Thermal coherent scattering real-time spectrum unfolding method based on stacked noise reduction auto-encoder

Abstract

The invention provides a thermal coherence scattering real-time spectrum unfolding method based on a stacked noise reduction auto-encoder, and relates to the technical field of magnetic confinement fusion plasma microwave diagnosis, and the method comprises the steps: 1, randomly dividing a thermal coherence scattering signal into a training set, a verification set and a test set, and training a stacked noise reduction auto-encoder 1 according to the training set, the verification set and the test set; 2, calculating a Spearman correlation coefficient between each output prediction error of the stacked noise reduction auto-encoder 1 and a main ion temperature prediction error, removing an output target with relatively low correlation with the main ion temperature, and training a stacked noise reduction auto-encoder 2 on the basis of the Spearman correlation coefficient; and step 3, randomly disrupting the arrangement of an input feature, calculating the influence of the process on the performance of the stacked noise reduction auto-encoder 2 until all features are traversed, removing the input feature which has small influence on the performance of the stacked noise reduction auto-encoder 2, and further training the stacked noise reduction auto-encoder 3 on the basis. The invention provides a real-time spectrum unfolding method for a thermal coherence scattering system for measuring the main ion temperature in magnetic confinement fusion.

Description

Technical Field

[0001] This invention belongs to the field of microwave diagnostic technology for magnetic confinement fusion plasma, and specifically relates to a real-time thermal coherent scattering despectrometry method based on a stacked noise-reducing autoencoder. Background Technology

[0002] Real-time feedback control in magnetic confinement fusion experimental reactors is a core technology for ensuring stable plasma operation and improving plasma confinement performance. Its successful implementation requires the integration of three core capabilities: high-speed diagnostics, real-time computation, and precise execution. Therefore, a series of methods have been developed to improve the accuracy and speed of diagnostic data inversion algorithms. However, in the field of spectral measurement, curve fitting methods based on physical models are computationally time-consuming and difficult to implement for high-speed diagnostics, typically only allowing for the calculation of the temporal evolution of physical quantities between shots. Artificial neural networks, on the other hand, based on large amounts of data, can directly and rapidly process spectral data and provide fast and reliable inversion results. Artificial neural networks have already been used to invert spectral data, including electron temperature measurement, ion temperature and plasma rotation velocity measurement, atomic spectrum noise reduction, and neutron energy spectrum analysis.

[0003] Therefore, in today's vigorous promotion of the commercialization of fusion reactors, how to achieve real-time spectral resolution of thermal coherent scattering systems has become an urgent problem to be solved. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a real-time thermal coherent scattering spectral decomposition method based on a stacked noise-reducing autoencoder, comprising the following steps:

[0005] Step 1: Randomly divide the thermal coherent scattering signal into a training set, a validation set, and a test set, and use these sets to train the stacked noise reduction autoencoder 1;

[0006] Step 2: Calculate the Spearman correlation coefficient between the prediction error of each output of the stacked noise reduction autoencoder 1 and the prediction error of the main ion temperature, remove the output targets that have a low correlation with the main ion temperature, and train the stacked noise reduction autoencoder 2 on this basis.

[0007] Step 3: Randomly shuffle the arrangement of an input feature, calculate the impact of this process on the performance of the stacked denoising autoencoder 2, until all features are traversed, remove the input features that have a small impact on the performance of the stacked denoising autoencoder 2, and further train the stacked denoising autoencoder 3 based on this.

