates to a method for monitoring deposition impurities on the first mirror of a tokamak based on a convolutional neural network. The method uses a spectrometer to real-time collect plasma spectral data of the sputtered and ionized surface material of the first mirror in a tokamak under low-pressure radio frequency conditions in a fusion device. According to the characteristics of plasma spectral data with different wavelengths and intensities corresponding to different sputtered materials on the first mirror surface, a convolutional neural network model is established and trained to be able to identify the spectral data of the composition and content of the sputtered materials on the first mirror surface. The spectral data obtained by plasma sputtering is input into a mature network model for analysis to judge the composition and content of the sputtered materials, so as to monitor the change of deposition impurities on the first mirror surface. The present invention obtains the radio frequency plasma spectral information generated on the first mirror surface through a spectrometer, and can identify the composition and content of the sputtered materials on the first mirror surface through a convolutional neural network model without damage, effectively and without affecting the experiment.

Description

Technical Field

[0001] The present invention belongs to the technical field of nuclear fusion spectral analysis, and particularly relates to a method for monitoring deposition impurities of the first mirror of a tokamak based on a convolutional neural network. Background Art

[0002] During the operation of a tokamak device, the optical diagnostic system is one of the important means for real-time acquisition and monitoring of the plasma state and ensuring the safety of the device. Due to the increase in the device size, a large number of mirrors need to be installed inside the device in all optical diagnostic systems to form a maze structure of the optical path. The mirror at the front end of the device directly faces the plasma and is called the first mirror.

[0003] During the operation of the tokamak, the first mirror is bombarded by high-energy ions, charge-exchange neutral atoms, etc., irradiated by various rays, and undergoes processes such as wall treatment element deposition and re-deposition of sputtered materials. Its optical reflectivity deteriorates sharply and its service life is rapidly shortened. The optical performance of the first mirror directly affects the working state and overall performance of the relevant diagnostic systems, and also determines whether various conventional optical diagnostic systems can be applied to large tokamak devices. By generating a radio-frequency plasma on the surface of the first mirror, sputtered deposition impurities can be produced, and spectral signals of the deposited impurities can be excited in the radio-frequency plasma. Since the spectral signals generated by impurities with different deposition compositions and thicknesses have different wavelength and intensity characteristics. Identifying the spectral characteristics can distinguish the deposition composition and thickness on the surface of the first mirror. Using the good learning ability of the convolutional neural network, the spectral data of the impurity deposition on the first mirror can be used for feature learning, and a convolutional neural network can be constructed. On this basis, the change state of the surface material composition and content of the radio-frequency plasma tokamak first mirror can be monitored.

[0004] Therefore, the present invention proposes a method for monitoring deposition impurities of the first mirror of a tokamak based on a convolutional neural network, that is, first training a convolutional neural network model with spectral data of different thicknesses and different deposition layer materials, and during the process of sputtering the first mirror with radio-frequency plasma, using the trained model to process the spectra of the sputtered materials during the cleaning process of the first mirror collected in real time, accurately monitoring the surface deposition layer materials of the first mirror, and accurately evaluating the optical performance of the first mirror. Summary of the Invention

[0005] The present invention aims at the situation that the state of impurity deposition on the surface of the first mirror cannot be monitored during the process of sputtering the first mirror with radio-frequency plasma, changes the previous situation that relevant experimental studies on the first mirror need to suspend the experiment and then conduct monitoring and analysis, solves the problems that the state of material impurity deposition on the surface of the first mirror cannot be known in time, and the time and human resource efficiency are low, provides a reference for the deposition of material impurities on the surface of the first mirror, and accelerates the research efficiency related to the surface deposition of the first mirror.

[0006] The object of the present invention is to provide a method for monitoring the deposited impurities on the first mirror of a tokamak based on a convolutional neural network. By utilizing the characteristics that the spectral signals of different deposited materials have different wavelengths and intensities during the process of depositing a layer on the first mirror by radio frequency plasma sputtering, and the good learning ability of the convolutional neural network, the monitoring of the surface material of the first mirror is realized. This method can monitor the surface material of the first mirror and evaluate its optical performance without opening the vacuum chamber and without changing the spatial position of the first mirror.

[0007] The technical solution of the present invention is a method for monitoring the deposited impurities on the first mirror of a tokamak based on a convolutional neural network, which includes the following steps:

[0008] Step 1: Use a spectrometer to collect the radio frequency plasma spectrum generated on the surface of the first mirror with a known deposition state, and preprocess the plasma spectrum data on the surface of the first mirror through a filter to obtain the original training data, meeting the requirements for training the convolutional neural network model.