[0008] Beneficial effects:

[0009] This invention provides a real-time thermocoherent scattering (TCS) spectral decomposition method based on a stacked denoising autoencoder. First, a stacked denoising autoencoder 1 is trained. By analyzing the correlation between each output and the host ion temperature, the dimensionality of the output data is reduced, and a stacked denoising autoencoder 2 is trained on this dataset. Then, the importance of each input feature is analyzed, and the input features with higher importance are selected to form a new dataset, thereby training a stacked denoising autoencoder 3. This method effectively compresses the dimensionality of the dataset, simplifies the network structure of the stacked denoising autoencoder, and accelerates the computation speed of the stacked denoising autoencoder, providing a general and efficient solution for real-time inversion in spectral diagnostic systems in magnetic confinement fusion reactors. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the real-time thermal coherent scattering spectral decomposition method based on a stacked noise reduction autoencoder according to the present invention;

[0011] Figure 2 The principle and training process of a stacked noise reduction auto-programmer;

[0012] Figure 2 (a) Thermal coherent scattering spectrum generated by simulation of stacked noise reduction custom encoder

[0013] Figure 3 The prediction results of the stacked noise reduction autoencoder 1 for the main ion temperature;

[0014] Figure 4 A performance comparison of stacked noise reduction autoencoder 2 and stacked noise reduction autoencoder 1;

[0015] Figure 5 Flowchart of the algorithm for determining the importance of permutation features;

[0016] Figure 6 The variation of the average relative error and the host ion temperature calculated using the substitution feature important algorithm. Thermocoherent scattering spectrum.

[0017] Figure 7 A performance comparison of stacked noise reduction autoencoder 3 and stacked noise reduction autoencoder 1; Figure 8 A comparison of the generalization capabilities of stacked noise reduction autoencoder 3 and stacked noise reduction autoencoder 1; Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0019] like Figure 1 As shown in the figure, an embodiment of the present invention provides a real-time thermal coherent scattering spectral decomposition method based on a stacked noise-reducing autoencoder, comprising the following steps:

[0020] Step 1: Randomly divide the thermal coherent scattering signal into a training set, a validation set, and a test set, and use these sets to train the stacked noise reduction autoencoder 1;

[0021] Step 2: Calculate the Spearman correlation coefficient between the prediction error of each output of the stacked noise reduction autoencoder 1 and the prediction error of the main ion temperature, remove the output targets that have a low correlation with the main ion temperature, and train the stacked noise reduction autoencoder 2 on this basis.

[0022] Step 3: Randomly shuffle the arrangement of an input feature, calculate the impact of this process on the performance of the stacked denoising autoencoder 2, until all features are traversed, remove the input features that have a small impact on the performance of the stacked denoising autoencoder 2, and further train the stacked denoising autoencoder 3 based on this.

[0023] In this embodiment, step 1 above, which involves randomly dividing the thermal coherent scattering signal into a training set, a validation set, and a test set, and using these sets to train the stacked denoising autoencoder 1, specifically includes:

[0024] First, the energy mean of the spectrum of a large dataset of noisy and noiseless thermal coherent scattering signals is normalized, i.e.:

[0025]

[0026] The dataset was then divided into training, validation, and test sets in a 6:2:2 ratio. During the pre-training phase, multiple denoising autoencoders were trained sequentially to extract spectral features from thermal coherent scattering, with the hidden layer of the last denoising autoencoder serving as input to the regression layer. After pre-training, all network layers were stacked together, and the entire network was fine-tuned at a low learning rate to optimize all weights and parameters.

[0027] like Figure 2 As shown, during the pre-training phase, the training of both the individual denoising autoencoder and the regression layer employs a classic neural network structure: the backpropagation (BP) neural network. A BP neural network is a feedforward neural network whose output is determined by the weights and biases of the preceding layers.

[0028]

[0029] in, It is the output of the k-th neuron in the l-th layer. It is the connection weight. It is the bias, and f(x) is the activation function.

[0030] The goal of training each layer of the neural network is to make the loss function...

[0031] Minimum, where N is the total number of samples. These are the parameter values ​​predicted by the network. It is the true value of the target parameter. λ is the sum of squares of the Frobenius norm, λ = 0.01 is the L2 regularization coefficient, and w j These are the weights of each target parameter. To highlight the importance of ion temperature, besides ion temperature... Other parameters w j =0.1.

[0032] To reduce the likelihood of the neural network getting stuck in a local minimum during training, gradient descent with inertial momentum is used to find the optimal connection weights and biases:

[0033]

[0034] Here, η = 0.01 is the learning rate, and γ = 0.1 is the momentum weight. It is the loss function with respect to the weights The derivative, The loss function is a biased loss function. The derivative, and These are the step sizes for the previous weight and bias updates, respectively.

[0035] like Figure 2 As shown, after pre-training, all network layers are stacked together during the fine-tuning phase, and the entire network is fine-tuned by reducing the learning rate η to 0.001. The mean relative error (MRE) is used to evaluate the accuracy of the ion temperature prediction.

[0036]

[0037] Where, N test It is the number of samples in the test set. It is the predicted value of the target output, y i,j This is the actual value of the target output. For example... Figure 3 As shown, the stacked denoising autoencoder has a prediction error of only 5.25% for the main ion temperature, with the vast majority of the prediction error concentrated in the range of -15% to +15%. This indicates that its hierarchical pre-training strategy can provide better initial parameters for the neural network, thereby achieving stronger performance.

[0038] In this embodiment, step 2 above: calculating the Spearman correlation coefficient between the prediction error of each output of the stacked noise reduction autoencoder 1 and the prediction error of the main ion temperature, removing output targets with low correlation to the main ion temperature, and training the stacked noise reduction autoencoder 2 based on this, specifically includes:

[0039] Using Spearman correlation coefficient

[0040]

[0041] This describes the correlation between the relative error of the main ion temperature and the relative errors of other variables. Where d i is the rank difference of the i-th data pair, and n is the total number of samples. The results show that the correlation coefficients between the outputs of other targets and the main ion temperature are all less than 0.02, indicating that they are almost uncorrelated.

[0042] Output targets unrelated to the main ion temperature are removed from the dataset, and a single-output stacked denoising autoencoder is trained on the new dataset. (e.g.) Figure 4 As shown, compared to stacked denoising autoencoder 1, the average relative error of the single-output stacked denoising autoencoder 2 only slightly increased from 5.25% to 5.35% (a relative increase of 1.9%), while the average computation time decreased from 8.1ms to 6.7ms (a reduction of 17.3%). This indicates that the single-output stacked denoising autoencoder 2 significantly reduces computation time without significantly increasing prediction error, demonstrating the feasibility of using correlation analysis to reduce the output dimension.

[0043] In this embodiment, step 3 above: randomly shuffling the arrangement of an input feature, calculating the impact of this process on the performance of the stacked denoising autoencoder 2, until all features are traversed, removing input features that have a small impact on the performance of the stacked denoising autoencoder 2, and further training the stacked denoising autoencoder 3 based on this, specifically includes:

[0044] like Figure 5 As shown, by randomly shuffling the permutation of an input feature, feature F is disrupted. i The relationship between the feature and the target output is used to quantify the impact of this feature on the performance of the stacked denoising autoencoder 2. If the target output is highly dependent on feature F... i Then the model's performance in feature F i The rearrangement will result in a significant decrease.

[0045] Figure 6 This demonstrates the variation in the average relative error calculated using the substitution feature importance algorithm and the main ion temperature T. i=200 eV thermal coherent scattering spectra. Clearly, the spectral features crucial for predicting the main ion temperature are concentrated in regions with small Doppler shifts (<300 MHz), which correlate with the highest ion temperature in the database. Here, frequency features with an average relative error change exceeding 0.1% are selected as key features to reconstruct the thermal coherent scattering spectrum database. Then, a novel simplified stacked denoising autoencoder is trained on the dimensionally compressed database.

[0046] like Figure 7 As shown, compared with stacked denoising autoencoder 1, the simplified stacked denoising autoencoder 3 only reduced the average relative error from 5.25% to 4.56% (a relative reduction of 13%), and the average computation time decreased from 8.1ms to 4.3ms (a reduction of 47%). This indicates that the feature selection scheme based on the permutation feature importance algorithm significantly improves computational efficiency while preserving computational accuracy. This method not only verifies the effectiveness of feature selection for simplified models but also provides new ideas for the design of broadband or multi-channel diagnostic systems.