[0009] Specifically, the wavelength range for collecting the plasma spectrum is 300 nm - 400 nm, which includes the plasma spectrum characteristic peaks of the first mirror material and the plasma spectrum characteristic peaks of surface impurity deposition.

[0010] Step 2: Train the initial one-dimensional convolutional neural network with the original training data, optimize and adjust the parameters until the loss value of the loss function converges, and obtain the trained convolutional neural network model as the final convolutional neural network model.

[0011] Specifically, the positions and intensities of the plasma spectrum characteristic peaks generated by the deposition of impurities on different surface materials of the first mirror are different. The composition and content of the deposited film on the surface of the first mirror can be known according to the changes in the spectral characteristic peaks. The plasma spectrum data collected in Step 1 is used to train the one-dimensional convolutional neural network.

[0012] Step 3: Real-time collect the radio frequency plasma spectrum data generated on the surface of the first mirror, and input the data into the trained convolutional neural network model to identify the deposition composition and thickness on the surface of the first mirror and output the results.

[0013] Specifically, when radio frequency sputtering the first mirror, use a spectrometer to collect the plasma spectrum of the unknown surface material of the first mirror, and output the classification category through the one-dimensional convolutional neural network model in Step 2 to monitor the composition and content state of the surface material of the first mirror.

[0014] The preprocessing in Step 1 uses a Savitzky-Golay filter to perform a reproduction fit on the plasma spectrum data of a specified number of 2n + 1 points.

[0015] In Step 2, specifically:

[0016] The one-dimensional convolutional neural network model mainly consists of an input layer, a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, a fully connected layer, and an output layer;

[0017] The data input into the input layer is the spectral data processed in Step 1, and it is read and operated in batches of every 128 data.

[0018] Each feature extraction layer contains a convolutional module and a pooling module. The convolutional kernel size of the convolutional module in the first feature extraction layer is 10×1, and the number of convolutional kernels is 64. The convolutional kernel size of the convolutional module in the second feature extraction layer is 5×1, and the number of convolutional kernels is 64. The convolutional kernel size of the convolutional module in the third feature extraction layer is 2×1, and the number of convolutional kernels is 128; in each convolutional operation layer, the LeakyReLU (a variant of the rectified linear unit) is used as the activation function; each pooling module uses max pooling to sample the feature maps generated by the convolutional operation.

[0019] The fully connected layer uses the Relu activation function and integrates the local spectral features extracted by the convolutional layer through a weight matrix to classify the material of the first mirror surface.

[0020] The output layer uses the softmax function. After normalizing the features output by the fully connected layer, it outputs the composition and content of the sputtering material on the first mirror surface corresponding to the input spectral data.

[0021] The expression of the one-dimensional data of the plasma spectrum is:

[0022] y = (∑a n *x + b)

[0023] In the formula, x and y are the input and output feature maps respectively, a n is the convolutional kernel used for the convolutional operation, and b is the bias of the feature map.

[0024] The ADAM optimizer and the cross-entropy loss function are adopted in the convolutional neural network model, and the radio frequency plasma spectrum data generated on the first mirror surface is input to train the one-dimensional convolutional neural network.

[0025] The composition and content status of the material on the first mirror surface can be monitored through the output of the model.

[0026] The present invention uses a convolutional neural network model to extract features from spectral data. Compared with other traditional neural networks, it improves the expression ability and convergence effect of the model, and has stronger generalization ability for the identification of the first mirror surface plasma spectra of different materials.

[0027] The present invention can monitor the changes in the surface material of the first mirror in the Tokamak in the device in real time by collecting the plasma spectrum generated by the first mirror radio frequency in the device, that is, the changes in the surface composition and content of the first mirror. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic structural diagram of the plasma spectrum acquisition device of the present invention;

[0029] Among them: 1 vacuum chamber; 2 first mirror sample; 3 gas tank; 4 radio frequency power supply; 5 optical window; 6 micro spectrometer; 7 vacuum pump group; 8 computer; 10 gas flow controller.

[0030] Figure 2 It is a flowchart of the method of the present invention.

[0031] Figure 3 It is a diagram of the convolutional neural network model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] The present invention will be described in detail below in conjunction with the drawings and examples in the embodiments of the present invention:

[0033] As Figure 1 shown, it is a schematic structural diagram of a method for monitoring the deposition impurities of the first mirror of a Tokamak based on a convolutional neural network. In a specific example of radio frequency plasma cleaning the first mirror, the first mirror sample 2 to be cleaned is placed in the vacuum chamber 1, and the gas flow controller 10 is installed on the gas tank 3. The radio frequency plasma is generated in the area near the surface of the first mirror sample 2 by using the radio frequency power supply 4 to clean the surface deposition layer. The spectrum of the sputtered material is collected by the micro spectrometer 6 through the optical window 5 and transmitted to the computer 8 for processing. The vacuum pump group 7 provides a vacuum condition for the device, and the vacuum pump group 7 and the gas flow controller 10 are connected to the computer 8 to control parameters such as the air pressure in the experiment in real time.

[0034] As Figure 2 shown, in a specific embodiment, a method for monitoring the deposition impurities of the first mirror of a Tokamak based on a convolutional neural network includes the following steps:

[0035] Step 1: Use a spectrometer to collect the radio frequency plasma spectrum generated on the surface of the first mirror with a known deposition state, and preprocess the plasma spectrum data on the surface of the first mirror through a filter to obtain the original data, meeting the requirements for training the convolutional neural network model.

[0036] The wavelength range of the plasma spectrum collection is 300nm - 400nm, which includes the plasma spectrum characteristic peaks of the first mirror material and the plasma spectrum characteristic peaks of surface impurity deposition. After the spectrometer collects the spectral data, the data is preprocessed, and the preprocessing is noise reduction processing to alleviate the influence of other factors in the detection device on the spectral signal collection of this substance.

[0037] Step 2: Train the initial one-dimensional convolutional neural network with the original data, adjust the parameters until the loss value of the loss function converges, and obtain the trained convolutional neural network model as the final convolutional neural network model.

[0038] Specifically, the positions and intensities of the plasma spectrum characteristic peaks generated by the impurity deposition on the surface of different first mirrors are different. The composition and content of the deposition film on the surface of the first mirror can be known according to the changes in the spectral characteristic peaks. The plasma spectral data collected in Step 1 is used to train the one-dimensional convolutional neural network.

[0039] Step 3: Real-time collect the radio frequency plasma spectral data generated on the surface of the first mirror, and input the data into the trained convolutional neural network model to identify the deposition composition and thickness on the surface of the first mirror and output the results.

[0040] Specifically, when radio frequency sputtering the first mirror, use the spectrometer to collect the plasma spectrum of the unknown material on the surface of the first mirror, and output the classification category through the one-dimensional convolutional neural network model in Step 2 to monitor the composition and content status of the material on the surface of the first mirror.

[0041] The preprocessing in Step 1 uses a Savitzky-Golay filter to perform a reproduction fit on the plasma spectral data of a specified number of 2n + 1 points.

[0042] After being processed by the filter, the adjacent spectral data points become smoother, and relevant features such as the position and width of the spectral peaks are retained.

[0043] As Figure 3 shown, in Step 2, specifically:

[0044] The one-dimensional convolutional neural network model is mainly composed of an input layer, a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, a fully connected layer, and an output layer;

[0045] The data input in the input layer is the spectral data processed in Step 1, and it is read and operated in batches with every 128 data as the batch size.

[0046] Each feature extraction layer contains a convolutional module and a pooling module. The convolutional kernel size of the convolutional module in the first feature extraction layer is 10×1, and the number of convolutional kernels is 64. The convolutional kernel size of the convolutional module in the second feature extraction layer is 5×1, and the number of convolutional kernels is 64. The convolutional kernel size of the convolutional module in the third feature extraction layer is 2×1, and the number of convolutional kernels is 128. In each convolutional operation layer, the variant of the rectified linear unit LeakyReLU is used as the activation function. Each pooling module uses max pooling to sample the feature map generated by the convolutional operation.

[0047] The fully connected layer uses the Relu activation function, and integrates the spectral local features extracted by the convolutional layer through the weight matrix to classify the material of the first mirror surface.

[0048] The output layer uses the softmax function. After normalizing the features output by the fully connected layer, it outputs the composition and content of the sputtered material on the first mirror surface corresponding to the input spectral data.

[0049] In the convolutional neural network model, the ADAM optimizer and the cross-entropy loss function are adopted, and the radio frequency plasma spectral data generated on the first mirror surface are input to train the one-dimensional convolutional neural network.

[0050] The composition and content status of the material on the first mirror surface can be monitored through the output of the model.

[0051] In the vacuum chamber of the fusion device, the spectral data of the sputtered material during the radio frequency plasma cleaning of the first mirror are collected in real time, and the data are input into the trained convolutional neural network model to identify the composition and content of the deposited film on the first mirror surface.