[0047] By comparing the generalization capabilities of stacked noise reduction autoencoder 1 and simplified stacked noise reduction autoencoder 3 ( Figure 8 The simplified stacked denoising autoencoder 3 was found to have slightly stronger generalization ability. This indicates that the feature selection scheme based on the permutation feature selection algorithm eliminates some frequency features that are more susceptible to noise, thereby enhancing the robustness of the model.

[0048] The above are merely specific embodiments of the present invention, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A real-time thermal coherent scattering spectral decomposition method based on a stacked noise-reducing autoencoder, characterized in that, Includes the following steps: Step 1: Randomly divide the thermal coherent scattering signal into a training set, a validation set, and a test set, and use these sets to train the stacked noise reduction autoencoder 1; Step 2: Calculate the Spearman correlation coefficient between the prediction error of each output of the stacked noise reduction autoencoder 1 and the prediction error of the main ion temperature, remove the output targets that have a low correlation with the main ion temperature, and train the stacked noise reduction autoencoder 2 on this basis. Step 3: Randomly shuffle the arrangement of an input feature, calculate the impact of this process on the performance of the stacked denoising autoencoder 2, until all features are traversed, remove the input features that have a small impact on the performance of the stacked denoising autoencoder 2, and further train the stacked denoising autoencoder 3 based on this.

2. The real-time thermal coherent scattering spectral decomposition method based on a stacked noise-reducing autoencoder according to claim 1, characterized in that, Step 1: Randomly dividing the thermal coherent scattering signal into a training set, a validation set, and a test set, and using these sets to train the stacked denoising autoencoder 1, specifically includes: First, the energy mean of the spectrum of a large dataset of noisy and noiseless thermal coherent scattering signals is normalized, i.e.: The dataset was then divided into training, validation, and test sets in a 6:2:2 ratio. During the pre-training phase, multiple denoising autoencoders were trained sequentially to extract the spectral features of thermal coherent scattering. The hidden layer of the last denoising autoencoder was used as the input to the regression layer. After pre-training, all network layers were stacked together, and the entire network was fine-tuned with a low learning rate to optimize all weights and parameters.

3. The real-time thermal coherent scattering spectral decomposition method based on a stacked noise-reducing autoencoder according to claim 1, characterized in that, Step 2: Calculate the Spearman correlation coefficient between the prediction error of each output of the stacked noise reduction autoencoder 1 and the prediction error of the main ion temperature, remove the output targets with low correlation to the main ion temperature, and train the stacked noise reduction autoencoder 2 based on this, specifically including: Using Spearman correlation coefficient Describe the correlation between the relative error of the main ion temperature and the relative errors of other variables; among which, is the rank difference of the i-th data pair, and n is the total number of samples; output targets that are not related to the main ion temperature are removed from the dataset, and a stacked denoising autoencoder2 is trained on the new dataset.

4. The real-time thermal coherent scattering spectral decomposition method based on a stacked noise-reducing autoencoder according to claim 1, characterized in that, Step 3: Randomly shuffle the arrangement of an input feature, calculate the impact of this process on the performance of the stacked denoising autoencoder 2, until all features are traversed, remove input features that have a small impact on the performance of the stacked denoising autoencoder 2, and further train the stacked denoising autoencoder 3 based on this, specifically including: By randomly shuffling the arrangement of an input feature, the feature is disrupted. The relationship between the feature and the target output is used to quantify the impact of this feature on the performance of the stacked denoising autoencoder 2; if the target output is highly dependent on the feature... Then the model's performance in features The rearrangement will significantly reduce the noise; after determining the truly important feature inputs for predicting the main ion temperature, the database is reconstructed using the more important input features, and then a stacked noise reduction autoencoder is trained on this database.

Images