[0052] When actually implementing this method, the principle of the interaction between the discharge radio frequency plasma and the impurity material on the first mirror surface is adopted to realize the cleaning and restoration of the reflectivity of the first mirror. When performing radio frequency plasma cleaning on the molybdenum mirror surface coated with an alumina film under argon conditions, at the beginning of the cleaning, only the spectral characteristic peak of aluminum is at 394 nm. After a period of time, the intensity of the aluminum characteristic peak weakens, and gradually the characteristic peak wavelength of molybdenum appears at 379 nm, indicating that the first mirror is being cleaned. When only the characteristic peak of molybdenum, that is, the characteristic peak of aluminum completely disappears, in the spectral data, it indicates that the cleaning of the first mirror has ended. After preprocessing the collected plasma spectrum, the spectral data features are extracted through the deep learning convolutional neural network. After training, a model that can classify the cleaning process of the spectral data can be obtained, and the spectrum of whether there are still impurities in the target spectral data is output, so as to obtain the cleaning process and avoid over-cleaning, which may damage the first mirror material and affect the optical performance of the first mirror.

[0053] The parts not elaborated in the present invention belong to the well-known technologies in the art. The above-described embodiments are only descriptions of the preferred embodiments of the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments described. Without departing from the spirit of the design of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for monitoring deposition impurities of the first mirror in a tokamak based on a convolutional neural network, characterized in that: It includes the following steps: Step 1: Use a spectrometer to collect the radio frequency plasma spectrum generated by the first mirror surface in a known deposition state, and preprocess the radio frequency plasma spectrum data generated by the first mirror surface through a filter to obtain the original training data, meeting the requirements for training the convolutional neural network model; Step 2: Train the initial one-dimensional convolutional neural network with the original training data, adjust the parameters until the loss value of the loss function converges, and obtain the trained convolutional neural network model as the final convolutional neural network model; Step 3: Real-time collect the radio frequency plasma spectrum data generated by the first mirror surface, and input the data into the trained convolutional neural network model to identify the deposition composition and thickness of the first mirror surface and output the results; In the said Step 2, specifically: The one-dimensional convolutional neural network model is mainly composed of an input layer, a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, a fully connected layer, and an output layer; The data input in the input layer is the spectral data processed in Step 1, and it is read and operated in batches of 128 data; Each feature extraction layer contains a convolutional module and a pooling module. The convolutional kernel size of the convolutional module in the first feature extraction layer is 10×1, and the number of convolutional kernels is 64. The convolutional kernel size of the convolutional module in the second feature extraction layer is 5×1, and the number of convolutional kernels is 64. The convolutional kernel size of the convolutional module in the third feature extraction layer is 2×1, and the number of convolutional kernels is 128; in each convolutional operation layer, the LeakyReLU variant of the rectified linear unit is used as the activation function; max pooling is used to sample the feature maps generated by the convolutional operation; The expression of the one-dimensional plasma spectrum data is: where x and y are the input and output feature maps respectively, a n is the convolution kernel used for the convolution operation, and b is the bias of the feature map.

2. The method for monitoring deposition impurities of the first mirror of a tokamak based on a convolutional neural network according to claim 1, wherein: The preprocessing in the first step described above uses a Savitzky-Golay filter to perform a reproduction fit on the plasma spectral data of a specified number of 2 n + 1 points.

3. A method for monitoring deposition impurities of the first mirror in a tokamak based on a convolutional neural network according to claim 1, characterized in that: The ADAM optimizer and the cross-entropy loss function are used in the convolutional neural network model, and the radio frequency plasma spectrum data generated by the first mirror surface is input to train the one-dimensional convolutional neural network.

4. A method for monitoring deposition impurities of the first mirror in a tokamak based on a convolutional neural network according to claim 1, characterized in that: The composition and content status of the material on the first mirror surface can be monitored through the output of the model.

5. A deposition impurity monitoring method for the first mirror of a tokamak based on a convolutional neural network according to claim 1, characterized in that: In Step 1, the wavelength range for plasma spectrum collection is 300nm - 400nm, which includes the plasma spectrum characteristic peaks of the first mirror material and the plasma spectrum characteristic peaks of surface impurity deposition.

6. A method for monitoring deposition impurities of the first mirror of a tokamak based on a convolutional neural network according to claim 1, characterized in that: The fully connected layer uses the Relu activation function, and integrates the spectral local features extracted by the three convolutional layers through the weight matrix to classify the material on the first mirror surface.

7. A method for monitoring deposition impurities of the first mirror in a tokamak based on a convolutional neural network according to claim 1, characterized in that: The output layer uses the softmax function, normalizes the features output by the fully connected layer and then outputs the composition and content of the sputtering material on the first mirror surface corresponding to the input spectral data.

Images